Add-host ridiculously fast: under 1.5 s local, under 4 s at 150 ms #189
Closed
opened 2026-10-12 14:03:20 +00:00 by kayg
·
581 comments
No Branch/Tag specified
main
j470b-server-mesh-rejoin
mesh-compat-382
j444b-d381-key-wins-flood
j461b-macos-utun-reader-thread
j475b-remove-host-motion
fix-reconnect-new-link
j481-d380
dev
j479-mesh-soak
fix/j476-host-facts
j475-motion-pill-split
j470-mac-node-join
root-upgrade-d376
j474-review-merge
latency-targets-120hz
nightly-suite-shards
gotatun-pin-j469
macos-utun-reader-thread
fix-flood-echo-stall
wip/j416
motion-fast-ease-out
mesh-addhost-j259
takeover-close-terminals-d368
kbd-instant-followups
fix-trust-key-double-click
nightly-suite-partial-report
appwin-icons-j207
rusqlite-slice1
swipe-content-first-d374
mesh-mtu-381
addhost-instant-open
glass-contrast-d365-mock
glass-contrast-d365
signout-followups-442
fix-startup-toast-red
research-gotatun-matrix-j317
themes-sync-j304
fix-load-flakes
fix-nix-zig-patch
fix-compute01-browser-baseline
ratchet-size-baseline-419
predict-auto-d130
revoke-coverage-d367
nightly-suite-queue-timeout
nightly-browser-suite
fix-rustsec-2026-0459
fonts-ioskeley-obel
log-retention-d375
fix-kbd-no-anim-427
fix-dev-red-specs-2
takeover-proof-305-j205
wip/j410-server-pump
ui-consistency-d372
fix-osc52-clipboard
libghostty-vt-1.0
forget-retired-host-d369
j404-merge-426
fix-setup-lock-race
fix-touch-select-426
revoke-access-report-ui
fix-predictive-echo-specs
fix-selinux-agent-exec
wip/j395-checkpoint2
wip/j395
j391-review-scratch
ui-consistency-j195-deferred-rebased
fix-rustc-paths-421
decisions-wording-341
fix-install-suite-red
fix-dev-rtt-specs
fix-pending-removal-align
fix-privileged-skip
fix-install-suite-404
review-j368
mesh-dns-j254
revoke-access-report-d356b
revoke-privilege-round5
fix-removal-confirm
review-j360
gotatun-rotation-367
deps-refresh-340
review-j354
review-J351
review-j350
addhost-takeover-d359
revoke-report-items-347
revoke-end-sessions-d362
review-j348
size-mesh-ssh-346
size-trim-337
revoke-access-diff-d356
j259-pre-rebase-b06c5a6
j305-ended-at-start-spec
addhost-sudo-consent
fix-layout-real-specs-411
fix-shell-open-411
fix-console-hosts-373
fix-shell-exit
fuzz-host-name-353
fix-host-details-410
macos-filter-j197
research-gotatun-tricks
revoke-p3-j216l
stale-state-pill
revoke-followups-j216k
revoke-progress-j216e
backup/macos01-20261106/lume/j197
backup/macos01-20261106/lume/appwin-icons-j207-stash0
backup/macos01-20261106/claude/stash-0
backup/macos01-20261106/claude/stash-1
backup/macos01-20261106/claude/J80-macos-s1
backup/macos01-20261106/claude/J80
backup/macos01-20261106/claude/J80-slice1-20261102
backup/macos01-20261106/claude/J16b-appwin-on-dev
backup/macos01-20261106/claude/appwin-launch-idle-146
backup/macos01-20261106/claude/J80-s1-run
wip/j259
trust-285-p3c
runtime-lock-j216f
mesh-ssh-p3-j175
tun-proof-p3b
trust-285-p3
terminal-polish-j271
tun-proof-p3
ssh-install-flakes-j216i
fuzz-fixes-j264
trust-285-a
trust-285-b
restart-oversized-fixture-j198
backup/macos01-20261106/lume/appwin-icons-j207
restart-bound-j216h
up-preflight-p3-j198
backup/macos01-20261106/lume/j251
up-preflight-j198
wip/up-preflight-j198-j245-p2
wip/up-preflight-j198-j245-p1
agent-fsync-offloop-j196
wip/up-preflight-j198-j245
settings-offscreen-j231
wip/up-preflight-j198-j226-inspection
wip/j103f
themes-j206
wip/up-preflight-j198-j226
flakes-j216c
backup/macos01-20261106/claude/j197
flakes-j216b
backup/macos01-20261106/claude/j197-red
wip/up-preflight-j198-review
flakes-j216
backup/macos01-20261106/claude/appwin-icons-j207
engineering-j208b
appwin-mock-j209
toast-stack-j165
engineering-j208
decisions-d351
decisions-d350
decisions-d349
appearance-presets-mock-j129
decisions-d348
docs-audit-spec
doc-coverage-baseline-raise
decisions-d347
decisions-d346b
decisions-d346
j88h-profile
ui-consistency-j195-deferred
ui-consistency-j195
engineering-practices-j194
preferred-login-ui-j166
signin-pools-263-j174
macos-node-265-s1
decisions-d345
decisions-d344
proxy-cgroup-j164
verify-quick-fuzz-329
proxy-start-io-j168
timing-flakes-j163
up-263-s1
mcp-events-j147
verify-quick-fuzz-j158
proxy-p3-j159
agent-resource-j157-detect
wip/j103d-final
remote-cancel-hang-j155
proxy-traefik-281-j62d
j149-verify-quick
proxy-traefik-281-j62b
j150-drop-unnamed
wip/j79d-red-r4
wip/j103d-d332
rootless-toast-j142
wip/j103d-head
wip/j103d
j125c-mcp
macos-build-j136
backup/macos01-20261106/claude/j136
wip/j79d-red-cli
wip/j79d-red-server
j121-link-liveness
j118c-tracer-failures
wip/j79c-head
release-desktop-j131
j124c-signed-macos
wip/j79c-red-p3
wip/j79c-red-p23
wip/j79c-red-p22
merge-j128-check
perf-deferred-j128
sidebar-toggle-260
j127-shortcut-changed-dot
j125-mcp
j118-tracer-failures
j126-account-summary
settings-sections-mock-j92
j120-pinch-lock
sidebar-292
backup/macos01-20261106/claude/j120
taste-2026-11-03
row-actions-242-215
backup/macos01-20261106/claude/stash-0-untracked
backup/macos01-20261106/claude/J80-265-s1-run
backup/macos01-20261106/claude/J80-macos-s1-20261102
j112-canvas-commands
ui-no-overshoot-scroll
fix-recovery-restart
backup/macos01-20261106/claude/stash-1-untracked
taste-2026-11-02
size-baseline-ratchet
machine-cards-structure-j108
j109-boot-copy-mockups
fix-dev-red
wip/up-263-s1
j95b-settings-rows
proxy-traefik-281
j95-option-rows
docs-human-j91
docs-keys-palette
fix-flood-1013-270-j38d
mesh-ssh-s1
mesh-ssh-s0
addhost-d291-tunnel
test-stability-load
mesh-relay-220
repro-296
shim-thp-off-24
mesh-pins-258-v2
mesh-dns
signin-d311-merged
docs-pages
ui-size-283
fix-ui-batch-specs
wip/j12b-edge-update
gate-on-arm
fix-rename-sync-293
trust-285-j60b
fix-focus-ring-clip
integ-ui-1101
edge-update-v2
signin-d311
taste-2026-11-01
layout-per-tab-250
ui-approved-splitopt-224
ui-approved-splitopt-224-on-dev
icon-text-centre-282
ui-approved-189-212-216
research-reverse-proxy
transfer-budget-320
review-mesh-pins-258-r2
sidebar-pill-259
review-mesh-relay-220
wip/j12-edge-update-v2
nix-package
status-crash-loop-237
fix-dev-test-failures
fix-fixture-socket-path
wip/mesh-relay-220-pre-dev-advance-rebase
fix-canvas-side-clip
research-head-to-head
frosted-real-256
review-mesh-pins-258
wip/mesh-relay-220-pre-final-rebase
review-layout-per-tab-250-r2
broad-suite-triage
nix-devshell
review-status-crash-loop-237-r3
wip/nix-devshell-pre-p3
taste-frosted-real
wip/code-serve-demo-2026-10-30
wip/code-serve-glass-2026-10-30
wip/code-serve-clusters-2026-10-30
wip/code-merge-release94-dev-2026-10-30
wip/rproxy-delta-scratch-2026-10-30
wip/stale-note-257-staged-2026-10-30
wip/d241-size-staged-2026-10-30
review-broad-suite-207
wip/mesh-pins-258-pre-split
wip/mesh-relay-220-pre-rebase
mesh-measure-220
plan-211-update
review-edge-autoupdate-r3
perf24-reconnect
review-stale-note-r2
release-prod-live
d241-shell-marker
review-mesh-pins-server-258
review-revocation-r3
review-setup-token-r2
review-sync-journal-r3
wip/review-mesh-pins-258-prune-review-1995fee46
frosted-scroll-256
sync-journal-r3
revocation-per-account-wip
review-tokio-asyncfd-0386
review-sync-p3s
wip/stash-8by64-stash0-2026-10-30
mesh-drag-r2
setup-token-data-wip
release-9-4
auth-busy-1013
mesh-view-overlap
sync-p3s
review-auth-busy-wake
review-readd-r8-wip
mesh-drag-bounds
audit-auth-93
app-windows-stream-start
review-authority-cap-r3
app-windows-r7-measure
readd-minimal-r6
readd-minimal-r5
app-windows-r7-soak
readd-minimal-r4
app-windows-r7
app-windows-r6
readd-minimal-r3
readd-minimal-r2
readd-minimal
merge-release8-dev
review-busy-page-liveness-r2
review-busy-page-liveness
mesh-view-trim
review-authority-read-retry-r2
review-remove-uninstall-r2
release-retry-reattach
release-retry
review-remove-uninstall
review-authority-read-retry
app-windows-ship
mesh-autoswitch-p2
mesh-ssh
mesh-relay
mesh-macos-node
screen-and-control
session-labels
review-real-logins-r5
review-real-logins-r4
review-real-logins-r3
shim-memory-r2
review-real-logins-r2
shim-memory
shadow-addhost-instant
shadow-real-logins
shadow-same-session-flood
shadow-prod-perf-fix
hosts-from-server
app-windows-real
app-windows-transport
release-prod-6w
mesh-policy-verify
quiet-close-toast
dev-merge-r4
review-mesh-agent-transport
review-app-windows-transport
settings-roundtrip
review-release-pipeline-r4
mesh-agent-transport
ended-strip-fix
review-audit-ca-update-d244
settings-rows
phone-width
signout-clear
fix-restart-tests
dev-merge-rp
signer-hardening
review-audit-ca-update
terminal-fidelity
review-audit-ca-update-wip
mesh-integrate
review-mesh-agent-transport-515d3712
review-mesh-dataplane-feec1988
mesh-dataplane
signout
research-canvas-node
wip/review-renderer-cost
wip/review-terminal-fidelity
padlock-glyph
review-release-pipeline-2e95b689
phone-polish
sidebar-polish
inbox-generic
review-mesh-dataplane-378d4055
prod-batch
ui-inbox
ui-ca-unlock
ui-machine-sections
review-mesh-dataplane-a5f744fb
tls-session-prod
session-bounds-prod
fix-dead-session
session-bounds
tls-session
palette-polish
shortcuts-d222
research-appmeta
phone-gutter
ui-centre-button
wip/ui-centre-button
wip/tls-session
review-addhost-breakit
ui-platform
renderer-cost
remote-agents
mesh-api
ssh-bootstrap
prod-hotfix2
prod-fixes
review-remote-agents
review-mesh-dataplane
session-life
research-browser
review-mesh-api
research-capture
review-mesh-api-round3
review-mesh-api-round1
gate-stability
review-gate-stability
review-relay-direct
review-mesh-api-findings
proto-sessions
font-picker
proto-addhost
proto-windows
mesh-core
latency-gate
research-bsd
research-discovery
keytray-mods
mock-core
ui-tweaks
fix-motion-sel
fix-motion
integrate2
fix-robust
t-web
t-server
vt-ghostty
fix-snap-echo
t-web-auth
review-vt2
t-agent
lane-protocol
lane-shim
lane-auth
s1-base
s1-contracts
s1-workspace
No results found.
No labels
launch-qa
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".
No due date set.
Dependencies
No dependencies set
Reference
kayg/infraternal#189
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Owner 2026-10-12: add-host takes 10 s+, the app is meant to be fast. Perf-VM breakdown (breakit2): 5.4 s local / 13.2 s at 150 ms RTT — failed 'none' auth 1.3 s, key auth 1.6 s, 19.9 MB uncompressed upload 3.0 s, join 1.2 s, three serial SSH commands for install/start/linger 1.6 s, terminal proof+commit 1.5 s, controller overhead incl. a fresh payload digest per add 2.3 s. Job: addhost-fast.
Mockup for the setup status moving into the machine pill (D255, TASTE "Progress lives where the thing is"). Branch
mockup-addhost-pill(pushed, not merged, not deployed), head53f99f2e, base release-prod-5 (5bf3f42e).URL (preview build, test double only; the real wizard is on screen, an autopilot picks the Ubuntu host
o2and presses Continue):/?mock=1&proto=clusters&scene=3&addhost-demo=pill&fail=join(install,start,readyalso work); "Try again" in the popover plays the stage again and succeeds&slow=3stretches all timings,&tickhold=3000holds the tick,&autopilot=0to drive it by handproto=windowsis the older window prototype without the canvas mock, so the demo lives onproto=clusters.Captures:
apps/tracer/captures/addhost-pill/: 1440x900, 393x852, 375x667, ink and paper:1-wizard-status,2-flight-1..6(screencast frames),3-pill-2of4,4-tick,5-os-icon,6-failed-arc,7-failure-popover(plus-closeupcrops), and*-happy-path.webmat 1440x900 and 393x852. Stills run 3x slower with a 3 s tick; the videos run at real timings.What it does: wizard progress for ~1.5 s, then
done(shell ready): the window opens with a prompt at once, a copy of the wizard's status glyph flies (280 ms, quartic ease-out, transform only, tracking the pill because the camera is still panning) into the pill mark, where it is a 4-arc ring in the same box as the system icon (pill width never changes). Arcs: pending track, running pulse, done fill, failed in the danger colour. All done: circle to tick (150 ms), hold 600 ms, crossfade to the system icon, ring removed. Failure: pressing the pill opens the shared confirm popover ("o2 couldn't join this server" / "Your terminal keeps working. Try again in a moment." with retry and dismiss icons; the retry glyph, not a tick). Reduced motion: no flight, no pulse, fades only.Event contract the UI expects (
apps/tracer/src/canvas/addhost/setup-stages.ts, header documents it,parseSetupEvent+ tests):doneevent means "the shell is attachable" (D255); it already carrieshost_id.runningthendoneordonealone;readydone completes.failedends the attempt; a retry sendsrunningfor the failed stage again.failed_step-> arc: connect/upload/install -> install, join -> join, start -> start, readiness -> ready.SetupSource.retry(hostId)(mock only today); the backend needs an action for it.Gates (worktree, load average 15-40 on the host):
bun run checkclean;bun test src180 pass (9 new in setup-stages.test.ts); production build +transfer-budget.ts: critical 294,964/307,200 gzip (+27 B), deferred 49,810/51,200 (+52 B), no mock code in dist (setup-pill is on-demand and nothing in the production graph imports it yet);TRACER_TEST_PORT=4989 systemd-run --user --scope -p CPUWeight=200 bun x playwright test test/addhost-pill.spec.ts: 4 passed (twice in a row). Measurements: script share 2.0-2.5 % of the run (idle canvas page 3.1-6.1 %); long tasks >= 50 ms from the status on: 0 and 3 in two runs, against 0 and 25 for the same page without the demo in the same minute, so headless long tasks here are the host's load, not the ring; worst frame gap 67-217 ms (headless, loaded).scripts/verify fast: Rust gate fails on this base (clippy FAIL 101, test FAIL 101, static builds FAIL: binary 18,109,872 B over the 13,817,816 B budget); nothing in Rust changed here.Note: the second spec asserts the pill width through ring, tick and icon; the session rename (
shelltoinfraternal) in the mock changes the pill's width once, which is mock inventory behaviour.Design questions (recommendation first):
Orchestrator decision (D206, reversible UI; owner approved the pill mockup 2026-10-12 "this is awesome! I love it!"): Q1 wizard ends at "Open a terminal on the host", stages live only in the pill; Q2 no "ready" toast when the pill carries it; Q3 camera pans only when the new window is off screen; Q4 arc fill stays an opacity fade (compositor-only); Q5 show the pill mark while setup runs on windows < 320 px, hide after. Build after the D255 backend.
Review of addhost-fast (
26ccb33a): MERGE, P0=0 P1=0 P2=2. Failing tests on origin/review-addhost-fast (821815e5), evidence/review-addhost-fast/README.md:The addhost-instant job (D255) builds on addhost-fast: cherry-pick both failing tests and fix forward there.
Starting D255 on addhost-instant at base
26ccb33ab4(origin/addhost-fast). Read worktree CLAUDE/DECISIONS/ARCHITECTURE/PROCESS, origin/dev D240/D247/D255 and #189/#169/#190 history and speed evidence. Preserve ownership/start-fresh/retirement admission and real shim sessions. Fold in R190-01 source binding and R190-02 closed-input receipt findings. Branch push authorized; no deployment or owner fixture changes. Measurements only on booked VM208 with shared lock.D255 contract proposal before code: publish an SSH-backed terminal only after verified payload staging and shim-owned shell launch. Keep host UUID, session UUID, execution uid, per-writer input cursor and output cursor unchanged when the existing terminal actor replaces SSH with admitted leg B. No process adoption outside the shim. Add replayable host setup events with host_id, session_id, window_id, attempt, sequence, stage install|join|start|ready and state running|done|failed. Window id is supplied by the canvas attachment, not fabricated as a server layout id. Host setup subscription outlives wizard closure; revocation/component-off still fences work. Before shell publication, close retains existing cancellation/rollback behavior; afterwards installation failure must preserve the working shell and explain failed setup. Logins compatibility: the early shim must enter the same per-person/account host login D240/D247 as installed sessions once that implementation lands, not reuse an SSH daemon login belonging to a different person or move an existing shell between logins.
Review repairs: preserve native OpenSSH source-binding/family/interface/IPQoS behavior by bypassing fd acceleration whenever those effective options differ from defaults; retain only canonically decoded named refusal receipts on input-write failure, never arbitrary stdout as diagnostics.
Completed review repairs:
b7486463retains only canonical setup refusal on closed invitation input; unchanged real-shell review regression red0/1 then process suite green8/8, strict SSH Clippy/fmt pass.4c58dba3preserves native source/interface/family dialing when configured; unchanged contact assertions red0/1 then green1/1, strict SSH/CLI Clippy on the repair passed. No host-key/authentication weakening. Raw red/green logs retained in job scratch for evidence commit. D255 bridge work uses the existing dispatcher and shim rendezvous, without a new framing/terminal engine.D255 primitives complete:
a63dfba1duplex SSH child lifetime; 09e8f51f readiness-driven standard pipes with CLOEXEC; 39f8bb0e additive SSH session interface; 6c0b1431 existing-master command builder; ef77dfae internal shim-first session bridge; a486d936 real-process handover oracle; 3c956c48 proposed event/login contract. Strict SSH/CLI Clippy/fmt pass; SSH suite passes, module tests26 pass/2 ignored. scripts/verify-ssh-session.py with real GNU binary passes: same session and even writer, pending input executes once after bridge replacement, output resumes. This is command-pipe proof, not real-SSH enrollment or latency acceptance.Remaining D255 integration: publish/route early sessions from bootstrap; background setup ownership after wizard close; persistent host setup events + test double; admitted transport handover/SSH retirement; ten fresh-target LAN/150ms measurements. Existing orchestration is deliberately not switched to an incomplete route. Final full gate and evidence retention proceed; no completion/latency claim.
Gate finding on source head5731f06d: fmt/VT fmt/workspace Clippy pass; server suite327 pass/1 fail/39 ignored. Unchanged auth::tests::socket_tests::mesh_calls_and_subscriptions_deliver_scoped_live_inventory fails socket_tests.rs677 (expected successful Result call_id3). Exact-artifact replay with matching target/state environment reproduces0/1; earlier direct replay omitted target environment and closed the socket, retained separately. This is an assertion, not a permitted timeout exemption. Comparing detached origin/dev5e383c3d in owned scratch; no cause/waiver claimed. Static builds continue. Browser first run29/1 (heading15s timeout), unchanged full replay30/30. Tracer check and171 source tests pass. Evidence paths .job-scratch/{verify-fast,mesh-replay-env,browser-replay,browser-final,tracer-src}.log; retained evidence will be committed. D255 controller/status integration remains unfinished.
Final branch-head scripts/verify fast output, verbatim part 1/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 2/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 3/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 4/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 5/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 6/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 7/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 8/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Final branch-head scripts/verify fast output, verbatim part 9/9. Head
c9644a85dd; exit1 only static-size caps. RUST_TEST_THREADS=2; unique TRACER_TEST_PORT=4997; unchanged RTT self-test on VM208/shared lock. D255 integration remains partial.Partial checkpoint pushed: addhost-instant head
c9644a85dd, remote SHA verified; base26ccb33a. No deployment, production branch update or issue closure.Completed: R190-01 native source/interface/family routing repair; R190-02 canonical refusal retained after closed invitation input; removable SSH duplex/stdio primitives; hidden shim-first command and real same-session/same-writer pending-input handover oracle; proposed D255 host event/login contract. Evidence and explicit continuation steps: evidence/addhost-instant/README.md and reviewer-checklist.md. No provisioning latency or full early-terminal completion claim.
Remaining: bootstrap controller publication/routing, background setup ownership after wizard closure, persistent host setup events/retry plus test double, admitted mTLS handover/SSH retirement, real-login context integration and10 fresh targets each at LAN/150ms RTT. Existing add-host orchestration is unchanged. The job's intended D255 outcome is NOT complete.
Final branch-head scripts/verify fast exit1 only at unchanged static-size caps;772 native pass/0 fail/79 ignored, no timeout exemption. Fmt/VT fmt/Clippy/cargo-deny PASS. X86 static20551632B (+76384 versus base20475248), cap16404816; ARM18309456B (+69040 versus base18240416), cap13817816. Both stripped/static/no PT_INTERP; x86 protected key-memory initialization passes under8MiB memlock. No ARM runtime claim, cap/profile/dependency changes or accepted CPU/RSS/latency measurement. Full209451-byte output quoted verbatim in preceding9 parts; SHA2561ced33361c228d98819c3ad171c35675014065630975c70166d7320fde0450ae. Final raw log retained locally at evidence/addhost-instant/verify-fast-final.log; committed milestone log retains earlier failures.
Browser unchanged full replay30/30, including Start fresh; first29/1 failure retained. Tracer check +171 source tests pass. Earlier mesh assertion repeats on milestone/exact artifact; newer detached dev5e383c3d passes. Final full head replay with2 test threads passes the mesh case and every native test. No causal waiver or asserted root cause. Historical failure-details capture restored; optimized test-double screenshot retained for independent review, no visual/physical-device claim.
Final Rust table, verbatim:
+---------------------------+--------------+
| Rust gate | Result |
+---------------------------+--------------+
| fmt | PASS |
| VT native fmt | PASS |
| clippy | PASS |
| test | PASS |
| cargo-deny | PASS |
| static builds | FAIL (1) |
+---------------------------+--------------+
GATE_EXIT=1
Owned scratch/build targets and VM input cleanup follows; fixture and process cleanup is limited to this job's own paths/PIDs. Owner live fixtures were never read or changed.
Checkpoint cleanup: detached dev checkout removed; owned addhost-instant build/baseline targets and fixture roots ahi/ahi-baseline removed; VM208 input script/directory removed after lock release. No listeners remain on this job's ports4997/4998. No shared process or owner live fixture touched. Branch remains pushed at
c9644a85dd; source unchanged after final gate. Committed evidence/README and local final raw log retained. D255 outcome remains incomplete, as recorded above; issue stays open.Continuing D255 round 2 on addhost-instant, checkpoint/base
c9644a85dd(speed base26ccb33a). Read checkpoint remaining integration 1–6, repository contracts, D255 on origin/dev and #189/#169 history. Finish staging/publication, background ownership, retained setup events, admitted handover, login compatibility and VM208 acceptance in that order. Compare both static targets against origin/release-prod-live9ab4329cusing the same recipe; caps unchanged. Branch push authorized; no deployment or owner fixture changes.Staging prerequisite completed:
1aad38acbounded duplex receipts; shared verified compressed/reused staging launches __ssh-session before enrollment (see branch commits). Installation real-file suite61/61; strict SSH all-target Clippy/fmtPASS. Regression requires protocol bytes unchanged and background setup/rollback admitted while bridge remains live. Discovered/rejected retained parent fd9: release the setup directory guard before bridge exec; transaction lease remains owned. No prompt latency claim yet; router/publication remains next.D255 integration refinement before retry code: published setup failures retain the live shim/rendezvous and host identity. Retry is authorized by a fresh owner capability and reuses the exact host/session id; it restores only the installation journal through existing rollback, with fresh-root deletion disabled, then restages the verified cache and reissues enrollment for that same host. No new shell, foreign directory adoption or process stealing. Retained per-host status carries window id supplied by each subscriber, sequence and attempt; bounded history and explicit resync. Disabling SSH retires provisional bridge transports and detached setup; admitted infraternal-agent sessions retain their independent component lifecycle.
Retained status module committed7f65a9da/f52ca728:16-event replay plus4 authoritative stage states per host, exact host/session ownership, monotonic attempts/sequences; module testPASS. Live retry module/tests committed on branch:62/62 real-file installation tests; strict SSH/server Clippy/fmtPASS. Join/start refusal cases preserve enrollment-state bytes, rendezvous inode/bytes and bridge traffic through retry and commit. One initial test incorrectly requested unpublished rollback after a successful published-session retry; corrected the new fixture to exercise the required commit path. Production admission/router integration is still being validated; no latency or full completion claim.
Early duplex diagnostics regression found and repaired: canonical foreign-root refusal was lost before TRANSFER. early-refusal-red.log fails Upload assertion; unchanged oracle passes after shared setup-step wrapper/refusal decoder. Installation63/63, strict SSH Clippy/fmtPASS; two atomic commits on branch. No arbitrary stdout/stderr diagnostics. Review continues on publication/wizard-close ownership and same-actor handover before full integration gates.
Continuing addhost-instant at head/base
88b6f9356e, with retained uncommitted integration work from the capacity interruption. Read #189/#169 history and checkpoint contracts; validate publication, wizard-independent setup, retained status/retry and same-actor handover before remaining real-engine and VM208 acceptance. Branch push authorized; no deployment or owner fixture changes. Static targets will be compared to production9ab4329cwith unchanged caps.Shadow review of
addhost-instant@c9644a85(D255)Independent review against D255, D240, D242, D243, D262 and
contracts/host-setup.md. Failing tests are on branchshadow-addhost-instant@11034a7c(three test-only commits on top ofc9644a85). Scope: the bridge primitives (duplex.rs,stdio.rs,ssh_session.rs,openssh.rschannel builder) and the #190 speed work underneath them (negotiated transfer, setup steps, socket connector, payload encodings). The controller, host-status events and pill are not built yet (the author says so inc9644a85), so nothing here judges the end-to-end 1.5 s / 0.3 s target.Control: the author's
scripts/verify-ssh-session.pypasses on my build of the node binary (--no-default-features --features remote,ssh).P1
headreads through stdio (BusyBox/Alpine, BSD/macOS).crates/plugins/ssh/src/install.rs:180: the payload is framed withhead -c N, and the join step then reads the private invitation from the same stdin (process.rswrite_inputwrites payload and invitation back to back). GNU head reads exactly N bytes. getc/fread-based heads pull a whole buffer from the pipe, so they consume the invitation andjoingets nothing. The step-separated path never had this framing, so this is a regression. In the fixture, GNU head passes. A buffered head delivers 0 of the 21 invitation bytes; the fixture's join does not check its input, so setup still reports success there, but a realjoin --enroll-onlywould fail.Test:
cargo test -p infraternal-plugin-ssh --test install negotiated_transfer_delivers_the_invitation_after_a_buffering_headPossible fix: frame without over-reading. Use
dd bs=1 count=N(slow), a reader that cannot over-read, a length-prefixed read done by the uploaded side, or send the invitation as a separate step after a receipt.P2
Every
__ssh-sessionstart launches a new shell, so a reopened bridge strands the old one.crates/infraternal/src/ssh_session.rs:27callssessions::launchbefore it reads anything from the channel. When the bridge is lost and opened again (SSH drop, retry after a failed stage, server restart), a second login shell starts and the first keeps running with no viewer. The same happens when the 5 s Hello deadline expires because the server went away. Nothing limits or retires these shells until an agent joins, which may never happen. This also contradicts the contract line "A retry starts a new attempt and keeps the existing session": the primitive has no way to bridge to an existing session without starting another. The reopened bridge's census does list the old session, soAttachTermcould reach it. The unwanted launch is the problem.Test:
python3 scripts/verify-ssh-session-rebridge.py --binary <node build>fails with2 live sessions.LogLevel=VERBOSEon the master puts server banners into the 8 KiB pre-authentication budget.crates/plugins/ssh/src/openssh.rs:213. OpenSSH prints the pre-authentication banner only at INFO or above, so at ERROR it never reached the diagnostics. Now a long legal notice (about 9 KiB) refuses the host with "The SSH diagnostics exceeded their safety limit." The banner text also reachesclassify()now, so a banner that contains "Permission denied" or "Host key verification failed" picks the failure category for a failure of a different kind.Test:
cargo test -p infraternal-plugin-ssh --test shadow_banner. The fixture asks realssh -Gfor the effective level and prints the banner exactly when OpenSSH would.Possible fix: keep ERROR on the master and get the method receipt some other way, or stop counting the banner against the budget.
Server start now waits for four single-threaded
-9compressions (no test; this needs the perf VM).crates/infraternal-server/src/bootstrap_runtime.rs:107awaitsprepare_payloads()insideinitialize, before listener admission. That is zstd -9 plus gzip -9 for both architectures, on every start and restart, so the restart outage grows.payload_encoding::encodedalso holds the globalCACHEmutex while the compressor runs, so on a cold cache concurrent add-hosts queue behind it. The cache can hold up to 4 entries in resident memory for the server's lifetime. Suggestion: warm the cache in the background after the listener is up, and do not hold the lock across the compressor.Design gaps that will block the handover, not testable until the controller exists:
systemd-run --userscope under the SSH account, not in the D240 per-person login. Until that lands, revocation (D240/D242) cannot end this shell, either before the agent joins or after, because nothing records which person it belongs to.bridge()usesUid::effective()as the session uid. A root (sudo) install would need the bridge to run as root to use root's state directory (private()requires that the owner equals the effective uid), and then the shell would be a root shell, not the user's. Run without sudo, it uses a state directory the root agent never discovers.terminal.openand pluginCalls) over SSH. That is fine for privilege, because SSH already grants the account, but component-off and revocation fencing for this channel must live in the server controller.P3
ssh_session.rs:27). The window's real size is known before launch, so the first prompt needs a resize and a redraw. Also on the path to the first keystroke:run_channelrunsManager::detect()again (a secondsystemctlspawn) and a fulldiscover()before it handles Hello.socket_connector.rs) tries resolved addresses one after another under a single 10 s budget, so a black-holed AAAA record uses up the whole budget. It also ignores a user-configuredConnectTimeout.evidence/addhost-instant/(30 files, a PNG among them) and otherevidence/additions. These must move to the issue before merge.Not verified
sshdon the build host. BusyBox and BSDheadare modelled by the fixture, not run.Handover review finding: refused_installed_reattachment_restores_the_ssh_writer failed 0/1 at remote_ssh_tests.rs:200 (no writer Ack after installed connection drops before replay acknowledgment). The actor consumed the replacement then awaited a future admission before restoring its retained SSH channel. Changed connect_with_fallback to restore retained SSH first when no replacement is pending; red log .job-scratch/fallback-red.log, green replay running. Production same-recipe baseline
9ab4329calready built: x86_64 20,333,104 B; aarch64 18,112,240 B. No cap change or completion claim.Real VM208 D255 oracle failed after an interactive prompt: readiness=false, Ready stage failed. Initial LAN diagnostic prompt 495.171 ms (GNU controller, static payload; exploratory, not final acceptance). Cause: canary hardcodes writer streams 2/4, colliding with the already-attached early writer2. Proposed fix before code: reserve the canary's two private writer ids from the existing terminal_writers SQLite AUTOINCREMENT sequence, without publishing a private terminal row or inventing another writer namespace. Keep those ids with cleanup ownership. Same allocator bounds and ordinary replay remain. Raw initial-lan.log retained in VM job directory. Also, concurrent native recompilation deleted a running self-exec test image, causing ENOENT rows; subsequent verification will run serially with matching artifact/state and umask022. Those failed runs are retained, not waived.
Integration steps committed through17ff79bc (source base88b6f935): retained wizard-independent bootstrap7557420f; same-session retry command/oracle04bfdf66; same-actor acknowledged handover0ea64857; private readiness writer collision fixd065c3f3; owner/uid-admitted publication66beb64b; published recovery28ff7002; retained status API5eacbded; real acceptance oraclee042b223; canonical canvas attachf31f56e2; provisioning-double replay/retry17ff79bc. Strict touched Clippy/fmtPASS; installation63/63, bootstrap46/46, remote18/18, guard2/2; real same-session SSH retry/pending input oraclePASS; tracer173/173. VM208 green one-target LAN diagnostic: prompt491.240ms, fully joined753.510ms; original writer2/session preserved,59 input bytes confirmed, wizard handle closed during setup, zero unconsumed tokens/attempts. Static controller + stripped static payload20743952B. Ten-target acceptance and size-reduction investigation remain; no completion/target-met claim.
Author shutdown check: component_off_retires_ssh_fallback_after_installed_admission fails0/1 before fix (remote_ssh_tests.rs209, retained SSH peer remains alive5s despite SSH off). Installed admission can coexist with the old unacknowledged SSH fallback. Fix sends an owned retirement command through the existing terminal actor; it drops only SSH transports and acknowledges after cleanup, preserving installed mTLS. Green replay running. Weighted production-build browser replay41pass/0fail/4capture-only skipped includes Start fresh and retained retry; initial38/5 failures retained, assertions unchanged. Historical screenshots restored after copying fresh captures to job scratch for independent review.
Round-trip refinement before code: use one duplex command channel for the existing bounded read-only check and subsequent verified payload staging/shim launch. The trusted server sends only a bounded public setup script length/body through that channel after current check, mode, ownership and invitation admission; no credential enters that envelope. Existing check/refusal/transfer parsers and shell quoting stay authoritative. This removes reopening another SSH command channel between check and upload; legacy adapters retain their buffered path. Start fresh remains explicit and may use another channel only for its confirmed recovery. Size experiment also shares large setup polling boundaries through the existing boxed Operation type; no dependency/profile/cap changes.
Shadow review fold-in: unchanged negotiated_transfer_delivers_the_invitation_after_a_buffering_head reproduces0/1, invitation0/21 bytes; unchanged long-banner oracle reproduces0/1 at the8KiB safety limit. Fix before private stdin: release invitation only after a verified payload-ready receipt; early duplex already separates payload and subsequent protocol/enrollment with receipts. Keep existing final canonical setup output unchanged. Restore ERROR on the SSH master and selectively request its local authentication function diagnostics using documented LogVerbose; banners remain suppressed. Preserve explicit same-session resume and port the rebridge census to its new --session API (still exactly1 live session). Investigate background cache warming and lock scope for startup finding4. Public source references: OpenSSH sshconnect2.c userauth/banner and log.c selective logging; no host trust/auth weakening or dependency changes.
Completed review fix
6590195f: verified PAYLOAD ready barrier precedes private invitation stdin. Imported shadow buffered-head oracle failed before the fix (0/21 bytes delivered) and now passes exact and buffered readers. fmt, strict SSH clippy and all SSH tests pass; no dependency or static-cap change.Acceptance audit correction: the
c2994e43controller was current, but automatic authentication reported bridge capability before method selection, forcing the legacy path. The 20 fresh-target logs have no time-to-prompt/session proof and are rejected as D255 evidence. Fixing the ordering and adding an explicit published-session assertion before repeating all 20 measurements.Completed review fixes:
247eefd3suppresses remote banners (ERROR + selective local method receipt; the 9 KiB red oracle now passes);788bc9acselects the checked bridge after automatic sign-in and rejects a legacy completion in the instant acceptance oracle (10 ordered-auth tests pass).0a453a85binds HTTP before cancellable payload warming and moves compression outside the four-entry global index (3 encoding tests, 46 bootstrap tests, both shipping-route regressions, 6 composition tests pass). Strict SSH/server clippy with remote and fmt pass.310aa0ccrecords these invariants in contracts/host-setup.md.73057509ports the independent reconnect census to explicit --session resume; real binary proof has exactly one live session.Rebuilding both stripped static architectures and release acceptance controller/payload from
73057509. No dependency, release profile, static cap or owner fixture changes. Real-login launch remains compatible by contract with pending D240/D247; this branch does not claim per-person PAM implementation. Acceptance data must include prompt/session/writer proof before it is reported.Additional lifecycle finding at
73057509: routing refusal after shim creation was treated as already delivered Done, leaving the wizard with neither terminal event. New tests/publication.rs fails 0/1 before the fix with “routing refusal left the wizard waiting without Done or Failed”. Separating the protected-session resource bit from acknowledged publication; preservation/rollback rules stay intact. Only the owned artifact-build PIDs were stopped so both architectures/controller can be rebuilt from one final head.Parser verification: existing ssh_check_decode instrumented target includes TransferChoice, canonical setup/refusal decoders and every-byte method-receipt split equivalence. VM208 under flock, 4 vCPU; ASan smoke campaign passed 100000 inputs with no finding. Relevant parser sources match branch head
71c97ac7. Fuzz-only build used CARGO_PROFILE_RELEASE_LTO=false after the normal nightly ASan/LTO link failed with undefined _sancov_gen symbols; repository/product release profile is unchanged. Fuzz artifact SHA256 c96e7f9de9716fa9322c533d6685bb40c4d2330e37d0ec9a9fa9236d3fdc1efa.71c97ac7fixes the additional publication-refusal hang: red 0/1, green 1/1; all SSH tests and strict SSH clippy/fmt pass. Final consistent-artifact builds and acceptance are running.D262 historical checkpoint archive evidence/addhost-instant/README.md, part 1/1; original SHA256 48f1e04050dc60401106b6c2f84a26aaf3d6d661a4962b216a48960b5050ae76, 6360 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
Full fast gate output and final disposition are recorded on #189. Static-size
caps, dependencies, release profile and owner fixtures are unchanged. Issue
remains open; no deployment or production branch update.
Source-head milestone gate (
5731f06d)Full scripts/verify fast exits1: fmt, VT fmt, workspace Clippy and cargo-deny
pass; test and static rows fail. Server327 pass/1 fail/39 ignored. The unchanged
mesh_calls_and_subscriptions_deliver_scoped_live_inventory assertion fails at
socket_tests.rs677, expecting a successful Result for call_id3. Exact-artifact
replay with matching target/state environment also fails. Earlier replay omitted
the target environment and closed its socket; that output is kept separately.
No timeout exemption or causal waiver is claimed.
Detached origin/dev5e383c3d passes that exact test, both before and after its
matching CLI build. This newer baseline differs substantially from the lane's
26ccb33abase; it does not establish which change causes the legacy failure.The failed predicate must be resolved before a green completion claim.
Browser first replay29/1: missing broken.lab failure heading at the original
15s assertion. Unchanged full replay30/30 passes, including Start fresh.
Tracer check and171 source tests pass. browser-failure-details.png is a real
optimized tracer build capture with the provisioning test double; no new UI,
visual verdict or physical-device claim. Historical capture restored.
The final branch-head gate is run again after this evidence-only commit and
quoted verbatim on #189. Its separate log is retained locally; the committed
milestone transcript remains reproducible source-head evidence.
D262 historical checkpoint archive evidence/addhost-instant/baseline-cli-build.log, part 1/1; original SHA256 100e0210727588e0fbe4f702ef19fd3c0c252ae11820943fe7bd5b6642e3d020, 2987 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/bridge-real-final.log, part 1/1; original SHA256 ad034b925e6a6564b8c20493f1e50911a1ba4500e0ff8db6569e1c5872a6e2c7, 95 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/bridge-real.log, part 1/1; original SHA256 ad034b925e6a6564b8c20493f1e50911a1ba4500e0ff8db6569e1c5872a6e2c7, 95 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/bridge-ssh-tests.log, part 1/1; original SHA256 e051c391b1a19d11ade392c225d608e237cb887fb312e620af4a3a84ae8e88dc, 17186 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
Historical production-build screenshot from checkpoint, before final integration: evidence/addhost-instant/browser-failure-details.png; 231266 bytes; SHA256 31e5cd0454d8d3aeade6ae9d85c905387767726cc856d065f7c4a468cd68c24c. Immutable archive:
. No author visual verdict. The current branch removes the archived evidence tree under D262.
D262 historical checkpoint archive evidence/addhost-instant/browser-final.log, part 1/1; original SHA256 3fde6647baa7243317e8fbb7f481ef8bd84b63ec3b87d5843516b8a5f6fccf70, 5391 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/browser-replay.log, part 1/1; original SHA256 8c0716106b3ef8f1d6773008c675a78d33704d7866bcc04c27edbf7a497bedb2, 6928 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/browser.log, part 1/1; original SHA256 62c4d861d2ebf87e67e96313f9f4e5224f59d5428f9b0bfe72b2ed1b0fe58e06, 1337 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/mesh-dev-baseline-with-cli.log, part 1/1; original SHA256 16cdf4acb6bbfd1df53a52eb4382cc36b460d3891fb71cf544144ce85d3b4f61, 3118 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/mesh-dev-baseline.log, part 1/1; original SHA256 c2abdbc1a105bd29d1a85dc3c582bb6069416e3e609d04c96121a06d6a909c1b, 12250 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/mesh-replay-env.log, part 1/1; original SHA256 229458ef5ccac8d0bc83b4815300c9301ae6bb422e9658ba103bec8319ce42c8, 721 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/mesh-replay.log, part 1/1; original SHA256 84763a3d4925be2b5ebc7c5b73e4e4f8de87cc6c0263651e34b0742aa5038dff, 640 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/primitives-clippy.log, part 1/1; original SHA256 5db5095ba1ea5565f9fa16564f10965b2dcae474cc0cf6e32857d3a143c4509f, 436 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/primitives-fmt.log, part 1/1; original SHA256 e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855, 0 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/primitives-test.log, part 1/1; original SHA256 e8e2e999309297a671e11ec591e9cfbf4dc75d2796607240efd03d056717cb21, 3117 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/receipt-clippy.log, part 1/1; original SHA256 e72f6833433ef9b3662afb9c3fa5a61a2dbef1e635ce9cd9a96cff4918ccda66, 3584 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/receipt-fmt.log, part 1/1; original SHA256 e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855, 0 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/receipt-green.log, part 1/1; original SHA256 9b3444c9dc0118abb2d8d5dbeb3a37419a8bf6214c8293e54a74cdbbb0a12cfb, 996 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/receipt-red.log, part 1/1; original SHA256 213ed1111859bd77e3a69996b39ef1ecd99db02977ccd15e51b761e375f6e044, 5122 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/reviewer-checklist.md, part 1/1; original SHA256 f92ded50500da111cf2b711e0a4ebba0a608ae97ec2a47374874308184328a79, 1799 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/sha256.txt, part 1/1; original SHA256 b0e3743b6cf58f3f1531e692d51f559d8507bc35ab593ddf90dea9aa6084cf4e, 2314 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/source-clippy.log, part 1/1; original SHA256 d93375b40ad17d6994177ccfd7126d184214aeb5dcc17b562727d8e6bb1557d3, 6557 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/source-green-final.log, part 1/1; original SHA256 907a3904c1ef436866ed9961237801ace2e6a3bfade59f61663c5d9cd39ddbe7, 1347 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/source-green.log, part 1/1; original SHA256 5c5d5a9715fab28d75ab15cc8178a7c362060fa3431cfdea28782397203ba536, 763 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/source-red.log, part 1/1; original SHA256 7621b68ffde1d2614ffa7ab6cdca07297c249698f13d3aac0169a83869e389e6, 7847 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/tracer-check.log, part 1/1; original SHA256 7e3a32d0a5836dd14dca0cd39671bb110f858de7bf7cc3fbfadae88324027b7a, 49 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/tracer-src.log, part 1/1; original SHA256 43072b0d7ee7fb8dc9ba98deea8c10a1e7766fe30dfc508dfe9fa8ba01ac7dfd, 16716 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast-module-order-red.log, part 1/1; original SHA256 a97673ea95b47886f0c5a506c22a0050059544af60804ff5c9bb8de09741a869, 3631 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 1/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 2/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 3/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 4/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 5/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 6/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 7/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 8/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 9/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
D262 historical checkpoint archive evidence/addhost-instant/verify-fast.log, part 10/10; original SHA256 9ea226dc6e5e15762a4f664d98c75bbde6ed15ea809f26814f04d00cb4125f0f, 218435 bytes. This is historical evidence, including then-failing gates, not final acceptance. Preserved on #189 before removing the branch evidence tree.
Acceptance milestone at
1195665a(production source71c97ac7; subsequent D262 docs-only removal): 20/20 real fresh targets pass, one original session/writer, 59 acknowledged input bytes before/after, zero live unused tokens/attempts, zero qdisc drops. VM208 4 vCPU; rootless manual user-key sign-in. Ten fresh targets per profile; p99 is nearest-rank maximum of ten.RTT 0 ms: prompt p50 373.501 ms / p99 414.674 ms; fully joined p50 639.770 ms / p99 697.955 ms; 1-minute load [0.0, 1.26].
RTT 150 ms: prompt p50 3562.458 ms / p99 3608.184 ms; fully joined p50 5343.458 ms / p99 5406.993 ms; 1-minute load [0.9, 1.31].
Targets remain unmet, especially the 150 ms prompt. First high-RTT sample: sign-in1264.083 ms; check/authority→Install1696 ms; published session3414 ms; working prompt3570.308 ms; fully joined5371.831 ms. The verified full payload is still on the critical path. Investigating a smaller bootstrap artifact built from the existing authenticated dispatcher and existing shim, with installed transports excluded until normal handover; this is a packaging/feature boundary measurement, not a different terminal/auth engine. No new foundational dependency or relaxed admission. Any additive contract change is recorded here first.
Static sizes, same recipe production9ab4329c vs branch:
No-growth remains unmet; caps unchanged. scripts/verify fast milestone exits1 only static-size caps: fmt, VT native fmt, clippy, test and cargo-deny PASS. Separate TS gate PASS (36 protocol tests /34462 assertions); prior exact tracer source production-browser functional41/41,4 capture-only skips. No completion or deployment claim.
1195665amilestone scripts/verify fast output, verbatim part 1/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 2/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 3/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 4/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 5/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 6/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 7/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 8/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.1195665amilestone scripts/verify fast output, verbatim part 9/9; exit1 only stale static-size caps. All other Rust rows PASS. RTT probe delegated unchanged to locked VM208; TRACER_TEST_PORT4997. Full output SHA256 6626f800b01c4f6481676a279356a1b74b56cbb311ae2669db189a7a00425c8f. This is a milestone, not acceptance completion: high-RTT prompt and no-growth targets remain unmet.Performance follow-up, source experiments not yet committed: existing CLI ssh-only profile is 3507760 raw /1535212 gzip bytes; dedicated entry point sharing the same shim arguments and SSH session module is 3307952 raw /1406022 gzip bytes. Both retain the real shim and same-session/writer pending-input proof against the full installed binary. Dedicated entry point rejected: only129190 gzip bytes saved; source was reverted and archived in job scratch. No engine/dependency change.
Current experiment shares the large authenticated dispatcher across channel implementations in crates/infraternal-agent/src/channel.rs and daemon.rs, retaining all original channel deadlines/retirement. Full x86 static20684944 bytes, saving72000 against1195665a but still351840 above production. Strict clippyPASS,67 remote testsPASS,39 ssh-session-only core testsPASS. Adds boxed I/O futures, so runtime cost must be measured before retention. The32-viewer×200-input existing VM harness initially refused the too-long job state path, then both baselines timed out on its assumed zsh prompt; the job-only harness now uses a real executed startup marker. Matched baseline/shared runs are queued under the shared VM208 lock with short state prefix/root/infraternal-bench/addhost-instant/s.
1195665ais pushed and its full milestone fast-gate output quoted verbatim above. Main D255 acceptance remains prompt150ms p50=3562.458ms /p99=3608.184ms, joined5343.458/5406.993ms. No target-met, no-growth, completion or deployment claim. Remaining nonblocking review details: initial80x24/second manager detection, connector address/timeout behavior (speed foundation), and actual per-person login implementation belongs to pending D240/D247; controller owner/off fencing and all high-priority uploaded-input/banner/reconnect findings are fixed.Experimental follow-up rejected and main source restored to1195665a. Sharing dispatcher polling saves72000B (x8620684944B; still+351840 vsproduction), but matched locked VM208 LAN16-viewer×200-input/3200-Ack runs show baseline medianp50/p99=.52/1.09ms and run=.185s; shared=.61/1.21ms and .216s. All three pairs preserve153600 delivered bytes and pass34ms. RSS baseline daemon5628–5632KiB, shared5632–5644KiB;4cores, load1m=.0405. The helper’s historical percentile call used unsorted samples; only the corrected sorted LAN results are evidence. Original32-viewer harness fails in both binaries (stream17 terminal shim disconnected), so32 capacity is not claimed. The executed startup barrier and paced attachment setup are job-only harness corrections; no repository gate/budget was changed. Performance priority rejects boxing for0.35% size saving.67 remote and39 ssh-only core tests passed the candidate, but it is not retained.
Packaging candidates: same-engine CLIssh-only3507760raw/1535212gzip; dedicated shared-entry3307952raw/1406022gzip, same-session/writer/pending-input proofPASS against installed full binary. No new engine/dependency selected; dedicated entry and optional feature experiment reverted. Existing BBR measurement candidate on one fresh150ms target: sign-in1282.076ms, prompt3591.120ms, joined5368.947ms, byte proofPASS; no improvement, not selected. All experimental source remains only in job scratch for audit.
Validated/pushed source1195665a unchanged; full20-target acceptance and exact production size deltas are recorded above. Prompt150ms and no-growth remain open. Re-running the exact head gate after restoring source; no deployment or issue closure.
1195665aaccepted functional measurement raw record acceptance-final/inputs.sha256; SHA256 963943bbd591cc86e35e8a109492a46d92f2677bf55eaec25dbac969b2468098. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-01.log; SHA256 9e17a67b0bd8f58a60abd8d88531ff221771239e6950be570560167f0c126ce2. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-02.log; SHA256 142123c4fee1cb5515b5ec8d05bc8ffb0b678a193b1a58fc26669e82c043d38a. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-03.log; SHA256 23274afe5f7d5eca4cd46644be62a3f13c3c8918838dc4433a28e5c4e49d06f1. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-04.log; SHA256 d4a708a99b8fcf1007051180d13b12dd59974d68764e48feb7a3a2489e452022. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-05.log; SHA256 d4727d3c1e98f13162a5b1ce0a39e08cd99c81733568b83f39fb9886299c6fe8. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-06.log; SHA256 695dceeb293bfee3ee721a57e1e4fe817a34d51aaac6cf0210276c7a46ad0a4a. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-07.log; SHA256 79c959c0131588c645cdb7e97b39b9f54c67a05af01134502266413a69cfacc6. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-08.log; SHA256 d17935bd6f5f960c5c1033685c7c4572c13dc30aa3a340dcac37d91e4d340599. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-09.log; SHA256 8406d70812eae91c6dd789bff66c246c451ad9e0bd50c55fd217e171723f3e68. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-0-10.log; SHA256 76a94d55a7fe75a57df34e1b3faec5a2fed79e79728422365f8ea80d2702a371. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-01.log; SHA256 575019ae4bb0d9c275dde143665a7ab86d75f33763acfa0542f87907a1cefc3d. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-02.log; SHA256 a7ff9113b18be5d6504ca4a675611915fed1a91c83d2e71c8a1f804e685d3d26. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-03.log; SHA256 019f50a27a49696c3710f701339cfe70c38a7417ff920d77dde398c88245124f. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-04.log; SHA256 eaa8c1069b17533319b332a6969f4cb0351d2b5039943ad981c97a49e9ad5c7d. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-05.log; SHA256 440580b3acf0943f60f2890ad2aed2e450d47fc898d5e343369c0edfc04baec6. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-06.log; SHA256 240c202985192583b83e04db21d8e8ac32593e289676c1cc30cc5eb3f5a27eb8. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-07.log; SHA256 fe0cce13eab93fa2312a184faff683c9655c29ac95d036fb7178096ac996c3fe. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-08.log; SHA256 04a2e5d617e26f1c722ff8e53a6f9bb059c342a8e6bf44d04c34fde49e20f5e1. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-09.log; SHA256 9e4b535745e41e6bf939e89b56990345825c9cca28cec1f8bfa45d888ce92e2c. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.1195665aaccepted functional measurement raw record acceptance-final/rtt-150-10.log; SHA256 084afcf4271aea3b3cdeb9c268310c5ee2b0f23400aa26052938e73d3d255d7e. Twenty fresh targets, locked VM208; timings remain above D255/D189 targets and do not establish completion.Resume point for remaining latency work: source restored to/pushed1195665a, no retained experimental source changes. At150ms the read-only check/public admission still adds≈432ms after key sign-in before Install, then verified full-payload staging/bridge publication1718ms and original attachment156ms in the first accepted sample. A concrete next RTT slice is to execute the existing allowlisted checked command as the foreground SSH master’s initial channel rather than -N followed by mux open/exec, preserving strict auth fallback, diagnostic bounds, credential erasure, and group ownership. This uses the existing OpenSSH engine; no new foundational selection is needed. Smaller same-engine feature payloads were measured above but are not integrated/chosen. Remaining acceptance stays open; final gate rerun is static-building, all earlier rows PASS.
Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 1/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 2/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 3/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 4/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 5/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 6/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 7/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 8/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Restored clean production source head
1195665ad6: scripts/verify fast output verbatim part 9/9, exit1 only static-size caps. Fmt, VT native fmt, clippy, test, cargo-deny PASS. RTT self-test unchanged on locked VM208; unique4997 browser port. Output SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Checkpoint remains unfinished: 150ms prompt and no-growth targets are open.Checkpoint result, not completion: addhost-instant head
1195665ad6, remote SHA verified. D255 integration/publication, retained background setup/retry, same-session handover and high-priority independent review fixes are committed atomically. No deployment/issue closure; owner fixtures untouched. Experimental source reverted; tracked tree clean.Acceptance VM208:10 fresh LAN +10 fresh150ms targets,20/20 original-session/writer and59-byte before/after confirmation proofs. Prompt p50/p99 LAN373.502/414.674ms;150ms3562.458/3608.184ms. Fully joined LAN639.770/697.955ms;150ms5343.458/5406.993ms. p99 nearest-rank=max of10; raw records/inputs hashes posted above. Targets remain unmet.
Production9ab4329c vshead, same stripped static recipe:
No-growth remains open: required guarded provisional transport, retained stage/retry API and lifecycle add code, with no new dependency or engine. The sharing experiment reduced72000B but regressed median LAN probe latency/run time, so was rejected under performance priority. No assertion that this growth is unavoidable or accepted.
Final scripts/verify fast output is quoted verbatim above in9parts, SHA256 d740fb261efe05600931b8adb1f6f56793feff9b7c6c0f7aa48bed78ee823c4b. Exit1 only historical static caps; other rows PASS. Separate TS static/bundle/protocol gatePASS, production-build browser41passed/4capture-only skipped, parser ASan100000 VM inputsPASS. Author checklist: UID/current-owner/host/certificate fences, exact-session resume, replay ACK fallback retention, failed replacement restoration, SSH-off retirement, bounded private stdin/diagnostics, parser refusal bounds and no duplicate UI/foundation were checked against real oracles. D240/D247 per-person login remains future-compatible by contract, not implemented here. Screenshots supplied for independent review; no author visual verdict.
Resume concrete next slice: initial checked command on the foreground SSH master to remove the remaining check/mux round trips; strict allowlist/fallback/helper lifecycle must remain. Build targets and scratch retained because the job is unfinished.
Final gate table verbatim:
Starting D255 round 3 (latency + size) on
addhost-instant, base1195665ad6(engineer: Claude). Read all #189 comments, D22/D116/D240/D243/D255/D256/D262/D263, PROCESS and job rules.Plan, in order:
9ab4329c(+423,840 B x86_64, +375,248 B aarch64), bytes per cut.Previous engineer's
.job-scratch/moved out of the worktree to$HOME/.cache/codex-scratch/addhost-instant/prev-job-scratch/(nothing deleted). Owner fixtures untouched; no deploys.Finding: round-trip accounting at
1195665a, 150 ms RTT (controller-side packet trace on VM208 docker0, fresh target, one sample; prompt 3,575 ms, sign-in 1,272 ms). Times are ms after the TCP SYN; the harness clock starts ~181 ms earlier.none+ publickey probe + signature (OpenSSH client)STEP upload+TRANSFER zstdbefore any payload byteSTEP start, shim launch,SESSIONreturnsAttachTerm→ prompt bytesSo after sign-in, 2,303 ms = 4 RTT of request/response + ~8 RTT of payload ramp + ~300 ms of local and remote processing.
Separate finding about the harness: in the 150 ms profile the first SYN waits for ARP, and netem delays the ARP reply by one RTT, because the fixture container is fresh (
base-1195665a-rtt150-arp-head.txt). Sokey_sign_in, and through ittime_to_prompt, include an extra ~150 ms that a routed WAN host would not pay. I am keeping the harness unchanged so the numbers stay comparable, and I also report time after sign-in.LAN (same build): sign-in → prompt 301 ms. Payload 90 ms; remote after the payload 124 ms (verify, publish, shim launch 57 ms); check 23 ms; upload guard + choice 27 ms.
Evidence:
base-1195665a-rtt150-timeline.txt
base-1195665a-lan-timeline.txt
base-1195665a-rtt150-run.log
base-1195665a-rtt150-arp-head.txt
Next, in order:
dd bs=1(exact, ~3–9 ms on the VM for 4–11 KB), so a bufferingheadcannot swallow payload bytes.Step done: latency commits pushed on
addhost-instant, base1195665a, headbc0761e6.8e99ea0eCAPS <zstd|-> <gzip|-> <cache sha256|->and which setup reader it uses (dd bs=1, exact, orhead). With an exact reader on a rootless install, the server declares the encoding and sends setup + payload without waiting forTRANSFER. Receipts are read concurrently: a refusal stops the writer, and a mismatched selection is refused. Without dd, without CAPS, or in root mode, the old ordering stays.7fc476cbfa299a62gzip/zstd -dc | tee payload | sha256sumplus an exact-size check, so the target does not re-read 11 MB after the last byte. A real-shell test refuses a truncated file and a corrupted stream before launch.7f245dce4d453b87bc0761e6VM208, one or two fresh targets per point (not acceptance):
1195665a8e99ea0e(declared transfer only)bc0761e6Gates per commit: fmt, strict clippy (ssh, server+remote, agent+remote), all
infraternal-plugin-sshtests (72 install tests), serverbootstrap(46) andremote_runtime(19), agent (67), andverify-ssh-session.py+verify-ssh-session-rebridge.pyall pass.Findings from the new traces:
SESSION→ attach, and prompt. The first 100 KB of payload take 407 ms and the first 300 KB take 639 ms (BBR startup from an idle connection). The last 3.7 MB take about 600 ms. Payload size therefore matters only logarithmically: zstd -19 (−9.6%) would save about 20 ms.systemctlprobe 6.5 ms,systemd-run --scopeabout 8 ms, shim and shell start). sshd's own post-auth work delays the channel confirm by about 20 ms at LAN; that is not ours.Next: binary-size cuts (D243), then the remaining latency options, then 10+10 acceptance.
Resuming D255 evening continuation on addhost-instant, base/head
f0d6a17f92(WIP checkpoint), prior reviewed sourcebc0761e6. Read required repository docs and #189/#169/#190 histories; continuing existing safety fixes, checkpoint size review, prompt-path optimization, then locked VM208 10+10 byte-integrity acceptance and final gates. Push authorized; no deployment or owner-fixture changes.Evening checkpoint review: fmt and strict SSH/CLI Clippy PASS; private helper parser tests 2/2 PASS. First full SSH run failed negotiated_retry_reuses_only_a_verified_cache_after_activation_refusal at tests/install.rs:439 (rollback status73); unchanged isolated replay passes. Cause unresolved; full serial-limited replay running. Reusing reviewed #198 native archive section cut (
722af2a3), caps unchanged. Its action-metadata constructor cut (5ae7cbf2) is inapplicable: this older branch has no action.rs/JsonSchema implementation, so no new framework is introduced. Evaluating existing Socket dispatcher reuse through a bounded temporary SSH pipe relay; acceptance must compare locked VM208 prompt/integrity before keeping it.D255 additive packaging/staging proposal before implementation: ship an optional architecture-matched session/SSH-only build of the same infraternal executable beside the existing full node payload (same shim, dispatcher, authenticated rendezvous and session UUID; no new engine/dependency). Trusted server paths only; absent optional artifact retains full-payload behavior. Upload/verify this build before the prompt, then, after protected publication, upload/verify the full node artifact over another command channel on the SAME authenticated master. Full upload uses existing ownership/transaction receipts and private invitation barrier; it writes a separate temporary inode and atomically replaces staged payload, never truncating the executable backing the running shim/bridge. Normal installation/enrollment and admitted reattach proceed afterwards. Retry preserves the original shim. Full and session artifacts have distinct cache identities and bounded warm-up entries. D240/D247 login invariants remain unchanged. This implements the existing D198 removable-feature packaging boundary, not a foundational engine choice.
Completed source steps:
93333258additive session-payload contract;d9491b31four-entry digest index;399e4340eight-entry codec index;2f14deffarchitecture-scoped optional artifact selection;53362222reviewed #198 native archive section cut;f18ca7c6real-file/process restaging tests. Strict SSH/CLI Clippy PASS; complete SSH suite PASS including installation74/74, path refusal and uninterrupted restaging; shim73 passed/13 ignored plus integration/doc suites PASS. Real Socket relay oracle and explicit rebridge census both PASS (same session/writer, pending input exactly once, output resumes).Findings retained: a second installation-only replay under heavier host contention failed72/74 (buffered-head result failed despite all21 invitation bytes arriving; cache retry returned Upload failure), while earlier full74/74 passed; not waived, reproductions pending. Browser run on49189 was stopped by its own runner PID after repeated boot/overall deadline failures;3 passed, no assertions changed, incomplete run retained. Session-only static build exposed the shared authenticated dispatcher being gated by remote. Factoring its existing UID/writer/cancellation boundary into one common entry, with a real-socket no-remote UID refusal regression. No new engine or dependency; no deployment.
Evening VM208 exploratory samples (not10+10 acceptance): session/full builds give LAN prompt228.631ms and joined727.784ms;150ms prompt3003.699ms, sign-in1263.110ms, joined6333.989ms; original session/writer and59 acknowledged input bytes preserved. Remaining after-sign-in≈1741ms includes two check/channel round trips, startup pacing of the1.3MB session upload, shim launch and attachment. Per-socket CUBIC candidate is being measured using the existing owned-fixture connector; no global setting or product congestion choice changed.
Size experiments: current full x86_6420477616B vsproduction20333104B (+144512); sharing the two bounded bootstrap WebSocket upgrade task types gives20467376B (−10240), all14 API oracles and strict Clippy PASS. Boxing cold setup saves96B, rejected; extracting the shim parser grows7104B, rejected and source restored. Measuring dead C-function/data section removal in existing vendored libraries; no provider, engine, optimization level or cap change.
Browser on VM208, frozen production dist + unchanged spec deadlines, port49189:30/31 pass; known broken.lab failure-heading case failed15s, unchanged isolated replay1/1 passes. Default host run remains failed/incomplete and retained. Focused negotiated-transfer functional replay under its own CPUWeight10000 scope passes2/2.
Source committed through
f0197700(release/test profiles); previous stepsf5f1324cprivate SSH-only entry,fc7ec012shared Socket bridge,94bde942cross-build real oracle,d388bd55shared admitted WebSocket upgrade, cceb77c6/f0197700 complete five-artifact release. All referenced targeted Clippy/fmt and functional gates passed; release oracle also refuses a bundle missing only the ARM session profile.Further size candidates rejected: vendored-C function/data section flags save0B (already effectively discarded); dispatcher constructor no-inline grows1216B; restored source. Full x86 source before those experiments20467376B, still+134272B vsproduction20333104B. I do not claim remaining growth is unavoidable or accepted. Default kernel congestion retained: three-per-profile owned-connector samples give after-sign-in median LAN159.754ms/default vs276.610ms/CUBIC,150ms1851.313ms/default vs1664.081ms/CUBIC; LAN regression rejects the socket change. No engine/library or global kernel change.
Freezing final controller/node/session artifacts, then10+10 fresh-target byte-proof acceptance and exact-head scripts/verify fast. First-prompt≈1.5s at150ms remains unproven; earlier session-first sample≈1.741s after sign-in. No deploy or owner-fixture change. Issue-attach helper is absent from PATH/worktree; requested its location while continuing, no credential investigation.
Pushed atomic source steps through
f0197700to origin/addhost-instant; no deployment. Final gate fmt/native fmt/strict workspace Clippy PASS, workspace tests running; independent TS fast gate PASS (final-ts.log). Final controller/node static build still linking; acceptance script requires20 fresh targets, the exact59-byte same-session/same-writer proof, and zero live enrollment tokens/attempts in every sample.Round-3 functional replay initially refused terminal creation because the chosen scratch directory made the AF_UNIX rendezvous pathname exceed108 bytes (round3-final-functional.log); rerunning unchanged oracle with the existing short cache root ahi3, without touching owner fixtures. Full31-case production-build add-host browser replay is running on VM208 with original deadlines and49189.
Functional milestones at
f0197700: frozen production-build add-host browser suite on VM208/49189 passes31/31 together (browser-vm-final.log, unchanged deadlines; includes Start fresh, root consent, cancellation, phone footer, retained setup). Round-3 real daemon restart oracle passes24/24 stop/restart iterations and exact4,710,396-byte input/output replay; no performance-counter calls (round3-final-functional-short.log). Long AF_UNIX scratch-path harness refusal is corrected by using existing ahi3 cache root, not changing production or assertions.Final artifacts shipped/sha verified: controller f2e4a1036271576edabc1df870f6d619cbc334fce7aa117d781cfd4c7bfebb14; full node5c9ababed3fc05cafe39ad189fd875fb99113f6500bf5380b33f0b904acbf3fc; session0adb61b90192e68b330061806ec40318177ffb53583d7a3f1a44d47908f35b3b. Twenty fresh-target acceptance now follows browser on the shared lock. Final Rust/static gate still running; no acceptance claim yet.
Author checklist at
f0197700: session-only/full builds share the exact UID/native-stream admission and rendezvous dispatcher; wrong-UID no-remote real-socket regression passes. Optional payload selection refuses symlinks/non-files and otherwise falls back only on absence. Full restaging uses the tested session-preserving rollback path (owned journal, no running-image truncation); original session/writer/replay cursors remain protected through cancellation, failed activation, retry and handover ACK fallback. Separate setup/bootstrap WebSocket admission retains original owner authentication, component activity/semaphore guards and32K/4K limits. No new engine/dependency, duplicate UI, prices, deployment or owner-fixture changes. Relay storage remains two16KiB buffers per temporary bridge (256KiB across8 active attempts); disappears when SSH ends. D240/D247 real-login integration is documented as the same per-person launcher for both builds, not falsely claimed implemented here.Final scripts/verify fast at
f019770069, verbatim part 1/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 2/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 3/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 4/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 5/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 6/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 7/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 8/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Final scripts/verify fast at
f019770069, verbatim part 9/9. Concatenate fenced contents to reconstruct original log; SHA256 0edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff. Exit1: unchanged historical static caps; all other Rust rows PASS. Separate TS gatePASS. D243 production comparison still fails (+134272B x86,+98240B ARM), not waived. Issue-attach helper unavailable; preserving required raw output inline.Acceptance measured at pushed head
f019770069; result, not completion. VM208 (infraternal-perf-test,4vCPU,8262770688B physical memory), owned fresh targets, lock released per target. Same-session/same-writer59 acknowledged input bytes plus output-before/after proof and zero live tokens/attempts:20/20. No assertions, RTT profiles, kernel congestion or budgets changed. p50=median;p99=nearest-rank maximum of10 samples.f0197700Both PT_INTERP absent/static, caps unchanged. The WAN prompt is~315ms above the approximate1.5s median target; size growth remains unaccepted. Final scripts/verify fast at this head is quoted verbatim in9preceding parts (214901B,SHA2560edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff): exit1 only static caps; fmt/native fmt/strict Clippy/workspace tests/cargo-deny PASS. TS fast gatePASS; browser31/31; round3exact4710396B replayPASS. No deployment or owner-fixture modification.
Raw acceptance/environment/artifact SHA files: /root/infraternal-bench/addhost-instant/r3/evening-final/acceptance onVM208; fetched archive final-acceptance-evidence.tar.gz in this job scratch. D262 issue-attach helper unavailable; no credentials searched. Continuing size work: shared formatter no-inline saves2112B only, restored; testing a bounded failure-only fingerprint scan future boundary. No new engine/dependency.
VM observed load ranges (1m,5m,15m): [(0.45, 3.88), (0.75, 2.7), (0.57, 1.7)]
Size finding against frozen
f0197700x86: boxed failure-only host-key scan builds20467984B (+608B); strict SSH Clippy PASS, rejected and restored. Formatter no-inline candidate20465264B (−2112B), also restored pending a larger justified reduction. Both measured under unchanged musl/release recipe, no cap edits, dependencies or engine changes.Now testing one common cold boundary for the existing Failure constructor and classification routines. Only compiler attributes/comments; no message, auth/trust, stage, refusal parser or cleanup semantic change. Success traffic does not call these failure routines. Acceptance remains the20-target
f0197700results/table above, not evidence for an uncommitted candidate.Verified checkpoint result: branch addhost-instant, head/remote
f019770069; tracked worktree clean; atomic steps pushed; no deployment, issue closure or owner-fixture modification. D255/D243 targets remain unmet; no unavoidability or waiver claim.Prompt total p50/p99: LAN332.408/408.406ms;150ms3122.818/3208.921ms. p99 is maximum of10 observations. All20 retain same session/writer,59 acknowledged input bytes, output before/after, and zero live tokens/attempts. VM load1m0.45–3.88 (4vCPU); exact environment and artifacts in the acceptance archive. Browser production build31/31 PASS; round3functional24/24 exact4,710,396-byte restart replayPASS; separate TS fastPASS.
Final scripts/verify fast (verbatim summary below; full214901B output quoted verbatim in9 preceding parts, SHA2560edecf6b78e8c978f180e860d2d5e31179fdddec3e1f10daabad666e7e3882ff). Historical caps unchanged; static-link/allowed-provider checks pass, static size caps fail. The separate production comparison also fails; it is not excused by stale caps.
Last source experiments restored: formatter boundary−2112B; boxed failure-only fingerprint scan+608B; shared cold refusal construction/classification+5696B. No rejected candidate remains in tracked source. Frozen final full binaries remain under $HOME/.cache/codex-scratch/addhost-instant/art/evening-final. Raw logs/backups/archive now under /home/kayg/.cache/codex-scratch/addhost-instant/evidence/2026-10-13-f0197700-final; acceptance archive SHA2567af9d09257d0e858fd1c930a325999d31166225135ae0db71c7a63145aebf071. VM raw files: /root/infraternal-bench/addhost-instant/r3/evening-final/acceptance. D262 issue-attach helper unavailable; required gate output preserved inline, no credential investigation.
Remaining work is WAN median≈315ms above1.5s and binary growth above production. Initial-channel investigation (not implemented): master.rs currently owns a foreground -N master with null stdin/stdout; command channels own separate child groups. A primary read-only check could overlap a round trip, but must retain the master through terminal proof/commit and isolate trusted local method diagnostics from remote command stderr. auth_evidence.rs deliberately keeps the last full receipt; sharing that pipe with command stderr would allow false method metadata. Preserve every shadow-review fence; do not use background ControlPersist as a shortcut. This issue contains the artifacts, measurements and gate output required to resume independently of this chat.
Continuing addhost-instant from clean head/base
f019770069. Read #189/#169 history and required worktree docs, origin/dev D255/D262. Current acceptance: WAN prompt after sign-in 1815 ms p50; joined 6680 ms p50; production size deltas +134272/+98240 B. Investigating joined regression and prompt round trips, preserving all shadow-review fences. Fresh 10+10 VM208 acceptance, byte integrity, production size comparison, iterate/final fast gates and evidence attachments remain required. Push authorized; no deployment.D262 raw evidence from the prior
f0197700run is now attached:Finding: 150-1.log reports four background command steps after the prompt: rollback 400.505 ms, payload_lookup 353.017 ms, upload 963.372 ms, setup 1192.319 ms. install.rs restart_setup uses stage + setup, bypassing the existing stage_and_setup compound transfer. Testing one compound restage that retains the exact journal cleanup before upload. This removes command round trips without concurrent payload traffic before the prompt or changing terminal ownership/proof fences.
Completed
31304db2: restore scripts/verify iterate on this older lane. Shares all existing source gates; only fast performs static cross-build/link/key-memory/size checks. No caps changed. bash -n passes; dispatch smoke retains fmt, VT fmt, clippy, tests, cargo-deny for iterate, plus static builds for fast. Full iterate run in progress.Restage regression red: 4 commands instead of 1. Green targeted safety 3/3: one compound restage, live bridge through full upload/commit, foreign journal refusal before upload. Earlier 75-test install suite green; final full SSH suite and strict Clippy rerun include the additional safety test. VM release controller build in progress; no new timing claim.
Completed
fec9ca11(compound session-preserving restage) andbf59777a(regressions). Full SSH suite passes, installation 76/76, strict SSH Clippy/fmt pass. Regression initially failed 4 commands versus 1; final one-command assertion, real-file foreign-journal refusal before transfer, and original bridge interaction before/after full restage/commit all pass. No shadow-review check removed. VM controller compilation continues; no timing claim yet. Restage uses one parent directory guard, cleanup in a fresh errexit child inheriting that guard, then the existing canonical negotiated transfer/private-invitation barrier.Restage code/test evidence for fec9ca11/bf59777a. Full SSH tests green, install 76/76; strict Clippy/fmt green. Foreign-journal refusal leaves journal/payload intact before enrollment. Real command bridge interactive before/after full restage and commit. One-command regression failed four versus one before fix.
restage-red.log
ssh-restage-final.log
restage-safety.log
restage-clippy-final.log
VM208 exploratory probe of
bf59777anative controller (SHA256 4d5499781535a60c433a0a017967965ec3597b133877cc3fb013da63e4635622) with frozenf0197700node/session payloads: prompt total 3022.121 ms, SSH sign-in 1265.039 ms, prompt after sign-in 1757.082 ms; fully joined 5409.467 ms. One compound upload_setup 1695.979 ms instead of prior rollback+lookup+upload+setup 2909.213 ms. Same session/writer, 59 acknowledged input bytes, output before/after, zero live tokens/attempts PASS. Single sample only, not acceptance; shared VM lock, 150 ms RTT, load in raw log.Next optimization proposal: preserve the check's rootless public zstd/gzip capabilities through immediate full restage. Declare the encoder so public bytes can follow the compound script without waiting for TRANSFER; require the exact matching receipt and keep PAYLOAD ready before all private invitation bytes. Unknown capabilities, root setup and retries retain negotiation. No codec/dependency/engine/trust choice changes. This removes one remaining 150 ms transfer-selection round trip. Keep every cleanup/proof fence.
restage-vm-probe.log
Completed
96a34ff3shared early/enrollment prelude;e6503d78additive input slot; e6936384/e7b7ec01 declared public streaming/private barriers and regressions; 522dff2e/dbdf186d checked restage encoder + bootstrap integration;7d79d064compressed real-file safety;21e9744dcontract. Each code slice passed fmt, strict SSH Clippy and touched tests. Full SSH suite green (install76/76; process10/10), targeted compressed original-bridge/foreign-journal safety3/3. Red-before-fix logs show payload stalled and mismatched choice could release synthetic private input when the newly added hint was still unimplemented. Final source requires exact choice and verified receipt before all private input.Completed scripts/verify iterate from
bf59777anative artifacts (SSH regression test source present at compile), RTT probe forwarded to VM208; Rust/TS source gates PASS. Its transcript is attached, not evidence for later native changes. Current optimized controller build continues. WAN single probe of earlier compound restage: joined5409.467ms versus previous final median6680.343ms; no acceptance claim yet.declared-red.log
declared-mismatch-red.log
declared-final-process.log
hinted-restage-final-safety.log
hinted-restage-clippy5.log
iterate-restage-final.log
Size experiment underway on top of pushed21e9744d: share setup-only poll boundaries for Installation transfer/stage/setup/session/restage and the process helper lifecycle. These operations are currently embedded in multiple initial/retry/authentication callers; boxed boundaries keep one generated poll function per operation. No parser, transport, engine, dependency, release flags, receipt or cleanup change. Terminal steady-state code is untouched. Full SSH tests and process10/10 pass; strict Clippy being rerun after explicit-lifetime lint repair. Stripped x86 static size build queued under unchanged recipe. Candidate will be retained only after size and VM comparison; no reduction claimed yet.
Further findings:
21e9744dcontroller VM208 WAN single probe: prompt total3010.731ms, sign-in1264.474ms => after-sign-in1746.257ms; fully joined5369.704ms, same-session/writer byte proof59B and zero ledgerPASS. Shared prelude/declared restage alone do not close prompt or joined targets. Full boxed-poll x86 experiment20474320B (+141216B vs production); not retained yet, exact unboxed21e comparison building. Functional SSH/Clippy tests pass but this is not an accepted size reduction.Next internal build optimization: optional session-payload Cargo feature (implies existing ssh), used only by the existing optional release session artifact, will link the fixed private SSH/session/shim entrypoints without the unused public operator CLI graph. Extract the existing shim Args schema/run implementation once and reuse it in full and session builds; no new parser/library/engine, no alternate terminal dispatcher or shim. Full/default public CLI stays the same. Existing same-shim cross-build and byte-integrity oracles will exercise the reduced image. This is an extension of the approved optional session image, not a new role or transport. Size/time measured before retention; D240/D247 identity/launcher seam unchanged. Contract proposal before wiring the release recipe.
hinted-prelude-vm-probe.log
boxed-full-size-x86.log
Measured private-profile candidates under unchanged release/musl recipe: original session3210992B; same-Clap entry3165328B (−45664B); fixed private invocation/shared Options runner2533392B (−677600B,21.1%). Full/private CLI strict Clippy PASS; full unit42 passed/3 ignored, including four launcher shapes compared against the full Clap schema and malformed/overflow/changed-order refusals. New fuzz target imports the real decoder, never starts a process; strict fuzz Clippy passes with the server SSH dependency enabled (its legacy remote-only build has unused bootstrap warnings). Smoke build in progress.
The fixed decoder is a bounded internal argv decoder added to the existing private startup interface, not a replacement protocol/terminal/CLI engine: public operator parsing still uses Clap, and UID/session checks and the shim runner are shared. Feature-removal profile skips inherited-fd/public operator parsing; installed full build supplies those modes. Same-session cross-build oracle remains required.
WAN exploratory private-image probe: prompt total2951.037ms; fully joined5337.332ms; all byte/ledger proofsPASS. Not acceptance; remaining prompt/join delay still needs reduction.
Exact boxed/unboxed21e9744d full x86 comparison: boxed20474320B; unboxed20473968B; boxed is +352B. Rejected/restored, no boxed candidate remains in tracked source. Unboxed production delta+140864B; no size waiver claimed.
thin-session-vm-probe.log
fixed-entry-build-valid.log
fixed-entry-final-tests.log
fixed-entry-final-session-clippy.log
unboxed-full-size-x86.log
Joined-delay finding: bootstrap.complete_setup currently serializes connected_facts (~356ms) then terminal_canary (~325ms), while the commit channel waits for both. Each RemoteBackend callback independently calls actor(), which requires current installed authority, owner capability and matching execution UID before RPC. Proposal: after an early terminal was published, run those two independent installed-link proofs together, still requiring both before the one-use commit signal; legacy adapters remain sequential. On either failure, finish only the registered private canary under the existing cleanup callback (bounded5s), preserving the user shell and withholding commit. Add red/green concurrency and no-commit/canary-cleanup tests. No proof, marker/status-zero exit, authority check, handover or replay fence removed. This overlaps roughly one or two WAN round trips without any traffic change before the prompt.
Completed/pending review heads:
ac5aa11aprivate session feature;644557a9shared shim schema/runner;eb85cadbreuse existing Clap dependency only in fuzz package;7f2c9a25fixed private release entrypoint;4f87abaereal-source decoder fuzz target. Full/private strict CLI Clippy and fmtPASS, full unit42/3ignored, private unit7/0. Real cross-build oraclePASS (same session/writer, pending input executes once, output resumes); explicit rebridgePASS (one original shell). Session static image2533392B vs3210992B,−677600B. Public full CLI retains existing parser and all-component builds retain operator modes. Release recipe wiring and ARM profile measurement remain pending.Parallel-proof regressions valid red2/2 (canary cannot begin while facts held; facts failure omitted private-probe cleanup). Green full SSH suite and strict ClippyPASS with both proofs required before commit and probe-only cleanup on failure. A further representation cut moves the same guard into setup_guard.sh so readonly checks no longer ship unused lease routines; gates in progress, no shadow ownership logic changed. Final controller rebuild/acceptance needed after these steps.
fixed-entry-cross-oracle.log
fixed-entry-rebridge.log
fixed-entry-session-tests.log
proof-red-valid.log
proof-final-green.log
The real shim argv decoder now passed instrumented nightly/libFuzzer smoke:10000 runs with five valid/overflow seeds and coverage enabled, no crash. The harness does not execute startup or touch state; this is functional parser evidence, not build-host performance evidence. Cross-build pending-input/same-session oracle and explicit rebridge also PASS. Parallel proofs + guard-only readonly check are committed at
d4afb515(32b62b72); native controller rebuilt from this source and WAN probe queued. Full x86/ARM production-size comparison build in progress, caps untouched. Release private-image recipe/doc wiring remains the next atomic step.Measured additional full-CLI shim dispatch factor:20478288B x86 versus20474608B before (+3680B). Strict Clippy/fmt and43/3ignored unit tests passed, but the size direction is wrong; rejected/restored. No candidate dispatch change remains. Current private session profile stays2533392B, with42/3ignored full tests and7/0 private tests plus cross-build byte/rebridge proof.
Author checklist at2b779884 (SSH/bootstrap/native integration):
-20 fresh VM208 targets,10 LAN/10 at150ms, all same-session/same-writer59B and output-before/after proofs and zero live token/attempt ledgers passed. FullWAN targets and production size deltas remain unmet; not waived. All accepted performance results from locked VM208, machine/load recorded per raw sample.
Acceptance result at source-equivalent2b779884 (native controller built
d4afb515; later commits only release recipe/docs). VM208,10 fresh targets each profile, lock released between targets. All20 retain same session/writer,59 acknowledged input bytes, output before/after, and zero live tokens/attempts. p50 median; p99 maximum of10.WAN joined p50 improved6680.343→5447.441ms (−1232.902ms); prompt after-sign-in1814.739→1757.436ms (−57.303ms). Neither WAN target is met; no acceptance/completion claim. Session image x863210992→2533392B and ARM2798544→2233584B.
Full stripped static comparison under unchanged recipe (PT_INTERP absent): x86 production20333104B/current20474608B/delta+141504B; ARM production18112240B/current18215584B/delta+103344B. Growth remains unaccepted; no unavoidability/waiver claim. Added D255 bridge/retained setup/adoption/proof code remains, with four size attempts rejected/restored (current round boxed+352B; shim dispatch+3680B). No shipping dependency/profile/cap changed.
Real round3 restart replay24/24 exact4710396B PASS; release checksums and complete/partial-session refusal PASS; instrumented parser fuzz10000 seeded runs PASS. Final gate and browser31/31 still running. No deployment/owner fixture change.
VM208,4vCPU; load ranges1m/5m/15m=[(0.9, 2.14), (0.67, 1.22), (0.64, 0.85)]
final-acceptance-raw.tar.gz
final-acceptance-summary.txt
final-size-table.txt
round3-final-functional.log
package-final-check.log
Final head2b77988400d045f4eba4a80968b9c97b29cfe78c is pushed to origin/addhost-instant. scripts/verify fast completed exit1: fmt/native fmt/clippy/test/cargo-deny PASS; static builds FAIL only existing cap checks (x8620474608B, ARM18215584B). No caps changed. Same-recipe production comparison still positive +141504/+103344B; this is independently unmet D243, not dismissed as stale caps. Entire unmodified gate output attached below.
Separate scripts/verify iterate exits0, including TypeScript/source gates. VM208 production browser addhost-instant+wizard31/31 PASS on owned port49327. Real-source instrumented shim argv fuzz10000 seeded cases PASS. Reviewer checklist attached. Exact-head full host payload acceptance re-run underway; initial20 proofs all passed but WAN targets unmet.
verify-fast-final.log
iterate-current.log
browser-final-vm.log
spawn-fuzz-instrumented-smoke.log
review-checklist.txt
Handoff for break-it review, not acceptance: pushed head2b77988400d045f4eba4a80968b9c97b29cfe78c (origin/addhost-instant matches). D255 prompt/byte-safe handover implemented; this continuation compounds deferred restage, predeclares only public transfer bytes, shares guarded setup, shrinks private session image and overlaps independent installation proofs. All prior trust/UID/journal/commit/replay safety fences retained. No deployment.
Exact-head native artifact set acceptance, VM208 4vCPU,10 fresh targets each LAN/150ms, lock released per target. p50 median, p99 nearest-rank maximum of10; prompt after sign-in subtracts that sample's key_sign_in.
20/20 same-session/same-writer proofs,59 acknowledged input bytes each, input and output before/after, zero live token/attempt ledgers. New payload hash a5b18967574a6e48ac46277f3419f73435b3e03addfa2f9eeb1c4ef49d088c69; prior run used frozen runtime-equivalent full payload and is retained separately, not conflated. Load ranges1m/5m/15m=0.01–0.91/0.38–0.60/0.53–0.60; differences between runs are not claimed as causal gains. WAN targets remain unmet (~1.5s prompt after sign-in; joined<5s).
Full static sizes unchanged: x8620474608B versus same-recipe production9ab4329c20333104B (+141504B); ARM18215584B versus18112240B (+103344B). D243 remains unmet, not waived or claimed unavoidable. Private session images2533392/2233584B (−677600/−564960B). No cap/dependency/link-profile change in shipping binaries.
Functional evidence: browser31/31 production-build VM replay; round3 exact4710396B at24 restart points; full SSH installation76 cases plus concurrent proof regression tests; shared full/private CLI schemas and cross-build adoption/rebridge;10000 seeded instrumented real decoder fuzz executions; release checksums/complete-or-refuse bundle. scripts/verify iterate exits0 incl TS. Author security checklist attached earlier.
Final scripts/verify fast exit1, entire raw transcript attached verbatim. Final table verbatim:
acceptance-head-raw.tar.gz
acceptance-head-summary.txt
acceptance-head-console.log
verify-fast-final.log
iterate-current.log
browser-final-vm.log
spawn-fuzz-instrumented-smoke.log
review-checklist.txt
Cleanup at unchanged pushed head2b779884: worktree clean; owned addhost-instant-target fixture removed after20/20 acceptance; owned42G Rust target and2.8G instrumented-fuzz target deleted. Raw evidence retained under /home/kayg/.cache/codex-scratch/addhost-instant/evidence-2026-10-14-2b779884 for resume. SHA attachment retried as .txt after Forgejo rejected .sha256 extension (HTTP422); same bytes and hashes also inside acceptance archive.
final-head-artifacts-sha256.txt
Stopped for requested break-it review; WAN latency and D243 deltas remain unmet, no acceptance/completion claim.
Independent verification starts on review-addhost-instant at target
2b77988400; merge base with local dev is93b00f5ab9. Read #189 history/report and parent #169; tracing #189 changes in the broader merge-base diff. Review-only: failing tests/evidence, no implementation fixes, push, deployment, owner fixtures or perf VM. Full scripts/verify fast is running on the unmodified target with owned scratch and TRACER_TEST_PORT=4879. Performance remeasurement is excluded by the explicit review brief; author reports unmet WAN and growth targets.P1 R189-02 confirmed: early publication is treated as fully ready by the product. Drive instant-fail.lab through the actual wizard: it closes and shows “instant-fail.lab is ready,” but sends zero host/window setup subscription frames; later Join:failed is invisible and has no product retry path. D255/contracts/host-setup.md require remaining progress and failure recovery in the machine pill. Existing raw-socket control passes; new product regression fails expected true/received false. Test: apps/tracer/test/addhost-instant-review.spec.ts, “review published host window subscribes to setup after the wizard closes”. Committed locally at
8b94495cf3on review-addhost-instant; no implementation change/push. Production-build screenshot: apps/tracer/test-results/addhost-instant-review-after-publication.png; log: /home/kayg/.cache/codex-scratch/addhost-instant-review/browser.log. Functional test double, no timing/physical-iPhone claim.Completed independent cleanup check: P2 R189-01 at bootstrap.rs:1040. After facts and private-canary callbacks start, stop the commit channel before proof. The outer try_join drops its proof sibling and never calls cleanup_sessions; no commit occurs. New ignored regression review_stopped_commit_channel_finishes_the_private_probe fails 0 PASS/1 FAIL, cleaned=false/committed=false, exit101. Committed locally as
4c1de592f9; log /home/kayg/.cache/codex-scratch/addhost-instant-review/r189-01.log. Existing 10-second private-probe ceiling bounds impact; no user-shell removal or unbounded process claim. Contract requires explicit probe cleanup on all exits. Total confirmed so far P1=1/P2=1. Clean target-head gate continues under umask022 with owned compact fixtures; no implementation changes/push/perf VM.Review checkpoint at local head
4c1de592f9: P1 R189-02 reproduces a second time; stable production-build capture /home/kayg/.cache/codex-scratch/addhost-instant-review/r189-02.png (ready toast, no setup state). Existing wizard 30/30 PASS, tracer source 172/172 PASS, review fmt/strict SSH all-target Clippy PASS, existing proof controls 2/2 PASS with 1 intentional ignored regression. Confirmed P0=0/P1=1/P2=1/P3=0. Clean target2b779884gate and private decoder fuzz still running. Branch contains only two test commits; no push/fix/deploy/perf VM.Completed exact-target native gate stage on detached
2b779884: fmt/VT fmt/workspace strict Clippy/test/cargo-deny PASS; 812 tests PASS, 0 FAIL, 79 ignored. Both static release builds remain in progress. Independent private profile 7/7 PASS; real private/full handover and one-session rebridge PASS; real-source ASan decoder fuzz 10,000 empty + 10,000 seeded mutations PASS. These are functional checks, not latency/whole-network enrollment acceptance. Local review head remains4c1de592with P1 R189-02/P2 R189-01; tests only, clean. No baseline exemption, implementation fix, push, deployment or perf VM.HOLD (P1 present) — merge after fixes. Counts: P0=0, P1=1, P2=1, P3=0.
Target:
2b77988400; merge base with dev:93b00f5ab9. Local review-addhost-instant head:4c1de592f9, two test-only commits, no push/implementation fix/deployment.P1 R189-02 (
8b94495c): after early Done, the actual wizard closes and says the host is ready; the product sends no retained setup subscription. Later Join failure and retry are absent from the machine pill. Browser regression fails twice; existing raw-socket control passes. Test: apps/tracer/test/addhost-instant-review.spec.ts. Stable capture: /home/kayg/.cache/codex-scratch/addhost-instant-review/r189-02.png.P2 R189-01 (
4c1de592): a stopped commit channel drops the sibling proof future before cleanup_sessions runs. The private probe is left to its independent ceiling. Explicit ignored regression fails with cleaned=false/committed=false; both old proof controls pass. Test: crates/plugins/ssh/src/bootstrap_proof_tests.rs, review_stopped_commit_channel_finishes_the_private_probe. P2 can be fixed forward under the owner verdict rule.Independent controls: exact-target native 812 PASS/0 FAIL/79 ignored; tracer source 172 PASS; existing production-build browser checks 31 PASS (30 wizard + raw setup control); private profile 7 PASS; real private/full same-session/same-writer handover and one-session rebridge PASS; real-source ASan decoder fuzz 20,000 mutations PASS, including five launcher/overflow seeds. Review fmt/strict SSH Clippy pass; cargo tree -d checked, no review dependency changes. Browser uses the test double; full network SSH/NetBird enrollment, physical iPhone and performance were not independently remeasured. Explicit no-perf-VM rule followed; shared-host functional transcript durations are not performance acceptance. Author's unmet WAN targets remain unwaived.
Exact clean target-head scripts/verify fast exit1 solely at both static size caps: x86 20474608 B; ARM 18215584 B; both stripped/static/no PT_INTERP. X86 key initialization=1 under 8 MiB memlock; ARM runtime not checked on this x86 host. Sizes match the author's head artifacts; production baseline deltas were not independently rebuilt. No cap/profile/dependency changes. Initial inherited-umask and local RTT gate attempts were retained; final gate uses umask022, owned compact fixtures, CPUWeight10000 and unchanged gate limits, with no baseline exemption.
Reviewer checklist: each finding has a committed failing test; existing controls pass; only tests changed; no new component/helper/library; no owner fixture/credential inspection; issue start/findings/checkpoints recorded. Findings: /home/kayg/.local/state/codex-jobs/infraternal/review-addhost-instant/findings.md.
Final gate summary, verbatim:
Full unmodified transcript follows in numbered parts, SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042, 241065 bytes. Build targets/temporary checkout cleanup follows; logs and screenshot remain in owned review scratch.
Independent target-head gate transcript, verbatim 1/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 2/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 3/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 4/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 5/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 6/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 7/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 8/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 9/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 10/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent target-head gate transcript, verbatim 11/11. Target
2b779884; SHA256 22b8c2651945e50d6e4906b4e243885dd6ae48ffa6a6474c85382efd78218042. Functional transcript durations are not performance acceptance.Independent review cleanup complete. HOLD (P1 present), merge after fixes; P0=0/P1=1/P2=1/P3=0. Local review-addhost-instant remains clean at
4c1de592f9; only two failing-test commits. Exact target2b779884verdict and complete raw gate output are posted above in 11 parts. Owned temporary checkout, GNU/private/fuzz targets, compact fixture root and generated corpus removed; logs and screenshot retained in /home/kayg/.cache/codex-scratch/addhost-instant-review/. No push, implementation fix, deployment, perf VM or issue closure.Final scripts/verify fast summary, verbatim:
Resuming backend review fixes on addhost-instant, base/head
2b779884(origin/addhost-fast ancestry). Read #189/#169 history and project contracts. First import R189-01 regression and cover cleanup for stopped commit/cancellation; then pin retained setup failure/retry/done delivery and publish exact contract for designer. Continue WAN prompt/size and 10+10 VM208 acceptance. No deployment or owner fixture changes.Backend refinement (D255), before code: retain /v1/bootstrap/setup/ws and existing setup envelope. Add boolean retry_available to every setup event: true only for a failed attempt with retries remaining (maximum eight attempts); false for running/done. Fresh authority, host ownership, enabled component and worker capacity are still checked when retry is requested. Mark saved attempt retryable before publishing failure, eliminating immediate retry race. Existing failed_step + fixed safe message carry reason; ready:done is fully joined/reattached. Add real WebSocket late-subscription/live-failure/replay/retry-completion coverage. No new endpoint/dependency or UI implementation.
Designer/backend contract at
8362ea67(contracts/host-setup.md; additive retry eligibility implementation/tests in progress):Connect /v1/bootstrap/setup/ws with existing exact Origin/current unlocked owner cookie. Send {"action":"subscribe","host_id":"","window_id":"","after":0}. Receive retained then live:
stage=install|join|start|ready; state=running|done|failed. Failure includes safe message + failed_step=connect|upload|install|join|start|readiness, retry_available=true if attempt<8. Running/done/exhausted=false. ready:done means installed transport reattached original shell. Host sequence strictly increases; new retry increments attempt with same session. Cursor replay: {"action":"replay","after":N}; overflow emits {"resync":true,"host_id":"","window_id":""}, then current stage states. Retry: {"action":"retry","host_id":""}; acknowledgement {"retry":""}, refusal {"error":""}. Runtime authority/worker admission still applies. Socket closure never cancels background setup. Wizard done means shell attachable, not fully joined. No endpoint change; UI regression
8b94495cremains designer lane.R189-01 imported at3165a8ae: explicit unchanged regression fails0/1, cleaned=false/committed=false (cleanup-red.log). Fix moves private-probe cleanup outside proof/commit select; published enclosing deadline/cancellation paths also invoke bounded cleanup. No user shell retirement. Additional backend ordering finding: bootstrap.rs emits Failed before pending_setup.running=false (initial and retry paths), so immediate status-driven retry can race. Order failed attempt state before event publication while preserving serialized retries. Real WebSocket late replay/live failure/retry completion test building; double eligibility red1/1 then green2/2.
Completed R189-01 fix at5c4b4461: imported red regression0/1, green3/3 proof tests, full SSH suite PASS (37 unit cases incl2 ignored; all non-opt-in integration suites pass), fmt and strict SSH all-target Clippy PASS. Cleanup now covers commit-channel stop, host retirement and enclosing published setup deadlines; callback targets only the private canary. Existing original-shell/no-commit fences retained. Raw logs retained in .job-scratch/cleanup-{red,green,clippy}.log and ssh-status-tests.log for attachment.
Backend retained stream complete through241ad169: status envelope6e552dc7, failure/retry serializationb70593d8, real WebSocket regressionb9220279, aligned double80c2d8e8/241ad169. New API test red0/1 (missing retry_available) then green2/2 with remote enabled; late replay, later live failure/reason, reconnect and same-session attempt2 completion covered. Exhausted attempt8 remains failed with reason and retry_available=false. Full SSH suite PASS; fmt and strict SSH/server all-target Clippy PASS; tracer check and double2/2 PASS. Exact designer contract posted at8362ea67. Milestone iterate running. UI product subscription is separate designer lane (R189-02 browser regression not imported here).
UI for D255 pill built in the product on branch ui-addhost-pill (head

1ea2db04, pushed, not merged). The pill follows the real stream /v1/bootstrap/setup/ws after the wizard closes (fixes the review P1; review spec passes). Screenshots (production build, ink/paper, 1440x900, 393x852, 375x667, 2x): running, failed, explained with Try again, system icon.Budgets (production, brotli): critical 245265 -> 245508 (+243 B, limit 256000); deferred 43980 -> 44219 (+239 B, limit 45056); on-demand total +3.8 KB in the pill chunk (largest chunk 13121/18432).
UI for D255 pill built in the product on branch ui-addhost-pill (head

1ea2db04, pushed, not merged). The pill follows the real stream /v1/bootstrap/setup/ws after the wizard closes (fixes the review P1; review spec passes). Screenshots (production build, ink/paper, 1440x900, 393x852, 375x667, 2x): running, failed, explained with Try again, system icon.Budgets (production, brotli): critical 245265 -> 245508 (+243 B, limit 256000); deferred 43980 -> 44219 (+239 B, limit 45056); on-demand total +3.8 KB in the pill chunk (largest chunk 13121/18432).
Next prompt experiment before source/recipe changes (D255/D243): measure size-oriented compiler optimization only for the optional private session image, keeping the full release/production comparison recipe, native VT archive, LTO, single codegen unit and panic=unwind unchanged. Current private x86 image2533392B / gzip1111554B dominates WAN payload ramp; three OpenSSH-compression probes improve only
50–60ms (prompt after sign-in1.66–1.68s), not the whole remaining~200ms. Build private candidate with release opt-level=s in an isolated owned target, then measure real same-session byte handover, prompt/join and early-session echo/flood against the existing image on VM208. No dependency/engine change. Retain a dedicated private build profile only if measured runtime results justify its size/upload improvement; otherwise discard.VM208 matched exploratory comparison,3 fresh150ms targets per variant, same frozen2b779884 binaries, shared lock per run: normal prompt-after-sign-in p50=1694.669ms, fully joined p50=5194.530ms; OpenSSH Compression=yes p50=1670.505ms /5159.530ms (24.164ms /35.000ms better). Controller/fixture child CPU from bash time: normal median(user+sys)=2.49s vs compression2.62s, +0.13CPU-s per fresh setup (~1.04CPU-s at eight setup workers; includes unchanged cold payload preparation). All six59-byte handover/ledger proofs pass. Small latency benefit does not close D255; reject shipping this candidate pending the private-image experiment. Raw compression/normal logs kept in named VM review-fix directory. No host-key/credential/payload changes.
Correction to CPU arithmetic above: compressed child totals2.63/2.62/2.64CPU-s give median2.63, delta+0.14CPU-s versus normal2.49, or+1.12CPU-s at eight workers. Latency medians and candidate rejection unchanged.
D243 experiment before implementation: share the existing Tokio Runtime::block_on(io::Result<()>) boundary across inherited-fd, SSH bridge and node join/run entrypoints. Box only the top-level future once per reactor (never a terminal frame/read/write), with one ordinary non-inlined shared function in infraternal-agent. Reactor choice, panic=unwind, authentication, UID, parser, cleanup and scheduling stay unchanged. This targets repeated generated reactor poll code; test both private/full startup and byte handover, compare stripped full sizes under the unchanged release recipe, and benchmark native echo/flood on VM208 before retaining. Reject if size does not decrease or runtime budgets regress.
Backend red/green evidence for committed241ad169 (166 SSH tests,2 real setup WebSocket controls,2 double controls; strict Clippy/fmt/tracer check pass):
cleanup-red.log
cleanup-green.log
cleanup-clippy.log
status-red-remote.log
status-green.log
status-clippy.log
ssh-status-tests.log
double-red.log
double-green.log
Milestone iterate failed: native OpenSSL clippy build cache reported build_dir Directory not empty after the owned gate restart; server tests334 PASS/2 FAIL/39 ignored, mesh.tag.create successful-result assertion and plugin_lifecycle next_hosted2s timeout. These are not waived. Replaying after the build queue; inspect actual refusal before attributing load. Other reviewed readiness/proof/start-fresh rows pass.
Private opt-level=s exploratory image2341680B/gzip1029262B versus2533392B/1111554B; one150ms VM probe prompt after sign-in1683.613ms, joined5199.418ms, same-session59-byte/ledger proofPASS. Too small an observed improvement to retain without further measurement. Reactor code-sharing candidate is uncommitted and strictly lint-clean; size/runtime comparison pending.
D255 prompt experiment before code: reuse the existing public decimal-length script envelope on OpenSSH duplex command startup, so the exec request contains only a small fixed POSIX reader instead of a multiply shell-quoted setup program. The same authenticated master/channel, trusted generated scripts, exact 65536-byte bound, payload/session receipts and private-invitation barrier remain. Reader must verify complete byte count before executing anything; dd is exact, head fallback is safe here because no later bytes are written until program receipts. Share the envelope writer with checked staging; no new engine/dependency, command reordering, trust or UID change. Real-process tests cover incomplete script refusal before mutation and preservation of bytes following the script; measure before retention.
Completed reactor reuse45d8f516/b5084c50/d1de93c7. Same full release recipe: x8620448816B vs checkpoint20474064B (-25248B), production20333104B =>+115712B; ARM18197872B vs checkpoint18215584B (-17712B), production18112240B =>+85632B. Full caps/flags/dependencies unchanged. CLI42 PASS/3 ignored, daemon44 PASS/7 ignored, strict Clippy/fmtPASS, real private/full original-session/pending-input byte oraclePASS. VM208 real shim/container bounded100 echo samples each: old node image p50/p99.260/.400ms; new full image.255/.706ms (p99 below5ms, no improvement claimed), load.44/.14/.19. One allocation per reactor, none per frame. Full-image growth remains unwaived. New public-script framing experiment is uncommitted; reader partial-program/no-overread controls PASS, 166 inherited SSH controls PASS. Controller rebuilding before VM comparison.
Framing experiment controls: inherited SSH suite PASS; four real shell/pipe controls cover incomplete program refusal before mutation, exact following bytes and UTF-8 byte count with inherited locale preserved. Strict SSH Clippy PASS. First VM probe did not execute: new GNU controller had a Nix dynamic-loader path, whereas prior controller was static-pie. Not performance evidence; retained log. Static musl controller building with existing target recipe before repeat. Source remains uncommitted pending latency comparison. Both prior mesh failures replay individually PASS (1/1 each); second full iterate currently running, no exemption claimed.
exploratory-raw.tar.gz
reactor-echo-vm-retry.log
reactor-cross-build.log
reactor-cli-tests.log
reactor-agent-tests.log
Rejected public-script stdin framing: static musl controller VM208 one fresh150ms probe prompt3125.869ms total minus sign-in1260.825=1865.044ms; fully joined5360.009ms;59-byte same-session/writer and zero-token ledger PASS. Adds approximately one RTT versus retained command path~1.70s after sign-in /5.19s joined. Source restored exactly to committedd1de93c7; no new reader/parser/locale/feature change retained. Experimental controls pass, but performance rejects it. Finishing exact retained-source acceptance and gate; remaining WAN/size misses stay unwaived.
Final gate attempt atd1de93c7: agent native67 PASS/1 FAIL/12 ignored. winning_quic_drops_the_unfinished_tls_dial_before_control_dispatch failed at remote_client_tests.rs167 binding UDP to an ephemeral TCP-selected port: AddrInUse(code98). Not a timeout exemption or caused-by-change claim. Full gate will rerun serially after artifact/static builds complete; preserve this transcript. Second iterate was stopped by its own process tree before overlapping the final full test run. Prior mesh assertion and2s lifecycle timeout replay1/1 each; second full iterate reached both PASS before stop. No shared process, port or owner fixture stopped.
Exact retained-source acceptance atd1de93c7dc6419244369eec40674ca9d289f09bf, VM2084vCPU,10 fresh targets/profile, lock released per target.
RTT=0ms n=10 prompt_after_signin_p50/p99=156.972/175.011ms prompt_total_p50/p99=219.713/245.440ms fully_joined_p50/p99=656.244/691.530ms
RTT=150ms n=10 prompt_after_signin_p50/p99=1698.997/1704.356ms prompt_total_p50/p99=2962.017/2973.517ms fully_joined_p50/p99=5198.458/5212.318ms
20/20 original-session/writer and59-byte before/after proofs; p50=median, p99=nearest-rank maximum of10. VM208,4vCPU; load in each raw log; lock released per fresh target.
WAN prompt~1.5s and joined<5s remain unmet (no waiver). Same source/artifact count and hashes in archive; no timing evidence from shared build host. Round3 functional4710396 bytes across24 stops PASS; final private/full same-session pending-input oracle PASS.
acceptance-raw-final.tar.gz
acceptance-summary-final.txt
acceptance-final-console.log
round3-final.log
cross-final.log
Production-build functional browser replay31/31 PASS, including retained setup retry, all inherited add-host/sign-in/failure/mobile controls and Start fresh (TRACER_TEST_PORT4992, IPv4-first localhost fixture resolution, outer120s harness timeout; assertion bounds unchanged). First run6/31 with one outer30s page boot timeout then localhost ::1 reset ECONNREFUSED; preserved, not timing evidence. Historical tracked capture restored after copying generated failure capture to scratch; no visual-quality judgement or product UI change. Final full native gate running serially after earlier UDP port collision.

browser-final-replay.log
Gate blocker identified as missing upstream test fix: origin/dev
e3d9a942(2026-10-13,#173/#118) already holds inventory socket runtime time stationary across real SQLite/I/O, with baseline7/200 unexpected Close failures and qualified200/200 PASS. Our lane predates it; serial head gate reproduces mesh.tag.create successful-result assertion again. Port upstream inventory test only using this branch's existing FrozenSocketTime (no progress() method; test fits unchanged90s hang bound). Production25ms authority-read deadline and all explicit revocation/freshness/stall assertions unchanged. No timeout waiver, API/runtime change or repeated isolated-pass claim.Completed older-lane test integration ateef5ac83: port origin/dev
e3d9a942inventory clock control using the existing local FrozenSocketTime. Branch gate pre-fix repeats mesh.tag.create successful-result failure; post-fix focused real socket1/1 PASS, server shipping-feature/all-target strict Clippy/fmt PASS. No production deadline, assertion or shipping code changes. Final full head gate running with1 native test thread. Shipping acceptance remainsd1de93c7artifact source (eef5ac83adds test-only lines), hashes and20/20 byte proofs already attached.Head
eef5ac83: full fast gate now passes the inventory test with upstream test-only clock control, but SSH install negotiated_retry_reuses_only_a_verified_cache_after_activation_refusal failed at install.rs:478 during second rollback, synthetic exit73. Focused exact replay passes1/1; full gate repeat underway. No load/baseline waiver applies or was used. Raw full output attached; fmt/clippy/deny passed, static size caps fail unchanged. WAN acceptance and production size deltas remain unmet; final handoff will retain these findings explicitly.Backend review handoff:
eef5ac83b7pushed to origin/addhost-instant. No deployment. P2 private-probe cleanup fixed with imported failing regression; retained setup stream includes stages, safe failure reason, retry eligibility, done and late-subscriber replay/live-failure coverage. Exact designer contract posted above. UI subscription is a separate lane.VM208 acceptance, 10 fresh targets each (p50 / nearest-rank p99, milliseconds):
D243 same recipe against production
9ab4329c, exact bytes:x86_64 production 20333104; branch 20448816; delta +115712.
aarch64 production 18112240; branch 18197872; delta +85632.
Positive production deltas remain unmet; no cap changes or waiver. Shared reactor boundary cuts x86 25248 B and ARM 17712 B from this round's earlier checkpoint. Compression/compiler-size/framed-stdin experiments rejected; raw evidence retained.
Functional checks: browser production build 31/31; TypeScript/protocol checks pass; cross-build private/full adoption passes; round-3 4710396 bytes across 24 stops passes. Upstream inventory clock fix is test-only; production authority deadlines unchanged. Earlier gate's second rollback synthetic exit73 remains an unexplained intermittent failure: exact replay and later full installation suite 76/76 pass. No load/baseline waiver used.
Final gate: server327 PASS/9 FAIL/39 ignored; nine failed assertions/closed sockets/elapsed control tests are retained in the attached diagnostic. No cause or baseline waiver claimed. Final scripts/verify fast summary follows verbatim; raw full output attached. Review HOLD: remaining WAN/size acceptance work is not complete. Issue remains open.
verify-fast-last.log
final-server-failures.txt
artifact-hashes-final.txt
reviewer-checklist.txt
verify-fast-final-head.log
install-retry-gate-replay.log
ts-final-target.log
final-size-receipt.txt
reviewer-checklist.txt
framing-rejected.txt
framing-probe-static-vm.log
acceptance-raw-final.tar.gz
acceptance-summary-final.txt
acceptance-final-console.log
round3-final.log
cross-final.log
browser-final-replay.log

Starting 2026-10-14 continuation on addhost-instant, base/head
eef5ac83b7. Read all #189/#169 comments and worktree contracts; fetched origin/devb182529f. First merge dev (restart/audit test stability, size reductions, #199), reproduce all nine prior server failures without waivers, then investigate prompt stages and D243 deltas. Retained setup-status contract remains stable. No deployment. Prior raw acceptance/gate evidence is already attached above.Merge investigation (origin/dev
b182529fintoeef5ac83), before commit:Integration findings, pending merge commit:
Merged-source server run: 432 PASS /1 FAIL /46 ignored. All nine
eef5ac83failures pass without any baseline/load waiver or deadline change (exact rows below). New failure: remote_runtime::tests::ssh_bridge::refused_installed_reattachment_restores_the_ssh_writer expects first replacement frame Ack; upstream connection_lost publishes TerminalOutput::Reconnecting before fallback restoration. Adapt test receive to skip only Reconnecting/Pong under one unchanged5s deadline; still rejects data/Close and requires Ack plus SSH PID/master lifetime. Focused green pending.test auth::tests::server_setup_terminal_echo_and_resume_preserve_offsets_once ... ok
test auth::tests::socket_tests::fresh_delete::review_mesh_delete_requires_fresh_passkey_tap ... ok
test auth::tests::socket_tests::plugin_lifecycle::mesh_lifecycle_withdrawals_survive_calls_but_never_end_a_replacement ... ok
test auth::tests::socket_tests::plugin_lifecycle::mesh_plugin_discovery_disable_and_restart_preserve_control_socket ... ok
test auth::tests::socket_tests::review_control::pending_mesh_writers_preserve_authority_pool_capacity ... ok
test auth::tests::socket_tests::review_liveness::review_mesh_watch_reports_liveness_expiry_without_another_edit ... ok
test auth::tests::socket_tests::scale::review_mesh_headroom_snapshot_does_not_close_shared_socket ... ok
test auth::tests::socket_tests::session_gone::an_ended_or_unknown_session_gets_the_final_close_without_losing_the_socket ... ok
test auth::tests::socket_tests::session_gone::the_infraternal_agent_saying_gone_reaches_the_browser_as_the_final_close ... ok
Transport integration red/green: bridge listener 0/1 red AddrInUse ->2/2 green; nondumpable SSH-child connector image 0/1 wrong executable ->1/1 green. Shared existing materialization admits once, descriptor-clones for prompt reuse, erases at authentication/drop. No dump-protection weakening.
Additional gate integration red/green: scripts/test-static-size.py searched removed max label and raised TypeError on aligned synthetic static ELF fixtures; now reads printed cap. Both exact boundaries, one-byte overflow, shrink-ratchet and unsupported-target controls pass; no cap changed.
D262: moved remaining lane evidence out of repo. Relocated two executable measurement fixtures (unchanged bytes) to docs/verification/addhost-breakit2 and repaired their harness references; fixtures that tests execute remain source, raw receipts stay on issue. Workspace Clippy still compiles native dependencies; merge not committed/pushed yet.
Merged origin/dev
b182529finto addhost-instant:b980dbae. All nine previous server failures pass (no waiver). Complete server run: 432 passed, 1 failed, 46 ignored; its remaining SSH restoration oracle now handles only Reconnecting/Pong notices within the original 5-second deadline, and all three SSH restoration tests pass. Workspace remote clippy + fmt pass; full SSH and TypeScript gates pass. Real-process regressions: temporary SSH control-listener collision 0/1→2/2; nondumpable connector-image execution 0/1→1/1. Incoming static-size diagnostic parser max→cap repaired; numeric ceilings unchanged. Whole-workspace tests are still compiling. Signed session selection/deferred full acquisition and node compile-out are separate following steps.nine-after-merge.txt
bridge-listener-red.log
bridge-listener-green.log
connector-red.log
connector-green.log
ssh-restoration-green.log
size-parser-red.log
size-parser-green.log
Completed signed bundle publication slice:
7783f76b. New publication requires six artifacts before network access and the unchanged signer includes both session images. Python 8/8 and Bun 5/5 pass; real publication admission observed six uploads, zero replacements. Old signed omission remains compatible, advertised image absence/tampering refuses. Resolver RED: original active tests 9/9 pass; all ten new session tests fail against defaultNone as intended. No deploy.publication-red.log
publication-green.log
red-direct.log
Completed
502c7a7b(signed optional session resolver),0d75f386(publish session before resolving/installing full image), and0679e4e1(release/setup contract documentation). Resolver: 19 passed, 0 failed, 2 ignored; initial RED had 10 new failures. SSH: 179 passed, 0 failed, 11 ignored. Added real duplex-shell test proving input/output before full-image acquisition finishes; late failure is retained and retries keep the session ID. Public setup event contract unchanged; GetX86/GetArm64 errors map to public install failure, retaining internal diagnostics. Same-architecture installed payload remains the running server executable (D225).Size finding: empty compiled node-plugin catalog still started a worker thread. Regression failed before the optimization; current candidate removes that idle worker and shares the existing empty registry semantics. Private x86 image: 2,641,520 -> 2,601,648 bytes (-39,872); gzip -n -9: 1,160,814 -> 1,141,757 (-19,057). Functional feature-matrix gates running. These are byte counts, not build-host performance claims.
All nine prior server failures passed again in the broad post-merge test run; its remaining ten failures were exclusively the new resolver RED tests, since fixed. Full current-source iterate and final static builds are running. Current first full x86 build was 22,749,584 bytes; attribution versus incoming dev and the identical-recipe production baseline is pending. No size cap changed. Acceptance will use the full shipping image, rather than the smaller node image used by earlier harness runs.
Benchmark integration finding: crates/infraternal-server/src/bootstrap_breakit2_tests.rs still constructed Bootstrap with BinaryPaths only. After the signed-source change, PayloadSource::session defaults to None (correct for legacy adapters), so this live fixture would measure full-image fallback rather than the initial-session-image path. Fixing only the fixture: explicit --session-payload input, distinct retained full shipping image, no production resolver change. Added a regression for the source actually passed into Bootstrap; RED run is in progress. Missing explicit session input now refuses before controller execution or Docker mutation. Instrumented current controller built; full current x86 image frozen at22,749,712bytes, incomingdev/production identical-recipe baselines queued. Perf measurements are waiting for the shared lock; no build-host timing substitutes.
Correction to previous benchmark finding: I misread Bootstrap::payload_for. With no configured PayloadSource, the existing legacy branch still calls BinaryPaths::session_payload (crates/plugins/ssh/src/bootstrap.rs), so the earlier fixture selected its sibling private image when present. The default PayloadSource::session(None) does not apply to that branch. Earlier prompt-image measurements are not invalidated by this.
The new explicit --session-payload fixture is an additive harness improvement: exercise the modern PayloadSource selection path and refuse a missing explicit image before target mutation, retaining the full shipping executable separately. Its selection regression is test-first feature coverage, not a reproduced production bug. The separate full-installed-image mismatch remains real: earlier acceptance sent the smaller node image; new acceptance will send the full shipping executable (D225).
Completed steps:
5d36c519removes empty-catalog private-session workers (54 unit +2 integration passed bare; 100 unit +2 integration passed default/remote; both strict Clippy configurations pass).0f0b33farestores node component compile-out (all five feature variants pass; default full graph unchanged; node/server refusal and strict Clippy pass).88982c49makes the benchmark images explicit, adds stage timestamps, and proves installed full-image digest equality; parser target-mutation regression RED→GREEN. No production setup-event contract change.Current complete server lib: 444 passed, 0 failed, 46 ignored. All nine prior failures passed again, without waivers. Current x86 full image 22,749,712 B versus immutable incoming dev
b182529fat 22,275,888 B (+473,824); exact production x86/ARM rebuilds are running. Private x86 shrank 39,872 B raw /19,057 B gzip. Full repository native iterate and final scripts/verify fast are running on head88982c49; RTT self-test and live measurements are waiting for the shared VM lock. Final acceptance remains pending.server-current-full.log
green-no-default.log
green-default-remote.log
live-fixture-red.log
fixture-admission-red.log
fixture-admission-green.log
Blocked on perf-VM access: the configured RTT self-test flock -w3600 and diagnostic flock -w1800 waits expired before entering measurement; no samples or performance claims produced. VM208 remains reachable. Current lock receipt shows PID597982 holding WRITE on /run/lock/perf-bench.lock, with waiting jobs (receipt retained). The booking holder/orchestrator needs to release that exclusive measurement lock for D255 acceptance. Final fast is waiting on the same self-test; direct native gate and exact production-size rebuilds continue locally. Will retry measurements when access is available.
Resuming disk-full interruption on addhost-instant at base/head
88982c4955. Tracked worktree clean (only generated Python cache); preceding atomic steps retained. Read #189/#169 history and merge/444-pass server receipts. Rebuild deleted native targets, rerun iterate/final fast, resume VM208 10+10 initial-session/full-installed-image acceptance and same-recipe production-size comparison. Prior perf lock wait expired; VM reachable now. Setup stream contract unchanged. No deployment.Fresh VM208 diagnostic (current full/session images, old controller legacy source; exploratory only): prompt total3179.234ms minus key sign-in1319.051ms =1860.183ms; check/channel467.714ms; protected publication at3020ms; deferred full-image upload+setup2331.268ms (10130659 compressed bytes); joined6075.147ms. Same-session/writer59-byte proof and installed full executable digest PASS. Queue loss zero; load3.70/1.45/1.02. Earlier acceptance used the smaller node payload, so full-image installation adds upload cost; final acceptance will use the modern explicit source/current controller. Native targets rebuild; readonly/guard scripts unchanged from
eef5ac83. No safety fence removed.Exact production9ab4329c x86 Cargo release rebuild completed:20015792bytes, sha2566fde07975bb7116b390b643749b3f43660ba87a5e49369d5558a6562b88287f5 (previous reconstructed GNU-strip baseline explicitly lacked equivalent provenance and is not used). ARM baseline and candidate boundary size build continue. No cap/profile change.
TS source/compiler/build checks pass, protocol differential timeout: Bun spawnSync cargo run waited behind concurrent native build-dir lock, hitting unchanged30000ms limit (the gate prebuilt its oracle, then another Cargo command acquired the lock). Retaining failure and repeating serially after native Cargo work exits; no assertion/deadline waiver.
Functional optimized-browser replay:48PASS/0FAIL/4capture-only skipped, including Start fresh, same-session setup retry, every sign-in/failure/mobile control (TRACER_TEST_PORT4997, NODE_OPTIONS IPv4-first, existing per-assertion bounds). Browser plugin unavailable; used repository Playwright. No visual-quality or performance claims.
Native iterate source check: fmt/nativefmt/CLI compile-out/Clippy/deny pass; test row stops at setup_token_command_reads_only_the_expected_socket_response: InvalidInput path must be shorter than SUN_LEN. Cause is my TMPDIR continuation-target/tmp plus the fixture suffix exceeding Unix socket108-byte bound. Correcting only environment to existing owned short .../addhost-instant/s (the state root already used it) and rerunning full iterate. Failure retained; no source/test weakening or waiver. Branch prior atomic commits now pushed through88982c49; single payload-preparation boundary experiment remains uncommitted pending linked size and performance comparison.
Rejected the upload-preparation future boundary:22750256B versus retained22749712B (+544B), patch/artifact retained in cache; source restored. Testing an independent narrower boundary in the signed full/session resolver, with existing signature/manifest/cache/fallback controls; retain only if linked size falls. No contract/cap/profile/dependency change.
SSH standalone suite exposed a second harness path error: .../addhost-instant/s/c- crosses CredentialSession root length+48>=108, deliberately selecting the production runtime fallback; fixtures expect their explicit root and report ENOENT. No credential access/auth failure occurred. Use newly owned0700 HOME-cache/a189 for both fixture root and TMPDIR, not the long job prefix; interrupted only the prior owned waiting gate PIDs439635/440222/440226. Full iterate and resolver controls rerun in this environment.
Signed resolver candidate:19 active tests PASS,2 measurement tests ignored; workspace strict Clippy and fmt PASS. Linked size still pending behind the static controller.
Fresh exact production9ab4329c Cargo release baseline complete: x86_64 20015792 B; aarch64 17810608 B; both static/no PT_INTERP. Source clean, default recipe/no profile or release-identity overrides, receipts/hashes retained.
Second full iterate passed CLI81 and 24-stop/queue/credit oracles, then agent99PASS/1FAIL/14ignored: daemon::gone_tests fixture binds a109-byte Unix address under a189 TMPDIR (36-character root +25-character fixture name +session suffix). No source failure/waiver. New exclusively owned33-character /home/kayg/.cache/codex-scratch/a keeps that address106 bytes; rerunning full workspace tests serially there. Test limits/assertions unchanged.
Rejected second representation experiment: resolver polling boundary22750064B versus retained22749712B, +352B despite19 signature/cache controls and Clippy passing. Restored exact88982c49 source. Both rejected patches/images retained; no unaccepted allocator/runtime change, new dependency, cap or profile. Final scripts/verify fast running on clean tracked88982c49 source.
Native queue diagnosis: workspace artifact initially returned resource_unavailable before saturation; the frozen executable repeated that result. Package-specific replay passes but has a different Cargo feature/artifact graph, so it is not an exact-artifact cure. Final actual workspace gate now passes the full queue oracle unchanged, alongside CLI81/0 and24-stop proof; server tests running. Root cause of the intermittent open refusal remains unclaimed; no118 waiver/assertion/deadline change.
Fresh packet-header diagnostic on VM208 (not acceptance): prompt3110.996 - sign-in1306.135 =1804.861ms; read-only channel/check451.630ms; invitation3.335ms; upload+shim to protected publication1181ms; publication→interactive prompt166.996ms (event timestamps integer-rounded). Trace: channel-open1117.509ms after SYN→confirm1268.106;11412 script bytes paced1268.379–1336.810; final check receipts1552.760–1559.464; upload starts1564.010;SESSION receipt2743.499; attachment sends2743.979. Captured4240 packets, zero kernel drops, duplicate veth observations excluded by docker0 filter. Residual is request/payload pacing, not an identified safety-required extra round trip. Current full image deferred transfer sends10130659 gzip bytes; fully joined5933.178ms. Raw headers contain packet metadata only.
Retained-source static controller rebuilt (sha2566e6ccd62318befcf9b5634e38244a740b9d30d49a2c3d49160b303c20e0e22f3); launching final10+10 fresh acceptance with explicit session/full fixture inputs and installed full-image digest proof. Congestion diagnostics remain experiments only; kernel/product defaults unchanged.
Final-head88982c49 native stage PASS:1141 passed,0 failed,140 ignored across113 result rows (includes29 CLI compile-out cases). Complete server444 passed/0 failed/46 ignored; every one of the nine earlier server failures is explicitly green in nine-final.txt. Full24-stop, queue saturation, agent100-case and SSH/readiness/status/retry/cleanup oracles pass without any118 waiver or changed test bounds. Fmt, native fmt, workspace Clippy and cargo-deny pass. Final fast is now building/validating both static targets; TypeScript final runs separately because a D243 refusal stops the dispatcher before TS. Acceptance has completed two LAN fresh targets so far; final20-target results remain pending. No deployment; tracked source remains unchanged from pushed88982c49.
native-final-prefix.log
nine-final.txt
browser-functional-replay.log
New slice-gate finding after the full1141-pass native run: standalone crates/plugins/ssh/tests/previous_attempt_shell.py fails all16 controls with exit127. It still concatenates setup_lock.sh (lease helpers), but the shared guard moved to setup_guard.sh. Shipping previous_attempt.rs already imports the correct guard; this is test composition drift, not a reproduced production recovery bug. RED retained in ownership-shell.log. Fix the oracle to import the same guard-only prelude and register a small Cargo integration wrapper so the full native gate cannot skip these real-file/live-listener recovery controls. No production code, protocol, size-cap or event-contract change.
Completed and pushed test-only2f021459: corrected standalone ownership/recovery guard composition and registered the16-case real-file/live-listener oracle in Cargo. Original16/16 failed exit127; corrected16/16 PASS; new Cargo wrapper1/1 PASS; fmt and strict SSH all-target Clippy PASS. Shipping code, setup event contract, dependencies and static artifacts unchanged from88982c49. Final scripts/verify fast reruns on this exact head.
Retained-artifact VM208 acceptance is complete,10 fresh targets per profile,20/20 session/writer/59-byte/input-output/zero-token and installed-full-image digest proofs PASS. p50 median/p99 nearest-rank maximum-of10:
WAN targets remain unmet. Raw acceptance includes load/artifact receipts. Cross-build private/full pending-input deduplication and one-session retry oracles also pass. Recovered the existing pinned real NetBird fixture from its build-host image; SHA matches the committed#169 receipt exactly(e1e75f2091a0a739fc8523713a2b2f27928e8d02c436e4e1cb2fcf94bfe19a1e). Additional real NetBird compatibility controls follow. A current node-only image is building solely as a D225 payload-cost diagnostic; production still installs the full running image. No engine choice or deployment.
HOLD / resume handoff, addhost-instant@2f021459206d767b2f2c7e5e497356aa0e195332 (pushed). No deployment. Dev merged previously; all nine original server failures pass on this head. Final native totals: 1,142 passed / 0 failed / 140 ignored; server 444/0/46. Source TS 192/0, protocol 37/0 (34,463 assertions), streaming 134/0 (11,439 schedules, zero divergences); production-browser functional 48 pass / 4 capture-only skip. Real NetBird early-shell handover LAN/WAN, legacy install refusal+retry and foreign-file refusal pass. Test-only independent review found no issues. Setup-status contract unchanged.
New commit
2f021459: standalone ownership oracle was not in native gate and all 16 cases failed exit127 after setup_guard extraction. Source composition corrected to match production;16/16 now pass and Cargo integration wrapper keeps it in gate. RED/GREEN attached.Acceptance VM 208 Linux 6.12.111+deb13,4 cores;10 fresh targets each, unchanged default TCP. One-minute load ranges LAN0.32–4.08, WAN1.27–2.77. Every 20/20 run proves 59 input bytes, output before/after, same session/writer, zero unconsumed authority and installed FULL shipping-image digest (D225). p50 median; p99 nearest-rank maximum of 10:
| RTT | prompt after sign-in p50/p99 ms | fully joined p50/p99 ms |
|0|210.883 /250.419|1050.827 /1342.872|
|150|1800.941 /1837.258|6000.746 /6150.483|
LAN passes; WAN goals remain unmet. Runtime artifacts from
88982c49; head adds tests only; final rebuild sizes unchanged. Earlier 5,198 ms joined used smaller node image, not the required full image: it is not comparable D225 acceptance.Header-only owned-target packet diagnostic: prompt after sign-in1,804.861ms = check/channel451.630 + invitation3.335 + upload/shim1,182.900 + attach166.996. Channel open reply150.597ms;11,412 check-script bytes paced across
68ms. Private initial upload is1.14MB gzip; full background upload10,130,659B gzip. Full upload/setup2,265.988ms, parallel facts360.245/readiness553.857ms, commit720.792ms. No proven safety-required round trip makes 1.5 s impossible; no claim of unavoidability. Two boxed code-sharing experiments increased full image by 544 B / 352 B and were restored.Fresh immutable
9ab4329cproduction rebuild, identical standard Cargo release command/compiler 1.99 / opt 3/LTO/one codegen unit/strip/unwind, no overrides:|Architecture|production B|branch B|delta B|
|x86_64|20,015,792|22,749,712|+2,733,920|
|aarch64|17,810,608|20,122,136|+2,311,528|
Incoming dev
b182529fx86_64 22,275,888 B; lane delta +473,824 B. Both static/stripped/no PT_INTERP, no staging authority; x86 protected key memory passes 8 MiB cap. Unchanged size caps exceeded by 545,408 B / 483,392 B. Positive deltas are not waived or claimed unavoidable. D243 remains HOLD.Final scripts/verify fast output verbatim (full raw log attached):
Evidence:


verify-fast-2f021459.log
acceptance-raw.tar.gz
acceptance-summary.txt
sizes-final.txt
netbird-raw.tar.gz
netbird-regression.log
packets-150.log
packet-headers.txt
ts-final.log
cross-build-final.log
ownership-shell.log
ownership-shell-green.log
test-only-review.txt
author-checklist.txt
native-final-prefix.log
nine-final.txt
browser-functional-replay.log
Prior-round raw acceptance receipts already uploaded (these are superseded by current D225 full-image acceptance):
acceptance-raw-final.tar.gz
acceptance-summary-final.txt
acceptance-final-console.log
round3-final.log
cross-final.log
Production build and commit-gate provenance:
x86_64-receipt.txt
aarch64-receipt.txt
production-build.log
ownership-commit-gate.log
Setup-status adapter remains /v1/bootstrap/setup/ws; retained then live host/window/session events, install/join/start/ready running/done/failed, safe reason + retry_available. Contract unchanged at contracts/host-setup.md. No UI-branch merge.
Controlled diagnosis completed: 3 alternating fresh targets each on VM 208 at 150 ms. Same controller and private session image. Full 22,749,712 B image: upload/setup p50 2,214.338 ms, fully joined 5,948.790 ms, prompt after sign-in 1,800.553 ms. Reduced 11,337,248 B image: upload/setup 1,816.416 ms, joined 5,467.050 ms, prompt 1,758.267 ms. All 6 integrity/digest proofs pass. Difference in upload medians 397.922 ms, joined medians 481.740 ms. This supports a payload-size contribution; compiled roles and per-sample load differ, so total joined difference is not attributed solely to bytes. Reduced image is diagnostic ONLY and cannot replace D225 full-image acceptance.
compare-raw.tar.gz
compare-summary.txt
node-diagnostic-artifact.txt
Cleanup: completed native/static build targets, clean owned baseline checkouts, short owned fixture roots and owned VM resume directory removed; frozen executables and raw evidence retained. Compare, acceptance and NetBird owned containers removed by their runners. Checkout clean, head pushed. No caps, decisions, production state or owner fixtures changed. Stopping for the requested break-it review; full branch remains HOLD, not accepted or done.
Resuming addhost-instant at
2f02145920; clean worktree, merge base with origin/devb182529fcc. Reading #189/#169/#190 history and retained receipts. Current brief: measured browser/server/target click-to-painted-prompt timeline first, then failing benchmark and one-flight/overlap reductions, <=16KB lane delta against merge base, unchanged setup stream and safety fixes; final 10+10 integrity acceptance and iterate/fast gates. No deployment.Timeline audit at
2f021459: existing receipt measures the native terminal actor, not painted browser output. No fresh latency/paint claim yet. Existing early path already keeps check channel (install.rs check_channel/stage_session_on) and sends declared payload before its transfer receipt; these are not remaining missing optimisations. Added diagnostic-only live-fixture composition of production bootstrap router, /v1/ws and shipping TerminalClient.onPaintedOutput; owner-cookie stdin only, fixed SSH-message/target-stage clocks and TCP headers only. Fresh controller build pending under current shared-host load; setup event contract unchanged. Fetched origin/devb0a3799c, read D272; implementation not yet adopted here and no size-cap change.Diagnostic-only browser route now works on VM208: real bootstrap WebSocket, leg-A /v1/ws, shipping TerminalClient paint callback, generated one-use SSH fixture. 59 input bytes acknowledged before/after handover; installed FULL image SHA256 matches. Debug-controller run: SSH sign-in end 1791982190067.519 ms; publication 1791982191693.964; painted prompt 1791982191998.300 (1930.781 ms after sign-in). Renderer open 1791982191817.600; a second /v1/ws upgrade occurred at 1791982191922.714. Investigating that reconnect; optimized controller is building for the required median timeline. These diagnostic numbers are not acceptance. Fixed collector's JSON whitespace assumption and excluded embedded data-URI asset bodies from resource-path logging. No shipping behavior or setup contract changed.
D255 baseline timeline before production changes, source head
2f021459. VM208, 4 CPU/7GB, optimized controller (standard release recipe; server opt=s), shipping full/session images unchanged. Ten fresh 150ms targets: painted prompt after authenticated SSH p50 1852.438 / p99 1948.455 ms; click→paint p50 3205.150 ms; fully joined p50 5753.700 / p99 5930.000 ms. All ten acknowledge 59 input bytes across handover and verify the installed FULL image digest. p50 is the median; p99 is nearest-rank maximum of ten. Five runs requested a browser 4000 liveness reconnect before paint. Budget 1500ms is RED in 10/10 runs.Representative run 7 is nearest p50 (1839.715ms after sign-in; 3183.100ms from start). B=shipping browser client/renderer, S=server/its SSH client, T=target shell. Shared epoch origin B start=1791985262572.500ms. Offsets below are milliseconds; overlapping rows must not be summed. Browser calls the shipping provisioning client directly, using the existing mux and TerminalClient; this isolates transport/rendering from the separate designer's wizard/camera animations. Leg A is local; only the owned target namespace has 75ms delay each way. SSH host key is pre-pinned in the generated fixture; no trust bypass.
After sign-in the measured sum is check/channel 447.392 + invitation 1.934 + initial setup/upload/shim 1156.296 + authority publication 0.984 + browser delivery 2.338 + renderer prep 58.900 + attach 154.200 + decode/paint 17.600 = 1839.644ms (rounded endpoint bookkeeping differs by 0.071ms). No safety-required extra RTT has been proved to make 1.5s impossible. Target-side SSH packet clock collection is queued as a supplement; the table currently has both directions at the SSH client and target local execution clocks.
Next cuts are evidence-driven: overlap renderer preparation with sign-in; measure removal of the unused multithread scheduler from the private CURRENT-THREAD image (cargo feature closure currently contains rt-multi-thread); measure the per-100KB payload effect rather than assuming a linear slow-start cost. Setup-status contract remains unchanged. Current x86 full-image lane delta versus merge base
b182529fis still +473,824B (22,749,712 vs 22,275,888), not waived; new diagnostic code is test-only.timeline-release-baseline.tar.gz
timeline-baseline-summary.json
Build-host resource finding: session-only scheduler cut failed writing serde_core dep-info/fingerprint with OS error28 (No space left on device); exact log retained at evidence/session-scheduler-cut-build.log. Follow-up df: root95%,24GB available, inodes13%; owned target/debug23GB, native release968MB, vt-tools1.2GB. Removed only completed owned native release output after preserving optimized/debug controller executables and receipts; retrying static session build. No other job files/processes, owner fixtures, size caps or decisions changed. The current candidate moves rt-multi-thread to its consuming roles/tests; private current-thread feature closure is RED before/GREEN after. Renderer resources now prepare through the existing TerminalClient alongside sign-in, using the existing selected font; stage contract unchanged. Neither cut is committed or claimed measured yet.
Current iterate found four newly reported cargo-deny advisories on the already-pinned rustls0.23.45: RUSTSEC-2026-0346/0347/0348/0349; every gate receipt prescribes >=0.23.46. Updating that existing dependency to the minimum fixed patch, separately from D255 changes; no crypto/transport engine substitution, advisory ignore or cap change. Original gate log retained. Native full-suite test row was stopped after two serial-auth cases stalled together; exact frozen-artifact isolated replays pass1/1 each (mesh lifecycle and 16-device recovery). Full unchanged assertions now rerun with one test thread; no118 waiver claimed.
Perf measurements are waiting on the shared VM lock. At 2026-10-14 15:03 UTC,
lslocksshows holder PID 935095 (flock, elapsed 01:07:00), with queued waiters including this lane. Both-end packet supplement, renderer/payload paired comparisons and final acceptance need a measurement slot; no foreign process was stopped and no build-host timing substituted. The detailed baseline table was posted first; native functional checks and static rebuilds continue independently. The orchestrator needs to coordinate release of the occupied measurement slot if these 3600-second waits expire.Completed resource preparation in two atomic steps:
bf8db8d1adds the shared terminal preparation API;aafc0da7starts it during add-host sign-in using the selected font and existing on-demand import. Both pushed. TypeScript check, production build, all 195 source tests (11,058 assertions), and JavaScript/font transfer budgets pass. No terminal/socket is allocated by preparation. Setup-status contract unchanged. Performance comparison is still queued on the perf VM; baseline 10/10 runs fail the 1500 ms budget (table above).Same-recipe x86_64 stripped static rebuild complete (rustls 0.23.46): merge base
b182529f22,275,888 B; current 22,751,216 B; delta +475,328 B. Previous lane image was 22,749,712 B, so the security patch changes it by +1,504 B. The <=16 KB merge-base target remains unmet. New timing diagnostics are test-only and absent from shipping images. The private current-thread session image is separately reduced by 136,488 raw /63,149 gzip bytes; full roles retain their multithread scheduler. Aarch64 rebuild and perf measurements remain running/queued.The both-end trace exhausted its 3600-second lock wait (exit 1, no run collected). At 15:10 UTC holder PID 935095 still owns the lock; its elapsed time exceeds an hour. Queued a new 3600-second wait. Paired comparisons and the gate RTT self-test remain queued. No other process/lock changed; no performance result substituted. Measurement acceptance remains blocked on the shared slot, with functional tests continuing.
Current native server run: 443 passed /1 failed /46 ignored. The remaining setup-terminal oracle reproduces 0/1 in isolation at auth/tests.rs:1488: after duplicate-input replay, logout reads a binary frame it expected to be Close. Instrumenting only public frame type/stream/Ack sequence to identify the queued frame; no timeout/assertion waiver. Mesh headroom completes and assembles all 10,000 nodes +30,000 rules across 88 pages, max frame64,612 B, total5,632,444 B, shared socket Pong=true; current isolated replay also passes1/1. Production-build browser run:33/40 pass; seven cases exhaust existing test/readiness deadlines (several before opening the wizard). Replaying those seven on the same unchanged source; retained setup/retry and both phone/desktop captures pass. Clippy workspace/all-targets and fmt pass.
Size experiments (not adopted): package-local infraternal-agent opt=s saves only2,112 raw /224 gzip bytes in the private image, rejected as insufficient. Package-local infraternal-plugin-ssh opt=s produces x86_64 22,550,256 B versus standard22,751,216 B (-200,960); merge-base delta remains +274,368 B. Standard aarch64 is20,124,600 B; matching
b182529faarch64 baseline is building. No cap changed. All first lock waits and the both-end retry expired without a measurement; old holder935095 has now exited, and matched current-controller comparisons are queued behind the next short run. No build-host timing substituted.Matched pre-change2f021459 preview build reproduces all seven browser failures under the same original deadlines (current replay also7/7 fail). Earlier frozen-asset baseline is not used: its preview recipe was not matched. Harness inconsistency: cases allow45 s for canvas/discovery readiness but their whole-test watchdog is30 s; initial view readiness also has an unrelated5 s poll limit. Fixing only the functional harness watchdog/readiness alignment, retaining all behavioral assertions and operation-specific waits. These are functional checks, not performance budgets; D255 measurements remain VM-only and unchanged.
timeline-both-ends-current.tar.gz
Both-end packet supplement to the first timeline. Current-controller run near the original median: after-auth paint1864.755 ms; click→paint3222.800 ms; click→fully joined5838.800 ms. VM208, fresh150 ms target, original session image, current standard full image. All59 input bytes and full-image digest verified. Epoch origin1791996223877.7998 ms is browser sign-in. Offsets below are milliseconds; packet filtering stores no payload/credential.
S local: sign-in19.994→1358.045; check1358.067→1835.422; invitation1835.516→1837.365. T admission1709.011→1754.792 (7609 B); probe1757.100→1759.492 (1596 B). T setup2129.287 (5629 B), 291.922 ms after invitation; payload decode/hash2131.233→2854.495 (723.262 ms), session2,601,648 raw/1,141,757 gzip bytes. Shim exec2856.530→S publication2987.825; publication ends2988.745, B receives2990.900; renderer open3054.600; paint3222.800; ready5838.800. Checks/hash/ownership are required; renderer can overlap; payload and initial-script pacing are avoidable bytes/work. Capability probe adds no separate RTT.
The short5629 B setup program waits291.922 ms before execution; the VM default pipe capacity is65,536 B. Inference to test: a short setup program can coalesce with payload into a larger authenticated SSH data packet, delaying execution during TCP pacing. Limiting only the retained prompt channel input pipe should expose the script earlier, preserving every trust/hash/ownership check. No benefit is claimed until measured.
timeline-both-ends-current.tar.gz
Pushed atomic steps:
f623a43creplay/logout oracle (actual WebSocket closure under unchanged1500 ms; ordered duplicate processing);077a5c4c lifecycle clock control using existing#118 helper;63915249 live heartbeat replies while input Ack is held. Auth/socket red137/2 to green139/0; workspace Clippy/fmt pass.04b135dfnarrows private Tokio feature closure (session raw-136,488 B/gzip-63,149 B);072a1955 minimum rustls security patch (all four cargo-deny groups pass);17873593 shares renderer initialization. Real renderer loads3→1 for three callers, and failure retry shares one successful later load;195 source tests pass. Functional harness alignment is committed separately:39/40 broad cases, remaining initial-canvas timeout passes in isolated fallback replay. Full native rerun is running. Git default connections stalled; fresh strict connections pushed these commits; no host key/config/credential changed. Setup event contract unchanged.Diagnostic steps committed/pushed separately:17da6403 Rust browser/controller clocks;75c939ef real browser paint +59-byte handover proof;eec51138 target local stages/program bytes;98bd6cfb filtered SSH clocks +trusted pipe experiment;8fab3aa6 both-end collector/budget;020ef1cf execution/evidence documentation. Both-end baseline raw archive/table attached above. Private scheduler feature closure04b135df, security patch072a1955, renderer initialization17873593, browser functional harnessd5aaaecb also pushed. Current standard size vsb182529f: x86_64 22,751,216 vs22,275,888 (+475,328 B); ARM20,124,600 vs19,704,712 (+419,888 B). SSH opt=s candidate saves200,960 x86 bytes but still+274,368; not adopted pending performance. Full native rerun and milestone iterate on head020ef1cf are running; VM comparisons remain queued between short runs.
Fresh production-build browser run on current source:40 passed /0 failed across wizard, retained setup/retry and phone/desktop start-fresh captures. All behavioral and operation-specific assertions retained; only the functional journey/readiness watchdog mismatch was repaired in
d5aaaecb(baseline comparison recorded above). Renderer single initialization17873593 is included. Strict TypeScript and195 source tests pass. Performance acceptance remains separate and VM-only; no browser runtime duration from the build host is claimed as latency evidence.First matched warm-original150 ms run (current controller/full image, original session): prompt after-auth1895.262 ms, click3318.800 ms, all59 bytes and full digest pass. Preparation startsB1791998230492.100, done0534.400, well before S sign-in end1915.638. Publication B3579.500→AttachTerm3627.600 (48.1 ms constructor/DOM work despite resources ready); socket close-request4000 at3645.000, new Hello3719.600, paint3810.900. This is a real pre-paint liveness reconnect, not missing payload proof. Five original baseline runs had the same reconnect. Current client uses measured RTT+50 ms liveness; constructor work is near that margin. Not changing that separate contract/budget here. Paired smaller-image and8KB prompt-pipe cases are queued; warm-up alone has not met1500 ms. Raw warm-original1 log/archive retained.
Native fixture fixes pushed:
f6b23e60consumes only the priming Reconnecting notice under the unchanged absolute 100 ms output deadline;05557392freezes watch read authority while advancing the stored observation expiry. The iterate reds were an unexpected Reconnecting frame and watch error/resync mismatch, not proven production latency regressions. Each exact regression now passes (1 test each); touched workspace clippy and fmt pass. A milestone fast run also passed all 444 server tests before these fixes. Full head gates are still pending; no failures waived. Setup-status contract is unchanged.Private connector cut pushed at4c4ed11f: raw2465160→2434024 B (−31136); gzip1078608→1060498 B (−18110). Private recognition regression RED before guard; full/private dispatch tests and clippy GREEN. No setup contract change.
150 ms RTT byte-cost experiment: three fresh matched trios, original image / runtime-cut image / same cut image with102400 B deterministic public incompressible trailer. Every run verified59 input bytes and installed full digest. Added-byte decode+hash streaming costs: 14.687 ms, 0.395 ms, 16.909 ms; median14.687 ms/102400 B =14.343 ms/100000 B. This is a local comparison in one slow-start region, not a linear cost guarantee. End-to-end prompt differences were 70.672 ms, -6.174 ms, -128.634 ms: not stable evidence of total prompt improvement. All9 warmed runs also trigger client4000 reconnection before paint; publication→paint spans184.7–495.5 ms. Added control/attach clocks will distinguish renderer stalls from socket serialization before any liveness change. Current1500 ms budget remains RED.
timeline-matched-trios.tar.gz
matched-payload-cost.json
private-connector-size.json
Milestone fast invocation started at020ef1cf (before the two fixture fixes and later private-image work). All444 server tests passed; Rust fmt, VT fmt, compile-out, clippy, tests and cargo-deny passed. Static-size row failed; no failure waived and caps unchanged. Shipping standard artifacts: x86_64 22751216 B, aarch64 20124600 B. Same-recipe merge-base
b182529fcc: x86_64 22275888 B (+475328), aarch64 19704712 B (+419888). These remain above the requested16 KB delta and are not accepted. A final head gate is still required after the ongoing cuts.Verbatim milestone summary:
D243 private-image candidate before implementation: the Rust VT boundary is explicitly callback-free (docs/S1-VT-GHOSTTY.md:48–52), yet the retained native log sink can format warnings for a callback never installed by this executable. Measure an explicit session-image archive mode that omits logging callback setter/convenience exports and drops log formatting. Keep the ordinary archive API and logging behavior unchanged, preserve the pinned engine, parser checks, initialization/resize errors and panic/unwind behavior. Include mode/source in native cache identity. This is private-image packaging, not an engine choice; adoption depends on size and existing real-engine fidelity/recovery tests.
D243 next measured packaging candidate: set only native lib.root_module.strip=true for ReleaseFast, matching the existing stripped final Rust release image. Keep Debug/ReleaseSafe unchanged; do not alter omit_frame_pointer, unwind_tables, parser checks, engine, or Rust panic=unwind. Exact diagnostic tradeoff: native Zig panic still prints its message and terminates as before, but emits debug-info-stripped instead of trying its raw-address frame-pointer fallback or external-symbol lookup. The current final ELF already has no .debug_* sections, but those fallback paths can still be useful; they are not claimed to be dead. D5 containment remains Rust unwinding, unchanged. Measure before adoption and verify native fidelity/recovery and retained unwind sections. Private callback-free candidate independently saves195488 raw/57138 gzip B; not yet accepted pending those tests.
c3811c3cpushed: SESSION prefixes the dispatcher first write, rather than publishing before the dispatcher can greet. Two tests RED before fix; three GREEN afterward (including40000 binary bytes across reads); private all-target clippy GREEN. Three150 ms fresh targets verify59 input bytes/full digest; prompt1916.683/1869.271/1880.120 ms remains RED. No contract or liveness-budget change.Native candidates are uncommitted pending tests: callback-free private archive raw2433896→2238408/gzip1060426→1003288; ReleaseFast native strip then raw2076808/gzip930812. Diagnostic tradeoff as proposed above; .eh_frame and .eh_frame_hdr remain. Release tests/full shipping builds ongoing. Shared-source install functional suite76/0 and clippy passed before the last environment-source cleanup line; VM source comparison preflight caught an omitted remote feature in the diagnostic controller and ran no measurement. Rebuilding with remote. Independent read-only review found no correctness/security defect in shared guards, greeting prefix, or native modes; it does not substitute for gates/acceptance. Full head iterate/fast and final10+10 acceptance remain outstanding.
Extending the measured callback-free archive candidate to the full CLI image as well as its small session image: both use the same documented callback-free Rust VT boundary and declare no native configuration setters. The full executable is also the supported fallback session payload. Direct standalone shim archives still retain ordinary upstream callback APIs unless the image feature is selected. This removes only unavailable callback plumbing/formatters from executable images, not parser effects or errors. The mode remains explicit and keyed in the build cache; test its native ABI export distinction and all real-engine recovery/fidelity paths. This is a candidate until shipping sizes/tests/VM acceptance pass.
D255 follow-up, HEAD
c3811c3c; candidate changes not yet accepted. Three single-run VM comparisons (not acceptance percentiles): old private image 2,433,896 B / gzip 1,060,426 B => post-auth prompt 1,982.282 ms; native packaging candidate 2,076,808 B / gzip 930,812 B => 1,783.281 ms; same candidate with fixture-only SSH compression => 1,773.229 ms. All three verified 59 acknowledged input bytes and the installed full-image digest. No compression default changed.Both-end/browser clocks identify an additional renderer stall: old-image run renderer DOM open took 133.8 ms; a control Ping immediately before publication received its Pong after that synchronous stall, correctly triggering D265 reconnect. Native-cut run DOM open was 22.4 ms. Preparing a disposable renderer during sign-in is under measurement. Review found two vendor document-listener leaks; pinned source patch + rebuild fixes them. Real browser RED: 2 retained document listeners. GREEN: 0 listeners, 1 renderer open, 1 WASM load for 3 callers, 0 sockets, focus and body children preserved; retry after synthetic failure also green. Reproducible functional oracle: docs/verification/addhost-breakit2/renderer-prepare.mjs (explicit unique TRACER_TEST_PORT). No liveness thresholds changed.
Full shipping sizes, shared-source script comparison, final acceptance and final head gates remain pending.
renderer-listener-red.log
renderer-listener-green.log
renderer-listener-retry-green.log
image-comparison-summary.json
timeline-image-comparisons.tar.gz
Completed native packaging step
0c78d3f3(pushed): explicit executable-image VT mode omits uncallable logging configuration exports; ordinary archives retain them. ReleaseFast strips native diagnostics consistently with the final stripped executable; D5 unwind behavior, native panic message/termination and parser checks remain. Private image is 2,076,808 B, gzip 930,812 B; full x86/ARM deltas are still building and are not claimed.Per-step validation: workspace all-target clippy and fmt passed; 79 shim unit tests and 2 detachment tests passed; compiled archive API check passed; release-image public mouse probe passed. The initial integration mouse invocation lacked Zig in PATH; rerun with the required tool passed. Supplemental independent allocation probe reproduces the four documented upstream failure points (2–5) identically in ordinary and image modes; existing Rust retirement/reconstruction tests pass. Three VM handovers above all verified 59 input bytes and full-image digest.
native-archive-api-green.json
private-release-mouse.log
private-release-allocation.log
ordinary-release-allocation.log
Completed renderer steps:
238a8411fixes vendor document-handler cleanup and rebuilds pinned artifacts;d05fa019prepares a disposable inert renderer during sign-in. Both pushed. Standalone real-browser oracle: old artifact retains 2 document listeners; new artifact retains 0 after ordinary disposal AND after injected late opening failure. TypeScript + 195 source tests/11,058 assertions pass. Resource oracle: 3 callers share 1 open/load; no sockets, focus loss, DOM residue or listeners; resource refusal retries successfully. Read-only code review reports no remaining lifecycle findings.Three fresh 150 ms VM runs, smaller image + actual renderer preparation, post-auth prompt: 1,747.412 / 1,745.299 / 1,801.995 ms (still RED vs 1,500). All 59 input bytes and installed full-image hash verified. First two published-renderer DOM opens were 2.7 / 4.4 ms; no pre-paint 4000 retirement. Shared-source script/controller comparison starts next; final 10+10 acceptance and shipping size deltas remain pending. Setup-status contract unchanged.
timeline-renderer-warm.tar.gz
renderer-disposal-partial-green.log
renderer-disposal-baseline-red.log
renderer-prepare-reproducible.log
Rejected candidate, no commit: exporting public lock/cache helper source reduced the continuation script (5,629 to 3,241 B), but grew the initial exec request. Two independently warmed paired VM batches do not justify it. Batch 1 post-auth prompt old median 1,801.363 ms / shared-source 1,876.992 ms; batch 2 with identical explicit-probe candidate old median 1,813.879 ms / shared-source 1,854.228 ms. Batch 2 check phase old median 474.493 ms / new 518.867 ms. All twelve handovers verified 59 input bytes and full-image digest. Reverted the uncommitted source-sharing change: it moves bytes to a more expensive early flight.
The explicit pre-render probe candidate also showed no stable prompt improvement and was removed; D265 code, throttle and deadlines are unchanged. Browser-side remaining optimization is avoiding disposable warm-up when a real terminal already opened; this is under functional regression testing (the saved pre-warm baseline passes the three cases that failed after introducing redundant warming). No functional test timeouts were extended. Shipping static artifacts are being rebuilt without the rejected script candidate.
Completed frontend regression step
bade12eeand timeline clocks978815fd(pushed). Three failure-screen browser checks failed with redundant disposable warming; the saved pre-warm source passes all three, and the reuse fix passes all three with identical assertions/deadlines. A successful real terminal now marks its font warm in the existing singleflight map and supersedes optional work waiting on resources. Full 40-case rerun is in progress on the accepted source.Source checks: 194 functional tests pass. The broad initial source command also included an existing graph timing assertion; that row is now run separately on the locked perf VM and passes. Shared-host timing is not acceptance evidence.
Iterate finding: remote_client::tests::tcp_proxy_preserves_spacing_between_packets_already_in_flight fails its wall-clock interval assertion under build-host load. No waiver: replacing the interval measurement with a paused-clock assertion against the actual queued flight scheduler plus explicit consumption barriers. The existing real-socket bounded flood/half-close test remains. Native gates and final static builds continue.
renderer-final-browser.log
renderer-baseline-three.log
renderer-existing-three.log
renderer-functional-source-tests.log
mesh-source-timing-vm.log
Completed test fixture step
435ae868and SSH release profilee91c18af(pushed). Proxy flight spacing now uses actual relay consumption barriers and a paused clock, preserving exact 20 ms model spacing. A deliberately serialized negative control fails at 150 ms; after restoration all 100 native library tests pass, with the 14 existing opt-in diagnostics ignored. Crate all-target clippy and workspace fmt pass. No shipping proxy behavior changed.Accepted x86 full image: 22,194,224 B; same-recipe merge-base
b182529fcc: 22,275,888 B; delta -81,664 B (within the <=16 KB requirement). Accepted session image: 2,077,064 B. One fresh LAN preflight using these images: post-auth painted prompt 236.322 ms, all 59 input bytes acknowledged and installed full-image hash matched. Not the final 10+10 distribution. ARM build, rebuilt final controller, final acceptance and final head fast gate are running/pending. Caps remain untouched.Rejected script experiment receipts are attached: moving helpers shrank the continuation 5,627→3,241 B but enlarged initial exec 10,175→11,908 B and worsened both matched median check/prompt batches. It is absent from accepted source.
proxy-virtual-clock-negative.log
proxy-virtual-clock-restored.log
full-final-size.json
session-final-size.json
rejected-comparisons-summary.json
timeline-rejected-comparisons.tar.gz
Head fast gate finding (
e91c18af): two real-time QUIC cases failed on the shared build host: healthy_mesh_probes_do_not_inherit_the_previous_paths_rtt and component_down_activates_the_warm_fallback_without_a_new_connection. The first checks a 700 ms healthy interval; the second requires a healthy parked path while the production liveness guard runs. No waiver or deadline increase: the owner requires physical timing checks on the designated perf VM, so the final gate runner will execute these exact cases there under the shared lock and aggregate their exit codes with the other native cases. A zero-test selection or missing VM proof will fail the gate. The fixture directory will accept the existing INFRATERNAL_TEST_STATE_ROOT so VM writes remain under this job's designated bench path. Runtime and protocol code remain unchanged.Final-series finding at
e91c18af: the fresh sequence stopped at WAN target 6/10 because the real setup readiness step failed (669.504 ms), before integrity proof. This attempt is retained, not counted as a successful sample. Earlier 5 WAN + 6 LAN targets proved 59 acknowledged bytes and installed full SHA-256; LAN #6 was slow (587.289 ms post-authentication). Added fixed browser-boundary retention on failure to distinguish setup refusal from renderer/socket failure without recording terminal contents, request bodies or exception text. Investigation continues.Both same-recipe static binaries now shrink against merge base
b182529f: x86_64 22,194,224 B vs 22,275,888 B (-81,664 B); aarch64 19,633,384 B vs 19,704,712 B (-71,328 B). The final fast run still fails two physical QUIC timing cases on the shared build host plus the unchanged static-cap ratchet, which rejects even smaller binaries. Caps remain unchanged as instructed. The two exact physical cases will run on the locked perf VM with their existing assertions; failures remain fatal.Steps pushed:
b1f6e8e0retains only safe browser-boundary clocks on diagnostic failure;389ef212routes the two exact physical QUIC cases to the locked perf VM, with executable hash proof, one-case/nonzero-sensitive result checks and aggregate failure. Both VM cases pass unchanged; the remaining 98 pass from Cargo’s package working directory. A manual invocation from the checkout root initially failed 10 fixtures that require Cargo’s package cwd; the corrected run is attached and the final Cargo gate preserves its ordinary cwd. Runner counterexamples refuse changed image bytes, zero/11 tests and false child success; a review found and corrected hash parsing before commit. Clippy/fmt pass.Production browser fixtures on the locked VM: 1 instant + 2 start-fresh capture + 37 wizard = 40/40 pass, assertions/timeouts unchanged. Screenshots are attached for independent visual assessment. Static sizes are -81,664 B x86_64 / -71,328 B ARM against
b182529f. Final head fast is running; caps remain unchanged.final-static-sizes.json


functional-vm-accepted-addhost-instant.log
functional-vm-accepted-addhost-start-fresh-captures.log
functional-vm-accepted-addhost-wizard.log
native-local-cargo-cwd.log
native-timing-runner-oracles.log
Fresh-series failure is not waived: final-150-06 refused readiness before byte proof. Independent code review identified a matching race: remote_runtime.rs admit_inner queues replacement and publishes installed ledger authority before embedded.rs run_actor has consumed/greeted it; bootstrap_remote.rs actor currently checks only the ledger flag. A readiness open can still go to SSH, or be refused by retain_fallback while the queued installed link takes over. Successful fresh retries do not discharge that refusal.
Proposed additive internal fix: query the actor’s currently greeted transport through its existing bounded command queue before returning it for readiness/facts. The query performs no remote action and creates no session. Old SSH returns false; installed transport returns true only after ordinary greeting. Keep the existing 30 s admission bound/50 ms wait and exactly one readiness open. Preserve fair select scheduling and refuse later real transport failures; do not replay an unknown open. Add an original-behavior failing test before the fix. No setup event contract changes. D255.
The final native run moved the two physical cases successfully to the VM, but another real-time case parked_fallback_recovers_mesh_and_keeps_control_actions_usable timed out at remote_server_path_tests.rs:818 on the shared host. Run the physical server-path suite on the designated locked VM unchanged; retain nonzero outcomes, rather than widen timing assertions or waive a case.
8043e77cpushed: ledger admission no longer returns an actor still using SSH for facts/readiness. The actor’s existing bounded command queue answers a local active-route query; no remote frame, session, retry or idle task is added. The original ledger-only behavior deterministically fails with “ledger admission released the old SSH route”. The fixed regression, three SSH/TLS/fallback cases and six private-readiness cases pass; workspace clippy/fmt pass. Independent review found no blocking issue. Event contract unchanged. Release controller/static rebuilds and iterate are running; fresh acceptance will repeat on those artifacts.The prior artifact series is preserved, not accepted: 10 LAN + 11 WAN attempts, with one WAN readiness refusal. The other 20 verified all 59 bytes and installed digest. Percentiles below include only successful proofs; failure remains in denominator. p99 is nearest-rank maximum with ten successes.
Both prompt targets remain red. A zero-delay browser yield is rejected: paired three-sample WAN medians 1789.981 old / 1797.147 yield, all byte/digest proofs pass. A diagnostic Pong-wait experiment is still running, absent from production. Its first sample avoids reconnect but output still follows Attach/Credit by ~150 ms; therefore reconnect time overlaps the original SSH attachment flight, rather than constituting an additional serial RTT. No D265 deadline/throttle change.
8df91c25pushed: all ten nonignored physical QUIC server-path cases run unchanged under one short locked VM run; 10/10 pass, local remainder 90/90 passes. Every exact case and exit is checked, including negative runner proofs. This resolves execution-platform instability without waiving cases.fresh21-and-yield6-raw.tar.gz
final-acceptance-summary.json
native-suite-placement.log
native-suite-runner-oracles.log
admission-route-red-exact.log
admission-route-current-tests.log
admission-route-clippy.log
603a47be(pushed): native Snapshot repaint regression proves the old CRLF output predicate rejects correctly painted rows.5f2be250uses the existing screen reader only after paint, exact complete rows, unchanged 59-byte acknowledgement and installed-image SHA. Faithful repaint passes; corrupt/extended rows refuse. Earlier live timeout remains failed, not retroactively waived. Target capture now stops after the three TCP handshake packets; partial-start/nonzero-child cleanup oracles pass.ITERATE gate passed. This is not the final gate: run scripts/verify fast before reporting done.Fresh current-artifact acceptance stopped on audit-0-06: zero input acknowledged and no prompt; initial attach was sent, then client watchdog retired the socket, reconnect closed without Welcome. Investigating; all attempted cases stay in the denominator. Prompt budgets on completed cases still fail. Final fast is running.
real-engine negative control, green corruption checks, capture cleanup, iterate verbatim.
Timeline update — current production fix
8043e77c, diagnostic head603a47be. Acceptance is not green: audit-0-06 has no prompt/zero input acknowledged; investigating. This table is the median of the first five WAN150 integrity successes (audit-150-03), not a substitute for the requested final 10+10 distribution. Epoch origin/browser click = 1792011728067.700 ms. All processes share the VM clock. Each target is fresh, lock held only for that case; 4 vCPU VM, actual per-case load is in the attached raw log. The fixture invokes sign-in directly after page load and an established browser socket; wizard animation and initial page TLS are outside the click metric.This case is 1,890.516 ms after sign-in / 3,334.400 ms after click. The post-auth decomposition is check 515.520 + invitation 2.118 + private continuation/image/shim until publication 1,154.160 + publish 1.261 + browser delivery 3.737 + publication-to-paint 213.600 ms. It is not a 1.5 s result.
Size versus
b182529fcc(same stripped-static recipe): x86_64 22,200,816 versus 22,275,888 = −75,072 B; aarch64 19,635,704 versus 19,704,712 = −69,008 B. Private-image compression fell from approximately 1.141 MB to 932,008 B. Three paired +102,400-byte incompressible trailers changed the streaming interval by +14.687/+0.395/+16.909 ms: median 14.343 ms per 100,000 additional compressed bytes near this window, not a universal raw executable-byte conversion. Flight boundaries can cost a whole RTT.Rejected matched experiments remain in evidence: common shell-helper source sharing made first admission slower; task-yield before opening did not improve median; awaiting the next Pong removed reconnect but changed median only 1,800.301 → 1,797.317 ms (three pairs), because original attach/recovery flights overlapped. None shipped. Remaining channel-open and continuation scheduling are still architectural optimization candidates; this evidence does not establish a security-required 1.5 s lower bound.
Current-artifact 10+10 acceptance (
8043e77cproduction,603a47bediagnostic) is RED: LAN 9/10 exact 59-byte/full-SHA proofs, WAN150 10/10. audit-0-06 never painted a prompt, zero acknowledged bytes, first attach sent then1005 after Resume; later upgrades fail before Hello. No attempt excluded. Successful-only post-auth prompt p50/p99 LAN487.824/697.183 ms; WAN1874.720/2058.861 ms. Joined LAN1536.771/2007.784; WAN4654.489/4716.317 ms. p99 nearest rank (maximum at n10); LAN n9 successes is explicitly conditional, not acceptance.Independent inspection found periodic authority reads suspend an operation that may hold the fixture sole pool connection; a controlled virtual-clock regression is being added before fixing. A cancellation destroying the single-connection in-memory database could explain later HTTP refusals, but raw receipts do not prove that part. Adding fixed schema/session-existence flags on failure; no SQL errors, content or authority logged. Existing deadlines remain unchanged.
All20 raw receipts, all attempts and statistics. Detailed median-run table was posted above and attached.
Further controlled finding (uncommitted candidate, not shipped): polling a connection-owning operation alongside its pending authority read removes the self-deadlock, but the inner wait must also prioritize committed revocation and expiry. Operations include socket sends; rejecting a later return cannot undo transmitted bytes (contracts/keywrap.md286–300). Two new deterministic virtual-clock cases fail the first candidate with “periodic authority read polled a send-like side effect after authority interruption” (0 passed2 failed). No deadlines changed.
The tightened candidate freezes operation polling on revocation/expiry, finishes the SAME bounded read, then reads fresh authority before resuming. Its six-case regression/security rebuild, clippy and release rebuilds are running. No candidate commit until those pass. Prior four-case green alone is insufficient and is not acceptance.
Two side-effect negative controls, original self-deadlock negative control. Prior fast native rows completed without failures; static step cancelled because source changed, so it is not the final gate. Final fast will run on committed HEAD.
b0789b4apushed: fix the controlled periodic-authority self-deadlock while preserving revocation/expiry priority and the same25ms read cap. Original wrapper:1 failing connection-owner case. First uncommitted candidate:2 failing send-side-effect interruption cases. Final:6/6 deterministic virtual-clock cases pass; fmt + all-target server clippy pass; independent security review no further findings. Existing clock reused once, no new dependency/pool/timer/task. Event contract unchanged. Rebuilding optimized controller and both stripped static images; acceptance will run again on these artifacts.Six-case proof, final clippy. The earlier audit20 result remains19/20 integrity proofs and RED prompt budgets; no retroactive waiver or attribution of the full live failure to one controlled regression.
Measurement finding/proposed additive fixture correction before implementation: timeline_ssh.py interposes a Python interpreter (json/pathlib/subprocess imports) before every mux SSH command, and ADDHOST_FAST_STEP_PROFILE adds target shell/date wrappers. These are absent from the shipping transport. The profiled traces remain evidence, including every failed prompt budget; their observer cost must not be called a security/protocol lower bound.
Add explicit capture=0 mode to the same collector: keep the identical production browser first-paint timestamp, exact59-byte/painted-row proof, installed full SHA proof, same budgets and server authentication epoch. Use the normal generated strict SSH wrapper and unwrapped target shell, assert those paths, and retain controller output. Run matched profiled/unprofiled cases, then a separately named fresh10+10 production-path series. Keep every existing attempt and report both modes; do not retry/filter functional failures. Detailed TCP/target traces remain capture=1. No product/event/SSH trust contract change; no new engine.
Step committed/pushed:
1a424f9d(production code/artifacts unchanged fromb0789b4a+ safe fixture diagnostics). Detailed tracing remains default; capture=0 verifies actual SSH and shell paths and rejects all five profiling/transport candidate settings. Independent read-only review: no remaining findings. All five refusal controls PASS.Fresh WAN150 production-path control: 59 bytes acknowledged, before/after painted rows and installed full-image SHA PASS; prompt 1,743.131 ms after completed sign-in / 3,071.900 ms after click, budget FAIL. Breakdown: check 426.289 ms, invitation 1.909 ms, continuation/image/shim to publication 1,114.833 ms, publication-to-paint 198.967 ms. Removing observers does not establish a 1.5 s result. Full production-path 10+10 and separate detailed traced 10+10 are queued; all attempts retained. Final scripts/verify fast now running on
1a424f9d.Current same-recipe stripped sizes versus merge base
b182529fcc: x86_64 22,208,432 - 22,275,888 = -67,456 B; aarch64 19,641,640 - 19,704,712 = -63,072 B. Both satisfy the requested <=16 KB delta. The unchanged absolute gate caps are lower by 4,128 B / 2,896 B respectively; do not infer a green static-size gate. Private image remains 2,077,064 B / gzip 932,008 B. Exact TypeScript gate function PASS, unchanged assertions; full output attached.capture-mode-negative.log
native-150-proof.log
guarded-x86_64-artifact.json
guarded-aarch64-artifact.json
ts-gate-6212a129.log
Remaining-latency investigation: the retained check uses exact one-byte dd reads for the 5,627-byte continuation. This preserves payload framing but makes roughly 11,254 read/write syscalls. The audit case has 299 ms from invitation completion to target continuation start; that interval also includes delivery/pacing and cannot be assigned to dd from existing clocks. Independent read-only review agrees that fullblock count=1 could safely optimize the exact reader only after a local public capability/boundary canary and fragmented/truncated/sentinel regressions; count=1 without fullblock is unsafe. I am first measuring the two exact readers on the locked VM; no product change is justified by this hypothesis yet.
Rejected latency hypothesis: five paired exact-reader runs on the locked 4-vCPU VM (load1.05/1.03/1.04), 5,627-byte body plus a 4,096-byte binary sentinel. Every sentinel digest matches. Median whole-process time: bytewise dd11.153ms, fullblock4.978ms, difference6.175ms. This cannot explain the observed299ms continuation gap, so no reader/protocol change shipped.
A separate frame-coalescing hypothesis is being tested with three matched WAN150 pairs using the existing fixture pipe-capacity experiment, same guarded controller/full/private artifacts. audit15003's script starts only after approximately32KB of encrypted channel data can arrive, although the continuation is5,627B. A smaller pipe may make SSH expose that script in a smaller data frame. This is a hypothesis, not an assigned root cause or a shipped optimization; all integrity assertions and failed results remain retained.
dd-reader-perf.log
Rejected compiler candidate: only the CLI crate switched to release size optimization through a command-line Cargo override; no manifest or engine change. Private image increased from2,077,064 to2,420,312 B (+343,248 B); gzip9 from932,008 to978,453 B (+46,445 B). It was never staged for acceptance. Restoring the normal command reproduces session-final byte-for-byte (SHAe14bed0bbe94d676c3c24d7621487d855b21cab23534371adb651819812d97e2), confirming current-source private image identity. Current shipping artifacts and branch remain unchanged.
session-cli-size-s.json
session-default-repro.json
author-checklist-1a424f9d.txt
Per-stage finding, first matched WAN150 pipe pair (same guarded controller/full/private SHA, all59 bytes and installed full-image digest PASS):
This supports frame coalescing as an explanation of the late start marker, not an additional sequential299ms cost removable from the total. Smaller frames expose the script170ms earlier but move that time into the ongoing payload interval; end-to-end did not improve. No pipe tuning shipped. The remaining two pairs are still queued; this single pair is not the complete comparison.
frame-first-pair-raw.tar.gz
frame-first-pair-stages.json
Completed three matched WAN150 pipe pairs, same artifacts and retained target traces. All6/6 input/output/full-image proofs PASS; all prompt budgets FAIL. Default median1,787.022ms versus8KiB1,821.363ms (+34.341ms). The earlier continuation marker moves into the payload interval rather than reducing total transfer time. Reject tuning; production source/SSH pipe capacities unchanged. All six attempts and raw target/client clocks attached.
frame-three-pairs-raw.tar.gz
frame-three-pairs-summary.json
Milestone1a424f9d gate: native tests, fmt, clippy, compile-out and cargo-deny PASS; static builds FAIL only unchanged absolute size ceilings (x86+4,128B / ARM+2,896B) and the shipping size-fixture check. Complete output attached. The exact TypeScript gate passes using one unchanged graph timing test on the locked VM and191 unchanged functional cases local; source digests, exact counts and child statuses required. Host graph timing is not used as performance evidence.
Next isolated size candidate: share while_authorized_operation across its three result types, retaining the current one-box allocation payload for the frequent unit operations. A typed result adapter is needed only at handshake and terminal opening; stored values may escape only after the common authorization guard succeeds, and denial/cancellation must drop them. No authority ordering, deadline, timer, external contract or private session image change. Independent read-only review supports this seam. The current gate has finished before any edit. First validate typed return/disposal behavior, then measure stripped x86/ARM size; retain only a candidate that reduces size and passes safety tests. Current20+20 runs remain baseline receipts and will not be relabeled as a new artifact's acceptance.
fast-1a424f9d.log
ts-gate-physical-vm-1a424f9d.log
Test step committed/pushed:
f5c41b2b. Extended the existing four connection-owner/watchdog fixtures to carry a session-sized owned value through the original generic public wrapper, with exact return and destruction witnesses. The original implementation passes all6 cases (including the two no-side-effect interruption cases), fmt and strict server all-target clippy. This is the behavior baseline before changing result-type sharing; no production behavior changed in this commit.guard-value-baseline.log
guard-value-baseline-clippy.log
New functional failure retained: guarded-150-08 (unchanged guarded artifacts, before the uncommitted type-sharing candidate). Both BEFORE/AFTER complete rows paint, ready arrives, command sequence advances to59, but acknowledgment remains30 for120s. Traced baseline stops at16 attempts, one failed; no replacement retry or successful-only acceptance claim. Full failed trace attached.
Two independent read-only investigations found the same concrete loss path: shim InputStateRequest can subsume a pending InputAck, and emitting State(written,accepted) advances ack_queued. The daemon liveness handler turns that State into Pong only, discarding an advanced written cursor. Thus the shell can execute input while no later standalone Ack is emitted. Attribution of this exact run remains an inference until deterministic reproduction; no wire bodies were retained.
Independent test work now covers the daemon adapter (must fail before correction) and the real-PTY shim coalescing contract in both frame orders. A correction must forward only newly advanced WRITTEN input, never merely accepted input, and keep unchanged idle probes Pong-only. No relaxed integrity assertion, deadline or extra per-idle-session database writes is acceptable. Native production-path baseline series continues on its pinned artifacts; final acceptance must run fresh after the actual correction.
guarded-150-08.log
guarded-150-08-failure-raw.tar.gz
Steps committed/pushed:
c91bf3b4shares the authorization watchdog without changing its policy or frequent unit allocation path;94c6f815pins real-PTY state/input coalescing in both orders. Same-recipe size-only artifacts forc91bf3b4: x86 22,193,680 B (82,208 B belowb182529f, 10,624 B below unchanged cap); ARM19,634,008 B (70,704 B below base, 4,736 B below cap). These pre-Ack-fix artifacts are NOT final acceptance artifacts. Obsolete controller compilation was stopped by exact owned process identities after preserving both sizes; final artifacts will be rebuilt after the correction.Ack regressions: both active frame orders reproduce lost Ack59 while Pong survives (0/2); parked-Ping recovery independently reproduces it (0/1). The corrected shared progress emitter makes all4 liveness tests pass, acknowledges written59 rather than accepted88, keeps equal idle state Pong-only and rejects a regressing58 state. The shim's8 sequenced-input tests and strict agent(remote)/shim all-target clippy pass. Independent final read-only review found no missing state-return path. The feature-enabled regression rerun and final artifacts/gates remain pending; no fresh acceptance claim yet.
Proposed additive contract clarification before editing docs: existing InputState may subsume InputAck. Its written cursor remains acknowledgment progress even when requested for liveness, including parked output; the adapter forwards newly advanced written input through existing Ack before correlated Pong, never merely accepted bytes. Unchanged fences remain Pong-only with no repeated Ack/storage work. No wire layouts, stage names, retained host-setup stream or designer-facing event contract change.
shared-x86_64-artifact.json
guard-value-shared-tests.log
shared-guard-clippy.log
shared-guard-review3.log
ack-state-red-build.log
ack-state-regression-green.log
shared-aarch64-artifact.json
ack-state-all-green.log
ack-state-detached-red.log
ack-state-shim-final-tests.log
ack-state-shim-clippy-fixed.log
ack-state-agent-clippy.log
Transport experiment proposed while final Ack-fix artifacts/gate build; shipping source remains frozen at
ee2487b5. Current production-path WAN prompt remains roughly1.7–1.8s, and the private transfer is dominated by TCP window ramp. Test whether32KiB of fixed PUBLIC stdin, sent on the retained SSH channel while its read-only admission runs, warms that window without another request/response cycle or target mutation. Remote code consumes a bounded public line before the ordinary setup-length frame. Authentication, host-key pins, admission, session-image verification and all byte proofs remain unchanged.Use an explicitly experimental SSH stdin proxy in both controls (one newline) and padded cases (32KiB), identical final artifacts and three fresh WAN150 pairs under the perf lock. Proxy overhead is matched; these are not production-path acceptance samples. Cost candidate:32KiB additional public traffic per fresh target, bounded temporary line/proxy state; no new library/engine, timer, untrusted command or global network tuning. Keep only a measured end-to-end improvement, never a shifted start marker. No product/contract change is justified yet; real implementation would require its own safety tests and final native-path acceptance.
Ack correction committed/pushed as
9a5add222; additive writer-fence documentationee2487b57(current final-build head). Feature-enabled rerun passes all4 liveness adapter regressions; default4, real-PTY shim8, strict remote agent/shim clippy and fmt pass. No retained setup-status contract change.Perf-VM shipping GNU gzip confirms old private image2,077,064B /931,900 compressed; new2,078,216B /932,439 compressed (+1,152 raw /+539 compressed). Earlier Python gzip metadata used a different encoder and must not be read as exact wire length. Final full sizes, final-head native20 and full gate remain pending.
Production-path pre-fix baseline completed20/20 byte/full-install proofs: prompt after sign-in p50/p99 LAN369.167/2,201.120ms; WAN150 1,748.014/1,798.049ms. Joined LAN1,028.192/2,303.020ms; WAN4,466.255/4,592.137ms. These are baseline only, not acceptance for the Ack-fixed binaries. All10 WAN prompt budgets fail; retain the LAN outlier. Trace baseline stopped16 attempts with one real Ack-loss failure; not replaced.
ack-state-remote-green.log
private-gnu-gzip-vm.log
native-acceptance-summary.json
guarded-acceptance-summary.json
Current traced-baseline reconciliation, guarded-150-04; clock origin/browser click 1792016462124.200ms. This is the median successful WAN trace among the eight attempted traced cases (seven integrity successes, one retained Ack failure), not final acceptance. Separate actual-production-path20 baseline passed all20 integrity proofs, WAN prompt p50/p99 1748.014/1798.049ms. Tracing changes scheduling; do not combine populations. First page TLS and initial browser mux finish before click; the diagnostic bootstrap socket opens after click, independently. SSH itself has no TLS handshake. All timestamps share the perf VM clock; raw load recorded per case.
This sample paints 1803.246ms after sign-in; decomposition: check 477.296 + invitation 3.371 + continuation/image/shim to publication 1132.254 + publication/browser/render 190.325ms. Private receive/decode/hash 687.743ms. No evidence establishes1.5s as a security lower bound.
Correction to the earlier timeline: exact shipping GNU gzip private size is931,900B, not Python gzip932,008B. New Ack-fixed private size932,439B (+539); actual source integrity correction retained. Three matched additional102,400-compressed-byte trailer trials gave median14.343ms per100,000 compressed bytes near that window; not a universal conversion for raw binary bytes. A crossed congestion flight can cost a whole RTT.
guarded16-native20-raw.tar.gz
warm-proxy-functional.log
Warm-flight experiment artifact clarification: the final x86 full binary is still linking. To avoid idle VM time, these six explicitly experimental trace/proxy cases use fixed final
ee2487b5controller+private image and the already measuredc91bf3b4full image (full-shared, SHA ef33f3a77e49cee153b0e132b0e8bc1abb8dbb80c9f4088631234e32321d1602). That full image is used only after the prompt and is identical in both arms; its source predates the Ack correction, so byte-integrity failures still stop the series and all attempts are retained. No final acceptance claim or mixed artifact relabeling. Final acceptance waits for all ee artifacts.All shipping artifacts now built/staged from
ee2487b571; identities verified before final acceptance. Same-recipe stripped static sizes versus mergebaseb182529fcc, verbatim:Both satisfy the <=16KB mergebase delta and unchanged absolute caps. Final full gate and20 native-path fresh targets remain in progress; size success is not prompt-budget acceptance.
final-size-comparison.json
ack-fixed-controller-artifact.json
ack-fixed-session-artifact.json
Retained baseline LAN outlier native-0-10 investigated (prompt2,201.120ms after sign-in; all59 input/full-SHA proofs pass). After reattach is sent at epoch1792018950275.340, first browser Output arrives1865.400 (
1,590ms later); full upload/setup ends1832.900, only32.5ms earlier. Render adds36.3ms. Source bootstrap.rs:1027–1060 starts full upload after terminal publication, without a first-output fence; openssh.rs channel_command sends both distinct command channels through the same master.socket. No shared Rust pipe mutex serializes them.Independent read-only review agrees shared-transport/target contention is plausible, but capture was disabled: this run cannot prove when the PTY generated bytes or where they waited. The early watchdog retirement explains about141ms, not the remaining1.59s. Preserve the budget failure; do not turn correlation into a shipping change or waive it. Final-head fresh cases will show whether it recurs; stronger attribution needs shim-output/bridge/server-receipt clocks. Complete raw baseline attached previously.
Final-head
ee2487b5scripts/verify fast exited1. Native:450 server passes, one failure browser_input_overload_retires_only_the_terminal_stream at review_input_credit.rs373 receives terminal.reconnecting Event rather than expected terminal Close/Pong;46 ignored. This is outside allowed#118 exceptions: no waiver. Reproducing exact gate ELF and inspecting fixture/production race. fmt/VTfmt/compile-out/clippy/cargo-deny pass.Static FAILS despite both binaries below cap: rust.sh:85–90 requires a downward cap ratchet on every shrink; test-static-size.py:31 consequently rejects shipping fixtures. Brief explicitly says do not change caps. Preserving that instruction: caps/checker/artifacts unchanged; no padding or bypass. Sizes x8622199536 vs cap22204304(-4768),ARM19635832 vs19638744(-2912); mergebase deltas-76352/-68880. Cannot claim a green gate with this conflict. Continue native correction and remaining acceptance.
Verbatim gate summary:
fast-ee2487b5.log
Native finding: exact failed gate ELF passes isolated rerun1/1; default-feature variant also1/1. No waiver. Independent review found no observation-channel deadlock and no proof identifying which reconnect emitter expired. The overload test is an event-order oracle, but currently depends on learned50ms physical liveness while it transfers two2MiB frames and performs SQLite reads under shared build load. Production correctly reconnects if processing crosses that deadline.
Use the existing auth/socket_clock.rs FrozenSocketTime helper for only the two browser_credit_case tests, on start_paused current-thread runtimes after SERIAL admission. This is the same real SQLite/socket time-domain control already used by restart_open_liveness_review and authority tests. Keep every original data/queue bound, Close-before-Pong ordering, unexpected-Event rejection, control witness and production deadline unchanged. Existing clock helper retains its90s hang watchdog. No production change or claim that physical stalls are solved. Gate RED is retained; per-crate strict clippy/fmt and affected module tests will precede a separate atomic test-harness commit.
Atomic test-harness fix committed/pushed
b1b2fb0efb. All4 input-credit cases pass, strict server(remote) all-target clippy and fmt pass; independent review found no missing assertion or deadline change. Reuses existing FrozenSocketTime for exactly two functional ordering cases. Original full-gate failure retained; no claim physical-stall behavior changed.Final fast rerun on
b1b2fb0estarted; exact TypeScript gate also started separately because the known static ratchet conflict stops fast before TS. Shipping x86/ARM/private identity rebuilds will verify whether this cfg(test)-only commit leaves the already pinned ee measurement artifacts byte-identical. Native20/proxy6 remain queued under the perf lock; no completed acceptance claim.overload-clock-module.log
overload-clock-clippy.log
overload-exact-gate-reproduce.log
Rebuilt at
b1b2fb0efb: x86,ARM and private shipping artifacts are byte-identical to the pinnedee2487b5artifacts (same full hashes and sizes). Only cfg(test) input-credit scheduling changed. Native20 acceptance retains its fixed ee measurement-controller provenance and these exact shipping artifact identities; no shipping change is hidden by the test-only commit.VM availability currently blocks the remaining measurements and full-gate timing prelude: shared lock holder PID1981650, cwd /root/infraternal-bench/ppu-soak/r4long, not this lane. Our commands wait with flock -w3600; no other process/booking changed. Warm proxy control1 complete, padded1 and native0-01 still queued. Static cap/ratchet conflict already recorded remains unresolved; no cap/checker changes.
test-clock-artifact-identity.json
While the perf VM remains in the other lane's soak, one independent size-only candidate: CLI package opt-level=z for the private session-payload build, same toolchain/dependencies/features/LTO. Prior opt-level=s grew this image and was rejected. No tracked edit or final artifact replacement; use a command-line profile override in the private target only, save candidate identity/GNU-gzip bytes, then restore and require byte-identical normal output. Low-priority one-job build; no build-host timing claim. A smaller image alone cannot justify shipping without functional/VM validation; reject growth outright.
Rejected private opt-level=z: 2430000 raw /976642 GNU-gzip bytes versus2078216 /932439 normal (+351784 raw /+44203 compressed). Never staged or used in acceptance. Normal recipe restored and exact a6dbb7dc… private SHA reproduced. No tracked source/profile change. This is file-size evidence only; no host timing claim or candidate runtime claim.
private-z-size.json
private-z-normal-restore.json
Retained guarded-150-04 WAN trace, post-sign-in clock anchor1792016463594.154ms (not a new acceptance run):
The first exec spans68.546ms at the sender before target receipt; setup/image transmission spans894.968ms. The latter has no source-data gap above20ms. Target verification finishes1462.930ms, shim exec1465.604ms, publication1612.921ms and paint1803.246ms. Thus continuation-marker delay and subsequent image-receive interval overlap a continuous stream; they are not separate added protocol RTTs. Transfer includes producer/pacing/congestion work; packet headers alone cannot attribute all time to a kernel algorithm.
Exact image in this trace remains931900 GNU-gzip B;938580 includes setup/SSH framing. Counts of captured records are not physical packets because offload aggregation can occur. The20ms gap threshold only groups the display. Warm-flight controls remain separately pinned; no unpaired before/after claim. Complete raw trace attached earlier.
guarded-150-04-tcp-bursts.json
Standalone full server preflight retained449PASS/2FAIL/46ignored. The independent runner omitted CARGO_TARGET_DIR: signed_identity_and_exact_bytes_are_required unwraps this required variable; release staging falls back to relative target/, making its deliberately retained relative symlink appear dangling. This is a runner-environment mistake, not a shipping regression; normal scripts/verify already exports the absolute target root.
Corrected scratch runner now uses the same absolute CARGO_TARGET_DIR, Cargo manifest directory and package working directory. Both exact failing cases pass with the correct environment. Original runner/log retained. No leftover target/bucket-* fixture directories and no tracked changes. Rerunning the complete451 functional cases under the corrected environment; this is not the final fast gate or performance evidence. All4 input-credit cases passed in the prior standalone run.
server-functional-b1b2fb0e.log
release-fixture-cargo-env-targeted.log
Head
b1b2fb0efb(pushed, clean). Complete changed default+remote server functional suite now passes with the correct Cargo fixture environment:All 451 functional cases ran, including the earlier overload failure; normal ignored cases remain ignored. This is functional evidence, not build-host performance evidence. The original standalone harness environment mistake and its two failures remain attached separately.
Final artifact acceptance ack-0-01 and proxy experiment warm-150-padded-1 both exhausted their full 3600-second flock wait before execution. Neither VM case log exists; the experiment also correctly refused its missing applied-prefix receipt. No fresh target sample was collected or replaced. The final 10+10 acceptance and warm-prefix comparison remain blocked, with no claimed candidate improvement. At 00:25 UTC the designated shared lock still belongs to PID 1981650; it was not interrupted. Final current-head fast/TypeScript gate waiters remain queued on their original configured waits. No deployment.
server-functional-b1b2fb0e-cargo-env.log
server-functional-b1b2fb0e-cargo-env-identity.json
ack-acceptance-console.log
ack-0-01.log
warm-150-padded-1.log
HOLD / resumable handoff; D255 is not complete. Head
b1b2fb0efbis pushed to addhost-instant; no deployment. Timeline and per-stage explanation are posted above; all shadow safety fixes and the retained setup-status contract remain intact. The Ack/State coalescing regression is fixed, with RED-to-GREEN daemon tests and real-PTY coverage. Final changed server functional suite: 451 PASS / 0 FAIL / 46 normal ignored. Separate TS functional subset: 191 PASS / 0 FAIL; one physical graph case excluded only from that separately labeled subset.b182529fmerge-base bytesCurrent-head rebuilds reproduce the attached exact artifact SHAs; private session 2,078,216 raw / 932,439 GNU gzip -n -9 bytes. No cap changed. Prior completed gate's exact-size ratchet still requires lowering a ceiling when a binary shrinks, conflicting with the explicit no-cap-change instruction; neither padding nor a gate bypass was introduced.
Acceptance blocked: both ack-0-01 and warm-150-padded-1 exhausted 3600-second lock waits before execution, with no target logs/result receipts. No final fresh sample ran; no proposed warm-prefix change was shipped. Earlier actual-production-path baseline (not final artifacts): WAN post-sign-in prompt p50/p99 1,748.0135/1,798.049 ms, joined 4,466.2545/4,592.1372 ms; LAN prompt 369.167/2,201.120 ms, joined 1,028.192/2,303.020 ms. All 20 baseline byte/full-SHA proofs passed; budget failures and outlier retained. The earlier traced Ack-loss failure also remains retained.
Current-head scripts/verify fast was run; it exhausted its physical latency self-test lock wait and exited before Rust rows. Exact output, verbatim:
Separate exact TypeScript gate source/compiler checks reached the physical graph case; its 3600-second lock wait expired too:
Needed from orchestration: a measurement slot on the designated perf VM; another lane's lock was never interrupted. Resume from retained packaged artifacts/drivers under /home/kayg/.cache/codex-scratch/addhost-instant and VM /root/infraternal-bench/addhost-instant/tl. Preserve lock-timeout-20261015 receipts. Use new evidence namespaces for the matched warm-prefix pairs and final 20 production-path cases; existing warm-control-1 is a completed labeled proxy control and must not be overwritten. Require both native-painted witness rows, all 59 acknowledged bytes, unchanged session and installed full artifact SHA. No security lower-bound or candidate improvement is established. Rerun current-head final gates after a measurement slot is available. The static-ratchet conflict remains separately recorded; no green claim or waiver.
fast-b1b2fb0e.log
ts-gate-b1b2fb0e.log
ts-functional-b1b2fb0e.log
author-checklist-b1b2fb0e.txt
Cleanup: Removed only addhost-instant target/, session-target/, native-baseline/ and ui-baseline/ after all test/gate processes exited. Packaged artifacts, raw receipts, scratch drivers and VM-staged artifacts retained for resume. No other job processes, worktrees or VM lock changed.
Resuming #189 on addhost-instant at
b1b2fb0efb; merge base with origin/devb182529fcc. Clean worktree. Read issue/parent history and latest handoff. Current private image 2,078,216 raw / 932,439 GNU-gzip bytes. First step: isolate only the shell-owning shim and bridge without changing verification, ownership, replay or retained setup contract. Heavy commands use build-slot; VM runs use short fair lock holds with owner receipts. Native overload correction already exists; will reproduce final gate and compare any remaining failure against dev. No deployment.D255 implementation choice before code: make the first-prompt artifact a dedicated internal executable, sharing the existing shim startup decoder/runner and SSH session bridge rather than retaining the operator entrypoint. Release filenames, signed manifest/hash verification, wire frames, ownership checks, retained setup stream and installed-image fallback remain unchanged. No alternate terminal engine. Measure the existing VT engine with size optimization independently; retain only if functional/replay fidelity and perf-VM results justify it. <=250KB remains a target, not a claimed result. Native source baseline is frozen separately while its size-attribution build waits for a slot.
Before-change checks retained on exact
b1b2fb0eshipping artifact identities: private image 2,078,216 raw / 932,439 GNU-gzip bytes; new 250,000B size oracle FAIL. Two fresh production-path VM controls (not acceptance percentiles): post-sign-in painted prompt WAN150 1753.996ms, LAN 459.632ms; both budgets FAIL. Both verify all 59 input bytes, before/after output and installed full SHA a3c3c77f57c4bab6d5f5c9bf1dd462b1d2fe2f363a255133018e041c66c28d54. Short per-case locks now write owner/start and enforce <30min holds. Source candidate remains uncommitted pending build-slot availability and functional/size validation; setup stream unchanged.minimal-budget-red.log
minimal-baseline-150-01.log
minimal-baseline-0-01.log
Completed/pushed
7af0bdd22: native timing runner writes the perf lock owner/start, caps the held subprocess group at 1750s + 10s termination grace, and cleans the owner receipt before unlocking. Real local-flock functional regression 1/1 PASS (owner visibility, exit 7 preserved, receipt removed, lock available); workspace fmt PASS. No shipping artifact changed. Baseline bloat attribution is now available; testing native size optimization next. Bloat attribution is heuristic, not a shipping-size identity claim.Next bounded overlap experiment, before shipping changes: replace the large shell-quoted initial exec program with a small fixed POSIX reader and send the same trusted public program immediately on the retained SSH channel stdin. A bounded line envelope read by shell builtins must preserve all following binary/protocol bytes, including when coalesced in one write; no buffering head and no additional acknowledgment/RTT. Keep host-key/authentication, ownership, signature/hash checks and admission order. Validate real sh/bash coalesced binary sentinels and malformed envelopes, then measure matched VM cases; reject if no end-to-end gain. Baseline first exec was 10,175B in the attached median timeline; no before/after gain claimed. Setup-status event contract unchanged.
Completed/pushed
1e2e22f1fe: shared canonical shim option record/decoder/runner now lives in the shim crate; CLI parsing is an optional adapter. No default, identity, retention or accepted argv change. Existing exact-shape agreement tests 2/2 PASS; real shim/VT library 80 PASS / 13 existing opt-in ignored; strict touched-crate clippy and fmt PASS. The fuzz harness imports the same parser and drops its direct Clap edge.Size-only experiments, not acceptance: dedicated target 2,076,456 raw / 913,570 GNU-gzip B. Matched bloat-build + strip experiment: baseline 2,078,208 / 932,066 B versus same VT engine ReleaseSmall 1,832,128 / 841,011 B (-246,080 raw / -91,055 compressed). The bloat-build baseline differs from the pinned shipping ELF (2,078,216 / 932,439); it is not an identity reproduction. The 250,000B target remains unmet. Native correctness/throughput and VM prompt evidence still decide whether the compiler change is kept.
Next contract-compatible trim: specialize the existing dispatcher for the dedicated first-prompt image to attach only the original authenticated session and omit plugin workers/operator opens; full installed/fallback dispatch stays unchanged. Refuse an unavailable operation without disturbing the working session. No replacement engine or state-transfer protocol chosen. Initial-exec stdin framing remains an independent experiment: real sh/bash/dash exact-source + coalesced-byte and malformed-header controls pass 30/30; no latency gain claimed.
Adjacent short comparisons are queued as one <30min bounded measurement batch, with no builds/copy/analysis under the lock. An earlier queued single-case waiter was cancelled by its exact owned PID before execution (no target/log/sample), solely to regroup the samples. All actual baseline failures remain retained.
startup-extraction-gate.log
minimal-baseline-size.json
minimal-vt-small-size.json
minimal-dedicated-size.json
exec-envelope-shell-functional.log
Perf acceptance/comparison queue has waited approximately 45 minutes across the original queued comparison and its pre-execution regrouping. Current owner receipt at 2026-10-15T01:44:55Z:
Continuing to wait; no other job/process/lock touched. The bounded 21-case comparison batch has not started (console only WAITING_FOR_BOUNDED_COMPARISON_BATCH). Native source work continues independently. Latest pushed source step
36eda557: SSH bridge extraction; strict clippy/fmt, 2 internal command tests, 3 bridge tests, real same-session pending-input handover and single-session rebridge all PASS. No final acceptance or final-gate claim.Completed/pushed
9c92dbf1f(private original-session-only runtime), then merged current origin/dev at7d9f996df. No setup event contract change. Private dispatcher advertises no node action capability, refuses terminal.open without disturbing the original session, and cannot attach a different same-UID session. Real two-shim oracle RED on the old dispatcher, GREEN on the new image; real pending-input handover and rebridge PASS. Strict clippy/fmt PASS; image 2 PASS, agent 57 PASS / 7 normal ignored, bridge 3 PASS. Final isolated image is still 2,092,264 raw / 923,586 GNU gzip bytes: 250,000B target NOT met.minimal-scope-gate.log
private-scope-red-validuid.log
The bounded 21-case perf-VM comparison batch completed; all byte-integrity/full-SHA proofs PASS, all 21 prompt budgets FAIL. These n=3 comparisons are not final acceptance percentiles.
Reject the initial-exec stdin mechanism: worse end-to-end. Neither VT optimization nor padding closes the gap; neither is shipped. This disproves an assumed proportional reduction from removing 91KB compressed at the current cold-window threshold. No security lower bound is established. Need a substantially smaller actual first-prompt runtime; no alternate engine or lossy snapshot/state transfer has been chosen.
minimal-comparison-summary.json
minimal-batch-console.log
Raw receipts for all 21 VM comparisons (per-case logs, target/server/browser timelines and the incremental result table), retained including all prompt-budget failures:
minimal-comparison-raw.tar.gz
Current work: dedicated-image release packaging; functional streaming prefix test withholding the compressed tail until staged bytes and hash input are observed. No profile/compiler experiment shipped. The next attribution build removes the bridge entirely to quantify the existing shim/VT contribution; that diagnostic is explicitly not a usable add-host artifact. No alternate engine, state-transfer mechanism or relaxed snapshot guarantee selected.
Completed/pushed
6da5ec743: release packaging now builds infraternal-session for x86_64 and aarch64 and publishes the existing signed session-image filenames. Exact-copy checksums, executable/static checks and missing-architecture refusal PASS; release script suite 8/8 PASS. Both static images built: x86_64 2,092,264 B; aarch64 1,776,288 B. Host-setup contract describes original-session-only namespace and shared startup/bridge; retained setup-status stream unchanged. This is packaging integration, not a size or prompt acceptance claim.package-dedicated-gate.log
Functional prefix overlap oracle also executes 1/1 PASS: withhold a nonempty compressed tail until decoded staging bytes and hash input exist, finish exact-file verification and same-channel traffic. Full touched-test suite + clippy queued before its separate test commit. Build-host elapsed values are not performance evidence.
streaming-prefix-functional.log
Completed/pushed
8ec074b10: functional streaming-progress oracle. Touched SSH strict clippy/fmt PASS; complete real-files install/recovery suite 77 PASS / 0 FAIL. The hasher receives prefix input while a nonempty compressed tail remains withheld; exact final staging bytes and retained-channel traffic pass. This confirms the existing upload/decode/hash pipeline overlaps; no performance result is inferred from build-host test elapsed times.ssh-streaming-suite-gate.log
Size attribution (diagnostics, NOT shipping images or a theoretical minimum):
The VT-only diagnostic uses the existing ReleaseFast session-image archive, pinned by SHA in the receipt. Exact reset-state snapshot and approximate snapshot controls pass; plain C-initialized state correctly returns the existing unrepresentable-state refusal. Source attached as .txt because the attachment endpoint refused .c (HTTP422). No alternate engine or weakened snapshot path is shipped. The no-VT diagnostic is not usable and must never be packaged. D94 and contracts/shim.md require exact terminal state; simply dropping that engine is not the requested implementation.
vt-api-only-size-v2.json
vt-image-attribution.c.txt
shim-floor-with-vt-size.json
shim-floor-without-vt-diagnostic-size.json
session-isolated-budget.log
Current small cut under validation: reuse the full binary's compact relocation-table linker recipe in the private executable. Full CLI linker arguments remain identical on all four checked platform/architecture combinations; no profile/ASLR change. Native metadata from the diagnostic copy collided in the shared Cargo target (link lacked VT symbols); selectively cleared only this job's shim artifacts and refreshed its build-script timestamp before restoring the production recipe. No affected artifact is staged or accepted. Baseline overload check and final gates remain pending; no completion claim.
relocations-linker-equivalence.log
Completed/pushed
2eb6562d7: dedicated image now shares the existing full-CLI relocation compaction. Strict touched-target clippy/fmt PASS; image 2/2 tests and canonical startup 2/2 PASS; both static architectures built; real original-session restriction, same-writer pending-input handover and one-session rebridge PASS.No ASLR, release-profile or parser change. The 250KB compressed target remains FAIL; default full-binary delta versus pinned dev is still pending.
session-relocations-gate.log
session-relocations-size.json
Completed/pushed
77be7d26e: two proven fair-lock cleanup defects fixed in this lane's measurement helper. Forced termination skipped the timed shell's EXIT cleanup; a detached child inherited the lock descriptor and retained the lock after the measurement ended. Cleanup now runs outside the timed process group, still under the lock; flock --close prevents inheritance. Default deadline/grace unchanged and cannot be increased through the helper. Child status propagates. Both regressions RED before their fixes; final three real local-flock cases PASS, including proof that the detached child remains alive while the lock is available, then pidfd-targeted fixture cleanup. No other lane's process, helper or lock changed.measurement-stop-owner-red.log
measurement-inherited-lock-red.log
measurement-owner-green-final.log
Baseline exact overload case is compiling on pinned dev
1b23ef311in the shared build slot; branch comparison follows. Final current-head gates and new 10+10 acceptance have not run yet. No deployment or completion claim.Exact earlier overload regression now executes and passes on both pinned dev
1b23ef311and branch77be7d26e(one named case on each; no zero-test selection). The failure did not reproduce in these runs; no waiver is taken. Final complete gate remains pending.overload-origin-dev.log
overload-origin-branch.log
Same-recipe pinned-dev x86_64 static build: 22,277,392 B, SHA 09504f5f23f2852777a5db1beda0aaf4f2d709d20808b91878950fa6d47eb7ba. Baseline ARM and both branch images follow in the one shared build slot, then current controller, iterate and fast. No cap changed.
Current-image real-SSH qualification (not final acceptance: current private image against retained b1b2 controller/full peer) is staged and waiting. At 04:10:16 UTC both target logs were absent; the owner receipt was process-per-user ppu-soak-r5-raw-echo 2026-10-15T04:09:06.015735+00:00. Public lock metadata attached; no conclusion about other lanes' lock ownership is inferred, and nothing belonging to another lane is changed. Continue waiting while builds proceed. Final 10+10 and five paired public 100KB image-cost controls are prepared to start only after the complete current-head artifact manifest exists.
perf-lock-qualification-diagnostic.txt
Measured default-image regression at
77be7d26e; this checkpoint is NOT accepted. Same recipe against pinned dev1b23:No exception requested or cap changed. Likely retained second dispatcher future/drop paths in the async bridge's const-boolean branch. Replacing it with static callback selection, keeping the same shared dispatcher, protocol and ownership. The red measured size check precedes the fix.
default-size-growth-red.log
Retired only the owned obsolete artifact supervisor after its current Cargo compile exited normally; cancelled its manifest waiter before any of the planned20 measurements. No sample dropped or overwritten. Both completed default images and logs preserved. Old-head iterate/fast did not run; corrected-head validation will replace them. The already queued2-case real-SSH qualification keeps its place.
That qualification has waited45+ minutes. Current public owner receipt:
perf-qualification-wait-45m.txt
Continuing to wait; no other lane's process or lock touched. Source fix and touched gates proceed through build-slot. No deployment or completion claim.
Step completed:
d69829c60(pushed) selects the SSH dispatcher before building the async state, eliminating the unused dispatcher drop paths while preserving the shared bridge. Private x86: 1,937,544 B raw / 871,456 B gzip (-94,048 / -46,592 B); ARM: 1,687,936 / 799,553 B (-88,352 / -43,491). Private scope + real shim handover/rebridge PASS; ordinary full-binary handover/rebridge also PASS. Touched fmt, strict clippy, bridge 3 and CLI 2 tests PASS. Full x86 is 22,341,008 B vs pinned dev 22,277,392 (+63,616); ARM still building. This is above 16 KB and needs further cuts; no size exception or cap edit.Previous two-sample real-path qualification (old 918,048 B compressed image, not this artifact): LAN prompt 496.427 ms, 150 ms prompt 1766.849 ms after sign-in, both budget FAIL. Both retain original session, all 59 acknowledged input bytes, painted output before/after, installed full SHA. These are n=1 each, not acceptance percentiles.
session-static-dispatch-x86_64.json
session-static-dispatch-aarch64.json
full-static-dispatch-functional.log
qualification console qualification raw
Size finding: completed default
d69829c60builds are x86 22,341,008 B (+63,616 vs pinned dev), ARM 19,780,632 B (+75,520). ARM exceeds the no-owner-exception ceiling; not accepted. Testing typed empty/original session seeds to remove retained async ownership state, then node-resource types if needed.Nonshipping compiler experiment: same pinned VT engine, Rust opt-level=z, native ReleaseSmall, panic=unwind. Private x86 1,580,136 B raw / 726,539 B gzip (-144,917 gzip vs default). Real scope, original-session handover/input-once and rebridge proofs PASS. This remains above 250 KB. Four fresh 150 ms comparison runs queued on perf VM; no latency conclusion yet. No profile change shipped or new engine/library chosen.
full-static-dispatch-sizes.json
private-size-candidate-artifact.json
private-size-candidate-console.log
Rejected size experiments (source patches retained, all reverted): empty session seed alone saves x86 192 B / ARM 1,536 B; typed node resources plus constant greeting capability save full x86 6,624 B (22,334,384 B, still +56,992 vs dev) and private gzip 7,006 B (864,450 B). Default 61 tests, remote dispatcher 27, strict clippy and 3 real private-process proofs PASS; size benefit does not justify 100+ implementation lines.
Next controlled difference: the bridge moved from the CLI (opt-level=3) into SSH (package opt-level=s). Testing one shared bridge library in the existing session crate to preserve its original compiler profile. This rewires existing workspace edges only, introduces no third-party package/version, and preserves all wire/startup code. Not committed until measured and gated.
typed-seed-sizes.json
typed-seed-source.patch
typed-seed-gate.log
typed-node-work-full-size.json
typed-node-work-private-size.json
typed-node-work-source.patch
typed-node-work-gate.log
Perf VM comparison waiting >45 minutes; no new timing results yet. Continuing to wait, no resource or timing waiver. Current owner receipt:
profile-comparison-45m-owner.txt
Local bridge-library trial remains queued for a host build slot. No deployment or completion.
Confirmed capability regression: real SSH pipe sends Hello caps=0; RemoteLink::ssh hard-codes server_paths=true. Added test fails exactly at that assertion (1 failed, 497 filtered). Fix reads the existing Hello flight before admission, keeps all coalesced bytes for the actor, and preserves the existing frame bound. Added full-capability/coalesced-frame coverage; touched clippy and handover/admission/setup-stream checks queued. No setup event-contract change; no latency claim before measurement.
Rejected further size trials: moving the shared bridge back to opt3 through the existing session crate increases full x86 by44,256 B to22,385,264 (+107,872 vs dev), private gzip by21,536 to892,992. Reverted. Native diagnostic-only candidate preserves message+SIGABRT and Rust unwind but saves only3,398 gzip bytes; not retained.
Perf comparison waited one hour and timed out without entering the measurement. Verified zero profile-size case files on VM; retained output and restarted the same four cases. Continue waiting, no timing waiver. Owner observed at06:09 UTC: process-per-user ppu-soak-r5lazy-soak started05:42:41Z.
ssh-capability-red.log
bridge-library-full-size.json
native-panic-candidate-artifact.json
native-panic-signal.json
profile-size-comparison-lock-timeout.log
Step completed and pushed:
7940d9f6c6. SSH admission now derives node-path support from the existing peer Hello instead of transport kind. Preserves every coalesced byte, existing wire limit and one greeting deadline; no request/reply flight added.RED real-pipe caps=0 regression retained. GREEN: strict server all-target clippy (remote+ssh), 6 SSH greeting/lifetime tests (including unsupported greetings, full capability and byte preservation), 1 installed-route admission test, 2 retained setup-stream replay/failure/logout tests. The event contract for the designer is unchanged. Two original follow-up filters matched zero; corrected by actual test names and reran 1+2 tests, outputs attached.
Additional nonshipping abort profile with the selected engine and size optimization reaches679,915 gzip bytes, still above250KB. Not selected; shipped unwind behavior is unchanged. Default image still871,456 gzip bytes. Milestone builds and exact iterate/fast are queued on committed head; current-head20+10 cost controls will stage only after all7 artifact identities complete. Acceptance and binary-size budgets remain open; no deployment.
ssh-capability-gate.log
capability-followup-abort-console.log
private-abort-candidate-artifact.json
Current head
7940d9f6static artifacts built with pinned recipe. x86: base 22,277,392 B; head 22,344,624 B; delta +67,232 B. ARM: base 19,705,112 B; head 19,783,928 B; delta +78,816 B. Both exceed the size ceiling; no exception claimed. Private images unchanged: x86 871,456 gzip B, ARM 799,553 B. All 7 source/artifact identities complete; current-head acceptance 10+10 and five paired +100 KB controls staged and queued. Exact iterate is progressing; VM timing self-test passed. Final fast remains pending.Completed profile qualification on the retained b1b2 controller/full image, not final-head acceptance: 150 ms RTT, n=2 per image. Default prompt 1819.760/1766.856 ms, median 1793.308; smaller profile 1735.553/1730.732 ms, median 1733.1425. All 4 prompt budgets FAIL; all 4 original-session / 59-byte / painted-output / installed-full-SHA proofs PASS. No linear transfer-cost claim: compiler/VT flags also differ. Raw receipts attached.
Queued captured pipe-capacity diagnostic (default / 8192 / 16384 bytes, two each) on current pinned artifacts to test continuation packet batching. Separate from production-path acceptance; no production transport change yet.
profile-size-comparison-summary.json
profile-size-comparison-console.log
profile-size-comparison-raw.tar.gz
milestone-static-artifacts.json
7940d9f6production-path acceptance population completed: 10 fresh LAN +10 fresh 150 ms RTT, perf VM4vCPU, load recorded per case. All20 original-session / typed+painted before/after /59 acknowledged input bytes / installed-full-SHA proofs PASS. All20 prompt budgets FAIL; this is not accepted. Nearest-rank p99 (n=10 means maximum).Fully joined is back below5s at150ms; prompt and size remain open. +100KB paired controls still queued.
New LAN attribution from case0-05: sign-in ends2509.368; publication2725.147 (+215.779ms); browser sees2730.9; Attach/Credit2760; socket close4000 at2784.1, reconnect Welcome2868.5; first paint2919.9 (+410.532ms). Another close4000 at2933.9. Harness does not request a resume probe. Connection core has two4000 reasons (opt-in probe or liveness expiry); checking the reason before assigning cause. D265 lease remains unchanged. This delay occurs after publication and cannot be attributed solely to payload ramp. No UI/event-contract change shipped.
accept-7940d9f6-summary.json
current-head-20-raw.tar.gz
accept-7940d9f6-0-05.log
Head
7940d9f6: diagnostic follow-up (not acceptance). Same pinned images/controller; designated VM only, bounded fair-lock holds.Paired image diagnostic: +100182 gzip bytes adds 35.312 ms by the difference of medians (small sample; not a universal linear transfer rate). All 10 cases retained 59 acknowledged input bytes, before/after paint and exact installed full-image digest. Pipe sample is insufficient to ship a change; no production pipe tuning applied.
LAN diagnostic confirms local
connection unresponsivecloses at epoch 1792050472376.4 and 1792050472598.6. First is before prompt paint 1792050472568.600; sign-in ended 1792050472048.593 (520.007 ms to paint). A 57 ms browser long task starts 1792050472532.9, after the first close; causality remains unproven. No D265 deadline relaxed. First instrumentation attempt could not launch because staging outside the package selected Playwright 1.64.1 and an absent browser; corrected script location uses the existing package/browser and completed the functional proof. This diagnostic used a 1500 ms harness bound on LAN, so its PASS is not a 300 ms acceptance pass.Full iterate is still running; no gate verdict yet. Full-size deltas remain +67232 B x86 / +78816 B ARM against pinned dev base, over the allowance. Testing executable-local shared startup derivation next, after this job's existing heavy wrapper exits.
current-cost-raw.tar.gz
pipe2-diagnostic-raw.tar.gz
liveness2-case.log
liveness-case.log
Head
7940d9f6milestone:scripts/verify iteratepassed in full (Rust: 1164 tests, 0 failed, 140 ignored across 126 completed binaries; fmt, VT fmt, CLI compile-out, clippy, cargo-deny all PASS; TypeScript gate also completed). Final line verbatim:The checkpoint
fastinvocation was deliberately interrupted during its VM timing preflight, before native/static checks, to measure the pending size candidate. It is NOT a final gate result. A complete fast rerun remains required on the selected branch head.Browser scheduling experiment, two controls/two one-frame yields per RTT, same native artifacts and unchanged D265 deadline: median prompt ms LAN 387.755 control / 524.149 yield; 150 ms 1776.460 control / 1801.608 yield. All eight 59-byte, painted-row and installed-image proofs pass; all prompt budgets fail. Discarded; no production browser change.
Private bridge still invokes Manager::detect/systemctl, although initial admission already observes the manager capability. A functional witness confirms that call (attached expected-failing hypothesis oracle). Unconditional detached launch is NOT a correct replacement: it does not give a managed shim scope independent of the SSH login lifetime. Any optimisation must reuse authenticated checked manager information and retain managed launch, failure handling, and D240/D247 compatibility. No launch selection change made.
open-diagnostic-raw.tar.gz
iterate-7940-complete.log
private-init-probe-red.log
Head
7940d9f6: discarded further candidates; checkout remains clean.Current production-path median-nearest 150 ms case accept-7940d9f6-150-03 (sample p50 1758.161 ms; this case1757.921):
These are application anchors, not a replacement for the packet-level timeline already attached. Check-to-publication remains1122.621 ms; publication-to-paint200.724 ms. Dispatcher size experiment is queued for this job's next build slot. Full final fast still pending, as previously recorded.
startup-codegen-x86-qualification.json
pong-fence-diagnostic-raw.tar.gz
current-median-stage-anchors.json
private-dispatch-budget-before.log
Proposed internal optimisation for D255's signed private image (event/wire contract unchanged): reuse admission's existing manager prerequisite.
Check::parserefuses non-systemd;authorize_moderefuses rootless without user_systemd and binds root mode to root or consented sudo. The private image runs only after those checks. Select the existing System scope for effective uid0, User scope otherwise; if that scope cannot start, fail rather than silently fall back to a detached shell. Generic full-image SSH keeps Manager::detect and its existing compatibility behavior. No new manager/engine, no environment hint, no shell adoption or login movement. D240/D247 still requires replacing the shared launch boundary with the original person's real login when that lane lands.The attached earlier private-init-probe witness demonstrates the redundant systemctl call before a shim-owned prompt. Preparing a nonshipping callback candidate with the existing scope/reattach/replay oracles and a failure case for scope withdrawal. No latency saving claimed before designated-VM measurement; no production edit yet. The static dispatcher/typed-event trial remains queued for this job's sole build slot.
Waiting on the host-wide build slot since08:29 UTC (>45 min at09:16). Sole queued wrapper: build-private-dispatch.py; it has not entered its Python producer or started Cargo. All four build-slot receipts remain occupied. Need the orchestrator-managed slot to become available; continuing to wait, no slot bypass or other job interruption. Root7940d9f6 remains clean.
Prepared:4 private-dispatch size variants, real scope/handover/rebridge checks, managed-scope withdrawal RED oracle, VT-profile attribution and every add-host functional browser spec. Full iterate alreadyPASS; final fast and new acceptance still pending. No completion/latency/size waiver.
Build blocker cleared09:35 UTC. The sole queued wrapper acquired one of the official four slots and is compiling the first dispatcher candidate. No slot-limit change, extra heavy wrapper or other job interruption. An inotify watcher woke only this helper's exact owned sleep on a cleared slot receipt, so the existing helper could retry before its20-second poll. Watcher exited after producer start. Remaining size/functional/full-head gates still pending.
Nonshipping dispatcher measurements completed; Root7940 remains clean.
All4 variants pass real private scope, same-session/writer input replay and reopened-bridge checks (12 proofs). Manager variant also passes both no-repeat-probe and withdrawn-scope refusal cases; earlier withdrawal oracle fails on current head. Current setup-event contract is unchanged.
Clippy rejects writer specialization (104/100 lines); it also increases gzip by53 B, so discarded rather than adding a lint allowance. Follow-up candidate removes it while retaining result guards, typed events and checked manager, then reruns proofs, touched clippy/tests, both full static size comparisons and add-host browser flows. Sole wrapper queued since09:48; no production edit/commit yet.
VM qualification (not acceptance; n=2/variant/RTT, same7940 controller/full image): LAN prompt median453.101 ms control /422.729 candidate; WAN1501739.849 /1807.401. Candidate WAN cases1717.315 and1897.487 ms; sample does not establish a speed improvement. All8 native-painted-row/59-byte/exact-installed-digest proofs pass; all prompt budgets fail. The842263 B image remains above250000 B. No performance or size waiver.
private-dispatch-sizes.json
private-dispatch-console.log
private-dispatch-writer-clippy-red.log
dispatcher-qualification-raw.tar.gz
Unapplied selected-candidate patch attached for resumption/review; git apply --check passes on7940. Dependency graph confirms70 normal packages and no openssl/openssl-sys/tokio-openssl/server/mesh/remote-plugin/host-facts/clap in the private target.
After discarding writer specialization: private1888584 B /gzip842288 B, five real scope/replay/managed-refusal proofs PASS, strict touched clippy PASS. Agent unit run failed10 fixtures before exercising product code because the archived source lacked its ignored ../../target parent. Linked archived source/target to this job's existing single target directory; rerunning complete suite (no waiver), then SSH suite, both full static artifacts, VT attribution and all add-host browser flows. Sole wrapper queued since10:25. Root7940 remains clean; final fast and final acceptance remain outstanding.
private-dispatch-unapplied-candidate.patch.txt
session-dependency-boundary.json
Archived selected candidate now passes all touched functional gates after the fixture-parent fix: infraternal-agent 61 passed / 0 failed / 7 ignored; SSH session-bridge 41 passed / 0 failed / 2 ignored; strict touched clippy PASS. All five real private scope, same-session/writer replay, rebridge, no-repeat-probe and withdrawn-scope checks PASS. Both full static size comparisons and add-host browser flows are still running in the sole build-slot wrapper. Root7940 remains clean. No latency/size acceptance claim; final fast still outstanding.
private-dispatch-agent-tests.log
private-dispatch-ssh-tests.log
private-dispatch-clippy.log
Further scope finding: the same withdrawn-manager oracle is RED against the immutable7940 full x86 binary, not only the private image. All production __ssh-session callers are the admitted setup launcher/rebridge (install.rs571,981); probe.rs86 rejects non-systemd and authorize_mode96-110 checks the applicable user manager/root identity. The ordinary daemon remains init-independent and uses Manager::detect separately.
Refining the unapplied candidate: select System/User directly in the one shared SSH helper after those existing admission checks, rather than adding a second generic callback solely to preserve an unneeded hidden-helper fallback. A withdrawn scope must refuse in both images; this also removes the repeated probe. Add positive managed launch and withdrawn-scope checks against both full/private binaries. No argv, wire or setup-event contract change, no node-daemon behavior change. Root7940 remains unchanged; candidate not yet selected.
full-managed-withdrawn-red.log
Selected archived candidate x86 full static22329968 B, SHA7bad1c4dcdf82b3ff2736e41ef0c3a3a207b6a6bc2bbce448def70589b8cbd26: -14656 B against7940, +52576 B against pinned merge-base1b23ef31 (22277392 B). ARM comparison still compiling; no cap waiver.
ELF section attribution against that baseline: .text -6956 B; .rodata -16384 B; .eh_frame +34368 B; .gcc_except_table +32356 B; .eh_frame_hdr +5400 B. Unwind/exception tables total +72124 B, while text+rodata are23340 B smaller. Remaining work is to reduce shared async cleanup growth rather than remove terminal behavior or change unwind isolation. Root remains7940, clean.
full-x86-section-comparison.json
Both archived-candidate full static builds complete with the pinned recipe:
Both remain above the requested16 KB growth. No cap change/exception.
Selected VT API-only diagnostic (no PTY, Rust bridge, or usable session host): ReleaseFast942272 raw/368433 gzip; ReleaseSmall696224 raw/277448 gzip. Both pass empty-state exact refusal, reset-state exact output and approximate-output controls. These are actual probe artifacts, not a theoretical minimum or a proof250 KB is impossible. No compiler-profile or engine change selected. Browser functional suite remains green so far; finishing58 cases.
private-dispatch-full-sizes.json
vt-api-profile-attribution.json
Production-build add-host functional browser suite on unchanged7940 completes:50 passed /8 conditional capture skips /0 failures. Includes both phone/desktop Start fresh captures, challenge continuation, background setup/retry, host-key confirmation, root consent and signed payload refusal. Unique TRACER_TEST_PORT5043; raw receipts outside checkout. These functional tests are not timing evidence.
The archive candidate pipeline is complete, exit0 (both full static sizes, VT attribution, touched native tests/clippy and real session proofs). Next sole build-slot producer queued: three separate gated/pushed changes (impossible result handlers, typed notices with wire-boundary tests, shared admitted-scope refusal with regressions for both image kinds), then committed-head artifacts/controller. Root7940 remains clean until it gets a slot. New final acceptance and fast remain outstanding.
addhost-functional-current.log
Completed result-guards:
a50defeb0a, pushed. Private artifact 1933352 B raw / 869411 B gzip; SHA256 c7ceb2f7a927a63f9cf6c746e006b26eea13bd34edc04204b85189d6dd6e8cc5. Touched fmt/clippy and real same-session scope/replay checks PASS.test result: ok. 61 passed; 0 failed; 7 ignored; 0 measured; 0 filtered out; finished in 12.30s
root-dispatch-result-guards-fmt.log
root-dispatch-result-guards-clippy.log
root-dispatch-result-guards-tests.log
root-dispatch-result-guards-verify-ssh-session-scope.log
root-dispatch-result-guards-verify-ssh-session.log
root-dispatch-result-guards-verify-ssh-session-rebridge.log
No setup-event contract change. Final full gate, both architecture comparisons and new final acceptance remain outstanding.
Resumption state: addhost-instant@a50defeb0 (pushed; first atomic step green). Two uncommitted files implement borrowed typed terminal notices with cursor/optional-field boundary tests. Initial compile caught a helper/field name collision in the derive; helper renamed serialize_session_ids. No commit of the failing code. Corrected touched gates queued in the sole build-slot wrapper resume-root-landing.py. The attached patch matches the current worktree (git apply --reverse --check PASS). Then shared admitted-manager fix/regressions, immutable final artifacts/controller, unwind attribution, fresh20 acceptance and final fast remain. No compiler/linker shipping flag change.
root-pending-typed-notices.patch.txt
Production-build Start fresh screenshots for independent UI review (functional suite50 PASS/8 conditional capture skips):
Completed terminal-notices:
6123b4037d, pushed. Private artifact 1891336 B raw / 842939 B gzip; SHA256 05d781178b9a744e3226fb97c6ed5c7ac08108a3c29bd2ac846aa0297acfd3f2. Touched fmt/clippy and real same-session scope/replay checks PASS.test result: ok. 63 passed; 0 failed; 7 ignored; 0 measured; 0 filtered out; finished in 12.83s
root-dispatch-terminal-notices-fmt.log
root-dispatch-terminal-notices-clippy.log
root-dispatch-terminal-notices-tests.log
root-dispatch-terminal-notices-verify-ssh-session-scope.log
root-dispatch-terminal-notices-verify-ssh-session.log
root-dispatch-terminal-notices-verify-ssh-session-rebridge.log
No setup-event contract change. Final full gate, both architecture comparisons and new final acceptance remain outstanding.
Completed admitted-scope:
8b90da9ce6, pushed. Private artifact 1889480 B raw / 842123 B gzip; SHA256 3671cd32f3b92df658b3de2c16f5f1c4b1a1d2b72c024ee44b1d88f22a9bf212. Touched fmt/clippy and real same-session scope/replay checks PASS.test result: ok. 41 passed; 0 failed; 2 ignored; 0 measured; 0 filtered out; finished in 0.23s
root-dispatch-admitted-scope-fmt.log
root-dispatch-admitted-scope-clippy.log
root-dispatch-admitted-scope-tests.log
root-dispatch-admitted-scope-verify-ssh-session-scope.log
root-dispatch-admitted-scope-verify-ssh-session.log
root-dispatch-admitted-scope-verify-ssh-session-rebridge.log
root-dispatch-admitted-scope-private-probe.log
root-dispatch-admitted-scope-private-withdrawn.log
root-dispatch-admitted-scope-full-probe.log
root-dispatch-admitted-scope-full-withdrawn.log
No setup-event contract change. Final full gate, both architecture comparisons and new final acceptance remain outstanding.
Scope regression refinement landed with8b90da9ce: authenticate the fixture shim Hello/session and verify its actual kernel cgroup is infraternal-session-.scope. This is RED against both immutable7940 full/private binaries and GREEN against both new images. Both withdrawn-manager refusal cases also PASS.
The initial full-image zero-probe assertion counted the ordinary dispatcher background warmup for future node actions (daemon.rs493-504). That warmup remains supported; the private image still requires zero repeated probes. The full image now proves the launch property directly, not by counting an unrelated background probe. No production exemption, daemon behavior change or setup-event contract change.
Committed root head8b90da9ce is clean/pushed. Exact x86 full22327472 B (+50080 vs pinned dev22277392); private1889480 B/gzip842123 B. ARM/controller and combined nonshipping size-profile/unwind diagnostics building through the sole slot. Existing D240/D247 compatibility contract remains contracts/host-setup.md95-104; both images retain the shared launch boundary. Fresh20 acceptance and final fast pending.
corrected-managed-scope-full-red.log
corrected-managed-scope-private-red.log
Committed-head8b90da9ce fresh20 acceptance completed on the designated perf VM under the bounded owner-file lock. All20 real native-painted-row,59 acknowledged input bytes, same-session handover and exact installed full-SHA proofs PASS. No interposed transport. n=10/RTT, p99 nearest rank (maximum).
Prompt targets NOT met. Joined remains below5 s at150 ms. Private image842123 gzip B still exceeds250000 B. Both full same-recipe deltas vs dev1b23ef31 remain+50080 B x86/+67216 B ARM; no size waiver/cap change.
Independent optional node-result guards made gzip412 B larger; rejected. Nonshipping same-engine/Rust-unwind size profile is715062 gzip B, passes5 real session proofs, now awaiting eight-case VM qualification with the same new full/controller artifacts. No shipping profile/linker change and no speed claim for it. Unwind diagnostics still running in sole build-slot wrapper. Final fast pending.
admitted-8b90-20-raw.tar.gz
admitted-8b90-20-summary.json
root-dispatch-final-artifacts.json
Nonshipping size-profile qualification complete:8 fresh cases,2/variant/RTT, same8b90 full/controller and observer. All8 native row/59-byte/full-digest proofs PASS. Medians: LAN423.379 ms control /316.192 candidate;150 ms1715.589 /1686.820. This small sample does not establish a sufficient WAN improvement; targets remain missed.715062 B gzip candidate still exceeds250000; no profile selected and no bulk-runtime performance claim.
Unwind diagnostics verify all1606 baseline source blobs. Paired GNU-stripped diagnostics (8 B below actual recipe) give head/base22327464/22277384 B; safe identical-code folding22310440/22260360 B. Both shrink17024 B, so delta remains+50080: reject recipe change for this job. No shipping flag/panic/engine change. Frame-record attribution is led by daemon+10652 B, sessions+5328, SSH bootstrap+4244; JSON includes unresolved-record accounting and is not compressed-size attribution.
Current-head full iterate/fast now running through the sole official build slot (fresh checkout packages after archive probes); no issue closure/deploy.
small-profile-8b90-8-raw.tar.gz
small-profile-8b90-8-summary.json
combined-small-artifact.json
root-unwind-artifacts.json
root-unwind-by_module-delta.json
Checkpoint gate on clean/pushed addhost-instant@8b90da9ce66edf6ef8a836b3405086fea35e6fd8: iterate PASS; fast exits1 only in static builds. Both runs:1166 native tests PASS,0 FAIL,140 ignored. fmt/VT fmt/CLI plugin compile-out/clippy/cargo-deny PASS. Separate complete TypeScript gate PASS after fast stops at Rust. Final static binaries hash-match the immutable measured artifacts (x86 88d420b0d51141771be07197942ec34daa36963671dbbfa194fd7f8ee395f533; ARM8b93a3e2715f81e411927f171350162a83173975f311badc07297d6a1c38d641).
Ran baseline source checker against the already-built pinned dev1b23ef31 artifacts with its own unchanged gate: both FAIL the same stale caps. x86 baseline22277392 vs22204304 (+73088); ARM19705112 vs19638744 (+66368). Branch22327472/19772328 =>+50080/+67216 vs baseline; requested16KB growth still NOT met. Caps unchanged. x86 key-memory initialization passes under8MiB; ARM runtime key-memory remains unexecuted on this x86 host, as printed by the gate. No other gate waiver.
Fresh20 prompt budgets still fail (previous attached table); work is NOT complete. Next nonshipping source trial removes launch-only state from the original-session-only path and releases unused mode arguments before awaits; sole build-slot wrapper queued. No deployment/issue closure.
fast-8b90.log
iterate-8b90.log
ts-8b90.log
baseline-static-cap-x86_64.log
baseline-static-cap-aarch64.log
Full scripts/verify fast output, verbatim:
Non-shipping size experiment at
8b90da9ce: sharing the scheduler representation for three setup tasks saves 19,392 B x86_64 and 21,808 B aarch64. Same-recipe deltas vs dev merge base remain +30,688 / +45,408 B. Strict SSH clippy and 41 SSH tests pass (2 ordinary ignored). Cost: one additional allocation per setup task; terminal frame paths unchanged. Not selected pending measured runtime tradeoff. Testing the existing boxed cold-command boundary next; root remains clean at8b90da9ce.setup-task-erasure-artifacts.json
setup-task-erasure-tests.log
setup-task-erasure-clippy.log
Author checklist updated for committed head
8b90da9ce. Native gates and all20 fresh byte-integrity checks pass; unchanged static caps, image/prompt targets and merge-base size deltas remain explicitly open. No source change or deployment. Next isolated command/bridge-future trials are waiting for the official shared build slot.author-review-8b90.txt
Fresh20 capture attribution (same already-attached immutable8b90 artifacts): LAN9/10 runs request a code4000 reconnect before first prompt; publication→prompt median188.55ms. LANcase05: publication6335.2, Attach6353.9, input ACK6382.6, close request6390.6, new Attach6469.1, first terminal write6497.9, native-painted prompt6521.1 (all epoch-relative suffixes ms). Existing D265 liveness deadline remains unchanged. This is an observed avoidable reconnect delay, not yet a proven single cause. WAN pre-prompt reconnects occur during renderer preparation BEFORE publication; publication→prompt median199.25ms includes the actual150ms SSH attach round trip. Do not attribute the entire remaining gap to payload size. Raw capture source is admitted-8b90-20-raw.tar.gz already attached.
admitted-8b90-liveness-attribution.json
Rejected archived-source command-future erasure at8b90da9ce. Command-only x86full22,329,296B (+1,824 vs actualhead). Combined with setup-task sharing:22,309,520B x86 /19,750,808B ARM; these are1,440/288B larger than setup sharing alone. Strict SSH clippy and41tests pass (2ordinary ignored); no shipping change. Next measured candidate retains setup sharing alone; dispatcher experiment and actual server-side VM qualification are pending.
command-future-artifacts.json
command-future-tests.log
command-future-clippy.log
Allocation caveat for the unselected setup-task erasure experiment: pinned Tokio1.53.2 src/task/spawn.rs178–185 already boxes futures larger than BOX_FUTURE_THRESHOLD; src/runtime/mod.rs628–632 sets release threshold16384 B. Therefore one explicit erased box per setup task does not necessarily mean one additional allocation versus production. Earlier issue wording was a conservative cost bound, not an allocation-count measurement. Exact end-to-end runtime qualification remains required before selection.
Rejected archived dispatcher-future erasure at8b90da9ce: private image1,887,080B raw /841,879B gzip saves only244gzipbytes, while full22,327,888B grows416B raw /648B gzip. Seven real scope/refusal/handover checks, strict SSH clippy and41SSHtests pass. The candidate for actual VM qualification therefore uses setup-task sharing without the bridge or command-body changes.
bridge-future-artifacts.json
bridge-future-tests.log
bridge-future-clippy.log
Setup-only archived candidate is built on both static architectures and with an actual changed GNU server controller (29,938,096B SHA256 cbe4ad6b68ad3a6316f301406135337dc62f8ae66ef85e4b58b834767320573a). Seven real scope/refusal/handover checks and strict SSH clippy/41tests pass. Stage identity is pinned; preliminary eight fresh paired cases (2control+2candidate eachLAN/WAN, same liveness observer) wait for the bounded perf-VM lock. No shipping change; exact deltas remain+30,688/+45,408B vsdevbase. A separate cold-helper scheduler sharing size trial is compiling; no per-terminal frame task is changed.
cold-candidate-artifacts.json
cold-candidate-source-patch.txt
cold-candidate-tests.log
cold-candidate-clippy.log
Non-shipping six-task shared scheduler candidate reaches the x86 merge-base allowance:22,288,752B stripped, SHA2566258f70c2ff27b95a05425d32adc53001d5431318d6ff269a6807e5848ce3714; devbase22,277,392B =>+11,360B (limit16KiB). Saves38,720B fromactual8b90head. All1672tracked files compared: onlybootstrap/credentials/duplex/lib/master differ, plusnewinternaltasks.rs; engine, bridge, process runner andbuildrecipe unchanged. ARM andtouchedgates inprogress. Exact changedcontroller/runtimequalification is prepared; notselected orcommitted. Setup-only predecessor isalready staged andwaitsforVMlock.
cold-scheduler-artifacts.json
cold-scheduler-source-audit.json
Six-task scheduler size trial complete (nonshipping;8b90baseline):
Strict SSH clippy and41tests pass (2normal ignored). x86 meets16KiB relative allowance; ARM remains7,184B over it, not claimed unavoidable. No nativeengine/buildprofile/bridge/process-runner change. Exact server/private artifacts for this version are queued through the official build wrapper, then a paired designated-VM comparison. Setup-only predecessor already waits for VM lock; observed public owner16:21CEST was mesh-autoswitch. Root remains clean8b90; no deployment.
cold-scheduler-artifacts.json
cold-scheduler-tests.log
cold-scheduler-clippy.log
Preliminary setup-only VM comparison (2fresh samples pervariant/RTT; notacceptance):
All8actual-engine/59byte/installedfullSHA proofs pass; allprompt budgets fail. LANcandidate samples301.898/1296.613ms: slowcase03 publishes early, reconnects onliveness, then receives noTTYdata for~930ms nearfulluploadcompletion. Notattributing this toboxing fromn2.
New concrete regression: embedded.rs Command::Attach removes anexisting local route without sendingClose; daemon.rs shared dispatch rejects duplicateAttachTerm on anopenstream. New realUnixwire regression FAILS on8b90 beforefix; Close-before-Attach repair passes itsfocused test, same-session validation still precedes anyClose. Separate cross-session/no-wire-write security regression andbroad gates pending. Roothas onlythese2uncommittedserver files; schedulerchanges remainunmerged untilactualsix-task runtimecomparison.
cold-8b90-8-raw.tar.gz
cold-8b90-8-summary.json
cold-qualification-manifest.json
shared-writer-reclaim-red.log
Completed shared-writer reclaim repair
9accfc1b89, pushed independently of scheduler experiments. Real Unix regression RED before / GREEN after; another session cannot emit Close. Touched server fmt, strict all-target clippy and complete server library tests PASS. Replay/session ownership unchanged. Exact final-head artifact rebuild, 10+10 actual-engine acceptance and full fast still required; no speed improvement claimed from correctness proof.shared-writer-reclaim-red.log
shared-writer-reclaim-green.log
shared-writer-wrong-session-green.log
shared-writer-repair-fmt.log
shared-writer-repair-clippy.log
shared-writer-repair-tests.log
Six-task scheduler comparison completed on the designated VM. Preliminary only: 2 fresh samples per variant and RTT; both variants use the pre-repair 8b90 actor. All 8 native-row, 59-byte and exact installed-digest proofs pass; all prompt budgets fail.
LAN candidate samples 404.020/395.124 vs control 375.563/373.712. Stage attribution differs across read-only check, staging and publication-to-paint; n=2 does not establish cause. Treat the 38,720/43,648-byte static saving as measured, with runtime performance still requiring final-head acceptance. Prepared separate commits for each caller after 9acc milestone gate completes. The 9acc writer-reclaim correctness repair is independent; its speed has not yet been measured.
scheduler-8b90-8-raw.tar.gz
scheduler-8b90-8-summary.json
scheduler-qualification-manifest.json
Completed Share setup task scheduling to reduce static growth (#189):
c5befef773, pushed. Touched tasks.rs, lib.rs, bootstrap.rs. fmt PASS, clippy PASS, tests PASS; 41 SSH tests passed, 2 ordinary ignored. Actual paired VM comparison and final-head full gate/10+10 acceptance remain required; this is not completion.root-cold-scheduler-fmt.log
root-cold-scheduler-clippy.log
root-cold-scheduler-tests.log
Completed Reuse setup scheduling for the bounded credential helper (#189):
722288e1c3, pushed. Touched credentials.rs. fmt PASS, clippy PASS, tests PASS; 41 SSH tests passed, 2 ordinary ignored. Actual paired VM comparison and final-head full gate/10+10 acceptance remain required; this is not completion.root-cold-credentials-fmt.log
root-cold-credentials-clippy.log
root-cold-credentials-tests.log
Completed Reuse setup scheduling for bounded bridge diagnostics (#189):
1f516166cb, pushed. Touched duplex.rs. fmt PASS, clippy PASS, tests PASS; 41 SSH tests passed, 2 ordinary ignored. Actual paired VM comparison and final-head full gate/10+10 acceptance remain required; this is not completion.root-cold-diagnostics-fmt.log
root-cold-diagnostics-clippy.log
root-cold-diagnostics-tests.log
Completed Reuse setup scheduling for SSH sign-in diagnostics (#189):
b3fe6181e4, pushed. Touched master.rs. fmt PASS, clippy PASS, tests PASS; 41 SSH tests passed, 2 ordinary ignored. Actual paired VM comparison and final-head full gate/10+10 acceptance remain required; this is not completion.root-cold-master-fmt.log
root-cold-master-clippy.log
root-cold-master-tests.log
Milestone on pushed
9accfc1b89: scripts/verify iterate PASS. Native totals: 1168 passed, 0 failed, 140 ordinary ignored across 126 test-result binaries; all ten designated-VM native timing cases verified. TypeScript checks also pass. This is the milestone, not final fast. The four separately gated SSH scheduler commits are now pushed throughb3fe6181e4; actual final-head binaries, acceptance and full fast are in progress.iterate-9acc.log
author-review-9acc.txt
Committed head
b3fe6181e4rebuilt and pinned.Same-recipe base
1b23ef311c: 22,277,392 /19,705,112 B. x86 meets 16 KiB allowance; ARM remains 7,616 B over it, not claimed unavoidable. The 250 KB private-image goal remains open. Seven actual-image scope/refusal/handover checks pass. Exact final server controller SHA256 c6f47fd1fab62c45b0426004d01f50ef719b5be7e14b5106aaf38b479d77a0db is staged for 10+10 fresh-target acceptance; final gates are running. No deployment.final-scheduler-artifacts.json
final-scheduler-acceptance-manifest.json
Final-head fresh-target acceptance measured on the designated 4-vCPU perf VM:
b3fe6181e4, actual production transports and native-painted rows. 10 fresh LAN +10 fresh 150 ms targets, sequential and interleaved; 1-minute load 0.30–2.30. All 20 same-session/native-row/59 acknowledged-byte/installed-full-SHA proofs pass. No attempt discarded.p99 is nearest rank (n=10 means maximum). Prompt budget passes: LAN3/10, WAN0/10. Prompt after click: LAN503.600/564.000ms p50/p99; WAN3049.400/3210.000ms. Thus prompt targets remain unmet; fully joined stays below5s at150ms. Image842127B gzip and ARM+24000B delta also remain open. Compared with the previous immutable8b90 batch, prompt medians are31.398ms lower LAN and24.245ms lower WAN; these separate batches do not isolate cause. Current complete artifacts and digests were posted separately. Final-head gate sequence is still running; no completion/deployment claim.
final-scheduler-b3fe6181-20-raw.tar.gz
final-scheduler-20-summary.json
final-scheduler-environment.json
scripts/verify iterate PASS on current pushed head
b3fe6181e4. Native totals1168 passed,0 failed,140 ordinary ignored across126 test-result binaries; all ten designated-VM timing cases verified; TypeScript checks pass. Final10+10 receipts and actual static sizes already attached. Exact full scripts/verify fast is now running; this is not a final-fast pass. Image842127B, ARM+24000B and prompt-target misses remain open.iterate-final-scheduler.log
author-review-b3fe.txt
Review checkpoint at pushed head
b3fe6181e4. Stopping for the requested break-it review; D255 performance/image acceptance remains incomplete. No deployment and no issue closure.Changes in this continuation: same-session writer reclaim now sends the existing Close before AttachTerm (real Unix regression RED before / GREEN after; another session cannot emit Close), and six bounded SSH helper tasks reuse scheduler/drop code. Each logical source step was gated and pushed separately. Setup event contract and real-login compatibility remain unchanged.
Base:
1b23ef311c, same recipe. x86 meets16KiB allowance; ARM remains7,616B over it, not claimed unavoidable. Session images: x861,889,480B raw /842,127B gzip; ARM1,646,448 /772,697B. The250KB goal is unmet. Fast-built full binary digests match the immutable acceptance manifest on both architectures.Designated4-vCPU VM,10 fresh targets per RTT, sequential/interleaved,1-minute load0.30–2.30; p99 nearest rank=n10 maximum. All20 native-row/same-session/59 acknowledged-byte/installed-full-digest proofs pass. Prompt budgets: LAN3/10, WAN0/10; no attempt discarded. Fully joined remains below5s WAN. Seven additional real-image scope/refusal/rebridge/handover proofs pass.
Final-head iterate PASS. Final-head fast:1168 native tests passed,0 failed,140 ordinary ignored; fmt/nativefmt/compile-out/clippy/test/cargo-deny PASS. Static builds FAIL only on unchanged caps: x86+84,352B and ARM+90,368B over cap; pinned dev baseline also fails those same caps (+73,088/+66,368B), raw baseline receipts linked below. Caps were not changed or bypassed. x86 linkage/trust/key-memory checks pass; ARM key-memory runtime is explicitly untested on the x86 host. Separate full TypeScript gate PASS. Author checklist attached; independent break-it review is still required.
Evidence:
final-scheduler-artifacts.json
final-scheduler-acceptance-manifest.json
final-scheduler-b3fe6181-20-raw.tar.gz
final-scheduler-20-summary.json
final-scheduler-environment.json
fast-final-scheduler.log
ts-final-scheduler.log
final-fast-artifact-identity.json
author-review-b3fe.txt
baseline-static-cap-x86_64.log
baseline-static-cap-aarch64.log
Exact scripts/verify fast output, verbatim:
Review checkpoint cleanup complete at
b3fe6181e4. Checkout clean; all five source commits pushed. Removed this job’s build target and temporary source archives, including the invalid no-VT diagnostic source. Retained raw receipts, immutable measured binaries, baseline artifacts and scripts. Full fast output is quoted verbatim in the preceding checkpoint comment. No deployment or issue closure; prompt/image/ARM requirements remain open for review.review-cleanup.json
Delta verification started: target
b3fe6181e4, previously reviewed2b779884, branch review-addhost-instant. Preserved prior review as review-addhost-instant-prior-4c1de592; replaying R189-01/R189-02, checking writer reclaim/scheduler and production036bc2bb0comparison on locked perf VM. No implementation fixes, push or deployment.Delta review: R189-02 remains P1 at target
b3fe6181e. Original product regression on review-addhost-instant@94c9ce4f fails: expected retained setup subscription=true, received=false; same production-build run raw setup WebSocket control passes (1 PASS/1 FAIL), port 5179. Wizard closes after publishing instant-fail.lab but product sends no host/window-bound /v1/bootstrap/setup/ws subscription, so later failure/retry remains unavailable. Evidence: ~/.cache/codex-scratch/addhost-instant-review/evidence/browser-replay.log and addhost-instant-review-after-publication.png. Issue history puts UI repair on separate ui-addhost-pill@1ea2db04; that repair is absent from this target. No implementation changes. Writer-reclaim controls 2 PASS and SSH bridge controls 6 PASS on digest-pinned final controller. Perf rerun/full gate ongoing.Independent final-head VM208 rerun complete (4 vCPU; recorded per-attempt load; 10 LAN +10 at150ms; no discarded attempts). All20 same-session/59-byte/exact-full-SHA proofs PASS. After-sign-in p50/p99 ms: prompt LAN508.922/614.132, WAN1909.305/2019.318; joined LAN1435.751/1730.132, WAN4620.785/4732.245. Later D255 targets remain missed; these are not merge blockers under the approved step1 brief. Production036bc2bb0 was NOT measured: strict-improvement acceptance remains unverified. These20 successful transfers do NOT establish target-stop/network-cut/slow-disk failure acceptance.
The copied harness completed all20 records, then refused its stale cleanup-name assertion; removed only authenticated review-owned container8b4272b177bf manually and generated summary from the20 retained records without rerun/discard. Raw console/recovery preserved.
ARM ELF independent attribution: +24000B vs pinned dev1b23ef31; .gcc_except_table +30796, .eh_frame +26020, .eh_frame_hdr +4424, .text -23860, .rodata -16384. Unwind/exception tables +61240B dominate growth. Another7616B must be removed to meet16KiB; no measured cut demonstrating that limit yet. No recipe/cap/implementation changes. Full fast builds continue through a build slot; native latency queued on VM.
perf-20-summary.json
perf-20-raw.tar.gz
Review checkpoint: HOLD (R189-02 P1 present), target
b3fe6181e, local review head94c9ce4f2. Test-only commits863af0cf9/94c9ce4f2; no implementation fix/push/deploy. All20 fresh success-path VM proofs and native100-sample echo/flood gate PASS; raw receipts already attached. Four new managed-scope/withdrawn-manager checks also PASS, both full/private images. R189-01 source repair/regression is present, but its fresh full-gate row has not executed yet. Full scripts/verify fast: fmt PASS, VTfmt PASS, CLI plugin compile-out PASS (29tests), currently rebuilding workspace Clippy in fresh owned target. TS check waiting for build slot. Output ~/.cache/codex-scratch/addhost-instant-review/evidence/verify-fast.log; final full output/verdict still pending. Required production036bc2bb0 comparison and target-stop/network-cut/slow-disk matrix remain unverified, not green acceptance. Deferred image/prompt targets are not blockers under step1 approval.Gate checkpoint at unchanged review head94c9ce4f2: first fresh-target full-fast attempt reached Clippy, no reported compiler/test errors, but remained very slow. Preserved entire partial output as ~/.cache/codex-scratch/addhost-instant-review/evidence/verify-fast-shared-cache.log; stopped only that attempt’s owned compiler subtree. Restarted exact source with a verified job-local sccache socket/cache (shared cache untouched). Replay is now waiting for a host build slot; TS check also queued. No gate pass/implementation-failure or timeout waiver claimed. HOLD remains R189-02 P1. Final full output still pending.
HOLD (P1 present) — merge after fixes. Confirmed current counts: P0=0, P1=1, P2=0, P3=0. Target
b3fe6181e4; local review-addhost-instant head94c9ce4f2a7f9ccdb35c82f0fd70da650ff1ac74.R189-02 still FAILS: production wizard closes after publication and announces readiness (wizard.ts:986), but product sends no host/window setup subscription. Later failure/retry is unavailable. Original browser regression1FAIL; raw setup control1PASS. Test commits863af0cf9/94c9ce4f2. Separate ui-addhost-pill repair is absent from this target; verify the composition before release.
R189-01 repair and normal nonignored regression are present. Fresh independent replay did not execute; do not declare it independently resolved. Correct replay now omits --ignored: cargo test --locked -p infraternal-plugin-ssh review_stopped_commit_channel_finishes_the_private_probe.
20 fresh success-path VM proofs and native echo/flood budget PASS; TS checkPASS; targeted writer/admission/replay/namespace/managed-scope controlsPASS. Production036bc2bb0 head-to-head and target-stop/network-cut/slow-disk matrix remain UNVERIFIED. Strict improvement and complete D128 failure acceptance are not certified. Deferred D255 image/prompt targets are not used as blockers. ARM+24000B is dominated by unwind/exception growth; another7616B cut to16KiB is unproven. Measurements/attribution are in the attached verdict and preceding receipts.
Independent scripts/verify fast is INCOMPLETE. First attempt passed formatting/native formatting and CLI compile-out29tests; I interrupted the slow Clippy build to change cache/scheduling, releasing the build slot. The replay never obtained another slot and was cancelled after exceeding the intended review window. No full native/static gate pass or waiver claimed. Complete partial output is quoted verbatim below. Author's exact-target final summary is quoted separately and remains author evidence.
Cleanup: only owned queued replay/compiler subtree stopped; owned targets/cache/fixture state and VM review directory removed; raw receipts retained. Checkout clean at94c9ce4f2. No implementation fix, push, deployment, owner fixture change or issue closure.
verdict.md
verify-fast-shared-cache.log
receipt-sha256.txt
Independent incomplete attempt, verbatim:
Author
b3fe6181final gate summary, verbatim (NOT independently reproduced):Resuming #189 on addhost-instant at
b3fe6181e4; merge base with fetched origin/dev1b23ef311c, dev51abed0d5f. Clean worktree. Read latest review HOLD at94c9ce4f2: R189-02 product subscription absent, production head-to-head/failure matrix and R189-01 replay pending, ARM +24000 B. Rebase first, preserve existing safety fixes, integrate existing approved setup consumer, replay regression and complete measured acceptance/full gate through build-slot. No deployment.Rebased with merge structure preserved onto origin/dev
51abed0d5; headab75cfbcfa, pushed with an exact old-head lease. Imported unchanged R189-02 regression863af0cf9+ external-capture94c9ce4f2 and composed the approved ui-addhost-pill commits through1ea2db04. Conflicts preserve renderer preparation and embedded-host affinity. Existing host-setup event contract unchanged. TypeScript checkPASS; setup/identifier units16PASS/0FAIL. Full production-build browser regression/failure-retry/legacy suites queued through the sole build-slot wrapper. Stopped only my still-queued duplicate red replay (PID1151199); supplied review red remains evidence, no fresh red claim. Captures use the shared external evidence helper. Native/perf acceptance still pending; no deployment.Recovery matrix finding on immutable
b3fe6181artifacts, perf VM208, fresh150ms target: stopped the target-owned gzip decoder during deferred full upload at2064384 public image bytes. Original shim remains live (one running session scope); browser painted prompt and BEFORE marker. Retained failure arrived; one same-host retry six seconds later failed again. Controller0PASS/1FAIL; fault/lock exits101, no discarded attempt. Source install.rs968 resumes .bootstrap-lock/payload even when that file is the incomplete full-image write. Preparing a real-file regression and rebridge from the structurally owned cache, with digest verification against the admitted session/full images. Preserve original session/writers, signature and ownership checks. This is a preliminary pre-rebase diagnostic, not final-head acceptance.resume-fault-target-stop-diagnostic.log
Partial-upload repair candidate on pushed UI composition
ab75cfbcf(native changes not committed yet): real-file regression RED0/1 at resume_session before fix; GREEN1/1 now. Checks original bridge traffic, reopening after a partial write, foreign transaction refusal and no execution of altered cache bytes. Additional verified-full-image rebridge1/1PASS. Two shared immutable digest slots are populated only from trusted source-image hashes; retry hashes the structurally owned cache before executing and retains host/session/transaction checks. Buffered and compound transfer paths both pin the full image. No new dependencies, engines, wire or event changes. FmtPASS, strict library lint in progress; complete browser/SSH gates queued through one official build-slot wrapper before commit. Full gates, exact-head VM matrix and production comparison still pending. Patch attached so resumption does not depend on this chat.rebridge-repair.patch
partial-upload-red.log
partial-upload-green.log
full-cache-rebridge-green.log
R189-01 replayPASS: cargo test --locked -p infraternal-plugin-ssh --lib review_stopped_commit_channel_finishes_the_private_probe; 1passed/0failed/0ignored. Strict SSH library ClippyPASS after the cache-rebridge repair. Focused partial-cache and verified-full-cache testsPASS. Full all-target Clippy/SSH suite and production-build browser regression remain queued in the sole official build-slot wrapper (PID1163947, started17:56UTC, still waiting18:29UTC). No final gate claim. Production comparison will bundle the same test adapter against each version's own provisioning, connection and renderer source, with legacy publication handled explicitly; the baseline lacks renderer preparation and keeps its actual behavior. The earlier prebuilt-client candidate will not be used as final comparison evidence.
resume-r189-01.log
rebridge-lib-clippy.log
Committed and pushed partial-upload recovery as
3a823ad84e; checkout clean. Fmt and strict SSH library ClippyPASS; partial-upload/foreign-journal/altered-cache regression1PASS and verified-full-image rebridge1PASS. R189-01 normal replay1PASS/0FAIL/0ignored. The commit records these scoped gates accurately; complete SSH all-target Clippy/tests, production-build browser suite and final iterate/fast are still queued, not claimed passed. Retained setup event contract unchanged. Same-session cached rebridge uses at most two admitted image digests and repeats the cache hash before exec; no partial staging-file execution. Comparison builds will use this committed head and each version's own client source. Raw red/green/patch receipts linked in preceding comments.Build-capacity blocker at18:41UTC: the sole official build-slot wrapperPID1163947 has waited45minutes since17:56UTC. Needs an orchestrator-managed slot to become available; continuing to wait. Head3a823ad84 is pushed and clean. Focused recovery regressions, library lint and R189-01 replay pass. Full browser/SSH/static/native gates and exact-head VM acceptance remain pending.
Queued pipeline: /home/kayg/.cache/codex-scratch/addhost-instant/resume-browser.sh (browser→captures→SSH gates→resume-artifacts.py→resume-gates.sh). Raw output: evidence/composed-browser.log in the same cache. Artifact builder emits RESUME_COMPARISON_ARTIFACTS_READY when the first5immutable artifacts and both version-specific client bundles are ready; then stage-comparison.py runs40fresh comparisons in4short fair-lock segments. Remaining ARM builds/gates proceed independently. Recovery diagnostic/fix code and receipts are attached above.
build-slot-wait-45min.txt
Capacity cleared around18:58UTC: sole wrapperPID1163947 acquired an official build slot after about1hour. Production-build add-host suite is running on unique port5291; first functional casesPASS, no whole-suite verdict yet. The release-aware wake helper only woke this wrapper's owned sleep and exited after the official helper acquired its own lock.
Documentation step committed/pushed as
5b50c44604: add required module-role headers to the3imported client test modules; TypeScript check and16focused unitsPASS. Production code is unchanged from native fix3a823ad84. Checkout clean. The pipeline will continue with SSH gates, pinned comparison artifacts, both ARM/static comparisons, iterate and fast.R189-02 product regressionPASS in the ongoing production-build browser suite: the published host/window sends its retained setup subscription after the wizard closes. Raw retained-API controlPASS; actual failure/retry, clean completion while typing and reduced-motion failure/retry flowsPASS. Full suite still running.
Measurement correction: the broad addhost-*.spec.ts selection also ran imported addhost-pill.spec.ts, which contains long-task/frame/script-time assertions, on the build host. Those timing outputs are INVALID evidence and will not support acceptance. Rerun the unchanged four timing specs on the locked perf VM using staged prebuilt preview assets; do not build on the VM or change assertions/deadlines. Functional state results above remain functional observations only. Full native/static/performance comparison and failure matrix still pending.
Current pushed head46d85ba4cb0546574b8b5f77e79650080f0df597; clean. The inherited network wizard case failed only at its old ready-toast assertion (enrollment/Console/sidebar checks passed). Updated the assertion to require no toast per the approved #189/D255 pill flow, preserving authentication/publication/cleanup checks; focused production-build replay1PASS and TypeScript checkPASS. Separate test-only commit46d85ba4. The initial broad suite had57PASS/1obsolete-assertionFAIL/14opt-in-capture skips. Of57passing cases,4timing cases ran on the wrong host and are excluded as timing evidence;53functional passes plus the corrected1/1replay cover54functional cases. R189-02 subscription, later failure/retry and reduced-motion flowsPASS.
Correct timing replay on perfVM208/infraternal-perf-test, initial load2.51/1.77/1.32: all4unchanged pill timing specsPASS. Assets built on the build host and SHA-verified after staging; source/spec identity in attached provenance; no VM build or assertion/deadline change. Long tasks >=50ms:0 during status, control0; script share6.5% versus6.2% control; worst observed frame gap66.7ms (reported by the spec). These are pill/test-double measurements, separate from real host provisioning acceptance. Both explicit timing/lock exits0.
Six desktop/phone capture runs are in progress; complete SSH/native/static gates, production comparison and recovery matrix remain pending.
pill-timing-vm.log
pill-timing-provenance.json
readiness-browser-replay.log
Head46d85ba4c: complete SSH package tests182passed/0failed/11environment-dependent ignored (correcting the commentary count183); strict all-target SSH Clippy and fmt pass. Production-build captures completed6/6 (1440x900,393x852,375x667 ×ink/paper). Selected raw captures attached for independent visual review; no self-awarded visual verdict. Static artifacts, comparison and final whole-branch gates still pending.
rebridge-clippy.log



rebridge-ssh-tests.log
Checkpoint20:00UTC, clean pushed head46d85ba4cb0546574b8b5f77e79650080f0df597, dev merge base51abed0d5. Browser functional regression/captures, SSH182pass/0fail and all-target Clippy remain complete as attached above. No new compile failure. The sole official build-slot wrapperPID1699027 runs resume-after-browser.sh; current first x86 static build is in final project compilation. Raw master log evidence/composed-browser-replay.log under /home/kayg/.cache/codex-scratch/addhost-instant/.
Resumption: resume-artifacts.py builds immutable images/controllers and emits RESUME_COMPARISON_ARTIFACTS_READY before remaining ARM/baseline builds and controls; resume-gates.sh then runs iterate+fast. continue-measurements.py waits for that marker, then stage-comparison.py (40fresh comparisons, both versions own shipping clients) and stage-fault-matrix.py (three disposable target faults). collect-comparison.py fetches only public raw receipts and computes nearest-rank sample p99; size-comparison.py compares both full images to dev51. Run these from the worktree, not another checkout. Final current-head production/fault/native/static results remain pending; no acceptance claim.
Head46d85ba4c x86 buildsPASS. Full22343344B (gzip9:10001328B); dev51baseline comparison pending, so no cap-relative delta claim yet. Private session1889480B (gzip9:842448B), SHA2565fd1655184d2e9ea3c9a4df23535190f90d92edefebb96ca8f4f4681bb7be470, identical to the reviewed
b3fe6181private image. Recovery repair therefore adds no first-prompt image bytes. This is still above the earlier250KB compressed goal; do not describe that deferred image target as achieved. Actual transfer encoding/bytes and final prompt/join timings will come from the exact-head VM comparison, not build-host timings. Same panic/unwind and static build recipe; no caps changed. Current stage: release test controller.Release controller at46d85ba4c stops at5test-only compile errors after the dev rebase:3obsolete TerminalOutput::Pong matches and2manual Command::Attach constructors missing monitored. No normal-library build error (full/private static buildsPASS). Dev51removed forwarded Pong in favour of actor-owned route proofs; use RouteExpired as a refusal/outcome and keep raw/native regression attachments monitored:false, matching EmbeddedAgent::attach. Patch touches4test modules only. Baseline dev51server --tests check, strict all-target server lint and affected writer/SSH/control regression replays are queued through the sole official build-slot wrapper; no native gate or commit claim yet. VM comparison still pending.
resume-46d85ba4-controller.log
rebased-test-api.patch
Build-capacity blocker20:55UTC: sole queued official wrapperPID2062153 has waited45minutes since20:10UTC. Dev51baseline --tests check, strict server lint,9affected regressions and subsequent artifact/gate pipeline have not started. Need an orchestrator-managed slot to become available; continuing to wait. Current pushed head46d85ba4c;4test-module API-alignment patch (6added/4removed lines) attached above remains uncommitted until scoped gates pass.
Queued command: fix-rebased-tests.sh → commit-rebased-tests.py (only after passing baseline/fmt/lint/2writer+6SSH+1control checks; atomic commit/push/issue update) → resume-artifacts.py → resume-gates.sh. Current measurement watcher uses ADDHOST_BUILD_PIPELINE_PID2062153 and will start only after RESUME_COMPARISON_ARTIFACTS_READY. A release-event watcher wakes only this wrapper's own sleep; the official four-slot flock remains authoritative and no second wrapper/extra slot is started. Logs: evidence/rebase-test-pipeline.log and rebase-waiter-wakeup.log under /home/kayg/.cache/codex-scratch/addhost-instant/.
rebase-slot-45min.txt
Capacity cleared20:56UTC after about47minutes. Official wrapperPID2062153 acquired its own managed slot; release watcher woke only its child sleepPID2458993 and exited after fix-rebased-tests.sh started. Pinned dev51server test compilation is now running, then repaired-head all-target lint and9affected replays. The single wrapper retains the remaining artifact/gate pipeline; no second slot wrapper, capacity override or foreign process change. No native verdict yet.
Committed/pushed test-only dev API alignment
4872432dbc. Pinned dev51 server --tests checkPASS; obsolete output matches/manual attachment constructors are additions from this lane. Fmt and strict all-target server ClippyPASS; affected replays2writer+6SSH+1controlPASS/0FAIL. Route expiry still fails early-session/writer oracles; no deadline or assertion relaxed and no production-code change. The controller can now be rebuilt. Existing46d85ba4c browser/SSH/image receipts remain source-equivalent outside these4test modules. Continuing immutable comparisons, fault matrix, size baseline and iterate/fast in the same sole official slot.rebase-test-baseline.log
rebase-test-clippy.log
rebase-test-resumed_writer.log
rebase-test-remote_ssh_tests.log
rebase-test-saturated_native_control_queue_does_not_hide_primary_terminal_output.log
Comparison harness finding, caught BEFORE measurements: Bun onLoad ignored resolveDir, so initial production/branch entry hashes were identical despite different client sources. Those bundles cannot certify the comparison. No attempt JSON or active batch existed; paused only local stagePID2817064, stopped and inspected this job's queued flockPID3984126 (no children), then cancelled it. No foreign lock/process changed and no VM measurements discarded.
Fixed the measurement adapter with explicit absolute imports for each version's provisioning/terminal/connection/protocol/renderer modules; keep each version's previously built shipping assets. Rebuilt only the small instrumented entries, no product-code change or extra heavy build. Production entry21cdd129… versus branchd5c64586… now differ; provenance includes source-file hashes. Runner asserts different bundle hashes before staging. Corrected namespace comparison-4872432d-own-clients-v2; same-head native/ARM/gate pipeline continues independently in the sole official slot. Performance acceptance and recovery matrix still pending.
comparison-client-provenance-invalid-resolver.json
comparison-client-provenance.json
Reproducible comparison inputs: the explicit-import entry builder and production test-only fixture patch. The .ts attachment was rejected with HTTP422; the same source is attached as plain text. Corrected run still waiting for the fair VM lock; no performance result claimed.
comparison-client-builder.txt
production-fixture.patch
Corrected, version-specific VM comparison: first8attempts have59-byte/full-image/normal-transport proofs; production036bc2bb0 LAN attempt3 fails the unchanged acknowledgment oracle. Both BEFORE_HANDOVER and AFTER_HANDOVER rows paint, input sequence59, but client acknowledgment remains30 until the integrity timeout. Current dev-schema/session authority flags remain present; raw baseline failure attached. This is an immutable production baseline result, not a branch pass or a discarded attempt. No headline comparison yet (only2valid samples per profile).
Resume retains every existing row/log, skips already recorded attempts without overwriting them, stops on any new branch failure, and preserves/counts this diagnosed baseline failure. A final short segment adds fresh attempts only where necessary for10successful samples. Report success-conditioned timing distributions together with all attempts/failures; no production-code or byte/time assertion change. Current branch4872432db and full/private hashes unchanged. ARM/dev builds and final gates continue in the single managed host slot.
production-036bc2bb-0-03.log
comparison-segment-0-attempts.json
Exact-head production comparison on perfVM208/infraternal-perf-test (4vCPU), branch4872432dbcda7916c05dd10e49ca259159908a80 versus production036bc2bb0. Both versions use their own shipping clients/renderers and full static images, the same measurement adapter, and fresh disposable targets. Fair locks released between segments; per-attempt machine/load receipts and all raw logs attached.
Each timing row contains10successful samples; p99 is the nearest-rank sample maximum, not a population-tail guarantee. All20branch attemptsPASS with59acknowledged input bytes, complete before/after painted rows, exact installed-full-image SHA and normal transport paths. Total43attempts:40successful,3retained production LAN acknowledgment timeouts (both output rows painted, ACK30/59). These failures are not passed or discarded; fresh extra baseline attempts supplied10successful baseline LAN samples. Timings are success-conditioned and reported with those failure counts. All prompt/join p50 and p99 comparisons are faster on the branch, including click-to-prompt/join (in JSON).
Targets remain explicit: branch first prompt363/490ms LAN and1756/1815ms at150ms RTT still miss approximately300/1500ms; minimal image842448B gzip9 still exceeds250KB. This continuation fixes the review/retry issues and records the completed production comparison, not those deferred target claims. Fault matrix, ARM sizes and final whole-branch gates remain pending.
Fresh x86 size versus dev51: branch22343344B, dev22329104B, delta+14240B (within16KiB). The unchanged cap itself fails on dev51:22329104B versus22204304B, delta+124800B; production trust passes and PT_INTERP is absent. No cap/profile/unwind change.
comparison-summary.json
comparison-public-receipts.tar.gz
dev-x86-static-cap-control.log
Exact-head4872432db recovery check PASS: stopped target-owned gzip PID337 when deferred full-image write reached2064384B, after BEFORE_HANDOVER painted. Initial failure is retained; one retry40seconds after failure completes. AFTER_HANDOVER paints, all59input bytes acknowledged, installed full SHA71b68620… and private SHA5fd16551… match. This is a transfer-process stop preserving the original shim, not a host power-off claim. It reproduces the old partial-stage scenario now repaired by3a823ad84. Raw intervention and browser/server receipts attached. Network-cut and slow-write cases remain queued/in progress; do not count the matrix complete yet.
Dev51 ARM baseline built19752920B. Its unchanged static cap also fails independently: cap19638744B, delta+114176B; production trustPASS and PT_INTERP absent. Branch ARM is building; its delta/unwind breakdown remains pending. No caps or unwind flags changed.
fault-4872432d-target-stop.log
fault-4872432d-target-stop-summary.json
dev-arm-static-cap-control.log
Fresh same-recipe stripped static sizes, branch4872432db versus dev merge base51abed0d5 (pinned Rust1.99.0, normal release profile):
x86 meets the16KiB allowance. ARM remains8800B above it; this is the measured exception case requested in the brief, not a16KiB compliance claim. ARM unwind/exception deltas: .gcc_except_table+30636B, .eh_frame+26172B, .eh_frame_hdr+4456B, total+61264B. Its .text falls22708B and .rodata falls16384B. x86 has a similar unwind increase61444B but a larger text cut33836B, explaining most of the architecture difference. Net ARM growth is25184B (below64KiB).
Preserved production unwinding and crash-report information, panic isolation, dependencies and caps. Reject dropping unwind/exception sections or panic=abort just to meet size. No verified code-sharing candidate in this continuation proves the remaining8800B cut, so do not label this growth unavoidable or owner-approved. The section-level measured tradeoff is retained for review. Current cache repair adds no private first-prompt image bytes (same5fd16551… x86 private image as the reviewed head). Full/private ARM build and artifact ledger identities attached; final native/static gates and remaining recovery cases still pending.
resume-size-comparison.txt
resume-size-comparison.json
resume-artifacts.json
Precision on the preceding size post: that table compares the full shipping binaries. The private ARM image is still compiling; its final identity and the later real-binary controls are not yet complete. The attached ledger contains both full ARM identities and both x86 images/controllers at that checkpoint. Whole-branch iterate/fast and the remaining recovery cases stay pending.
Head
4872432dbartifact milestone complete: all 9 immutable image/controller records built; seven real x86 private/full binary controls PASS (scope ownership, same-session/writer handover with replay, rebridge, and both private/full manager probe + withdrawn-manager refusal). Fresh R189-01 normal replay: 1 passed, 0 failed, 0 ignored. Full ledger and verbatim control logs attached. ARM images were built, not executed on x86.Network-cut recovery PASS on VM208: target eth0 down for 35 seconds, then restored; initial failure, one retry, both complete painted rows, all 59 input bytes acknowledged and exact installed image hash. Transfer-decoder stop already PASS. Slow-write case is queued; matrix is not yet complete. scripts/verify iterate has started in the same official build slot and is waiting on its unchanged, source-verified VM RTT self-test. Final fast remains pending. No overall gate claim.
real-binary-controls.txt
resume-artifacts.json
fault-network-cut-console.log
Exact-head
4872432dbVM recovery matrix COMPLETE: all 3 cases PASS, with initial failed status, one explicit retry after the bounded recovery window, ready completion, both complete painted rows, all 59 input bytes acknowledged and exact installed full-image SHA.Each uses a fresh target and its own short fair-lock interval. Retry is sent40seconds after failure, allowing the35second link interruption to end; no product retry or byte assertion changed. This is transfer-process termination and a bounded network outage, not host power-off coverage. Raw target/browser/server receipts and all matrix summaries attached. No lost-byte claim beyond the tested before/after rows and59-byte acknowledgment oracle; independent real-binary replay/duplicate-input controls also pass as attached above.
All performance comparison and failure-matrix acceptance requested in this continuation is now recorded. Final iterate/fast remain running/pending; no overall gate or merge verdict yet.
recovery-matrix-summary.json
recovery-matrix-public-receipts.tar.gz
Final iterate at4872432db exposed pre-existing UDP/TCP fixture port collision before the capability assertion. Original reservation sequence is byte-identical on dev51. A real held TCP socket reproduces AddrInUse on both dev51 and branch: each1test,0pass,1fail,0ignored; receipts port-collision-baseline-red.log and port-collision-branch-red.log. Earlier commands omitted feature remote and selected0tests; their logs retained with -zero-selector suffix and never counted. Bounded test-only paired reservation repair now applied; other bind errors stay fatal, native/fallback assertions and timeout unchanged. Strict all-feature/all-target lint and two actual tests running through build-slot; whole gates follow.
Committed/pushed test-only paired-port repair
5bce43b627. Real held-TCP collision reproduces RED on both dev51 and branch before fix; bounded reservation retry now GREEN, original native/fallback capability assertion GREEN. Fmt and strict all-target agent crate lint PASS. One test module changed; production source and all measured image/controller code are unchanged from4872432db. Re-running iterate and final fast on this new head in the same managed slot; no gate waiver. Raw red/green/lint receipts in job evidence/port-collision-* and paired-fixture-*.log; attach with final receipts.Raw paired-port controls for pushed5bce43b62 (each actual test1passed/0failed/0ignored after fix):
dev RED, branch RED, strict lint, collision GREEN, original capability GREEN, first failed iterate, preserved.
Head5bce43b6277366a9cf152c267230dc30b9d98e4e: scripts/verify iterate PASS through the official build slot. Every Rust row PASS; the10physical transport cases and exact graph timing case PASS on VM208 under fair locks. Remaining206canvas/unit cases and37protocol cases PASS locally. Full log below; final fast is now running on the same head.
resume-iterate.log
Head
5bce43b627: final scripts/verify fast completed through build-slot, exit 1. Verbatim Rust summary:Static failure is the unchanged size caps (including the boundary self-test rejecting that same oversized input). Both shipping binaries built: x86 22343344 B versus cap 22204304 B; ARM 19778104 B versus cap 19638744 B. Production trust PASS and PT_INTERP absent on both; native x86 key locking/dump exclusion PASS. ARM key-memory execution explicitly not checked on x86. Both pristine dev cap controls already attached above. No cap/unwind change or waiver.
After the static failure prevented fast from reaching TS, ran its unchanged ts_verify_fast separately: exit 0, exact graph timing case on locked VM and remaining 206 unit plus 37 protocol cases PASS. Complete iterate had already passed on this head. Full verbatim fast/TS logs and identical 529-name server-controller/workspace coverage attached.
Provenance check: ARM and both private images byte-match measured
4872432dbartifacts; full x86 gate image has equal size but a different hash. Cargo recorded two distinct unit configuration hashes with identical feature lists/rustflags. Rebuilding final full x86 with the exact explicit-linker comparison recipe before asserting byte equivalence. Final reporting/cleanup pending this check.resume-fast.log
resume-ts-gate.log
server-feature-coverage.json
Artifact provenance follow-up at unchanged head
5bce43b62: rebuilding x86 with the exact explicit comparison linker/compiler environment still yields 42b91c5640fb1bb56f5461c36b7af16f6f6a8349a7e6d90142caa81fe53140b5, not the measured4872432dbfull image 71b68620…. Both are 22343344 B with identical section sizes; byte differences remain. ARM and both private images match previous identities exactly. No production source changed after4872432db(one agent test module only), but do not claim full x86 byte equivalence or attribute the old full-image measurement to this new hash.Keep all earlier 43 attempts and three recovery receipts. Building current release controller and snapshotting final artifacts through build-slot, then repeat the matched production comparison (10 fresh successful targets per version at LAN/150 ms, retain every failure) and three scoped recovery faults against the actual final images. Client source/bundles unchanged; both own shipping clients remain in use. Gate status stays: iterate PASS, final fast exit1 only unchanged size caps, separate unchanged TS gate PASS. No new commit or production change. Final handoff follows new exact-artifact acceptance.
Final artifact snapshot complete at
5bce43b62. Current release controller rebuilt successfully and byte-matches the previous controller: 30323648 B, SHA 9ebb79fcab6cac720f48f6cc8ce48fb128ef02aebab4eac27cf001e907173496. Final full x86 is 22343344 B, SHA 42b91c5640fb1bb56f5461c36b7af16f6f6a8349a7e6d90142caa81fe53140b5, gzip9 10001292 B (36 B smaller than the previous full image). ARM and both private images unchanged byte for byte; private x86 gzip9 remains 842448 B. Ledger preserves all old records and adds five final records.New exact-image comparison on VM208 is running in comparison-5bce43b6-own-clients-v2. First segment: all six branch attempts PASS; one production LAN ACK timeout retained. No failure discarded or counted as a pass. R189-01 also passes in final fast (bootstrap::proof_tests::review_stopped_commit_channel_finishes_the_private_probe). Full repeat comparison and scoped fault matrix still pending.
resume-artifacts.json
final-source-equivalence.json
canonical-final-build.log
gate-measured-elf-section-diff.json
New immutable production control failure in final-image repeat: production-036bc2bb-0-05. First prompt and complete BEFORE row painted, but inputAcknowledged stayed0 through the120-second integrity deadline while Ping/Pong continued. AFTER was not sent: unchanged adapter addhost-timeline-entry.ts:80 waits until inputAcknowledged >= inputSequence before enqueueing AFTER. This differs from the earlier ACK30/both-rows timeout, so the driver stopped for diagnosis. Case remains child_exit101 / functional_proof=false and contributes no successful latency sample. No attribution to server versus client beyond these observations.
At that stop:17attempts,15strict successes,2production LAN failures; all8branch attempts passed. Preserved complete logs/records. Resuming only unexecuted cases with the same staged comparator SHA and manifest, unchanged59-byte oracle and fresh targets; no failure is overwritten or converted to pass. Older
4872432dbreceipts remain separate.production-036bc2bb-0-05.log
comparison-segment-1-attempts.json
Found the cause of the third final-image control failure (production LAN08): FIXTURE schema_users_present=false session_present=false; socket closed1005, subsequent reconnects1006, no prompt/typed input, integrity timeout0. This is fixture state loss, not evidence of product latency. The live measurement fixture calls enrolled(), which uses max_connections(1) and sqlite::memory:. Replacing a cancelled/retired connection loses that database. Production uses a persistent file pool. The earlier LAN03 ACK30 and LAN05 ACK0 failures both report schema/session present, so do not conflate them with this loss.
Stopped repeat at30attempts: all15branch attempts PASS; three failed production LAN controls retained. Correct the live fixture for BOTH versions using existing enrolled_with_pool/fixture_with_store and a private disk-backed eight-connection SQLite pool; no production code or timeout/oracle change. First add a real connection-retirement regression that fails with the old memory fixture. Then scoped gates, atomic commit/push, final fast and a new matched exact-image acceptance. All prior receipts remain diagnostic, never silently relabelled. Fixture source: crates/infraternal-server/src/bootstrap_breakit2_tests.rs:451 and remote_authority.rs:202. Failed raw case: evidence/comparison-5bce43b6/evidence/production-036bc2bb-0-08.log in job cache; attach with fixture controls.
Persistent fixture fix committed locally at 2a0e26b945aa778fc2d83a584e6a7fe7ebdec0dc (two files,44insertions/1deletion). Real connection-retirement regression RED on both version fixtures before fix, GREEN on both after; fmt and strict remote-enabled all-target server lint PASS. Reuses production open_database through cfg(test) wrapper; no shipped behavior/contract/dependency/unwind change.
PUSH BLOCKED by Forgejo object storage, not authentication:
Forgejo operator/owner must free its storage; job cannot access/change that host under the brief. Remote branch remains5bce43b62; do not claim this commit pushed. Continuing local final gates and corrected-fixture artifact/VM acceptance; will retry push after verification. Prior successful timing sets remain diagnostic and separate.
Local commit 2a0e26b945aa778fc2d83a584e6a7fe7ebdec0dc: actual connection-retirement regression1failed/0ignored on both old fixtures,1passed/0ignored on both corrected fixtures; strict remote-enabled all-target server Clippy and fmt PASS. Patch and complete red/green/lint receipts attached. Whole gates are now running on this head; both corrected release controllers follow in the same build slot.
Push retry remains rejected: "remote unpack failed: unable to create temporary object directory". No authentication error; no credential investigation or access to the Git host. Operator storage repair still needed. Patch attachment makes the concrete commit reviewable while the remote branch remains5bce43b62.
persistent-fixture-2a0e26b9.patch
persistent-fixture-production-red.log
persistent-fixture-branch-red.log
persistent-fixture-production-green.log
persistent-fixture-branch-green.log
persistent-fixture-clippy.log
Operational checkpoint, local head2a0e26b945aa778fc2d83a584e6a7fe7ebdec0dc; remote still5bce43b62 because Git storage rejects pushes. Exactly one owned build-slot wrapper PID735022 runs cache/post-persistent-fixture.py; do not launch a second heavy wrapper. It runs resume-gates.sh then prepare-persistent-acceptance.py in the same slot. Logs under /home/kayg/.cache/codex-scratch/addhost-instant/evidence/: resume-iterate.log, resume-fast.log, post-persistent-gate-driver.log and post-persistent-fixture.log. Currently waiting for the VM RTT self-test lease; last observed owner process-per-user ppu-soak-r8-tail-aged, started02:31:03UTC.
On gate success (fast retains the unchanged cap failure), driver builds both release controllers with the same production runtime disk-pool fixture and snapshots actual head images into the artifact ledger with controller harness=runtime-disk8. Cached stage-comparison.py now requires that harness for BOTH variants. Then run its five short separately locked segments, collect-comparison.py, stage-fault-matrix.py and collect again, attach all new public receipts, update author checklist, quote final fast verbatim and retry push. This new acceptance must replace the memory-fixture results for final reporting; preserve previous4872432db and5bce43b62 datasets as diagnostics. The5bce repeat stopped at30attempts, all15branch attempts passed,3production LAN failures (ACK30, ACK0, then fixture schema loss). No failure was converted to pass. Fixture patch and both deterministic red/green controls are already attached above.
Raw diagnostic for the non-prompt production LAN08 failure and the stopped30-attempt set: schema_users_present=false/session_present=false. This is retained evidence for the persistent-fixture correction, not final performance acceptance. The two earlier ACK stalls with schema/session present remain separate observations.
production-036bc2bb-0-08.log
comparison-summary.json
Audited all463tracked Rust files in production036bc2bb0 against the corrected controller archive. Differences are confined to bootstrap_breakit2_tests.rs, bootstrap_browser_tests.rs and the appended cfg(test) runtime pool wrapper. The complete normal runtime body is byte-identical. This preserves the production implementation while using its actual persistent pool for measurement. Exact replay patch SHA d93e9ef544277c0aac54a059351518ee1b3c290b30ce20555773536c98d06c6c and audit attached. New-head whole gates still waiting for VM lease; no acceptance numbers claimed yet.
production-runtime-disk8-test-only.patch
production-runtime-disk8-source-audit.json
Head2a0e26b9 iterate stopped at the remote timing runner: no remote case receipt/output arrived. It correctly returned failure despite96remaining local cases passing (14existing ignores,10remote cases filtered from the local run). No remote case is counted as run/passed; no waiver. Preserved complete failed log as evidence/head-2a0e26b9-missing-vm-receipt-resume-iterate.log. Last observed VM owner was reconnect-release service, started03:17:13UTC. Cause is not established; do not label this a code/test timing failure or an auth failure without evidence.
Re-running unchanged oracle through a cache-only transport observer that emits exit code, output/error byte counts and fixed error categories; never raw SSH diagnostics. It calls the checked-in runner unchanged and still requires every real remote receipt and every remaining local case. Auth/host-identity refusal stops without credential investigation. One new owned build-slot wrapper PID1147737 runs post-persistent-fixture.py; previous735022exited. Same pending sequence: full gates then both persistent-fixture release controllers/actual image snapshot, matched fresh comparison and scoped fault matrix. No new source commit; local2a still blocked from push by external Git storage.
Before final corrected-fixture acceptance, origin/dev advanced to9e84426c4 (merged size-ratchet-r2 and agent-reconnect). Preserve local2a0e26b9 in job-cache pre-9e84426c-2a0e26b9.bundle. Rebase this lane onto the current approved dev cuts/reconnect fixes; keep all reviewed setup/ownership/cache fixes and stable event contract. This adopts owner-approved upstream work, not a foundational engine choice or lane cap edit.
Latest2a iterate: transport receipts10/10PASS, remaining local native96PASS, server482PASS/1FAIL/47existing ignores. Failure auth::tests::server_setup_terminal_echo_and_resume_preserve_offsets_once closed WebSocket(None). The failing auth test and runtime source have no upstream diff; create detached dev control under cache/baseline9e84426c4 and replay at same load window before classifying. No waiver. Raw2a failed iterate preserved as head-2a0e26b9-auth-close-resume-iterate.log. Previous VM missing receipt remains a separate failed run. Full fast/controllers were not reached on2a. Forgejo push still storage-blocked; code patch already attached. New final head/gates/size comparison will be reported after rebase and replay.
Rebased lane onto origin/dev9e84426c48f3818cebaa99acede14fec1d24ab4e; new local head
aef934bce. Preserved early shim, retained setup status, reviewed cache/cleanup fixes and persistent measurement fixture. Reconnect service manager behavior was applied to the shared setup script (Restart=always, RestartSec=2, StartLimitIntervalSec=0); all new dev size caps and decisions are inherited unchanged. Next: actual auth-resume control against detached dev in the same load window, whole gates, rebuild same-recipe images/controllers, corrected-fixture acceptance. Prior timings remain diagnostic.Retained the two distinct failed2a iterate runs before rebase: (1) missing VM measurement receipt, correctly failed; (2)10remote cases passed, server WebSocket(None) failure. Neither is waived. New head
aef934bceauth-control + whole gates queued via one owned build-slot wrapper PID1521061.head-2a0e26b9-auth-close-resume-iterate.log
head-2a0e26b9-missing-vm-receipt-resume-iterate.log
Actual auth-resume control complete at aef934bce:3/3branch and3/3detached dev9e84426c4 passed, each exactly1test/0failed/0ignored. Alternating run order, machine infraternal-dev, load7.42/7.37/7.01 recorded in every raw log. Earlier WebSocket(None) close did not reproduce in this control; it is NOT waived or declared fixed. Full iterate/fast now running on rebased head through the same owned build-slot wrapper1521061. All prior failed receipts retained. Rebased lane patch attached while remote Git storage remains blocked; no decision/cap edits by the lane.
rebase-auth-controls-summary.json
rebased-aef934bce-lane.patch
rebase-auth-controls-raw.tar.gz
Rebased head
aef934bce4: atomic history preserved in attached bundle; git bundle verify PASS. Its prerequisites are contained in dev9e84426c4 history. Patch already attached. Actual push retry follows validation; the last real push was rejected by Forgejo object storage. Current iterate: VM latency-proxy required receipt PASS; fmt, native fmt, compile-out and strict workspace all-target Clippy PASS; native test binaries compiling. No final gate/performance acceptance claim.rebased-aef934bce-history.tar.gz
rebased-history-provenance.json
Operational checkpoint: clean local head
aef934bce4, base9e84426c4. One owned build-slot wrapper PID1521061 runs rebase-auth-controls.py -> rebased-milestone.py. Auth exact control3/3per version PASS; first VM proxy receipt PASS; fmt/nativefmt/compile-out/Clippy PASS; CLI, real daemon restart/backpressure, connector refusal, vt_credit and reconnect snapshot tests PASS so far. Agent physical suite staged with SHA d02bcd0d247c5c98d78e6a5480ec17d17461c1ec500c640f54f267a718bd45bd; now waits for fair VM lock (last observed owner process-per-user ppu-soak-r8c-dev started04:23:10UTC). No final iterate/fast result yet. Logs: evidence/resume-iterate.log, rebased-gate-driver.log, rebase-auth-driver.log in job cache.On gate completion this same driver builds fresh private images, both runtime-disk8 release controllers, both own-client browser bundles, dev9e static baselines, real-binary controls/R189-01 and functional/capture browser checks. prepare-persistent-acceptance.py uses current base9e and preserves ledger. Then run rebased-size-report.py explicitly (the later on-disk driver edit may not be loaded by the active interpreter), stage-comparison.py/collect-comparison.py for fresh 10per version/profile, then stage-fault-matrix.py and collect again. These scripts require corrected persistent fixture controllers and actual current image hashes; previous487/5bce memory-fixture metrics remain diagnostic only. Full current receipts, binary sizes, final fast and real push retry still pending; do not report done. No deployment. Bundle/patch and control receipts linked above.
Head
aef934bce4, base9e84426c4: complete scripts/verify iterate PASS, no waivers. All six Rust rows PASS; server483passed/0failed/48existing ignores;10physical transport cases passed with VM receipts plus106remaining local cases; exact VM graph case plus206local canvas/UI and37protocol cases PASS. Earlier auth-resume failure did not recur; R189-01 and retained setup/retry regressions PASS.Verbatim final line:
Delayed setup API checks2passed/0failed/0ignored; persistent authority-retirement1passed/0failed/0ignored. Upload retries now succeed after the HTTP500 interval; Git push storage status remains unverified. Complete raw iterate/API receipts below.
One owned build-slot wrapper1521061 continues final fast. Its relay/fmt/compile-out/Clippy and early daemon/credit/reconnect tests passed;10transport cases now wait for the VM lock (last owner process-per-user ppu-soak-r8c-dev started05:51:31UTC). Static sizes and fresh corrected-fixture production comparison are NOT complete.
Next acceptance uses unchanged comparison segments0–4 within one hard-bounded1700-second fair lease, then unchanged fault oracles within a separate1000-second lease; staging/builds outside locks, owner cleanup/kill grace unchanged, release between suites. Dry controls verified exact case order and refusal propagation; no new measurements. Fresh browser receipts must also show typed echo before joined and new input after handover. Previous memory-fixture timing sets remain diagnostics. No cap edits or deployment.
head-aef934bce-iterate.log
rebased-setup-stream-smoke.log
rebased-persistent-authority-smoke.log
rebased-setup-smoke-summary.json
Git publishing recovered: rebased atomic history pushed to addhost-instant@aef934bce4b4f7d07e858e8ae393bb9ae3b8f2be; subsequent ls-remote matches. Explicit lease expected5bce43b6277366a9cf152c267230dc30b9d98e4e, so no unseen branch update was overwritten. Complete iteratePASS receipts and delayed API checks are linked in previous comment. Final fast still waits fresh10-case physical suite; no static-size or corrected-fixture acceptance claim yet. No cap edits, merge or deployment.
Push output verbatim:
For the new target-killed matrix row, SIGKILL the actual remote setup shell during the full-image transfer, while the independent session shim remains alive. The earlier diagnostic row killed only the decoder. New injector selects the decoder parent, verifies fixture uid1101 and shell identity, and pins it with pidfd before signalling; the oracle requires an explicit setup_shell_stopped receipt as well as59acknowledged bytes, before/after painted rows, full-image hash, failed setup -> explicit retry -> ready. This covers loss of setup cleanup rather than a decoder exit that can still run cleanup. The separate35-second link cut and controlled slow writes remain. This is a disposable-target setup-process kill, not host power-off. No product code/contract changes; no new injection run yet (images and final fast pending). Job-cache target-fault.py/fault-batch.py recorded for final replay archive.
Injector detail from install.rs:155: the decoder pipeline runs inside streamed=$(...), so killing only its immediate shell parent would leave the enclosing setup shell able to run cleanup. The new target-killed row pins and verifies the same-uid shell ancestry up to the non-shell SSH boundary, then SIGKILLs those exact setup PIDs. It leaves the independent session process outside that ancestry. No process argv/environ or credentials read; only comm/status in the job-owned disposable target. Receipt setup_shell_stopped records actual selected PIDs; existing byte/hash/retry proofs remain mandatory. This is preparation only; no new injection measured yet.
For the upcoming dev9e static size control, use the exact same approved compiler/archive/linker implementations at the head worktree canonical paths. Verified byte-identical SHA256 for rust-lld.sh, musl-cc.sh, musl-cc-avx512.sh, musl-ar.sh and the AVX512 patcher on both sources. Only the temporary baseline .cargo/config.toml tool paths differ; Rust/native source, compiler flags, features and profiles unchanged. This preserves the same recipe and avoids a second tool-path cache namespace. Earlier GNU auth controls used the pristine checkout before this musl-only path overlay. Raw overlay and hash audit retained for final evidence. No branch source/cap change or new measurement.
Operational checkpoint: pushed/clean
aef934bce4, base9e84426c4. Complete iteratePASS already attached. Current final fast: relayVM receiptPASS, fmt/nativefmt/compile-out/Clippy/test/cargo-deny PASS; fresh10physical transport receiptsPASS, local106PASS, server483PASS/0FAIL/48existing ignores (including auth-resume, R189-01 and retained setup/retry). Static release phase is rebuilding x86 crypto with active object-file updates; no current static size/crypto/key-heap result yet. No waiver or cap edit.One owned build-slot wrapper1521061 (session35376), cache rebase-auth-controls.py -> rebased-milestone.py, logs evidence/resume-fast.log, rebased-gate-driver.log, rebase-auth-driver.log. Do not start a duplicate wrapper. After gates it runs functional/capture browser checks, private images, prepare-persistent-acceptance.py (guards exact gated head, both current ELF/size receipts, NIST/key-memory/stub receipts; round3 functional oracle; both runtime-disk8 release controllers), own-client bundles, R189-01/real-binary controls and dev9e static baselines. Run rebased-size-report.py explicitly after baseline builds. Baseline has only documented tool-path overlay (.cargo/config.toml), verified identical tool implementations; flags/features/profiles/Rust/native sources unchanged.
Then fresh production036 versus actual new images through stage-comparison.py (one hard-bounded1700-second comparison lease), collect-comparison.py, stage-fault-matrix.py (separate1000-second matrix lease), collect again, verify-working-prompt-order.py, plus stage-pill-timing.py for physical UI tests. Kill row now SIGKILLs verified/pinned setup-shell ancestry, not merely decoder; link cut/slow disk unchanged. All per-case59-byte/hash/transport/retry proofs remain; raw records retained. No new performance acceptance yet. Old memory-fixture datasets are diagnostic only. Update final replay archive/checklist/issue with exact raw tables/gate; no deployment; stop for break-it review after acceptance.
First static result at
aef934bce: x86 stripped static image21,459,440B, current dev cap21,422,096B, delta+37,344B. Actual same-recipe dev control still pending; this is a cap delta, not yet the rebuilt-baseline delta. Production trust/static linkage/AES-128-GCM NIST vector/key memory initialization=1 allPASS. ARM still building; gate continues. Immutable pre-cut image + SHA and section table saved under job cache pre-poll-aef934bce-full-x86_64 / evidence/pre-poll-aef934bce-x86*.Code inspection identifies a candidate: rebase retained upstream remote_service panic/bounded-shutdown guard, but its generic Runtime::block_on duplicates the reactor poll already provided by infraternal_agent::block_on_io. Measure both current architectures and dev first, then test reusing that existing erased poll boundary while retaining all unwind info and watchdog behavior. Candidate not applied or claimed measured yet; no source/cap change during active gate.
x86 cap delta+37,344B warrants the existing-poll-loop cut before fresh acceptance. Current fast continues unchanged through both architectures and all required checks. Then its same owned wrapper keeps browser functional/capture receipts and private images, snapshots both actual gated full images, builds dev9e static controls with verified identical canonical tools, and runs rebased-size-report.py. It will explicitly stop with PRE_CUT_SIZE_PROBE_COMPLETE / exit75 before release-controller and acceptance preparation on aef. This is a workflow stop for the measured size cut, not a hidden test waiver or product failure. All actual gate results remain preserved. Next: apply/test the existing poll-boundary reuse, scoped atomic commit/push, final fast on the new head, actual new-image acceptance; unchanged baseline images may be reused only after hash/size identity validation. No source change during the current gate.
Intermediate head
aef934bce4: scripts/verify fast exit1, only static size row fails (both caps plus boundary self-test refusing the same oversized x86 input). All other Rust rows PASS. Actual stripped static bytes: x86=21,459,440, dev cap21,422,096, delta+37,344; ARM=19,648,824, dev cap19,594,856, delta+53,968. These are cap deltas; rebuilt dev controls remain pending. Production trust/static linkage/NIST AES-GCM vector/x86 key memory initialization=1 PASS. ARM kernel key-memory runtime explicitly not checked on this x86 host. No cap change or waiver.Verbatim Rust summary:
Complete raw output, immutable pre-cut image hashes, ARM section table and identical-tool baseline overlay below. Separate TypeScript gate is running (fast stops before it on a Rust failure). Same wrapper then finishes browser/private-image checks and measures dev baseline sizes; controller/benchmark prep explicitly stops for the existing-poll-boundary cut. No new acceptance metrics.
head-aef934bce-fast.log
pre-poll-aef934bce-static.json
pre-poll-aef934bce-aarch64-sections.txt
baseline-canonical-tools.json
baseline-canonical-tools.patch
Head
aef934bce: separate TypeScript gate PASS after the fast size-row failure. Exact VM graph receipt1PASS plus206local canvas/UI cases and37protocol casesPASS; compiler/vendor checks/build/transfer budgetPASS. Final Rust fast remains exit1 only for measured size caps/boundary check; no waiver. Browser functional checks are now running in the same owned wrapper1521061 on TRACER_TEST_PORT5291, followed by captures/private images and actual dev size control before the explicit pre-cut stop. Complete TS raw output retained as evidence/head-aef934bce-ts-gate.log. No new timing acceptance or source change.head-aef934bce-ts-gate.log
Pushed head
aef934bceproduction-build browser functional run:54passed/0failed/8capture-only skips, uniqueTRACER_TEST_PORT5291. Includes independent R189-02 published-window subscription, later failure/retry with the original session, success/reduced-motion flow, signed-payload refusal, Start-fresh recovery, phone layout, consent and host-key confirmation. Start-fresh desktop/phone captures also produced. Physical pill timing specs were excluded from this build-host functional run; their separate VM replay remains pending. Six current-build pill capture runs now active; no self-judgment of visual quality. Private images and actual dev size control follow before the explicit pre-cut stop.rebased-addhost-functional.log
aef934bce: all 6 production-build pill capture cases passed (desktop 1440×900, phones 393×852 and 375×667, both themes). Raw log and selected status/failure/success frames attached for owner/reviewer visual assessment. These are functional captures, not performance evidence.head-aef934bce-pill-captures.log





Same-recipe controls complete:
aef934bcevs merge base9e84426c(Rust/native sources pristine; documented canonical-tool-path-only config overlay).Both dev controls exactly match the existing caps; caps remain untouched. Immutable full/private head images and hashes preserved. The previous wrapper intentionally stopped after these controls (prep exit75) before preparing obsolete acceptance inputs; that stop is not a gate waiver. Applying the existing shared block_on_io boundary inside remote_service panic fencing now; one startup allocation moves from caller to owner, bounded shutdown and unwinding unchanged. Scoped fmt/clippy, all three panic/shutdown regressions and both static builds running through build-slot; no size-saving claim until measured.
pre-poll-aef934bce-vs-dev.json
pre-cut-baseline-x86_64.log
pre-cut-baseline-aarch64.log
Committed and pushed
265f06980(shared enrolled-link reactor, existing panic recovery and bounded shutdown retained). Scoped fmt PASS; strict CLI all-target clippy PASS; exact remote_service regressions 3 passed/0 failed.Unwind sections retained; no release-setting or cap change. The saving is smaller than the hypothesis, so no larger saving claimed. Testing a separate single-module reuse of the same boundary in component command startup next; complete CLI unit tests plus fmt/clippy and both static builds through build-slot. Full final-head fast and fresh VM acceptance remain pending.
poll-cut-candidate.json
poll-cut-vs-dev.json
poll-cut-clippy.log
poll-cut-tests.log
poll-cut-build-x86_64.log
poll-cut-build-aarch64.log
Committed and pushed
38b2201611: reuse the shared reactor for component command startup. Scoped fmt PASS, strict CLI all-target clippy PASS, complete CLI unit suite 85 passed/0 failed/1 pre-existing ignored optimized signature-verification benchmark. One allocation per explicit component command; terminal frame paths unchanged. This second measured cut saves 6,080 B x86 /4,352 B ARM. Both cuts together save12,224/9,328 B versus aef.Measured D243 exception case for review (NOT approved; no cap/profile/unwind-setting edits):
38b220161bytesARM remains28,256 B above the16 KiB allowance; x86 remains8,736 B above it. The added unwind/exception sections exceed the entire net increase; other image bytes are smaller than dev. Retain production unwinding so the additional setup/replay/ownership futures and the #225 service panic boundary can clean up and recover. Removing unwind tables, panic=abort or weakening those paths is rejected. This is a measured case for the allowed exception route, not a claim that all further optimizations are impossible. Exact section deltas, image/source hashes and build logs attached.
Continuing the authorized remaining acceptance with caps untouched: unchanged full scripts/verify fast through build-slot on38b220161, current-build browser checks and exact R189-01/real-binary/round3 controls, runtime-disk8 branch/prod036 release controllers, fresh ≥10/version/profile comparison and stronger three-fault recovery matrix. Separate physical pill timing replay has been staged against identical UI source and is waiting for its fair VM lease. No new latency numbers or final gate result yet; no deployment.
component-poll-cut-candidate.json
post-cut-exception-case.json
component-poll-cut-clippy.log
component-poll-cut-tests.log
component-poll-cut-build-x86_64.log
component-poll-cut-build-aarch64.log
Resumable acceptance checkpoint: pushed/clean38b220161, source frozen. ONE owned build-slot wrapper3577192/session76327 is running the unchanged full fast (post-cut-milestone-driver.log -> rebased-gate-driver.log/resume-fast.log), then head-pinned browser/private-image/round3/real-image/R189-01/controller preparation. A separate VM-only pill replay3579007/session38195 is staged and queued. Current fair VM owner receipt attached; waiting normally, no lock bypass. No final gate or new acceptance timing result yet.
Attached replay bundle contains exact helper sources + hashes, test-only prod036 runtime-disk8 fixture patch and audit, dev canonical-tool overlay audit, artifact ledger, exception case, and complete remaining sequence/strict oracles. Pill prebuilt assets have identical UI source to38 but build metadata fromaef; that origin is explicitly recorded. Current-head assets will also be rebuilt by the milestone. Old memory-fixture benchmark datasets remain diagnostic. Do not start a duplicate wrapper or edit the source mid-gate.
head-38b220161-acceptance-replay.tar.gz
head-38b220161-acceptance-runbook.md
post-cut-perf-owner-first.txt
A lightweight acceptance coordinator is now waiting: PID3643589/session43591, cache acceptance-after-inputs.py, log evidence/acceptance-watch-driver.log. It holds no build slot or VM lease while waiting. It starts only after the owned milestone3577192 prints the exact38b220161 immutable-inputs-ready marker, then runs comparison -> collection/nonregression checks -> three faults -> collection -> before/after interactive-order proof. It stops if the input wrapper dies without readiness, HEAD/clean state changes, a branch comparison regresses, or any fault/byte/order oracle fails. It reports target misses explicitly. No new measurements yet. Do not launch duplicate acceptance while this coordinator is alive. The existing build-slot wrapper and separate VM pill stage3579007 remain active.
head-38b220161-acceptance-scheduler.py.txt
Physical UI timing replay complete on the designated4vCPU/7GB perf VM (infraternal-perf-test), fair-locked: unchanged pill spec4passed/0failed; PILL_TIMING_EXIT=0 and PILL_LOCK_EXIT=0. VM load0.87/0.78/0.91. Status lifecycle0long tasks >=50ms; reduced-motion status0long tasks (its baseline page1); worst observed frame gap50.0ms. Page script share2.6% (159ms over6056ms), baseline page3.4%. These are observed measurements, not a claim about every frame or a statistically established CPU improvement. Failure/retry and reduced-motion cases both pass.
Provenance: spec/current UI source matches38b220161 exactly; assets are the real production build made ataef, before the two backend-only cuts. Their build-date/commit metadata remainsaef, explicitly recorded along with311asset file hashes and an empty UI source diff. No claim they were built at38. Complete console/perf JSON archive attached; six production screenshot capture cases were already attached separately for owner visual review.
Full38 fast has now passed its separate physical RTT relay gate and is running Rust tests; no final fast result yet. The input-gated acceptance coordinator remains idle until all actual images/controllers and controls are verified. No fresh host-prompt/comparison/fault measurements yet.
head-38b220161-pill-timing-vm.log
head-38b22016-pill-vm-evidence.tar.gz
pill-timing-provenance.json
post-cut-ui-source-identity.json
Head38b2201611d7404f22ca8ed3389f3afc479c86e5: full unchanged scripts/verify fast through build-slot completed, exit1. Only static size fails, including its boundary self-test refusing the same oversized x86 image. All other Rust rowsPASS; no test waiver. Server483passed/0failed/48existing ignores;10physical native cases each explicit exit0, plus106local/14existing ignores; CLI85passed/0failed/1existing ignored signature benchmark; auth echo/resume and R189-01 bothPASS in the full gate.
Actual stripped static bytes match the measured candidates: x86=21,447,216 vs dev/cap21,422,096 (+25,120); ARM=19,639,496 vs19,594,856 (+44,640). Production trust/static PIE/noPT_INTERP bothPASS; AES-GCM NIST vectorPASS; x86 key memory initialization1/locked/dump-excludedPASS. ARM native kernel key-memory check is explicitly not performed on an x86 host. Caps/profile/unwind remain unchanged; exception case remains unapproved.
Verbatim Rust gate summary:
Because fast exits at the Rust size failure, the complete separate TypeScript gate was also run: exit0; exact graph1physicalVM case plus206local canvas/UI cases,37protocol cases, compiler/vendor/build/transfer checksPASS. Full raw outputs attached. Same owned wrapper is now running production-build browser functional/capture cases, followed by guarded image/controller/real-engine preparation. Input-gated fresh comparison/fault coordinator remains idle until readiness. No new host-prompt measurements yet.
head-38b220161-fast.log
head-38b220161-ts-gate.log
head-38b220161-gate-driver.log
Head38b220161 current production-build browser results:54functional passed/0failed/8capture-only skips, plus all6explicit pill capture cases passed. Independent R189-02, later failure/retry with original session, Start-fresh, consent, signed-image refusal, responsive layouts and reduced motion pass. Selected new-build desktop/phone frames attached for owner visual assessment; no author visual-quality verdict. Unique local test ports5291/5292.
Both actual full/private static architectures now snapshotted under immutable head filenames with hashes, only after final-fast trust/ELF/size/NIST/key-memory receipts were verified. Full x86 compressed9,845,849 B; ARM9,658,846 B. Private x86 gzip842,448 B and ARM771,898 B, same private hashes as aef. The two CLI cuts reduce full payload bytes; they do not claim a first-prompt-image reduction. The real shipping-binary round3 functional oracle passed with performance-counter calls prohibited. Release controller compilation has begun (branch first, then prod036 runtime-disk8). Comparison coordinator still waits for all preparation/real-binary controls to finish. No new host startup latency or fault-matrix results yet.
head-38b220161-browser-functional.log



head-38b220161-pill-captures.log
head-38b220161-round3-functional.log
head-38b220161-static-images.json
Immutable acceptance inputs READY for head38b220161; host build-slot wrapper3577192/session76327 exited0 after preparation (actual full fast remains exit1 solely for the documented size row; TS0, not relabeled). Exact R189-01 replay1passed/0failed; all7real shipping/private binary SSH/session controls passed. Round3 functional counter-guarded oracle already passed.
Both release server controllers use runtime-disk8 and are immutable:
production: 29763872 B, SHA256 21d680510cc8a734ba474d6104d0f70a68aa2bf443d4b13d3aad906675861715
branch: 30230944 B, SHA256 bca0cc6b5d3f1923f7166a63d7d869371c964849c9878ec933e702daa9b67226
Distinct browser entries were built against each version's actual shipping clients/renderer; first-party source hashes and adapter hashes attached. Production source036bc2bb0 retains its normal code; fixture-only observer/runtime-pool audit is attached earlier. Both dev size baselines reused only after hash/size validation.
Coordinator3643589/session43591 is now running the staged comparison (child36612), all artifact/source identities verified; awaiting its fair-locked VM receipts. Do not duplicate the measurement. It next enforces >=10successes/profile/version, no branch p50/p99 regressions, three strong faults, and before/after input order. No new host startup numbers yet.
head-38b220161-r189-01.log
head-38b220161-real-binary-controls.log
persistent-acceptance-provenance.json
comparison-manifest.json
comparison-client-provenance.json
head-38b220161-vs-dev-size.json
Fresh comparison stopped at production LAN sample7:25attempts,24functional passes (all12branch attempts passed),1production failure. No final >=10/profile table yet. Failed log production-036bc2bb-0-07 retained verbatim: first prompt and complete BEFORE_HANDOVER row painted; input acknowledgement stays0, so AFTER is never enqueued; browser integrity timeout, child101. Both persistent schema and owner session existence flags remaintrue, user serviceactive, one session scope. This is distinct from the eliminated memory-fixture schema loss; underlying acknowledgement-stall cause is NOT established.
The immutable baseline failure stays failed and is excluded from successful-sample quantiles; it is not waived or reported byte-complete. Resuming only to gather the required fresh successful samples, keeping every attempt. The helper now recognizes this precise healthy-schema production-LAN ACK0 symptom alongside its existing completed-rows ACK30 category, and validates previously recorded failures too: unknown errors, schema loss and ANY branch failure stop. Nine pure classifier controls pass, including rejection of an earlier ACK0 when the actual timeout value differs. Functional success still requires child0,59ACK, both complete rows, full-image hash and actual transport; no success assertion or application/binary changed. Raw log hashes of all first25attempts saved for append-only verification after resume. Source remains clean38b220161 and the completed final-head gate remains as recorded.
production-healthy-schema-ack0-finding.json
production-036bc2bb-0-07.log
comparison-summary.json
comparison-ack-classifier.patch
comparison-ack-classifier-controls.json
comparison-first25-raw-log-hashes.json
Comparison resume is active under coordinator PID199960/session51332; prior coordinator3643589 exited on the recorded strict failure. No host build slot is held. Same immutable38b220161 binaries/controllers/browser entries; source remains clean and pushed. The first25raw logs and original failed-driver/console outputs were preserved before resume.
Updated driver bundle attached: precise failed-baseline classification plus unknown-retained-failure refusal,9classifier controls, immutable-input-ready resume after its producer exits, and dependency reuse after matching Playwright package metadata (no nested re-copy). Per-case59-byte/hash/transport success oracle is unchanged. The production ACK0 record remains false; fresh samples fill the success count. The comparison is staged and awaiting VM receipts; no final statistics or fault-matrix result yet.
head-38b220161-resume-drivers.tar.gz
head-38b220161-resume-driver-manifest.json
Fresh own-client production comparison complete at38b220161 vs036bc2bb0 on infraternal-perf-test (4vCPU/7GB), fair-locked, normal fresh targets,1minute VM load range0.06–2.29.41attempts/40functional successes: branch20/20passed59-byte/complete-row/full-image-hash/actual-transport oracles; production20/21 with the single previously recorded healthy-schema ACK0failure retained and counted. No branch p50/p99 regression in either profile for any of the four startup metrics. At n10, nearest-rank p99 is the observed maximum, not a high-confidence population tail estimate.
After SSH sign-in (server key_sign_in end -> actual painted prompt / fully-joined event):
From browser Sign-in click:
The1.5s after-SSH prompt target remains missed by228.9ms at p50 (LAN300ms target by28.2ms). No claim these targets are met. Fully joined at150ms is4,476.9/4,577.9ms p50/p99 afterSSH. Private image remains842,448B gzip onx86.
All20branch browser logs independently prove complete BEFORE output before fully joined and AFTER input enqueued afterward; each has59ACK and complete output. The first25rawlog hashes are unchanged after the append-only resume. All41rawattempt logs/manifests attached, including failed production sample7; no failed sample is a success or in successful-sample quantiles. The strong three-fault matrix is now staged in its separate bounded VM lease; no fault result yet.
head-38b220161-comparison-summary.json
head-38b220161-comparison-raw41.tar.gz
comparison-20-branch-working-order.json
comparison-first25-raw-log-hashes.json
Strong failure matrix completed on the fair-locked perf VM at38b220161:3/3passed, FAULT_LOCK_EXIT=0. Target-stop SIGKILLed pidfd-pinned owning setup-shell ancestry[289,290,342] at2,064,384public image bytes, not merely the decoder; independent shim survived. Network cut restored after35s. One-shot slow public writes exercised the real deadline. Every case emitted failure, accepted retry, reached ready with the original session, acknowledged59inputbytes, renderedcompleteBEFORE/AFTER rows, matched the full shipping-image hash and used actual transport. Each finishedone session scope/two infraternal processes.
Independent event-order proof now23/23:all20branch comparison cases plus3faults echo BEFORE before join and send AFTER input after join. First25comparison rawlog hashes remain unchanged. All41comparison attempts remain recorded; production ACK0failure stays failed.
Fresh post-build provenance audit:467tracked Rust/build-setting files examined against full production commit036bc2bb067a231736f219f32897aaa7d8942b70,465identical; only the gated measurement test module and exact cfg(test) runtime-wrapper append differ, plus one new test-only browser observer. Normal production code and build settings unchanged. Requested review checks are complete; explicit prompt-target misses and the unapproved measured size exception remain. Rawfaults/order/audit/checklist attached. Finishing source/remote identity checks and owned scratch cleanup before final handoff; no deployment.
head-38b220161-strong-fault-matrix.json
head-38b220161-strong-fault-raw.tar.gz
working-prompt-order.json
head-38b220161-comparison-session-lifetime.json
head-38b220161-production-post-build-source-audit.json
post-cut-acceptance-target-report.json
head-38b220161-reviewer-checklist.json
Finished this review-follow-up at pushed head
38b2201611, base9e84426c48. Source and remote SHA match; checkout is clean. Stopping for the requested independent break-it review. No deployment, merge or issue close.The published-window retained setup/failure/retry regression is verified in the current production build. Exact R189-01 replay passed; round-three real-binary functional oracle and all seven private/full-image SSH/session/manager controls passed. New atomic commits
265f06980and38b220161 reuse the existing reactor boundary while retaining panic recovery, bounded shutdown and production unwind information. No new dependency, release-setting or cap change.Fresh comparison: 41 attempts, 40 functional successes, ten successful samples per version/profile. All20branch runs passed. Production20/21 includes its retained healthy-schema ACK0failure; that failed attempt is not counted as a success or in successful-sample quantiles. All first25raw logs remained byte-identical after resume. No branch p50/p99 regression in either profile for any of four startup metrics. Values below start at SSH sign-in completion; p99 at n10 is the observed maximum. Full click-origin metrics and all raw attempts were attached earlier.
The prompt targets remain missed: LAN p50 is28.2ms above300ms; at150ms RTT p50 is228.9ms above1500ms. No claim those targets are met. At150ms, browser-click-to-prompt is3074.2/3151.2ms and click-to-joined5823.3/5940.4ms. Private x86 image remains842,448B gzip.
Strong fault matrix 3/3 passed: owning setup-shell ancestry SIGKILL during transfer,35s target network cut/restoration, and slow public writes hitting the real deadline. Each failed, retried, and reached ready in the original session with59input bytes acknowledged, complete BEFORE/AFTER output, actual transport and matching full-image hash. All23branch/fault logs independently prove BEFORE echo before join and AFTER input afterward. Each ends with one session scope/two infraternal processes. No lost input/output in these tested cases.
Size (same-recipe rebuilt dev controls; stripped static bytes):
The two cuts save12,224/9,328B versus the pre-cut head. Unwind-section deltas are+60,684/+61,332B; net other file bytes are-35,564/-16,692B. The measured exception case is posted and unapproved. Caps and unwind settings remain unchanged.
Final
scripts/verify fastthrough build-slot: actual exit1, static size/boundary only. Every other Rust row passed, including483server tests/0failures,10physical native exits0 and106remaining native passes. Production trust/static PIE/noPT_INTERP/NIST/x86locked-and-dump-excluded key initialization passed. ARM native kernel key protection was explicitly not tested on an x86 host. Verbatim Rust summary:Complete separate TypeScript gate: exit0 (one physical geometry case,206local canvas/UI cases,37protocol cases, compiler/vendor/build/transfer checks). Current production browser suite54passed/0failed/8optional capture-only skips; six explicit screenshot captures passed. Four physical VM pill cases passed; their assets were built ataef with UI source identical to38, metadata origin explicitly recorded. Current-head screenshots were separately rebuilt and attached for owner assessment. No author visual-quality verdict.
Complete raw fast/TS outputs, comparison/fault logs, images, manifests, audits and status contract were attached in preceding comments. Final replay bundle/checklist below include precise helper sources and hashes, test-only production fixture patch, source audit, failed-attempt policy and all documented limits. Owned57,051,987,590-byte Cargo target and three temporary baseline sources removed; evidence and immutable images/controllers preserved. Shared short-state directory and owner's live fixtures untouched. No coordinators remain running.
head-38b220161-final-replay.tar.gz
head-38b220161-final-replay.md
head-38b220161-final-replay-manifest.json
head-38b220161-reviewer-checklist.json
head-38b220161-cleanup.json
Delta verification started: review-addhost-instant-r3 at
38b220161, baseb3fe6181e. Both requested earlier test commits are already included in target (aec341876/060f49d00); empty cherry-picks skipped. Replaying earlier findings, checking only delta, and independently verifying functional gates and perf/size claims. No implementation changes, push or deployment.Delta step: confirmed R189-03 P2, committed failing test
d8b88bb93(setup-feed.test.ts, three empty open/close cycles evade two-reconnect bound; expected lost, got []). Raw reconnect-test.log: 6 pass/1 fail. Earlier subscription and cleanup replays are queued through build-slot. Independent SHA-verified ELF inspection confirms x86 +25120 B / ARM +44640 B vs9e84426c4; complete ARM file accounting in independent-size.json. Fresh five-run/profile/version comparison staged in unique VM review directory and waiting on required flock; no new timing result yet. Full fast still waiting for a host build slot; no author gate relabelled as independent.Checkpoint at review head
a592a43ef: added R189-05 P3 failing JSON-only parser regression (TypeError from malformed state/failed_step); parser controls 9 pass/1 fail. R189-03 P2 reconnect regression committedd8b88bb93. Actual final-controller retained-status controls 2/2 PASS; separate tracer TypeScript compiler check PASS. Client source/bundle identities verified against both refs. ARM exact file accounting matches +44640 B: unwind/exception +61332, data/relocated tables +2960, text -7476, rodata -12288, alignment/header +112. Current ARM symbol-level attribution remains unavailable in stripped supplied artifacts. Browser, original cleanup replay and full fast remain queued through official build slots; fresh perf comparison is still waiting on the designated VM lock. No implementation edits/push/deployment.Independent replay is resource-blocked at this checkpoint, not a source failure. All four official build slots remain occupied; browser regression, R189-01 replay and full fast have not started. VM lock changed from process-per-user ppu-soak-r8e-aged to mesh-unreachable-review scale pair=2 at 12:12:36 UTC; fresh comparison and native latency leases remain queued with flock -w 3600. Need scheduler/lock owners to finish their runs; no owner fixtures, processes, slots or lock files changed. Existing confirmed evidence is committed at
a592a43efand attached. Continuing the required waits; no independent full-gate or timing pass claimed.Fresh independent perf VM208 comparison: 5 successful attempts per version/profile, 20/20 functional proofs PASS (same session, 59 acknowledged bytes, complete output rows, actual transport, installed full-image SHA). REVIEW_PERF_EXIT=0. Each version uses its own hash-verified shipping client sources/bundle and supplied final runtime-disk8 release controller. New run IDs and unique target/port; no reused attempt records. p99 below is only the observed maximum at n=5. Values start at SSH sign-in completion.
No measured startup regression. Deferred prompt targets remain missed. Per-attempt machine load and every attempt retained in fresh-perf-raw.tar.gz/fresh-perf-summary.json. Native echo/flood still queued on separate VM lease; browser/full-fast remain queued on build slots. Narrow single-crate --lib R189-01 replay compiling under the explicit small-check exemption, one build job. No independent gate pass yet.
Earlier R189-01 independently RESOLVED at exact target38b220161: cargo test --locked -p infraternal-plugin-ssh --lib review_stopped_commit_channel_finishes_the_private_probe -- --nocapture, normal nonignored run, exit0. Verbatim result: test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 39 filtered out. Narrow single-crate check used one build job under the explicit small-check exemption. R189-02 browser replay and full fast are still queued through official build slots; native latency waits on VM lock.
Earlier R189-02 independently RESOLVED on target38 production-build UI: original published-window subscription regression PASS; ordinary post-prompt failure/retry PASS; clean completion PASS; reduced-motion failure/retry PASS. Original R189-01 also freshly passed.
New R189-04 P2 confirmed and committed614b0cc4f: attempt8 with authoritative retry_available=false still offers one enabled Try again button (expected0/received1). New status parser drops retry_available; pill always builds a retry action. Existing server cap remains eight, so this advertised retry cannot succeed. Browser test only projects the documented exhausted status; real wizard/status connection and UI remain in use. Raw browser.log and error-context.md retained. New confirmed counts P0=0/P1=0/P2=2/P3=1; R189-03
d8b88bb93and R189-05a592a43efalready attached.Broad functional wizard suite still running in the same production build. Exact-target full fast remains queued; native echo/flood waits on designated VM lock. No full-gate pass or final verdict yet. No implementation fixes/push/deployment.
Production-build browser replay completed at review head614b0cc4f: 41 PASS/1 intentional new regression FAIL. Original R189-02 subscription, ordinary failure/retry, clean completion, reduced motion and all selected inherited wizard cases PASS. Sole failure R189-04 exhausted retry (expected0 enabled retry buttons, received1). Unique port5289; functional run, no timing claim. Final TypeScript compiler check PASS; git diff --check PASS. Raw browser.log, failed context and current publication PNG attached. Exact-target full fast has started through build-slot, passed initial artifact checks, and now waits on VM RTT self-check; native echo/flood is separately queued. No full-gate outcome yet.
Independent native echo/flood gate PASS on designated VM208, locked short lease, supplied exact target38 full static image, RTT0, 100 samples/phase. Echo p99 baseline1.25 ms, other-session flood2.16 ms (growth0.91), same-session flood9.21 ms (growth7.96); both growth values below16.67 ms. REVIEW_LATENCY_EXIT=0. This is native real-PTY coverage, not full browser input-to-photon acceptance or a production echo comparison. Initial invocation refused before testing because my fixture state path exceeded AF_UNIX bounds; that diagnostic is retained separately, and the unchanged oracle was rerun with a shorter owned path. No failed measurement relabelled.
Exact-target full fast passed physical RTT self-check/fmt/VTfmt/CLI compile-out and is running strict workspace Clippy. No final gate outcome yet.
Correction to previous checkpoint: CLI compile-out had NOT passed in the first full-fast invocation. My CARGO_TARGET_X86_64_UNKNOWN_LINUX_GNU_RUNNER used a relative Python path; Cargo executes tests from the crate directory, so the runner refused with ENOENT before the CLI test binary ran. Wrapper now resolves that one known runner to its absolute exact-target path. Independent direct replay of the generated no-default-feature CLI test binary:29 PASS/0 FAIL, exit0. This is a reviewer harness failure, not an implementation finding or waived gate row. First full run continues to preserve remaining results; a complete unchanged fast replay is required afterward. Full outcome not yet known. Raw first invocation and cli-compileout-replay.log retained. Native echo/flood actual PASS remains valid and separate.
Known-invalid first fast stopped during Clippy before native/static work; preserved as verify-fast-invalid-relative-runner.log. Corrected full replay queued through official build-slot on clean detached38b220161 source; runner path is absolute, host and VM fixture roots are short and owned, CARGO_BUILD_JOBS=2/RUST_TEST_THREADS=2. No source/oracle/deadline/cap changes. Full outcome remains pending. Independent direct CLI control29PASS, original R189-01PASS, original R189-02PASS, browser41PASS/solenewR189-04FAIL, native echo/floodPASS and20fresh startup proofsPASS remain valid. Additional analysis of all10fresh branch receipts proves BEFORE echo before join and AFTER input after join (fresh-order-proofs.json). Review branch614b0cc4f clean; no implementation edits/push/deployment.
Additional focused recovery finding R189-06 P2: after retained retry_available=true failure, bounded status reconnects are exhausted. The pill intentionally keeps the failure/retry explanation, but feed end deletes the host entry and explicit retry becomes a no-op (no new socket or request). New unit regression review explicit retry resumes after the setup feed stops FAILS: expected4 sockets, got3, after confirming lost callback. Automatic retries must stay bounded; deliberate user recovery is separate. Committed as one atomic test-only change. Current confirmed counts P0=0/P1=0/P2=3/P3=1. Earlier findings remain independently resolved. Full corrected fast still queued; no implementation fixes/push/deployment.
MERGE under the stated severity rule (no confirmed P0/P1). Independent full-gate verification is INCOMPLETE; this is not a passing release-gate certification. Target38b220161; prior targetb3fe6181e; evidence branchreview-addhost-instant-r3 at7590258029a13f6671c8eca294ca0352f4113f02. Implementation untouched. No push/deployment/issue closure.
Both earlier findings independently pass: R189-01 exact-target nonignored native cleanup1/1; R189-02 original production-build subscription and ordinary failure/retry browser controls. Earlier requested browser test commits were already in target, so empty cherry-picks were skipped.
Current confirmed counts: P0=0/P1=0/P2=3/P3=1.
d8b88bb93614b0cc4fa592a43ef759025802Every finding has a committed failing test. Unit controls:15PASS/3intentionalFAIL across the two setup files; browser41PASS/1intentionalFAIL. TypeScript compiler checkPASS. Actual final controller retained-status API controls2PASS; partial-upload verified-cache/foreign-transaction controls1PASS; generated compile-out CLI control29PASS.
Fresh comparison on VM208,4vCPU, normal fresh targets,1minute load1.06–1.87:5successful samples per version/profile,20/20functional proofs. Supplied immutable release controllers and images, each version's hash-verified actual shipping clients/renderer through the measurement adapter. p99 below means only observed maximum at n5. Values start at SSH sign-in completion.
All20passed same-session/59-byte/complete-row/full-image-SHA/actual-transport proof; all10branch records independently show BEFORE echo before join and AFTER input after join. No startup regression observed; deferred300/1500ms prompt targets remain missed. This is the matched shipping-client measurement adapter, not the complete PWA/browser latency suite.
Locked native real-PTY gate, RTT0,100samples/phasePASS: echo p99baseline1.25ms,other-session flood2.16ms,growth0.91; same-session flood9.21ms,growth7.96; growth budget16.67ms. Initial invocation refused fixture setup because reviewer state path was too long; diagnostic retained separately, unchanged oracle replay passes with short owned path.
Stripped static supplied images independently SHA-verified and inspected: x86+25120B (21422096→21447216), ARM+44640B (19594856→19639496). See size-review.md for byte-exact ARM accounting: unwind/exception+61332B, net all other file bytes-16692B. Current ARM symbol-level attribution remains UNVERIFIED: supplied artifacts are stripped. Historical x86 symbol diagnostics are not substituted for current ARM. Caps/unwind settings unchanged; measured exception remains unapproved, without adding a size-only P1 blocker.
Independent scripts/verify fast is INCOMPLETE. The first invocation used a relative Cargo runner and refused CLI test execution (ENOENT), then was stopped during Clippy; preserved as verify-fast-invalid-relative-runner.log. The generated CLI binary directly passes29/29. The corrected absolute-runner invocation remained queued on all four occupied host build slots for over20minutes, produced no gate output, and was cancelled at the end of this bounded review. Sources, oracles, deadlines and caps were unchanged. No independent full-gate pass or static-size-only outcome is claimed. The reviewer harness error and resource wait are both part of this limitation.
Author-reported exact-target fast output, verbatim from the job report, NOT independently reproduced by this review:
The author attributes that failure to the unchanged static size/boundary checks. Independent supplied-image byte inspection agrees with the reported sizes; that is not a fresh independent full gate or an owner-approved size exception.
Browser artifacts were built on the review branch with target production UI sources; only tests differ from target. No byte-identical build-metadata claim, visual-quality judgement, physical iPhone result or independently repeated strong fault matrix is made.
Owned temporary gate checkout, build target, local/VM short fixture roots, VM review directory and queued gate process group removed. Raw receipts and helper source preserved outside the checkout. Main evidence branch is clean; author branch/worktree and live owner fixtures untouched. No implementation fix, push, deployment or issue closure.
ARM static file accounting, target
38b220161versus dev9e84426c4. Artifacts were SHA-256 checked against their immutable ledger, then independently inspected with readelf -SW. This is artifact inspection, not a fresh baseline rebuild. NOBITS (.bss) is excluded from file-size accounting.All ARM file bytes are accounted for. +61,332 B of unwind/exception data is offset by -16,692 B elsewhere. x86 +25,120 B (21,422,096 → 21,447,216); its unwind/exception data grows +60,684 B and other file bytes shrink -35,564 B.
The final two sharing commits reduce ARM by 4,976 B and 4,352 B respectively (author candidate identities inspected; final candidate hash equals the supplied final image). Removing those duplicated reactor paths retains panic fencing and shutdown. No new measured candidate proves another safe cut; do not call the remaining growth unavoidable. No ARM symbol-level cargo-bloat report is available from stripped artifacts; that specific requested attribution remains unverified. Historical x86 symbol diagnostics are from an older head and are not presented as current ARM evidence.
Resuming addhost-instant at base/head
38b220161(clean worktree), issue #189. Read repository contracts, parent #169 and #189 history, and r3 findings. Replaying d8b88bb93/a592a43ef/614b0cc4f/759025802 before fixing R189-03/04/05/06. Then attribute current unwind entries, reduce shared-code duplication with panic recovery retained, and run final fast through build-slot using an absolute Cargo runner. No deployment; caps unchanged.Review tests cherry-picked and pushed: cba89932b/914150d8c/a2d7f1e42/78dcad6d7. Reproduced unit baseline 15 pass/3 fail (R189-03/05/06). Repairs pushed:
1efe94cf7bounds reconnects on useful status (5 focused passes);bd8d2509fpreserves explicit recovery after bounded status loss (8 feed passes, including subscribe-from-zero and disposal);bcb8a1478validates enum strings without coercion (10 parser passes). Each tracer TypeScript check passed. Exhausted-retry production browser reproduction is queued through build-slot (all four regular slots occupied); no browser outcome yet. Native source/profile/caps unchanged so far. Current-symbol attribution still requires a fresh static build; older symbol diagnostics are not current evidence.Raw red/green receipts for R189-03/05/06 (pushed head
bcb8a1478). Unit red: 15 pass/3 fail; repairs: feed8/8 and parser10/10. These are functional checks, not build-host latency claims.unit-red.log
reconnect-green.log
retry-green.log
parser-green.log
R189-04 reproduced in the production UI on
bcb8a1478: expected zero enabled Retry buttons after attempt8/retry_available=false, received one (1 failed test). The repaired adapter now preserves the documented boolean and the shared confirmation can disable its action; plain next step names Add host. Focused parser/feed19/19 and TypeScript check pass; six viewport/theme browser checks await the current static baseline builds in the same slot. Contract and server stream unchanged.browser-red.log
Current functional status: original R189-04 browser regression freshly fails as reported (enabled Retry with authoritative false); repaired parser/feed19/19 and TypeScript compiler pass. Three earlier repairs are pushed at
bcb8a1478; eligibility wiring remains uncommitted until its production browser replay. Current Rust sources are still identical to reviewed38b220161. The slot is compiling fresh default stripped static builds with sidecar LLD symbol maps, retaining production unwind and profile settings. Pinned dev9e84426c48 source archive verified against1630 Git blobs for subsequent symbol comparison; no old symbol report substituted for current ARM attribution. No cap edit or deployment.Resume checkpoint (not completion): pushed head
bcb8a1478; native source unchanged from38b220161. Pending R189-04 patch attached: shared disabled-confirm prop, retry eligibility adapter/guide/pill, six exhausted-retry viewport/theme captures with runtime-error assertions. Unit19/19 and tsc pass; production-browser green still pending (original red reproduced and attached). Current single build-slot wrapper runs r4/baselines.sh: before-x86/ARM stripped static images plus symbol maps, then47 functional browser cases. Local current receipts: ~/.cache/codex-scratch/addhost-instant/r4/evidence/. Verified dev9e84426c baseline source: r4/baseline. Next: browser-gate and atomically commit the pending patch, attribute current unwind entries, measure code-sharing cuts, final fast with absolute runner. Caps/profile/panic recovery untouched.exhausted-retry-pending.patch
Build tooling finding/recovery: ARM stalled at its first two crates behind sleeping shared compiler-wrapper clients. A separate sccache rustc -vV probe exited124 at its15s diagnostic deadline; the job-local instance exits0 with pinnedRust1.99.0. Stopped only own cache clients1690508/1690539 and own Cargo1690419, preserving the incomplete ARM exit143 log. Shared cache server and credentials untouched. Caching remains enabled with a job-owned Unix socket/config/disk cache (documented upstream mechanism: https://github.com/mozilla/sccache/blob/main/README.md). Resumed ARM + browser through build-slot. Completed stripped x86 baseline21447216B; all its unwind/exception sections byte-identical to the reviewed artifact (overall image SHA differs; no byte-identical-image claim). Fresh sidecar map now attributes29472FDEs and8342exception tables. Current setup_api group:22entries/3680B attributed unwind+exception spans; detailed ARM and before/after tables still pending.
R189-04 repaired and pushed in logical steps:
92834adfdshared disabled confirmation;6e48c0239authoritative retry adapter/guide;446fb4a26pill wiring and six viewport/theme regressions. All four r3 findings now have pushed repairs. Production browser47/47, source19/19, tscPASS. Original R189-02 subscription and ordinary failure/retry/completion/reduced-motion controls pass. Captures cover1440/393/375 ink/paper, with zero pageerror events. Assets were built with pending R04 sources before those UI commits (checkout metadatabcb8a1478); final exact-head rebuild/replay remains planned. Contract unchanged; captures are supplied for independent assessment.Fresh ordinary stripped static baselines match reviewed sizes: x8621447216B (+25120 vs verified dev9e84426c); ARM19639496B (+44640). Both have current sidecar maps, retaining all unwind/exception data. Current grouped attribution (spans include alignment, aliases use the linker-selected owner):
Correction: the earlier22 setup entries meant grouped symbol names; its actual FDE count is36. These are current costs, not an exclusive allocation of incremental growth. Full maps, attribution code, section/address data and verified dev artifact identities are attached. The candidate shares the remaining cold typed CLI reactors through the existing non-generic I/O runner; Clippy/behavior/size evaluation is running. Panic recovery, profiles, dependencies and caps stay unchanged.
browser-green.log






ui-units.log
ui-check.log
before-symbol-evidence.tar.gz
Completed/pushed reactor module step main:
80af812e22. Existing shared I/O runner retains typed outcomes and panic cleanup (D243/D255); no profile/dependency/cap change. Actual intermediate source: fmt PASS, CLI all-target Clippy PASS; test result: ok. 86 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 31.61s Output-contract controls4/4 PASS. Logs: ~/.cache/codex-scratch/addhost-instant/r4/evidence/atomic-main-*.log. Final head gate/size table still follows.Completed/pushed reactor module step restart:
a11a320547. Existing shared I/O runner retains typed outcomes and panic cleanup (D243/D255); no profile/dependency/cap change. Actual intermediate source: fmt PASS, CLI all-target Clippy PASS; test result: ok. 86 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 29.18s Output-contract controls4/4 PASS. Logs: ~/.cache/codex-scratch/addhost-instant/r4/evidence/atomic-restart-*.log. Final head gate/size table still follows.Completed/pushed reactor module step release:
7d6b4a5617. Existing shared I/O runner retains typed outcomes and panic cleanup (D243/D255); no profile/dependency/cap change. Actual intermediate source: fmt PASS, CLI all-target Clippy PASS; test result: ok. 86 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 27.93s Output-contract controls4/4 PASS. Logs: ~/.cache/codex-scratch/addhost-instant/r4/evidence/atomic-release-*.log. Final head gate/size table still follows.Completed/pushed reactor module step cleanup:
1c67acdcc1. Existing shared I/O runner retains typed outcomes and panic cleanup (D243/D255); no profile/dependency/cap change. Actual intermediate source: fmt PASS, CLI all-target Clippy PASS; test result: ok. 86 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 28.71s Output-contract controls4/4 PASS. Logs: ~/.cache/codex-scratch/addhost-instant/r4/evidence/atomic-cleanup-*.log. Final head gate/size table still follows.Completed/pushed reactor module step remove:
bd3eb352e0. Existing shared I/O runner retains typed outcomes and panic cleanup (D243/D255); no profile/dependency/cap change. Actual intermediate source: fmt PASS, CLI all-target Clippy PASS; test result: ok. 86 passed; 0 failed; 1 ignored; 0 measured; 0 filtered out; finished in 31.15s Output-contract controls4/4 PASS. Logs: ~/.cache/codex-scratch/addhost-instant/r4/evidence/atomic-remove-*.log. Final head gate/size table still follows.Committed/pushed native sharing steps:
80af812e2main/result adapter + ownership regression;a11a32054restart;7d6b4a561release;1c67acdccI/O-only cleanup;bd3eb352ehost removal + remote-only availability. Each actual intermediate source passes fmt, strict CLI all-target Clippy,86CLI tests/0fail/1existingignore and4output controls; server-only and remote-only checks pass. Final committed source hashes exactly match the measured candidate.Same-recipe stripped static measurements with sidecar symbol maps (final ordinary fast rebuild pending):
All crash-report/unwind information and panic recovery remain enabled; no profile/dependency/feature/cap change. ARM is within the requested16KiB growth bound. It still has+57808B unwind/exception vs dev, offset by-43280B elsewhere; this is measured accounting, not a claim that further cuts are impossible or that a size exception is approved. Source reuse removes reactor/drop duplication rather than removing cleanup coverage. New test also checks original panic payload, drop count, non-Send borrow/error outcomes and runtime reuse; all3existing service panic/bounded-shutdown controls pass.
Current final milestone on
bd3eb352e: fmt/VTfmt/plugin compile-out/workspaceClippy passed; workspace test build is running. Full fast/verbatim final output and final inherited browser/real-image controls still pending. Source frozen, no deployment.reactor-cut-evidence.tar.gz
direct-size-accounting.txt
Milestone at
bd3eb352e: iterate passed fmt, plugin compile-out, clippy and cargo-deny; server suite failed 482 pass / 1 fail / 48 ignored at server_setup_terminal_echo_and_resume_preserve_offsets_once (unexpected websocket Close). The exact isolated replay passed, and the complete exact-artifact replay also passed this case. No timeout waiver or source change. Final iterate/fast will rerun with backtraces.The full direct-artifact replay had two harness failures because I omitted CARGO_TARGET_DIR: the signed-manifest fixture requires it, and the staging fixture then built a broken relative symlink. Both pass with the required absolute CARGO_TARGET_DIR restored. Raw logs are retained; this replay is not reported as a green full gate.
Follow-up R189-04 finding: bootstrap_remote.rs rejects existing non-revoked host rows, so Add host alone cannot restart exhausted setup. Guidance now names the existing Remove host command-palette action before Add host; stronger guidance assertion failed before the text change, then 19 feed/stage tests and tsc passed. Production-browser validation pending.
verify-iterate-bd3eb352e-first-fail.log
echo-resume-exact-replay.log
guide-restart-red.log
Completed follow-up R189-04 in
348809c79(pushed): exhausted setup explains Remove host in the command palette, then Add host; the existing disabled Retry stays disabled. Regression red before correction, then feed/stages 19/19, bun run check, production browser 6/6 (1440/393/375, ink/paper). Six screenshots retained for independent visual review. Final iterate/fast now running on this head through build-slot with the absolute Cargo runner.guide-restart-green.log

guide-restart-check.log
guide-restart-browser.log
echo-resume-unfiltered-server-replay.log
server-replay-env-correction.log
Final-head iterate at
348809c79: the prior websocket case passed. One different server assertion failed, cancelled_invitation_retires_after_sqlite_busy_timeout, after its two-second polling wait; server total 482 pass / 1 fail / 48 ignored. The exact artifact replay with the proper environment passed. Root cause remains unproven; no test/production timeout changed and no failure waived. This repair round has no infraternal-server source changes. The required full fast gate is now running on the same clean head (absolute interpreter and Cargo runner script paths), rerunning the entire native suite.verify-iterate-348809c79-retirement-fail.log
retirement-exact-replay.log
Required full fast completed on clean pushed head
348809c793through build-slot; absolute /usr/bin/python3 and absolute Cargo runner script. Exit 1: unchanged exact-size gates only. Every native test passed (server 483/0/48 ignored), including both assertions that failed in prior iterate attempts. No timeouts or native tests changed/waived. TS verification then ran separately and passed: 210 local tests, 1 unchanged timing case on fair-locked VM208, 37 protocol tests; compiler and transfer budgets pass.Both ordinary final static outputs are SHA-identical to the sidecar-map candidate, confirming the size attribution applies to the actual shipping builds: x86 21,414,096 B (-8,000 vs dev); ARM 19,609,384 B (+14,528 vs dev, under 16 KiB allowance). Caps remain unchanged: x86 exact-cap ratchet refuses a decrease, ARM cap refuses any growth; static-size fixture likewise refuses x86. No non-size failure in final fast. Browser functional suite and locked VM pill controls still running.
Verbatim Rust gate summary:
Full unedited output and TypeScript continuation:
verify-fast.log
verify-ts-after-fast.log
author-checklist.md
Repair round finished at
348809c793, clean and pushed to origin/addhost-instant. Stop for independent break-it review; no deployment or issue closure.All four review tests were cherry-picked and confirmed red before their fixes. R189-03: empty opened sockets and old-sequence replays cannot restore the reconnect budget. R189-06: deliberate Retry survives exhausted status delivery, re-subscribes from zero and then requests retry. R189-04: the server's retry_available controls the shared disabled confirmation and callback; exhausted setup explains Remove host in the command palette, then Add host (the duplicate-host guard requires this). R189-05: malformed enum objects are dropped without coercion exceptions. Existing retained-status contract unchanged.
Atomic fixes and size steps were pushed throughout; last correction
348809c79. Native size cut shares the existing I/O reactor across cold server/restart/release/removal paths, retaining typed borrowed results and every panic-recovery boundary. Regression covers original panic payload, exactly-once cleanup and reactor reuse; disabled plugin configurations also pass.9e84426c4B38b220161BException/unwind section cuts: x86 .gcc_except_table -2972 B, .eh_frame -832 B, .eh_frame_hdr -8 B (total -3812); ARM -3196/-296/-32 B (total -3524). FDE records 29472→29471 x86, 29063→29059 ARM. ARM is within the requested 16384 B delta. Profiles, unwind information and caps unchanged. Final ordinary shipping images are SHA-identical to the same-link map candidate. The attached size review includes current new-scope attribution, per-function before/after reductions, aliases/alignment limits and exact file accounting; maps and parsing scripts are in the already-attached reactor-cut-evidence archive.
Validation: full fast native test row PASS (server 483/0/48 existing ignored); fmt, VT fmt, plugin compile-out, strict workspace Clippy and cargo-deny PASS. The two distinct earlier iterate failures are preserved on this issue; both exact replays and both cases inside final fast pass, without timeout changes or waivers. Full fast exits 1 solely because frozen exact caps reject x86's -8000 B decrease and ARM's +14528 B delta; the static-size fixture also rejects the x86 decrease. Do not read this as a green complete release gate.
TS continuation PASS: 210 local tests + 1 unchanged timing case on VM208; 37 protocol tests, compiler, bundle/transfer checks PASS. Production browser 60 PASS / 8 optional capture-only skips; 15 final-head screenshots attached for independent assessment. R189-01 normal nonignored replay 1/1 PASS. Seven real-image controls PASS: same session, pending input once, output replay, rebridge, private image scope and manager admission/refusal.
Unchanged physical pill specs on fair-locked VM208: 4/4 PASS, machine infraternal-perf-test (4 vCPU), starting load 0.16/0.05/0.05. Normal/reduced-motion status-induced long tasks 0; script share 2.9% vs 5.0% same-page baseline; worst frame gap 50 ms is retained in the raw receipt. Source/spec and prebuilt final-head asset hashes attached. Startup latency evidence remains the r3 comparison already on this issue; the deferred first-prompt targets remain open.
How to test:
Final scripts/verify fast summary verbatim (absolute Cargo runner; complete unedited transcript linked below):
verify-fast.log
verify-ts-after-fast.log
author-checklist.md
final-size-review.txt
final-size-accounting.json
final-size-accounting.txt
final-real-image-controls.log
final-r189-01.log
final-receipt.json
final-gates-console.log
final-browser.log



head-348809c7-production-captures.tar.gz
pill-timing-vm.log
pill-timing-provenance.json
head-348809c7-pill-vm-evidence.tar.gz
gate-target-hashes.json
reactor-cut-evidence.tar.gz
direct-size-accounting.txt
Starting r4 delta verification on review-addhost-instant-r4, base/head
348809c793; scope 38b220161..348809c79 only. Four requested r3 cherry-picks are already present in target as cba89932b/914150d8c/a2d7f1e42/78dcad6d7; attempted replay conflicts with existing copies, so retained target tests without duplication. Replaying all four, checking setup/replay panic recovery and supplied static sizes; no implementation changes, push or deployment. Evidence: ~/.cache/codex-scratch/addhost-instant-review/evidence/r4/.r4 completed source/browser replay at unchanged head
348809c79: feed+stages 19 PASS/0 FAIL; all four r3 findings now pass (R189-04 all six width/theme cases), production browser 9 PASS/0 FAIL including ordinary failure/retry, completion and reduced-motion controls. Independent read-only SHA256/stat/readelf checks of final supplied images match head ledger: x86 21,414,096 B (-8,000 vs pinned dev); ARM 19,609,384 B (+14,528). Both static/no PT_INTERP and retain .gcc_except_table/.eh_frame/.eh_frame_hdr. Full fast restarted before native tests with umask022 and absolute Cargo runner after detecting shell umask002; initial log preserved. Panic replay/native gate pending. No new confirmed finding, implementation change, push or deployment. Evidence: ~/.cache/codex-scratch/addhost-instant-review/evidence/r4/.Completed native CLI replay at unchanged review/target348809c79:86PASS/0FAIL/1existingignore. shared_reactor_preserves_borrows_errors_and_panic_cleanup PASS, polling_panic_reaches_bounded_shutdown PASS, successful_link_return_is_a_service_failure PASS, link_failure_does_not_wait_forever_for_blocking_work PASS. Original panic payload, exactly-once cleanup, borrowed/non-Send result/error and runtime reuse assertions pass. Profile/unwind and setup/task/replay boundaries remain unchanged in delta; supplied static images retain unwind sections. This executes the reactor/service controls, not a newly injected end-to-end setup/replay panic. Full fast now through plugin compile-out into workspace Clippy; pending. No new confirmed findings.
units.log

browser.log
independent-size-check.json
static-cap-inspection-copy.log
native-cli.log
panic-source-scope.txt
r4 full fast checkpoint, unchanged head348809c79: fmt PASS, VT fmt PASS, CLI compile-out PASS (29 tests), strict workspace/all-target Clippy PASS. Workspace test row executing: CLI86PASS/0FAIL/1existingignore again; first daemon/integration cases PASS. Static rebuilds have not started. Focused19source/9browser and native panic controls remain PASS, both supplied artifact sizes independently verified. No new findings/implementation edits. Full fast remains incomplete at this checkpoint; no passing full-gate claim. Raw owned log: ~/.cache/codex-scratch/addhost-instant-review/evidence/r4/verify-fast.log.
Independent r4 full-gate native rows completed at unchanged head
348809c79: fmt, VT fmt, CLI compile-out, strict workspace/all-target Clippy, unfiltered workspace tests (ten temporal transport cases executed unchanged on locked VM208) and cargo-deny PASS. Server 483 PASS / 0 FAIL / 48 existing ignores; shim descriptor/signal tests also PASS after their nested probe build completed. Both earlier author-intermittent server assertions passed in this run. Static release cross-builds now running. No source/oracle/deadline/cap edits or waivers; no new findings. Full fast not yet complete. Evidence: ~/.cache/codex-scratch/addhost-instant-review/evidence/r4/verify-fast.log.Fresh independent source build at review/target
348809c79: x86 static image 21,414,096 B, exactly -8,000 B vs pinned dev. SHA256 e904ca390bfeea27bb314dc51eb7a769ef6b57fab8a2a62cbb174450f2d67371 differs from supplied image; same byte count and byte-identical unwind/exception sections, no byte-identical-image claim. Production trust/static linkage checks pass; size checker refuses the decrease until the frozen ceiling is ratcheted. ARM source build running. All native rows PASS (1,205 passing results/0 failures/143 existing ignores). Ten unchanged temporal cases separately replayed on locked VM208 with explicit machine/load: infraternal-perf-test,4vCPU,start0.07/0.14/0.34;10PASS. Full fast pending static completion. No source/deadline/cap changes, new findings, push or deployment.MERGE (merge), under the requested no-P0/P1 rule. Remaining scoped findings: P0=0, P1=0, P2=0, P3=0. All four r3 findings pass; no new finding.
Target and review head:
348809c793ad9786bb450a6622b7b0d361a81df5, branchreview-addhost-instant-r4. Scope:38b220161..348809c79only. The four requested tests were already imported as cba89932b/914150d8c/a2d7f1e42/78dcad6d7; the duplicate cherry-pick conflicted and was aborted. Target copies were retained. No implementation edit, new test commit, push, deployment or issue closure.review reconnect budget survives sockets that open then close without eventsreview exhausted setup does not offer an enabled retryat 1440/393/375 px, ink/paperreview malformed setup fields are dropped without coercion errorsreview explicit retry resumes after the setup feed stopsFocused source units: 19 pass / 0 fail. Production browser: 9 pass / 0 fail, including ordinary retry, completion and reduced-motion controls. TypeScript compiler passes. Screenshots are retained; no aesthetic, physical-iPhone or new startup-latency claim.
Native panic controls pass: original panic payload, exactly-once cleanup, borrowed/non-Send typed result/error and runtime reuse in
shared_reactor_preserves_borrows_errors_and_panic_cleanup; all three service panic/error/bounded-shutdown controls pass. The release unwind profile, existing setup task boundary, service panic fence and replay boundaries are unchanged in this delta. Both supplied images retain unwind/exception sections. This checks the existing reactor/service boundaries; no new end-to-end setup/replay panic injection is claimed.Full native gate: fmt, VT fmt, CLI compile-out (29 tests), strict workspace/all-target Clippy, workspace tests and cargo-deny pass. Workspace/VM result rows record 1,205 passes / 0 failures / 143 existing ignores, including server 483/0/48. Both assertions that intermittently failed in the author's earlier runs pass here. Ten unchanged temporal cases pass on locked VM208; separate receipt records infraternal-perf-test, 4 vCPUs, start load 0.07/0.14/0.34. No deadline/oracle/profile/cap change or test waiver. No dependency manifest/lockfile changed; cargo tree -d retained.
Both supplied images were independently stat/SHA256/readelf checked against their exact-head ledger: static/no PT_INTERP, production trust check passes, unwind/exception sections retained. Fresh independent source builds reproduce BOTH byte counts exactly. Both fresh hashes differ from the supplied artifacts; no byte-identical-build claim. Fresh images also retain unwind/exception sections and no PT_INTERP. The supplied-artifact hashes and fresh-build hashes are recorded separately.
Evidence:
~/.cache/codex-scratch/addhost-instant-review/evidence/r4/. Initial pre-native gate invocation was stopped to correct shell umask002; corrected full invocation uses umask022, an absolute native runner and the checked-in VM temporal oracles. The initial log remains separate. Local gate durations are not performance evidence.Independent
scripts/verify fastcompleted on this exact head with exit 1: frozen exact-size checks only. x86 must ratchet its ceiling down; ARM exceeds its frozen ceiling by 14,528 B. The static-size fixture likewise refuses x86’s decrease. Production trust, static linkage and shipping key-memory checks pass. This is not a green full release gate; no cap exception is claimed.Verbatim final summary:
Cleanup complete: owned local Rust/supplied-image targets, fixture root, generated test target/UI build/cache and VM208 review directory removed. Evidence retained.
git diff --checkpasses; worktree clean at the same target/review SHA. No push, deployment or issue closure.Complete unedited gate transcript SHA256:
8417857eab9df70346e668ada333bc5521b7fee9ae5e485f8d79d66f55a8b062.verify-fast.log
fresh-static-check.json
temporal-vm-with-load.log
review-harness.tar.gz
units.log

browser.log
independent-size-check.json
static-cap-inspection-copy.log
Starting the requested add-host demo on branch
addhost-instant, head348809c793ad9786bb450a6622b7b0d361a81df5;origin/devis650472b987dbd1fe7192248dd1367d0766c5fb79, with merge-base9e84426c48f3818cebaa99acede14fec1d24ab4e. The worktree is clean. I will preserve this head without rebasing and use disposable targets. The referenced reconnect-release artifact directory is currently absent; locating the production binary and demo path before launching anything.Finding with evidence: the referenced production artifacts directory
/home/kayg/.cache/codex-scratch/reconnect-release/artifactsdoes not exist, and/root/infraternal-bench/rules.mdis absent on the VM.netbird ssh --no-browser root@10.69.69.208succeeded with strict host-key checking; VM hostnameinfraternal-perf-test, 4 vCPU, load0.25/0.32/0.41, no running containers, and cached disposable imageinfraternal-addhost-breakit2:debian(294 MB). I am rebuilding the exact production image from492e1505aand branch head348809c79on the build host, then will hold the VM lock only for the short demo run (under 30 minutes).Finding: production source
492e1505abpredates D255's retained machine-pill setup UI:apps/tracer/src/canvas/addhost/setup-pill.on-demand.tsandsetup-feed.tsare absent from that source tree. The production recording will therefore show its existing joined-before-shell flow; the branch recording will show the early prompt, retained setup status, and post-prompt network-cut Retry.Demo build receipt
2026-10-16; x86_64-unknown-linux-musl; static release profile
Production source:
492e1505abBuild: cargo build --locked --release --package infraternal --target x86_64-unknown-linux-musl
Size: 22,299,504 bytes
SHA-256: bebd5b8059d5d6ffe5c3c4c697a47339f5ba24c2be842219f443f1bdc4000746
Branch source: 348809c793ad9786bb450a6627b2b0d361a81df5
Build: cargo build --locked --release --package infraternal --target x86_64-unknown-linux-musl
Size: 21,414,096 bytes
SHA-256: eeba7f402b342e597d0cf0da12c01b5d69d244d2754e086f6801328e3d2fe3a2
Branch early-session payload:
Size: 1,889,480 bytes
SHA-256: 5fd1655184d2e9ea3c9a4df23535190f90d92edefebb96ca8f4f4681bb7be470
Both server executables are static PIE, stripped; readelf reports no PT_INTERP.
The brief's /home/kayg/.cache/codex-scratch/reconnect-release/artifacts directory was absent. The production release URL https://builds.infraternal.com/releases/0.1.0/SHA256SUMS returned HTTP/1.1 404. The production demo therefore uses a release-profile binary built from the exact production source commit and its local x86_64 payload sibling, with the production web build from that commit. This is not a byte-for-byte copy of the signed released artifact.
Finding on
348809c793: at 150.8 ms median RTT, the branch prompt appeared at 12,639 ms. Resetting the target SSH transport at 13,183 ms caused the backend to log upload failure (exit 255). The canvas showed no failed setup state or Retry after 30 s. A fresh authenticated subscription then replayedinstall:failed,failed_step=upload,retry_available=true, sequence 2. This points to the live canvas subscription losing the retained failure. Partial recording and sanitized logs are attached.Follow-up on the retained-status finding: the 150.8 ms RTT capture shows the pill status socket subscribed with valid host/window fields and received
install:running. The page explicitly called WebSocket.close 12,843.6 ms after flow start (238.6 ms after the prompt, before the injected reset at 13,135.3 ms). After the upload failure, a fresh subscription replayedinstall:failed,failed_step=upload,retry_available=true, sequence 2, while the visible host window had no setup state. Logs are attached. This narrows the loss to the live canvas subscription/window lifecycle.Demo finding: on branch
348809c79, the setup-status socket is disposed at 12,414.6 ms, 213 ms after the first prompt and before the injected network reset at 12,800.2 ms. The post-reset subscription still replays install:failed (sequence 2, retry_available=true), but its window ID differs from the remaining visible terminal window, so no failure pill or Retry is shown. Full run and cleanup stack: status-window-close-stack.log. I’m adding a regression that exercises clean terminal exit during pending setup before fixing the window lifecycle.Perf VM demo finding, branch
348809c79at 150.7 ms RTT: the remote host inventory replacement closed the new terminal window before the host had joined (window_close kind=terminal setup_bound=false, callerreplaceHosts). The late subscriber still received retainedinstall:runningtheninstall:failed(failed_step=upload, sequence 2, retry_available=true), but the window was gone so no pill or Retry appeared. Raw run: debug2-demo.log.I added a regression in
live.test.ts; before the fix it failed with 6 passing / 1 failing, where the empty snapshot removed the setup host. The fix keeps locally enrolled setup hosts through inventory snapshots until setup completes or the host is removed.Demo finding on addhost-instant at
348809c79: after first prompt (12,470 ms at 150.8 ms RTT), resetting the target SSH transport makes setup fail honestly (install, retry available). The window survives, and explicit Retry is sent; retained status then transitionsinstall:running→install:failedagain within the 55 s join timeout, with no joined event. Server log records upload channel exit 255 and install transport unavailable. Evidence: retry trace, server log, capture. I am treating this as a retry-path defect and will add a failing regression before any fix.Step complete: UI host-window retention committed as dbc7d23e7. The locally enrolled terminal stays attached across an early empty inventory snapshot and a clean shell exit until retained setup reports ready or the host is explicitly removed. Gates:
bun test src/canvas— 180 passed, 0 failed, 11,010 expects;bun run check— exit 0;git diff --check— exit 0. The regression reproduced before the fix as the shell window closing before the retained failure/retry reached its machine pill.The server retry defect now has a focused red regression in
crates/plugins/ssh/tests/publication.rs: after the published shell's first SSH transport is lost, the second retry attempt ends at Install instead of Ready. The failure is deterministic (1 failed assertion, 0 passed; expectedReady, gotInstall) and models preserving the same shell UUID while requiring a newly opened authenticated transport. Raw test output.Step complete: transport-loss retry fix committed as 20a3f9299. Explicit retry now first attempts same-master reattach; on
Unreachable, it re-resolves the stored credential reference with the live requester authority, opens through the strict SSH transport, and resumes the existing shim/session ID and transaction. The red regression is green; publication tests: 4 passed, 0 failed; plugin all-targets clippy: exit 0;cargo fmt --all -- --check: exit 0. Evidence: publication suite, clippy.Final
scripts/verify fastat 20a3f9299 fails only the static-build row: x86_64 is 21,415,952 B vs the 21,422,096 B cap (−6,144 B, which requires ratcheting the cap down); aarch64 is 19,611,048 B vs 19,594,856 B (+16,192 B). The static-size fixture also fails while the x86 cap is stale. I am investigating the ARM size and will keep both ceilings unchanged or lower them; no cap increase.The size refinement reuses the credential reference already held on the retry target and moves reattachment/reconnect into a non-inlined helper; host-key verification and the same owned shell ID remain required. Focused publication suite passed 4/4, plugin SSH clippy passed, and formatting is clean. I am measuring both static targets before deciding whether the cut is sufficient.
Final size accounting at 20a3f9299, measured against the frozen origin/dev caps:
ARM is 192 B inside the 16 KiB allowance and retains
.gcc_except_table,.eh_frame_hdr, and.eh_frame. No cap was changed.scripts/verify fastpassed every row exceptstatic builds(x86 decrease requires a downward ratchet; ARM exceeds the frozen cap by 16,192 B). I tried an isolated non-inlined reattachment helper; it increased ARM by 288 B, including +128 B unwind/exception sections, so I reverted it. The measured section table and full gate log are attached: size record, gate output.The real 150 ms demo found the remaining Retry gap. The branch shell prompt appeared at 12.66 s; the SSH transport was reset at 13.22 s. The explicit Retry did not join within 55 s. The server log records the reset upload as channel exit 255 and the retry as
The SSH terminal bridge pipes are unavailable.The old master returned an opened command channel that ended at the SESSION receipt, soresume_setupnever sawFailureKind::Unreachableand did not resolve a fresh SSH session.I added a regression that models this exact real channel-EOF path with a command that closes before the SESSION receipt. It fails before the fix: attempt 2 ends at Install/Failed with
opens=1, resolves=1, published_sessions=1instead of reopening SSH and completing the same shell. Receipts: screen recording, server log, red regression, demo console.EOF before receipt now reports
FailureKind::Unreachable; local pipe errors and oversized receipts keep their existing refusal path. The reopened-session regression passes, both duplex tests pass, and all four publication tests pass. SSH-plugin clippy is queued through build-slot; static gate and real demo rerun remain.Committed
e2d9d74426674f9060026abe281c078f44308547: receipt-channel EOF now enables the strict reconnect path. Focused checks: duplex 2/2, publication 4/4, SSH-plugin all-target clippy PASS, fmt and diff check PASS. The regression’s prior red output is attached above. Starting the post-fix workspace iterate gate now; full fast will run at the final head.Post-commit iterate finding at e2d9d7442: server test
remote_runtime::tests::standby::control_queue::saturated_native_control_queue_does_not_hide_primary_terminal_outputfailed once. Evidence:CONTROL_SATURATION active=true queued=1 native_primary_live=true shadow_output=0; terminal delivery ended at its 100 ms deadline. Server tests reported 482 passed, 1 failed, 48 ignored (exit 101); other iterate rows passed. Full raw gate log: https://git.kayg.org/attachments/eb333141-9a28-462c-a04b-38d60ea5d85d. The change in this commit is limited to SSH stdout-EOF classification; I am checking whether this timing-sensitive server test reproduces before proceeding.Follow-up on the one iterate failure: with the server
remotefeature enabled, the isolated case passed (1 passed, 0 failed;CONTROL_SATURATION active=true queued=1 native_primary_live=true shadow_output=0, finished in 0.13 s). This confirms the control saturation setup and makes a 100 ms full-suite scheduling miss the likely cause. The initial full log is attached above; isolated log: https://git.kayg.org/attachments/3716dc9b-17a5-4202-8efe-6a2bfe700441. No source changes were made for this unrelated case.Final fast at e2d9d7442 completed: fmt, VT fmt, CLI compile-out, clippy, workspace tests, and cargo-deny passed; static builds failed. x86_64 is 21,415,952 B (6,144 B below its existing 21,422,096 B cap); ARM is 19,611,048 B (16,192 B above its existing 19,594,856 B cap). The static-size fixture also rejects the x86 down-ratchet requirement. Caps are unchanged. Full raw gate output: https://git.kayg.org/attachments/446cb97c-2599-429a-bcb3-e3424cd957ef.
Demo finding on head e2d9d7442: at median 150.7 ms RTT, the branch produced a prompt at 12,580 ms. After the injected target SSH reset at about 13,250 ms, Retry did not reach joined within 55 s; the server log reports install failure with
unexpected end of file. The branch video and raw logs are attached. The production-vs-branch demo is incomplete until Retry succeeds.Attachments: run, server log, branch recording.
Retry regression isolated on the perf VM with 150.7 ms median RTT: prompt 12,551.9 ms; network cut 13,094.1 ms. Attempt 1 retained
install:failed,failed_step=upload,retry_available=true. Explicit Retry then emitted attempt 2install:runningfollowed byinstall:failed, with the public reasonSign in before opening the SSH terminal bridge.The host/window subscription matched on both events. This is the deferred automatic-auth session being used forstream()before the required read-only check. Regressionretry_reopens_ssh_after_transport_loss_and_keeps_the_published_shellfailed 0/1 before the fix with that exact message; after adding the check before reattachment it passed 1/1. Trace: retry debug.Continuing the requested demo/retry repair on branch
addhost-instant, starting ate2d9d7442(lane base348809c79). The perf-VM demo exposed a reconnect failure; the regression is now red before the retry authentication check and green after it. I am rerunning the full iterate gate, then I will rebuild and record the production-versus-branch demo with the network-cut Retry.Retry fix committed as
c250bd7f2onaddhost-instant. The deferred automatic-auth regression was red before the fix and passes 1/1 after it; fullscripts/verify iteratepasses, including workspace tests, Clippy, Rust formatting, and TypeScript checks. The gate transcript is attached: verify-iterate-auth-retry-green.log. I am running the final static gate next.Final gate finding on
c250bd7f2: server tests reported 482 passed, 1 failed, 48 ignored;responsive_socket_survives_a_held_terminal_open_with_an_existing_attachmenthit the 90 s socket-test watchdog. A filtered branch replay first failed at 90 s, then passed 1/1 in 1.13 s. The clean detachedorigin/devbaseline (c47a438be) passed the same test 1/1 in 1.13 s using a separate Cargo target. The socket-test source is unchanged between the baseline and branch; the changedFrozenSocketTimevisibility does not alter its behavior. I am rerunning the full gate with one Rust test thread to check the suite interaction.Evidence: fast gate, branch red replay, branch green replay, clean dev baseline.
Finding from the real 150 ms RTT demo: production completed at first prompt 20,915.8 ms / joined 20,416.2 ms. On c250bd7f2, the prompt appeared at 13,342.7 ms; after the target SSH transport was reset, setup reported
failedat stepuploadwith retry available. Clicking Try again emitted attempt 2 running, then failed with “Choose an SSH authentication method.” The target RTT median was 151.2 ms. This is either a retry-flow gap or a missing precondition in the demo; I’m tracing it before changing code. Raw event log and both recordings are attached: log, production recording, branch recording.Demo complete at head
c250bd7f23304cc1902b9ed21b6b98ab5ac89914; no code changed after this commit.492e1505ac250bd7f2The branch recording shows the target-to-server route cut at 13.50 s, a retained
join/failedstatus with retry available and the message “Joining the host to this server failed. Check the host can reach the server and resolve its address, then try again.”, route restoration at 19.57 s, then attempt 2 throughready/done. It kept the published SSH channel alive so Retry could use the existing authenticated session. The production and branch clips are sequential in a 60.52 s video: demo video; raw event log. Local video:/home/kayg/.cache/codex-scratch/addhost-instant-demo/evidence/addhost-instant-c250bd7f2-demo-final.webm.A prior full SSH transport reset used the wizard's one-use private key; Retry then correctly requested fresh SSH authentication. This disposable CLI server has no configured retained server SSH key, so the final acceptance demo cuts only the target-to-server route and restores it before Retry. The first recording and trace are attached in the earlier finding comment.
Final
scripts/verify fastoutput on this head (also attached as raw gate log):The one test failure was the intermittent 90 s auth socket watchdog
responsive_socket_survives_a_held_terminal_open_with_an_existing_attachment; an immediate branch repeat and a cleanorigin/devbaseline each passed 1/1 in 1.13 s. Static size check: x86_64 21,416,656 B (5,440 B below the unchanged ceiling); aarch64 19,611,560 B (16,704 B over the unchanged ceiling). No caps were changed.Starting merge preparation on branch
addhost-instantatc250bd7f23304cc1902b9ed21b6b98ab5ac89914. After fetch,origin/devis7884d13dcad71e8bc6f7d73fe1339c3f688e650b; the pre-rebase merge base is9e84426c48f3818cebaa99acede14fec1d24ab4e. Worktree is clean. I am reviewing the diverged history and will rebase the lane onto current dev, then run the requested final gate and exact-size checks.Rebase finding/resolution: the first replay (
0d7cbaac7, fixed ownership-check authentication) conflicted with current dev incrates/plugins/ssh/src/auth_order.rsandcrates/plugins/ssh/tests/auth_order.rs. I kept dev's cached exact-script allowlist and retained both test paths: authentication fallback for root/rootless ownership checks, plus refusal of appended mutations. The resolution is staged andgit diff --checkis clean.Rebase finding/resolution:
f28492adcoverlapped the existing dev regression incrates/infraternal-agent/src/readiness_exit_tests.rs. Dev already asserts the replay-complete fence's type, sequence and session ID before credit-funded output; I retained that stronger assertion and the new daemon-side source change.git diff --checkis clean.Rebase finding/resolution: the production HTTPS fixture in
crates/infraternal-server/src/bootstrap_breakit2_tests.rsconflicted between direct OpenSSL binding and the lane's shared production acceptor. I keptserver_tls::acceptorfor the provisioning listener so the fixture exercises the production TCP admission path and its TCP_NODELAY behavior;server_tls.rsand its real-socket assertion are included.git diff --checkis clean.Rebase finding/resolution:
crates/plugins/ssh/src/lib.rsconflicted where dev added the prompt image module and the lane added its owned-socket connector. Both are independent modules, so I kept both declarations; the connector implementation remains unchanged.git diff --checkis clean.Rebase finding/resolution:
cb414a4bconly adds__ssh-connectto the old hand-written CLI enum and dispatcher incrates/infraternal/src/main.rs; current dev uses the registered action tree and has no such enum. I am skipping this obsolete adapter commit. The later privatessh_internalentry path calls the retainedsocket_connectordirectly, which preserves dev's CLI architecture and the intended descriptor handoff.Rebase finding/resolution: the TCP_NODELAY SSH connector overlapped current dev's stricter SSH error boundary in
crates/plugins/ssh/src/openssh.rs. I retained the dev boundary mapping and attached the connector to the session;bootstrap_runtime.rskeeps dev's/proc/self/exeaskpass while pointing the connector at the resolved executable path. The connector remains confined to the configured host dial, before SSH's normal host-key and authentication checks.git diff --checkis clean.Rebase finding/resolution: extracting the SSH activation script into
setup_scripts.rswould have reverted current dev's service unit fromRestart=alwayswithStartLimitIntervalSec=0andRestartSec=2toRestart=on-failure. I copied the current dev activation body into the shared constant and kept the installer calling that helper.STARTandLINGERmatch dev's existing bodies;git diff --checkis clean.Rebase finding/resolution: the compressed-payload fixture's
file_bytesaccounting conflicted incrates/infraternal-server/src/bootstrap_breakit2_tests.rs. I kept dev's file-input accounting and added byte-input accounting only when the step is an upload, matching the compressed transfer path. The payload implementation and startup preparation metric remain applied;git diff --checkis clean.Rebase finding/resolution: the one-command setup change overlapped the current first-prompt payload selection in
crates/plugins/ssh/src/bootstrap.rsand the SSH error boundary inopenssh.rs. I keptstage(&payload)and its missing-binary receipt, then use the combined setup command for enrollment/start/linger; transport failures keep dev's sanitized boundary and preserve the named failed step from the bounded setup receipt.git diff --checkis clean.Rebase finding/resolution: the install-refusal injection in
crates/infraternal-server/src/bootstrap_breakit2_tests.rscollided with dev's compressed-upload byte accounting. I kept that accounting and the compound-setup classification, then retained the injectedFAIL installreceipt so the test exercises the production setup decoder after enrollment.git diff --checkis clean.Rebase finding/resolution: the remote plugin's optional UUID dependency conflicted with current dev's agent transport features and dependencies. I kept dev's
ring,zeroize,bytes,nix/net, andsocket-hostfeature declarations, added optional workspaceuuidto the transport feature, and merged its package entry intoCargo.lock. No engine or version choice changed;git diff --checkis clean.Rebase finding/resolution:
crates/plugins/remote/src/lib.rshad an export conflict between dev'sremote_bindingsand the lane's sharedenrollmentmodels. They serve separate modules, so I kept both behind the transport feature; the server adapters now use the shared model without pulling server storage into node builds.git diff --checkis clean.Rebase finding/resolution: the old server-removal commit conflicted with dev's registered action CLI and existing
plugin-serverfeature. I kept dev'smain.rs, lockfile and action-output test, then adapted the feature split:remote/sshuse weak server dependencies, whileplugin-serverenables the remote server half when present. I retained the shared enrollment model import so remote node builds do not depend on server storage. The default feature set still includes the server.git diff --checkis clean.Rebase finding/resolution:
crates/infraternal-agent/src/plugins.rshad a registry-helper rename conflict. Dev'sregistry_at(mode, state, installation)also registers the optional Mesh component, so I kept that state-aware API and its empty-registry behavior, carrying over the explicit lint reason.git diff --checkis clean.Starting the already-added host id change on
already-added-open, based onorigin/devb231301396227821a992692410cf55c50abe04ef. The worktree is clean. Reviewed #189 and parent #169 history; the existing already-added response has no host id for the client action.Rebase finding:
#189commit7557420conflicted with dev's signedPayloadSourcepath incrates/plugins/ssh/src/bootstrap.rs. Keptself.payload(...)as the selected, trusted payload for both the early shell and regular setup; early setup still stages the shim before the setup sequence and retains the resolved path for retry. This avoids bypassing release verification while keeping same-architecture resolution on the existing local executable.Rebase finding: SSH transport admission overlapped current remote routing changes in
remote_link.rs,remote_runtime.rsandruntime.rs. Kept the existing remote host/plugin bindings and canonical tunnel IP behavior, added the SSH link and generation notification, and bound bootstrap progress to the existing session inventory watch. The shared remote runtime test helpers remain available to current sibling suites; the SSH bridge oracle is registered alongside them.Rebase finding: component-off retirement overlapped dev's priming command and remote link identity helpers in
embedded.rsandremote_link.rs. Kept both actor commands, uses the current synchronousconnection_losthelper for an active SSH channel, retires only SSH-backed fallback/current channels, and preserves peer-address mapping for QUIC/TCP.RetireSshis acknowledged after cleanup completes.Rebase finding:
crates/infraternal-agent/src/daemon.rsnow warms init-manager detection in the background while preserving dev's async plugin admission worker and sealed-image detached mode. The greeting/census can proceed while the singleOnceCellprobe runs;terminal.openawaits that result before launching the shim.Rebase finding: the session-image split in
bootstrap.rsoverlapped dev's signed fullPayloadSource. Early staging now uses the selected session image, while background restaging and retained retry coordinates use the exact payload returned by the trusted source. When no separate session artifact exists, the existing setup path remains in place.Rebase finding: the check-time transfer encoder conflict is resolved by retaining dev's verified full
PayloadSourcepath while passing the authenticated check channel's rootless encoding choice torestart_setup_with. The background restage uses the exact trusted payload without changing the handover order.Finding:
bootstrap_remote.rsuses an owner-filteredremote_host_bootstrap/remote_hostsjoin but selects onlyEXISTS, discarding the matching host id. The new focused regression is red: expected00000000-0000-0000-0000-00000000002a, serialized lookup returnednull(0 passed, 1 failed). Raw outputThe requester fence remains unchanged:
RemoteBackend::authorizerevalidates the bootstrap session owner, and the duplicate query filtersh.owner_idto that owner plush.revoked_at IS NULL. Terminal session admission uses the same owner/live-host predicate (websocket.rsaroundterminal_actor); no other requester gains a host id.Contract proposal for #189 / TASTE row “Messages hand you the next step”: a failed bootstrap event may add optional
failure.host_id(UUID) only whenfailure.already_addedis true. The id comes from the existing duplicate match scoped to the validated bootstrap owner and a non-revoked host; all other failures omit it. This is additive for existing clients, and the UI remains outside this change.Clippy finding: the initial inline
Option<Uuid>made shared SSHFailure144 bytes and triggered 65result_large_errdiagnostics across the touched crates. I am changing the representation to reuse the existing UUID slot, since host-key challenges and duplicate-host ids are mutually exclusive; serialization will keep the existinghost_key_challengekey and emithost_idonly foralready_added. Raw Clippy outputRebase compile finding: the first x86_64 static build failed at
crates/infraternal-server/src/bootstrap_remote.rs:303because currentRemoteHostsexposesadmit_observed(authority, link, ip)rather thanadmit;crates/infraternal-server/src/remote_link.rs:589also lacked the SSH case inPrimingAdmission.connection. I am adapting both to current dev semantics. No size measurements yet.Resolved the Clippy finding without growing
Failure: its existing UUID slot now serializes ashost_key_challengeorhost_idaccording to the failure case. Touched-crate Clippy passes; focused Rust tests pass 4/4 (duplicate id, unrelated omission, foreign-owner omission, and unchanged host-key field). Clippy outputBackend and contract slice committed:
eae097700.The duplicate lookup now returns the matching live host UUID under the existing bootstrap-owner and non-revoked filters. Only that already-added failure serializes
host_id; cold-CA and foreign-authority failures omit it. Host-key challenge serialization remains unchanged. Focused Rust tests passed (server cases 3/3, compact UUID serialization 1/1), touched-crate Clippy passed, and formatting/diff checks passed. An addedOption<Uuid>madeFailure144 bytes and produced 65result_large_errClippy failures, so the implementation reuses its existing UUID slot.TASTE: “Messages hand you the next step.” #189
Receipts: serialization test, server tests, Clippy.
TypeScript slice committed:
cecfc219c.FailureDetailnow accepts optionalhost_id, andparseFailurekeeps it only when it is a canonical UUID on analready_addedfailure. Parser coverage includes the valid id, unrelated failures, and malformed ids; no UI code changed.bun test src/canvas/addhost/real.test.ts: 13 passed, 0 failed, 32 assertions.bun run check: passed.TASTE: “Messages hand you the next step.” #189
Receipts: Bun tests, TypeScript check.
Rebase integration finding:
cargo test --locked -p infraternal-server --lib --features sshfailed while compilinginfraternal-agent:daemon.rs:411calledrun_authenticated_with_plugins, but that function was gated behind theremotefeature. The SSH session-bridge dependency intentionally builds the agent withoutremote; I’m removing the unnecessary gate so authenticated dispatch remains available with an empty/minimal plugin registry.Rebase integration finding: the SSH-enabled server unit build now reaches
socket_tests.rs, wheremesh_calls_and_subscriptions_deliver_scoped_live_inventorybinds_clockat line 555 but callsclock.progress()at line 722.git blameattributes the binding to 0934babde3 and the call toe3d9a942e5on the rebased history. I’m correcting the binding so the paused-time test can compile and exercise the intended progress point.Rebase gate finding: SSH-enabled
cargo clippy --all-targets -- -D warningsfailed inbootstrap_api.rs:179: the smallSocketModeenum is passed by value but not consumed. The enum is matched once in the owned upgrade future, so I’m marking itCopyto express its value semantics and keep the shared upgrade path.Rebase gate finding: CLI clippy reported
remote_link.rs::read_transportat 102 lines, over the 100-line lint threshold after carrying the SSH transport arm. I’m moving the simple SSH prefix/read path into a private helper; the existing TCP mux and QUIC ownership logic remain in the transport dispatcher.Rebase gate finding: CLI clippy exposed stale tests in the rebased tree.
remote_output.rsuses a removedserverfeature name and undefinedbinary/refusalhelpers (the lines trace to 73fca2132f); the SSH action test reads removedActionDefinition.inputfields although the current shared schema isinput_schema(f676296901). I’m aligning the test cfg and executable lookup with the current plugin-server feature/CLI test harness, and asserting the current JSON Schema fields.Committed
e043354f2: SSH admission now uses the rebased optional-IP API, the minimal agent retains authenticated dispatch, and transport compatibility passes clippy. Verification: SSH-enabled server unit suite passed (298 passed, 0 failed, 22 ignored; 277.92 s); clippy passed forinfraternal-server,infraternal-agent, andinfraternalafter the final API fixes.Committed
07b75cda6: fixed the paused mesh socket test's clock binding. The SSH-enabled server suite had already passed this corrected test as part of its full run: 298 passed, 0 failed, 22 ignored (277.92 s).Committed
0bd0074fa: aligned the CLI integration and registry tests withplugin-server,CARGO_BIN_EXE_infraternal, andActionDefinition.input_schema.cargo clippy --locked -p infraternal --all-targets -- -D warningspassed with the corrected tests compiled.Committed
32070b2e0: applied rustfmt to the remote client observed-input fixture.cargo fmt --all --checkandgit diff --checkpass.Fetched
origin/dev; it advanced from the previous rebase base7884d13dto2a5302cc. Starting a second rebase ofaddhost-instant(32070b2e0) onto the fetchedorigin/devso merge readiness is measured against current dev.Current-dev rebase conflict at commit
27d72bd9f:origin/devhas a newer vendor artifact withlink-underline.patch, while this branch addsrenderer-disposal.patch. I’m keeping both source patches in order and rebuilding the checked artifacts against the pinned source so neither dev behavior nor the disposal fix is lost.Rebase complete:
addhost-instantnow has 245 commits on currentorigin/dev2a5302cc76ec; HEAD is23893be36983, and the merge base equalsorigin/dev. The WIP checkpoint is split into atomic commits. Resolved the vendor overlap by retaining link-underlining and disposal patches and rebuilding the pinned artifact: ESM 707.79 kB (199.97 kB gzip); all threeSHA256SUMSentries verify. Browser oracle results:{"refused":false,"partial":false,"documentListeners":0}and{"refused":true,"partial":true,"documentListeners":0}.Finding from the branch-head static gate: x86_64
infraternalis a stripped static-pie with PT_INTERP absent, but measures 21,300,272 bytes against the fixed 21,300,016-byte cap (+256 bytes). This is the known static-build size issue noted in the task brief; the gate is still running its aarch64 build. Base comparison is pending. Receipt: static size output.Branch-head
scripts/verify fastended with exit 1. All non-static Rust gates passed; static builds failed only because both stripped artifacts are 256 bytes over their current fixed caps. The full output is attached: verify-fast.log.Rust gate summary, verbatim:
The x86_64 artifact measured 21,300,272 bytes (cap 21,300,016); aarch64 measured 19,500,632 bytes (cap 19,500,376). Base comparison is pending.
Static size comparison using
scripts/static-cargo build --locked --release --package infraternalonorigin/dev2a5302ccand branch23893be3:Both deltas are below the 16 KB growth limit. I’m ratcheting the static gate caps to 21,090,448 and 19,316,264 bytes respectively.
Committed
f29a721c1:scripts/verify.d/rust.shnow uses exact caps of 21,090,448 bytes x86_64 and 19,316,264 bytes aarch64. Comparison to current origin/dev: −209,568 B x86_64 and −184,112 B aarch64.bash -n scripts/verify.d/rust.shandgit diff --checkpassed; the finalscripts/verify fastwill rebuild and verify both artifacts.Final gate finding: current
verify.d/evidence-size.pyrejects every changed path underevidence/; the branch addsevidence/addhost-fastas 690 files / 5.4 MB, soscripts/verify fastexits at the artifact gate before tests. Two runnable source fixtures in that directory are referenced bydocs/verification/addhost-breakit2/measure.py; I’ll relocate those sources underdocs/verification, preserve the raw receipts in the job cache/issue record, and remove the repository receipt copies to comply with the new gate.Committed
7eb58eb5f: removed the 690 repository receipt files (5.4 MB) while preserving them under/home/kayg/.cache/codex-scratch/addhost-instant/preserved-evidence/addhost-fast; relocated the two executable harness fixtures and updatedmeasure.py. Syntax checks passed, and the artifact gate now reports:repository artifact gate: PASS (221 changed path(s), merge base 2a5302cc76ec).Merge-base size comparison is queued for
b23130139622: all four regular build slots are occupied, and this job is waiting in the required build-slot wrapper. It will start when a regular slot opens.Final-gate integration findings from workspace clippy: (1)
crates/plugins/ssh/tests/mesh_ssh.rs:67initializesOpenSshwithout the current requiredtcp_connectorfield; (2)crates/infraternal-server/src/remote_runtime_tests.rsdeclaresssh_bridgeat lines 389 and 835, producing E0428. I’ll fix both feature/test integration issues and rerun the gate; no runtime check has been waived.Additional workspace clippy errors from rebasing onto current dev:
remote_ssh_tests.rs:295andbootstrap_admission_tests.rs:41,90still call removedRemoteHosts::admitinstead ofadmit_observed;remote_link.rs:573,582do not match the new SSH transport variant. I’ll update those fixtures withSome(ip)for socket transports,Nonefor SSH/no observed IP, and add the SSH no-peer/no-priming arms where appropriate.Further clippy-target findings after the first fixture fixes:
remote_link.rs:514has one remaining test-only match without the SSH case, andcrates/plugins/ssh/tests/install.rsdefinesinstalled_infraternal_agent_always_restarts_without_start_limittwice (lines 2041 and 2762). I’ll add the SSH readiness false arm and reconcile the duplicated regression test after this clippy invocation exits.Finished and pushed to
origin/already-added-open.Head:
cecfc219c870a261226e77264e5fbda6c7e31727(two commits: backend/contract, then TypeScript parser/tests).The duplicate lookup returns the matched live host id only within the existing bootstrap owner's non-revoked-host scope. Cold-CA and foreign-authority tests confirm other failures omit
host_id; existing host-key challenge serialization remains intact. Focused Rust tests passed (server cases 3/3, UUID serialization 1/1);bun test src/canvas/addhost/real.test.tspassed (13 tests, 32 assertions);bun run checkpassed. No UI code changed.Static x86_64 size against merge base
b231301396227821a992692410cf55c50abe04ef: 21,300,016 -> 21,300,272 bytes, +256 bytes (below D243's 16 KB threshold). Both static targets are 256 bytes over the fixed repository caps; this is the known static-build issue from the brief. Receipts: corrected size comparison, full final gate log.scripts/verify fastreturned exit 1. Its Rust gate summary, verbatim:Static size output, verbatim:
Remote-feature test finding:
remote_runtime::tests::ssh_bridge::refused_installed_reattachment_restores_the_ssh_writerreceivedTerminalOutput::Reconnectingbefore the retained SSH writer Ack (reproduced with--features remote; failure atremote_ssh_tests.rs:381). The notice is expected when the refused candidate closes; the test's existing single-deadline frame helper was unused. Wired the assertion through it so reconnect notices are accepted while data/close frames still fail the oracle.Rebase compatibility slice committed:
f3ec80883. Updated SSH/bootstrap fixtures for optional observed IPs and current API, removed duplicate declarations, and made the SSH handover assertion accept the expected Reconnecting notice before Ack. Results: remote SSH bridge 6/6, bootstrap admission 1/1, restart service test 1/1, workspace clippy PASS. The ignored mesh SSH test compiled; its Go-backed fixture could not run becausegois not installed (research/ssh-client/mesh-fixture/runexits 127).Break-it review (independent, Claude) started on branch
review-already-added, base54dc648c3c177791f07c0fb9b4acc868a38e6f43(= origin/already-added-open). Findings to follow; failing tests will be committed on this branch.Break-it review: already-added-open (#189), base
54dc648c3Review branch
review-already-addedhead78195942ea5f76860a05ca53b1e716d51ff55488(tests:109d02b22guards pass;fc5faa8aband78195942efail on purpose).Reviewer: Claude (independent). Branch review-already-added.
Checked and holding (no finding)
authorize()(bootstrap_owner == authority.owner) runs before the duplicate query; host enrollment is owner-only (users.owner=1, single owner per instance by unique index). Testa_non_owner_session_never_learns_the_owners_host_id.h.revoked_at IS NULL. Testa_revoked_host_never_names_its_id.uuid_referenceserializes ashost_idonly whenalready_added; onlywith_existing_host_idsets both;host_key_challenge()hides the slot on already_added so confirmation cannot consume a host id as a challenge.issue_attemptvia?(no.at()/at_ssh_boundary), log_tail empty; nothing diagnostic is lost. Leftovers/failed_step still render if present.Findings
LIMIT 1without ORDER BY, but SQLite's current plan returns the address match. Guard testthe_typed_address_decides_which_host_is_named_when_two_matchpasses. No finding; recommendORDER BY (b.address=?) DESCto make it a contract.P3-1 Shared UUID slot keyed by a public flag (latent)
Failure.already_addedispub;Serializeemitsuuid_referenceashost_idwhenever it is true. Setting it on a host-key failure that holds a challenge sends the challenge ashost_id→ browser offers "Open a terminal" on a non-host UUID. No production path today (onlywith_existing_host_idsets it). Failing test: crates/plugins/ssh/tests/already_added_breakit.rs (a_host_key_challenge_never_serializes_as_a_host_id), commitfc5faa8ab.P3-2 No route to the terminal when host_id is absent
New copy: happened "It's already in Console.", next "", Details hidden. With a server that predates
host_id(web-only deploy) there is no button and no instruction; previously "Open its terminal from Console or the command palette." Failing test: apps/tracer/src/canvas/addhost/already-added-breakit.test.ts, commit78195942e.P3-3 (no test) Button for a host the page does not list
Duplicate query matches any non-revoked host with a bootstrap row;
remote.host.listlists only hosts with a certificate. If the page's inventory lacks the id (refused list → replaceHosts([]), or an uncertified non-revoked host),win.open(uuid)makes a window titled with the raw UUID andnoAccess. Reachability is narrow (startup retires uncertified attempts; the in-memory attempt map answers "being added" first). Suggest offering the button only whensessions.hosts()has the id, else falling back to the palette copy.Verdict
MERGE — P0 0, P1 0, P2 0, P3 3.
Gate (scripts/verify fast at
54dc648c3, verbatim table)idle_browser_probes_do_not_renew_or_poll_session_authority,saturated_native_control_queue_does_not_hide_primary_terminal_output); neither file touched by the change; both pass in isolation (1/1 and 3/3) → load flakes on the shared host.delta: +256 bytesregistry-path issue (x86_64 21300272 vs 21300016; aarch64 19500632 vs 19500376).already-added host offers its terminal…at 1440 and 393: 2 passed.addhost-instant rebased onto origin/dev
393791e09, heada144c073a5(force-pushed with lease to addhost-instant).Gate on this head (verify fast, final run; Rust table verbatim):
The test row failed in that run on one ETXTBSY flake (restart::tests::deleted_path_review::relative_recovery_after_atomic_install_keeps_live_install_path, "Text file busy", host load 25-45 on 12 cores). Earlier verify runs hit other load flakes (uninstall_handoff socket test, component_down_proves_tcp_fallback, transport test binary "never executed", latency proxy RTT p50 350 ms). Each passes in isolation (uninstall_handoff 5/5, component_down 5/5, infraternal bin 5/5). Full
cargo test --workspace --locked --no-fail-fastwith the verify PATH: exit 0 on this head. All other gates PASS.Static sizes (exact caps, delta vs dev): x86_64-musl 21,090,448 B (dev 21,300,016: -209,568); aarch64-musl 19,316,264 B (dev 19,500,376: -184,112). Caps equal measured sizes (+0).
Extra fix this round: compile-out gate failed (crate::output unresolved with no features); fixed in its own commit.
addhost-instant rebased onto origin/dev
7392b925b(docs-only delta from the gated base), head53ef70af6c(force-pushed with lease to addhost-instant).Gate on this head (verify fast, final run; Rust table verbatim):
The test row failed in that run on one ETXTBSY flake (restart::tests::deleted_path_review::relative_recovery_after_atomic_install_keeps_live_install_path, "Text file busy", host load 25-45 on 12 cores). Earlier verify runs hit other load flakes (uninstall_handoff socket test, component_down_proves_tcp_fallback, transport test binary "never executed", latency proxy RTT p50 350 ms). Each passes in isolation (uninstall_handoff 5/5, component_down 5/5, infraternal bin 5/5). Full
cargo test --workspace --locked --no-fail-fastwith the verify PATH: exit 0 on this head. All other gates PASS.Static sizes (exact caps, delta vs dev): x86_64-musl 21,090,448 B (dev 21,300,016: -209,568); aarch64-musl 19,316,264 B (dev 19,500,376: -184,112). Caps equal measured sizes (+0).
Extra fix this round: compile-out gate failed (crate::output unresolved with no features); fixed in its own commit.
Starting release-addhost-instant: branch release-addhost-instant, base release-prod-live
393791e09, merged addhost-instant53ef70af6->7dde647ed(tree identical to addhost-instant). Building artifacts and staging proof next; no deploy.release-addhost-instant: artifacts built (branch head
7dde647ed= release-prod-live393791e09+ addhost-instant53ef70af6; tree identical to addhost-instant). Progress comment; staging handoff proof still running.Static x86_64-musl / aarch64-musl, per the release recipe (pinned Zig via scripts/auth-toolchain, own target dir):
Both sizes equal the caps recorded on this issue (dev 21,300,016 / 19,500,376: -209,568 / -184,112). scripts/package-release produced a byte-identical server binary.
Web: apps/tracer bun install --frozen-lockfile + bun run build + transfer-budget: 265 files, tar (sorted, mtime 0)
3614720B sha256 30a3d54f97a27cfcf5f07016c6bb18b1d2da8af4dc1c97641d5056cc400b8ed6; manifest web-SHA256SUMS sha256 6f86214e6744d1b4509337421a01670606b73932bf03f6f04b93b9b113c3794c.Paths: /home/kayg/.cache/codex-scratch/release-addhost-instant/artifacts/ (bundle/ holds the host payloads and session images from scripts/package-release x86_64).
Add-host end to end (disposable Docker target, real sshd fixture, exact payload bytes above, instant path with the session image; functional only, no timing claim; build host, controller on 172.17.0.1):
release-addhost-instant: finished. Branch release-addhost-instant head
7dde647ed(pushed). No deploy, nothing on monitor01 touched.Signed staging proof (scripts/release/staging-proof, perf VM, exclusive lock, run dir removed afterwards, no deploy units left): PASS, exit 0.
Caveat: the signed proof needs a per-run compiled test authority (feature staging-trust), so the proof server is the same source tree with that feature, not the release bytes below. The release bytes were proven in the add-host runs (payload) and are reproducible (package-release and the recipe build produced the identical server binary).
Add-host (disposable Docker target, real sshd fixture; build host; functional, no timing claim): clean add PASS; Retry (injected install refusal, then Retry) PASS; instant-mode post-prompt Retry covered by late_setup_subscriber_receives_failure_and_retry_completion (PASS); the live harness itself cannot retry after the prompt.
Artifacts ($HOME/.cache/codex-scratch/release-addhost-instant/artifacts/):
Gate: scripts/verify fast was not re-run; the tree equals addhost-instant
53ef70af6, gated above (load-only test timeouts, each green in isolation).Logs: staging-proof, add-host clean, Retry full path, instant-mode harness limit, retry tests.
Deployed to production 2026-10-17 13:23:57 UTC (18:53 IST) as
3048580c1(release-addhost-instant + production web line). MainPID 61562, healthz 200. Server sha256 28e3a819…, host payload x86_64 4033d6e0…, x86_64 session image 93ed80c7… (bundle from scripts/package-release, checksums verified before shipping). Backup /var/backups/infraternal/20261017-152353-3048580c1…. release-prod-live and dev =3048580c1. deploy-prod.sh gained --bundle to place host payloads + session images next to the server.already-added-open head
db376e626: test-only fix of the stale Details assertion (see commit). bun test src/canvas/addhost: 56 pass. addhost-wizard.spec.ts: 39 passed. Gate (scripts/verify fast), final table verbatim:clippy: 66 x clippy::result_large_err in infraternal-plugin-ssh (Failure grew in
3bea93b64, #189). test: 1 failed, restart_open_liveness_review::responsive_socket_survives_a_held_terminal_open_with_an_existing_attachment (load ~25 during run). static: size 21311216 vs cap 21300016 (+11200, known registry-path issue).already-added-open head
de1928b27: Failure uuids boxed (clippy clean, wire format pinned by test), size caps ratcheted (x86 21296496, aarch64 19500760; merge-base caps 21300016/19500376; release-prod-live measures 21090448/19316264, gap is stale-base drift). verify fast:test FAIL was the flaky review_full_input_writer_preserves_browser_control_progress (passes alone; the earlier flake was responsive_socket_survives_a_held_terminal_open). Remaining workspace tests with --no-fail-fast skipping that module: 0 failed.
Follow-up on
de1928b27: with --no-fail-fast the remaining workspace tests had 2 more failures. saturated_native_control_queue_does_not_hide_primary_terminal_output failed once under load and passes 3/3 when rerun alone (flake). faithful_snapshot_preserves_public_mouse_encoder_reports (infraternal-shim) fails identically at the merge baseb2313013, so it is environmental and not from this branch.release-2 started (production line + #234 D300 ping/FIFO fix + #189 already-added "Open a terminal").
Branch
release-2, base origin/release-prod-lived5a4188d3(Retry + faster add-host + #241).Merging origin/ping-fifo-fix
8f3945801with review tests origin/review-ping-fifo-fix-r2fda0c95f7(agent flake fixed withreader.shutdown(Shutdown::Both)), then origin/already-added-opende1928b27. Then caps, tests, gate, artifacts, staging proof on the perf VM, Docker sshd add-host. No deploy.release-2 finished. Head
103815e70(pushed torelease-2), base origin/release-prod-lived5a4188d3. No deploy; monitor01 untouched.Merged (each resolution explained in its merge commit)
5317c3b20: #234 D300 ping-fifo-fix8f3945801+ r2 review testsfda0c95f7. The flaky agent test now callsreader.shutdown(Shutdown::Both)before the drop. The review's two fork-race tests are now regression tests and pass (a_shut_down_peer_refuses_the_write_while_a_forked_copy_holds_it,the_unhealthy_handoff_case_survives_concurrent_spawns; 3/3 with the original).c4a814da7: link-doubt-canvase6387508e(P3-4, web only). link-doubt-canvas.test.ts passes.4d70ec609: already-added-opende1928b27(#189). Conflicts were only in the web app (AddHostSink keeps prepareTerminal, enrolled(hosts,status), setupInPill, openHost/hasHost; mock sink; spec watchdog + activeWindow). Rust auto-merged: the duplicate lookup runs before register(), so Retry's lock/refresh and the instant path are untouched.87676178a: remove-uninstall27a0a1e65(#235). Checked against Retry: runtime admission takes the shared parent flock only while admitting a start, so uninstall's bounded exclusive setup_guard cannot deadlock a running infraternal agent.6c3c5f02f: authority-read-retrya9db5526b(#240). websocket.rs merged cleanly (different functions from D300). The r2 failing P3 test review_authority_cap.rs was left out.d0422100d+103815e70: live harnessBREAKIT2_READD=1re-adds the host, checks already_added + host_id, then opens a terminal on that id.7cdbfc9fd: caps.Size (same target dir; release-prod-live reproduces its caps exactly)
d5a4188d3Total -704 / +6,288 B (≤ +16 KiB).
Artifacts (
/home/kayg/.cache/codex-scratch/release-2/artifacts, built at7cdbfc9fd; the later commit is test-only)3614720Staging
103815e70and rerun above.Tests
INFRATERNAL_TEST_STATE_ROOT=$HOME/.cache/r2and the static musl helper. With$HOME/.cache/r2rel,review_live_session_socket_in_newline_home_blocks_cleanupfails, because the socket path exceeds sun_path. That is the environment, not the code; the one ssh-plugin failure was the same test.cargo test -p infraternal-plugin-remotealone does not compile (tokio rt-multi-thread is only enabled through workspace feature unification). This is pre-existing; the workspace gate runs it.Gate (
scripts/verify fastat103815e70, RUST_TEST_THREADS=2, INFRATERNAL_TEST_STATE_ROOT=$HOME/.cache/r2, start load 3.9, 73 GB free; exit 0)x86_64 21080144 cap 21080144 +0; aarch64 19322648 cap 19322648 +0. TS protocol 37/0.
Evidence: instant-readd.log full-readd.log full-clean.log full-retry.log remove-rootless.log remove-rootless-readd.log remove-root.log remove-root-readd.log remove-root-mesh.log remove-root-mesh-readd.log staging-proof.log authority-stall-soak.log MANIFEST.txt addhost-wizard-playwright.log suites.txt verify-103815e70.log
Release 2 deployed to production at 14:08 IST: release-prod-live
103815e70, server sha256 8608623e…, MainPID 86026, /healthz ok. Backup /var/backups/infraternal/20261018-083808-103815e70…. CA is locked after the restart until the owner unlocks it.Closed by the orchestrator during the 2026-10-28 handover (#254), on the owner's request to close finished issues. Evidence: Add-host speed work (addhost-instant
53ef70af6) is in dev and prod; deployed to production as3048580c1and release 2 (103815e70). Reopen if anything is still missing.kayg referenced this issue2026-10-28 17:53:55 +00:00
Rebuilt on current dev for the owner's review. Branch
mockups-2026-10-31, head586cd2b13(base dev42ab33035, 2026-10-31). Test double only; nothing is real or saved. Not merged, not deployed. Old previews are gone; this replaces them.Preview (NetBird, bound to infraternal-dev-02-std; index of every mockup: http://10.69.69.117:47650/mockups.html):
How: The real wizard runs; an autopilot picks a host. Add
&slow=3to stretch timings,&fail=install|join|start|readyfor a failure (press the pill for the popover),&autopilot=0to drive by hand.Note: Already on dev as the
addhost-demo=pilldemo; nothing was merged for it. Captures are taken at slowed timings.Captures are Chromium on the production build of the mockup, taken by
apps/tracer/test/capture-mockups.ts; I have read them for wrong states, not judged them as final.Desktop 1440×900 @2x, ink
Desktop 1440×900 @2x, paper
Phone 393×852 @3x, ink
Phone 393×852 @3x, paper (first state only)
Update to the comment above: the branch head is now
57871651f(the preview at http://10.69.69.117:47650/mockups.html runs it). The captures above were taken one build earlier; the only difference is that the window terminals there hide their cursor until attached, a production change that was since removed from the branch. Gate output, what is left out, and the budget finding: #272.J56 started (Claude Sonnet 5.5). Branch
ui-approved-189-212-216, base origin/devf4b61cd6c. Building the owner-approved mockups (mockups-2026-10-31) into the product. Owner 2026-11-01: "add-host pill is approved. connection states is approved. version pill is mostly approved." plus the value-tap changes (in-place clipboard + Copied, contextual default action, one Updates key).Finding (#189): the add-host pill is already in dev and in the product:
setup-pill.on-demand.ts,setup-ring.ts,setup-attach.tsand the real wiring incanvas/product.ts(attachSetup, fed by the server setup stream). The mockup branch only adds theaddhost-demo=pillscript (test double). So "as approved" means: prove the real product path behaves like the approved mockup. I will run the real-server browser spec for it and attach screenshots, and change code only where the product differs from the approved mockup.Progress (J56). History is now atomic and pushed (
ui-approved-189-212-216head0176caf22): sidebar-card refactor, in-place copy + contextual press, About list on demand (keeps the critical transfer budget green: critical gzip 305441/307200, brotli 254299/256000), connection icon in the window pill (#212), version pill + panel (#216), TASTE rows. Browser specs on dev-02 (arm64): connection-pill 12/12, version-pill 13/13 after fixes, add-host pill specs 4/4 + stream 9/9 (the unmodified-dev baseline has 4 failures in canvas.spec/sessions-canvas on this VM that I reproduced there: touch grip 43.99px, footer Settings, Linux key glyphs, session badge). Final batch + captures running;scripts/verify fastis queued on compute01.J56: add-host pill, verified in the product (branch head
bd2671f8a). The pill is already in dev and wired in the product (canvas/product.ts→attachSetup, fed by the server's setup stream); the mockup branch only added theaddhost-demo=pillscript, so no pill code changed in this job. Evidence on the production build:addhost-pill.spec.ts4/4 andaddhost-pill-stream.spec.ts9/9 (the real wizard against the setup-stream double: ring fills, tick, system icon, failure popover, retry, reduced motion), and the captures below (addhost-pill-captures.spec.ts: 1440×900 @2x, 393×852 @3x). The window pill's status dot is now the connection icon (#212), and these specs pass with it.J56 finished. Branch
ui-approved-189-212-216, headbd2671f8abb5fa32ae5279e440cb309d8214bc96, base origin/devf4b61cd6c. Pushed, not merged.Commits:
9845b0ec7sidebar-card refactor ·9652ae0cein-place copy + contextualpress·5c15bd0c0About list on demand (D188 budget) ·1e9369200connection icon in the window pill (#212) ·22032d9baversion pill and panel (#216) ·bd2671f8aTASTE rows.Gate.
scripts/verify faston compute01 (x86), copy of this head: Rust half failed for lack of disk, not for code:No space left on device (os error 28)attarget/debug/deps(the VM was 100% full, 2.5 GB free, from other jobs; my target was 8.6 GB, now deleted). Table, verbatim:No Rust file changed in this branch. A re-run needs ≥25 GB free there (build-slot waits for it). The TypeScript half of
verify fast(its exact commands fromscripts/verify.d/ts.sh) was run on dev-02 instead:bun run checkclean; unit tests 394 + 105 + 17 pass, 0 fail; protocol tsc clean; transfer budgetcritical gzip 305448/307200 brotli 254318/256000, deferred 58585/61440, onDemand not limited (largest chunk 15456/20480); logTS_GATE_OK.Browser specs (dev-02, arm64, production build). Final batch: 107 passed, 5 failed: 4 fail identically on unmodified
f4b61cd6con the same VM (canvas.spec touch grip 43.99999 px < 44; canvas.spec sidebar footer Settings; canvas.spec Linux key glyphs; sessions-canvas badge); 1 version-pill timing flake (6/6 on repeat, assertion timeout raised). Add-host pill specs 4/4 alone (a run next to other heavy jobs hit 2 s long tasks; same spec passes on dev when alone).Not built / for kayg: (1) typed-text echo under a lost link (secret risk); (2) licence-key entry and release notes (no backend); (3) the product's update check is stubbed behind
TODO(#211)inapps/tracer/src/canvas/version/source.ts, so the row ends in "Couldn't check — try again" until #211 exists; (4) the pill number is the build's short commit id (the web app has no version number).exitreports code 127 and the window does not close #393