#1168 — Windows native co-presence health transport / Dashboard readiness
Date: 2026-10-14 Asia/Shanghai
Scope: runtime source and isolated Windows fake-Codex gate; no live node changes.

Failure retained: PR2711 old head73fb5628f, QA38443138147/job115454675894,
artifact11836996773. Dashboard POST409 node_degraded/tui session-missing,
health_observed_ms_ago93; the new inbox/steering path was not reached.
Final PR2711 headb868 later passed the unchanged suite. Timing-dependent
passes do not establish correct health reporting or original-device acceptance.

Causal source finding: cli configured tuiSession=alias for EVERY co-presence
bridge, including Windows, so codex-health used probeTmuxTui on a platform
whose launcher uses native processes. Suppressing the first onChange callback
does not suppress snapshot(): a later reportStatus still sends session-missing.
This explains the observed false tui layer; exact ordering of the old report
cannot be reconstructed from the retained log. It is not just a delay to retry.

Change: select tmux TUI health only for non-Windows co-presence. Windows
ongoing tui health stays absent/unknown, never fabricated healthy. App-server
failure and expired/revoked model authentication remain reported and refused.
Linux/macOS session selection and the Hub health guard are unchanged.
Native Windows fixture requires launcher PID/birth/socket attribution before
reading fresh exact-network/alias Hub app-server health. It rejects ANY tmux
TUI layer on Windows, then performs the original normal/high POST, same-turn
steer, no-new-turn, receipt, shared-HOME and single-bridge assertions. No force.
Isolated fixture health/status intervals1s expose the snapshot before POST;
production defaults are unchanged. PowerShell now explicitly propagates the
health-regression/native-E2E nonzero exit statuses.

Docker local: Bun1.3.14, frozen dependencies from test1163 independent image,
current source read-only mounted; codex-health33PASS83 assertions, ~86ms.
Docker runtime+real Hub HTTP health dispatch50PASS156 assertions,1.284s.
First combined local invocation omitted the required throwaway COMMHUB_DB;
the production-DB guard correctly refused before Hub assertions. Corrected
isolated invocation used /tmp/health1168.db. This was fixture setup, not PASS.
Docker exact-pair6PASS23 (~54ms); old Windows-selection mutation32PASS/1FAIL,
the named Windows bridge regression assertion fails (~83ms). An initial private
mutation shell command had a quoting error before assertions; corrected isolated
script reached the designated assertion. No product/main mutation was made.

SOURCE/core completed: PR2715 final head
4e2cfcfc09909f26d2a73c6c00da41ff6002528b, tested CI merge
6aa47248804bb48e7924609fb220ef3a2457aacc, main
d52418e614301143771f9ad40f7c40afd3777b1c (merged19:51:54Z).
All three trees5e0e706b07e389c4950b8250bda6ce242403ecd6 match.
Exact final head10required+7selected SUCCESS: Windows-native, Node typecheck,
exact runtime pair, doc-symbol pins, docs build, test-file coverage, mutation
integrity. Broad QA is separate:19/138 jobs completed without FAIL at snapshot;
only remaining merged-PR broad run38457313634 was cancelled after preserving
its status. Main broad QA continues; this is not a full-matrix PASS claim.

Windows job115503898763 actually SUCCESS; artifact report-test751-windows
downloaded and source_commit is the exact CI merge above. Native portable
suite25PASS68, health33PASS88; native E2E PASS Hub current app_server=ok,
startup TUI PID attributed, ongoing tui=unknown, model_auth=unknown. Then
normal/high Dashboard tasks steer the same human turn, receive individual
terminal receipts, preserve remote/thread/HOME, create no second turn and
leave one bridge. This is native Windows with fake Codex, not an actual model.

Formal Node144 run38458335261 and CLI178 run38458339458 dispatched against
exact merged main d52418e6, checked still main ancestor. Both sets of four formal
gates PASSED; full npm package verification is tracked in report-test1168-public.txt. Both
same four-gated archives are public at the official GitHub prerelease; Node144
and CLI178 npm are fully verified; the CLI workflow initial registry-timeout
FAIL remains recorded, followed by a successful complete registry re-download.
Hub139 is reused and not republished. Original-device adoption remains unknown.
Local tests capped at2minutes; broad/native coverage runs on GitHub Actions.

Recovery: no launch topology/ports/proxy/config/credentials/schema changes.
Use repository deploy/ and existing node installation/runbooks. This is runtime
source, not an adopted live version; software release requires exact merged
main, paired formal gates and immutable artifacts. Roll back to prior main or
previous paired version via the same runbook; Windows false-tmux gap returns.
Git rebuilds software only. Original login/private task/receipt state comes from
the existing private backups/credential authority, never copied from this gate.
No restart, fork, reset, new authority, model or credential replacement occurred.
Native ongoing TUI liveness beyond startup attribution remains a separate gap.
