Skip to content

Commit 6a0f5b5

Browse files
authored
fix(web-search): bound sidecar search deadline to 60s + graceful degradation (lidge-jun#398) (lidge-jun#416)
* docs(devlog): 260724 issue/PR cleanup roadmap (WP1) — 12 work-phases, docs-first WP1 of goalplan opencodex-open-issue-pr-cleanup-loop-hotl-12-pab. 000_plan + decade docs 010..110 from 5 Sol investigations; A-gate (Hilbert) GO-WITH-FIXES blockers=3 folded (WP2 error:string+sentinel, WP8 enriched clone, WP7 lidge-jun#394 conflict source corrected). * fix(web-search): bound sidecar search deadline to 60s + defensive degradation (lidge-jun#398) The hosted web-search sidecar deadline defaulted to 200s, so a limit-exhausted/unavailable search backend ran the full 200s and the client cancelled first (turn 499) or the forced-answer routed iteration failed (502). Lower DEFAULT_TIMEOUT_MS 200000->60000 (p90 ~43s) so the turn degrades to a no-result answer within ~1 min; cfg.timeoutMs still overrides. Add a defensive try/catch around the sidecar dispatch that degrades an unexpected throw to a failed tool result while re-raising a genuine parent abort as LoopError(499). Redact the anthropic executor's echoed HTTP body. Docs synced (en + zh-cn + structure/04). Regression tests: default 60s deadline + thrown-sidecar degradation with secret non-leak. * docs(devlog): WP2 lidge-jun#398 A-gate folds (60s deadline, post-await abort, docs sync, anthropic redact) * fix(catalog): advertise image input for noVisionModels custom rows (lidge-jun#349/lidge-jun#344) customModels bypassed the vision-sidecar modality augmentation, so a noVisionModels-tagged custom override was re-advertised text-only and the Codex app blocked images before the sidecar could run. Apply a narrow image-only augmentation to custom rows using the registry-enriched provider clone's noVisionModels — deliberately NOT the full applyProviderConfigHints pass, so an explicit custom contextWindow / inputModalities / reasoning override is preserved verbatim. Regressions: registry-enriched image add, non-sidecar row untouched, and explicit custom fields preserved. * fix(catalog): log provider discovery failures once per signature (lidge-jun#395) Anthropic-adapter providers whose baseUrl lacks /v1/models (e.g. Azure AI Foundry) return HTTP 404 forever; the 30s discovery cooldown re-probed and re-logged the same warning every ~30s, flooding the log. Add shouldLogDiscoveryFailure() (log-once-per-signature): a failure warns only when its signature (reason + httpStatus) differs from the last observed discovery status. The 30s re-probe cooldown and the observable discovery status (/api/providers) are unchanged; a first occurrence, a changed status, a recovery, or a cache clear all log again. Gates all five failure-branch warnings. Regression: tests/provider-discovery-log-suppression.test.ts.
1 parent eb01dd7 commit 6a0f5b5

24 files changed

Lines changed: 983 additions & 37 deletions
Lines changed: 51 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,51 @@
1+
# 260724 Issue/PR Cleanup Loop — 000 Plan (HOTL, 12 work-phases)
2+
3+
Goal: `opencodex-open-issue-pr-cleanup-loop-hotl-12-pab`.
4+
Branch: `codex/260724-issue-pr-triage` off `origin/dev` (tip cc7bb577).
5+
Sol medium subagents: unlimited (A-gate reviews + B-gate help).
6+
Merge authority: maintainer pre-approved the specific mutations named per WP.
7+
Never main/preview. Every merge requires CI-green + mergeable head on dev.
8+
9+
## Dependency-ordered work-phase map
10+
11+
Ordering rationale: the two files with cross-PR overlap are
12+
`src/server/responses/core.ts` (touched by #390 and #394) — so #390 merges
13+
before #394 and #394 rebases after. Investigations (no repo mutation) are
14+
cheap and unblock their code cycles, so each investigation cycle precedes its
15+
optional fix. GUI merge (#393/#340) is last because CI on that head is still
16+
finishing.
17+
18+
| WP | Unit | Type | Decade doc | Depends on |
19+
|----|------|------|-----------|------------|
20+
| WP1 | Roadmap (this) | docs | — | — |
21+
| WP2 | #398 sidecar graceful degradation + 2 follow-up issues | code + issues | 010 | — |
22+
| WP3 | #320 native-auth-expiry investigation | investigate + comment | 020 | — |
23+
| WP4 | #390 → #382 merge | merge | 030 | — |
24+
| WP5 | #389 model-switch visibility + ocx sync | investigate → merge/report | 040 | — |
25+
| WP6 | #397 openai-chat system-first merge | merge | 050 | — |
26+
| WP7 | #394 rebase on dev + merge | rebase + merge | 060 | WP4 (core.ts) |
27+
| WP8 | vision sidecar #349/#344 investigate + fix | investigate + maybe code | 070 | — |
28+
| WP9 | #338 slow-calls close + re-test request | comment + close | 080 | — |
29+
| WP10 | #396 close (Claude Desktop unofficial) | comment + close | 090 | — |
30+
| WP11 | #395 404 log-flood investigate + comment | investigate + comment | 100 | — |
31+
| WP12 | #340 verify + merge #393 | verify + merge | 110 | — |
32+
33+
## Scope
34+
35+
IN: `src/`, `gui/` (GUI only for #393/#340), `devlog/`, and GitHub issue/PR
36+
state for #398/#320/#390/#382/#389/#397/#394/#349/#344/#338/#396/#395/#340/#393.
37+
OUT: main/preview; provider-add PRs (#403/#385/#177/#178/#201); cursor items
38+
(#399/#402/#373/#376); upstream-tracking (#241/#92); other roadmap/features.
39+
40+
## Terminal outcomes per WP
41+
42+
DONE = named artifact exists (merged SHA / posted comment URL / created issue
43+
numbers / landed code+tests). NOOP if already handled. BLOCKED if a merge
44+
needs author-side conflict resolution. NEEDS_HUMAN if a mutation exceeds
45+
pre-approved scope. UNSAFE for security-boundary changes (auth/credential/
46+
release workflow) without explicit review.
47+
48+
## Verification
49+
50+
`gh pr/issue` state, `git log` SHAs, `bun run typecheck` + focused `bun test`
51+
output captured as checkOutput for any code cycle.
Lines changed: 160 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,160 @@
1+
# WP2 — #398 sidecar graceful degradation + 2 follow-up issues (010)
2+
3+
Issue #398: web-search/vision sidecar backend is fixed to openai/anthropic; when
4+
that backend's account limit is exhausted, the turn hangs ~200s then fails whole
5+
with 499/502; no graceful degradation to "continue this turn without search/vision".
6+
7+
## Findings (Sol explorer, read-only)
8+
9+
The branch ALREADY degrades ordinary sidecar HTTP/connect/timeout failures into
10+
error markers — executors are "never throws" (`src/web-search/executor.ts:84-100`,
11+
`anthropic-executor.ts:168-182`); `runSearchCall()` stores the failed outcome and
12+
forces the routed model to answer without search (`loop.ts:430-457`,
13+
`format-result.ts:30`). Vision similarly strips images with a marker
14+
(`vision/index.ts:197-204,334-340`).
15+
16+
**The real defect is latency, not missing degradation.** The default sidecar
17+
deadline is 200_000ms (`src/web-search/index.ts:134,145-157`). A hung sidecar
18+
consumes up to 200s; the client cancels first → next routed op maps abort to
19+
`LoopError(499)` (`loop.ts:320,341,377`), or the forced-answer iteration fails
20+
as 502 (`loop.ts:343,379-389`, `lib/errors.ts:240`). So 499/502 is a symptom of
21+
the 200s wait, not a direct sidecar emission.
22+
23+
Also: HTTP status/error-code is flattened into an untyped `error: string`, losing
24+
429 (exhausted) / 401 / 403 / 5xx and Anthropic `max_uses_exceeded`
25+
(`anthropic-executor.ts:49-55`), so we can't degrade *immediately* on a known
26+
exhaustion status.
27+
28+
## Plan (scope boundary IN)
29+
30+
No new `degradeOnSidecarFailure` flag (default-on is already the contract).
31+
32+
MODIFY:
33+
- `src/web-search/executor.ts` — KEEP the existing sanitized `error: string`
34+
(consumers depend on it) AND add optional structured metadata
35+
`failure?: { kind: "http"|"timeout"|"connect"; status?; code?; message }`.
36+
- `src/web-search/anthropic-executor.ts` — preserve HTTP status +
37+
`web_search_tool_result_error.error_code`; raw response bodies enter outcomes
38+
at `anthropic-executor.ts:168-172` — SANITIZE before storing (never persist raw
39+
body/token into the outcome/tool-result/SSE/log).
40+
- `src/web-search/loop.ts` — defensive try/catch at `runSearchCall()` dispatch
41+
(`~427-432`); on 429/401/403/5xx degrade immediately; preserve genuine parent
42+
abort as 499.
43+
- `src/web-search/index.ts` — lower default sidecar degradation deadline from
44+
200_000ms to a bounded 30_000–45_000ms; keep `webSearchSidecar.timeoutMs`
45+
override; recompute `stallTimeoutSec` (already at `:152-157`).
46+
- `src/vision/describe.ts` + `src/vision/anthropic-describe.ts` — structured
47+
failure metadata; keep marker degradation in `vision/index.ts`.
48+
49+
Activation scenario (C-ACTIVATION-GROUNDING-01): a test where all sidecar queries
50+
return 429/502/connect/timeout must produce exactly one failed search cell, a
51+
routed-model answer, `response.completed`, and NO `response.failed`; and an
52+
internal sidecar timeout must degrade while a genuine parent abort stays 499.
53+
54+
TESTS:
55+
- `tests/web-search.test.ts` — full-turn all-failed/timeout completion regression.
56+
- `tests/sidecar-abort.test.ts` — sidecar deadline degrades vs parent abort = 499.
57+
- `tests/vision-cache.test.ts` — timeout/429 marker, not cached.
58+
- `tests/web-search-anthropic.test.ts` / `tests/vision-anthropic.test.ts` —
59+
preserve error_code/type in structured outcomes.
60+
61+
## Consumer compatibility (A-gate blocker #1 fold)
62+
63+
The `error?: string` field is REQUIRED by existing consumers — do NOT replace it:
64+
- `src/web-search/executor.ts:32-40` (outcome shape)
65+
- `src/web-search/format-result.ts:27-30` (renders failed result)
66+
- `src/vision/index.ts:335-342` (catch → marker)
67+
Add `failure?` alongside; every consumer keeps reading `error`. New structured
68+
field is additive/optional.
69+
70+
## Secret-safety (A-gate blocker #1 fold, STRICT)
71+
72+
"reuse existing redaction" is NOT automatic on the Anthropic paths
73+
(`anthropic-executor.ts:168-172`, `vision/anthropic-describe.ts:166-170` put raw
74+
bodies into outcomes). Requirement: structured `failure.message` and `error`
75+
MUST be sanitized (status/code/short reason only, no raw body, no headers, no
76+
token). Add a sentinel-secret test: inject a fake token/secret into a mocked
77+
error body and assert it never appears in the outcome, tool result, SSE frames,
78+
or logs. `bun run privacy:scan` must stay green.
79+
80+
## Two follow-up issues to create (per user directive)
81+
82+
1. **feat(sidecar): add exa and other search providers as sidecar backends** —
83+
today backend ∈ {openai, anthropic} only; add exa (and pluggable search
84+
providers) so non-OpenAI/Anthropic accounts can back the search sidecar.
85+
2. **feat(sidecar): investigate/enable search-api-capable providers (e.g. gemini)
86+
as sidecar backends** — evaluate Gemini (and similar) native search APIs as
87+
selectable sidecar backends.
88+
89+
Then comment on #398 linking both follow-ups and summarizing the degradation fix.
90+
91+
## Security boundary
92+
93+
Touches sidecar auth-adjacent paths but no credential logging/serialization is
94+
added. Structured failure must NOT include tokens/bodies with secrets (reuse
95+
existing redaction). Not a release-workflow change.
96+
97+
Terminal: DONE = degradation code + tests green + 2 issues created + #398 comment.
98+
If B proves too large for one cycle, land the deadline-lowering + defensive catch
99+
first (the direct 499/502 fix) and split structured-metadata into a follow-up.
100+
101+
## P stale-check (WP2 cycle, current tree)
102+
103+
Re-read against current `src/web-search/*`:
104+
- `executor.ts:84-100` and `anthropic-executor.ts` ALREADY degrade HTTP non-2xx
105+
and timeout/connect into `{ error }` and ALREADY `redactSecretString(...)` the
106+
body (`executor.ts:88`). So HTTP 429/401/403/5xx already degrade FAST — they are
107+
not the hang.
108+
- The ONLY 200s source is the sidecar search deadline default:
109+
`index.ts:19 DEFAULT_TIMEOUT_MS = 200_000` → `index.ts:145 timeoutMs =
110+
cfg.timeoutMs ?? DEFAULT_TIMEOUT_MS` → `settings.timeoutMs` →
111+
`executor.ts:73 signalWithTimeout(settings.timeoutMs)`. Vision already uses
112+
45_000 (`vision/index.ts:16`).
113+
- `#398`'s observed `201,434ms` ≈ 200_000ms confirms the deadline is the cause.
114+
115+
### Scoped B for THIS cycle (direct 499/502 fix)
116+
117+
1. MODIFY `src/web-search/index.ts` — lower `DEFAULT_TIMEOUT_MS` 200_000 → 45_000
118+
(sidecar SEARCH deadline only; do NOT touch
119+
`DEFAULT_ROUTED_MODEL_STALL_TIMEOUT_MS`, which is the routed-model body
120+
inactivity budget, a different knob). `cfg.timeoutMs` override still honored;
121+
`webSearchStallTimeoutSec` recomputes from the smaller value automatically.
122+
2. MODIFY `src/web-search/loop.ts` — defensive try/catch around the
123+
`runWebSearch`/`runAnthropicWebSearch` dispatch (~:430) so any future executor
124+
throw becomes a failed `SidecarOutcome` (executors are "never throws" today,
125+
but this makes the contract enforced, not assumed). Re-raise genuine parent
126+
aborts (signal.aborted) so real client cancel stays 499.
127+
3. TEST `tests/web-search.test.ts` — assert the default sidecar deadline is now
128+
bounded (45s) and a thrown executor degrades to a failed cell + completed turn.
129+
130+
### Deferred to follow-up (NOT this commit)
131+
132+
The structured `failure{ kind,status,code }` metadata refactor (A-gate blocker
133+
#1) is DEFERRED — since executors already redact+degrade, it is an enhancement,
134+
not required for the 499/502 fix. Deferring it also removes the secret-safety
135+
risk from this commit (no new raw-body handling is added). It becomes a THIRD
136+
follow-up note in the #398 comment. This matches the "if B too large" clause above.
137+
## A-gate (Halley) folds — GO-WITH-FIXES blockers=3
138+
139+
1. **Deadline 60_000ms, not 45_000.** Per-search deadline; maxSearches=3
140+
sequential; hosted-search p90 ≈ 43s (commit 11cc822e). 45s leaves no tail
141+
margin → use `DEFAULT_TIMEOUT_MS = 60_000`. `cfg.timeoutMs` override intact;
142+
stall watchdog still ≥230s (index.ts:52, stall-timeout.ts:8).
143+
2. **Abort correctness.** Executors already catch parent abort and RETURN
144+
`{error}` (executor.ts:96, anthropic-executor.ts:179) — so checking
145+
`signal.aborted` only in a catch misses the fulfilled `{error}`. Check
146+
`signal.aborted` immediately AFTER the dispatch await (and in catch), and
147+
throw `LoopError(499, "client closed request during web-search")` to match
148+
loop.ts:320/341/377. Add a parent-abort-during-sidecar regression.
149+
3. **Docs sync (SOT-SYNC-01).** 200s default is promised in
150+
`structure/04_transports-and-sidecars.md:129`,
151+
`docs-site/.../reference/configuration.md:380` (+ `:450-451` example) and
152+
translations (`zh-cn/.../configuration.md`). Update the `timeoutMs` default to
153+
60000 there or users override the fix back to 200s.
154+
4. **Bonus security fold (Halley note):** `anthropic-executor.ts:168` places an
155+
UNSANITIZED HTTP body into `error`. Wrap it with `redactSecretString(...)` in
156+
THIS commit (small, real secret-leak) — matches `executor.ts:88`.
157+
158+
Final scoped B: `DEFAULT_TIMEOUT_MS 200_000 → 60_000`; loop.ts defensive
159+
catch + post-await abort→LoopError(499); anthropic-executor body redaction;
160+
docs sync (en + zh-cn + structure/04); regression tests.
Lines changed: 61 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,61 @@
1+
# WP3 — #320 native-auth-expiry-with-healthy-pool investigation (020)
2+
3+
Issue #320: with a healthy OpenAI/Codex pool and valid OAuth accounts, an
4+
expired local Codex CLI native `auth.json` still opens a ChatGPT/Codex login
5+
window when running `codex`, before the request reaches the OpenCodex proxy.
6+
Secondary: after a Codex CLI npm update replaced the binary, the autostart shim
7+
disappeared until `ocx codex-shim install` was re-run.
8+
9+
## Findings (Sol explorer, read-only; cite file:line)
10+
11+
Root cause = **upstream Codex behavior**, not a proxy bug:
12+
- `codexAccountMode="pool"` is evaluated only AFTER a request reaches
13+
`/v1/responses` (`src/server/responses/core.ts:819`,
14+
`src/codex/auth-context.ts:100`). At that point OpenCodex picks + refreshes
15+
its managed credential and replaces the incoming native bearer
16+
(`auth-context.ts:142,181`, `account-store.ts:319`).
17+
- The login window happens EARLIER, inside the upstream Codex executable.
18+
Loopback injection deliberately keeps Codex's built-in `openai` provider and
19+
only rewrites `openai_base_url` (`src/codex/inject.ts:41,105`); dedicated
20+
mode sets `requires_openai_auth=true` (`inject.ts:87`). So Codex validates its
21+
native `$CODEX_HOME/auth.json` before sending any HTTP.
22+
- OpenCodex's own expiry check (`src/codex/main-account.ts:40`) only excludes
23+
the native main account from proxy-side rotation; it does not open the browser.
24+
- Pool can serve without a usable native account once a request arrives; tests
25+
isolate an absent native auth.json while exercising pool creds
26+
(`tests/codex-auth-context.test.ts:39,99`).
27+
28+
Shim-disappears-after-npm-update:
29+
- Install renames npm launcher to `*.opencodex-real`, writes wrapper at the
30+
original path, records both in `codex-shim.json` (`src/codex/shim.ts:1025`).
31+
npm update rewrites the launcher at that path, replacing the wrapper.
32+
- Already self-repairs: every ordinary `ocx` command auto-restores before
33+
dispatch (`src/cli/index.ts:64`; restore logic `shim.ts:1043,1092`;
34+
`ocx status` eligible `src/cli/codex-shim-autorestore.ts:18`; test
35+
`tests/codex-shim-autorestore.test.ts:127`). A direct first `codex` call after
36+
replacement can't self-trigger (wrapper no longer runs); next `ocx` repairs it.
37+
`ocx service install` avoids launcher dependency for proxy startup.
38+
39+
## Decision
40+
41+
No minimal safe in-repo runtime fix for the native-login gate (suppressing it
42+
would change Codex App/TUI account-gated semantics). Shim issue already fixed on
43+
this branch. → **WP3 = investigate + English root-cause comment** (documentation
44+
guidance), optionally a docs note that Pool creds are proxy-side and do not
45+
satisfy Codex's pre-request native-login prerequisite.
46+
47+
## Comment to post (English)
48+
49+
> Thanks — this is two separate layers. In Pool mode, OpenCodex selects and
50+
> refreshes its managed account only after Codex sends a request to
51+
> `/v1/responses`; Codex CLI validates its own `auth.json` earlier, so an
52+
> expired native login can open the login window before the proxy is contacted.
53+
> That startup gate is upstream Codex's auth-required OpenAI provider behavior
54+
> and can't be safely suppressed by the proxy without changing provider/UI
55+
> semantics. The npm-update symptom is a launcher replacement; current dev code
56+
> detects a previously tracked shim replacement and restores it on the next
57+
> ordinary `ocx` command (manual fallback: `ocx codex-shim install`). For
58+
> restart-safe operation independent of the launcher, use `ocx service install`.
59+
> We'll document that Pool mode does not remove Codex's native-login prerequisite.
60+
61+
Terminal: DONE = comment posted on #320.
Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,30 @@
1+
# WP4 — #390 → #382 merge (030)
2+
3+
PR #390 `fix(codex): clear stale weekly quota on WHAM/header refresh` closes
4+
issue #382 (30-day primary-only Codex Auth account showing stale weekly +
5+
30-day simultaneously).
6+
7+
State at plan: draft=true, MERGEABLE/CLEAN, all CI green (macos/ubuntu/windows +
8+
npm-global matrix + react-doctor pass; CodeRabbit skipped as draft).
9+
10+
Files: `src/codex/auth-api.ts` (+12/-25), `src/codex/quota.ts` (+120/-0 new),
11+
`src/server/responses/core.ts` (+7/-19), `tests/codex-auth-api.test.ts` (+39),
12+
`tests/rate-limit-reset-credits.test.ts` (+141).
13+
14+
Security note: touches auth/quota path (`src/codex/**`) — verify no credential
15+
logging/serialization, quota parse is fail-closed, no auth escalation.
16+
17+
Actions:
18+
1. A-gate: Sol reviewer reads the diff, confirms the stale-weekly clearing is
19+
correct for primary-only 30d accounts, no regression to weekly-bearing
20+
accounts, and the security boundary is clean.
21+
2. `gh pr ready 390` (mark ready so CodeRabbit + non-draft gates run) — or merge
22+
directly if maintainer pre-approval covers draft merge. Prefer ready first.
23+
3. Re-check mergeable + CI, then `gh pr merge 390 --squash` targeting dev.
24+
4. Verify: `gh pr view 390 --json state,mergeCommit`; #382 auto-closed or close
25+
with reference.
26+
27+
Ordering: #390 merges BEFORE #394 (both touch `src/server/responses/core.ts`);
28+
#394 rebases after.
29+
30+
Terminal: DONE = merged SHA + #382 closed. BLOCKED if CI regresses on ready.
Lines changed: 41 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,41 @@
1+
# WP5 — #389 model-switch visibility + ocx sync relationship (040)
2+
3+
PR #389 `fix(models): make switches reflect final visibility` (head
4+
`csa906:codex/fix-models-selected-visibility`, +677/-71).
5+
6+
## Findings (Sol explorer, read-only)
7+
8+
Real dashboard bug: a switch showed ON whenever a model was absent from
9+
`disabledModels`, even if a non-empty provider `selectedModels` allowlist still
10+
excluded it. #389 computes final visibility = `selectedModels allows &&
11+
!disabledModels blocks` (`gui/src/model-visibility.ts:36-55`), uses it for
12+
counts/ordering/bulk state (`gui/src/pages/Models.tsx:201-271,783-807`), and adds
13+
`PUT /api/model-visibility` updating both filters in one save
14+
(`src/server/management/model-routes.ts:132-228`). Earlier reviewer findings
15+
fixed in `b89c5399`/`792baab4`; regressions at
16+
`tests/model-visibility-management-api.test.ts:96-245`.
17+
18+
### ocx sync relationship
19+
20+
Shared downstream catalog regen path, but #389 does NOT change the CLI sync impl:
21+
visibility persists via atomic `saveConfig()` (`src/config.ts:775-786`);
22+
`syncCatalogModels()` applies `filterCatalogVisibleModels()` and writes the Codex
23+
catalog (`provider-fetch.ts:424-444`, `sync.ts:455-496`); `ocx sync` loads config
24+
→ `refreshCodexModelCatalog()` (`cli/index.ts:621-623`, `codex/sync.ts:30-77`).
25+
The new route immediately refreshes after save (`model-routes.ts:225-228`).
26+
Conclusion: no semantic conflict; `ocx sync` reproduces the same visibility; a
27+
simultaneous CLI sync is an existing last-writer race, not introduced by #389.
28+
29+
## Decision: merge-with-note
30+
31+
Logically correct, well-tested; no correctness blocker. #389 is GUI-touching and
32+
NOT in the user's named GUI-merge scope (only #393/#340 were). So WP5 = A-gate
33+
review + re-check full CI; if green, post a merge-with-note review and confirm
34+
with user before the GUI merge, OR merge if maintainer scope clearly allows.
35+
36+
Actions:
37+
1. `gh pr checks 389` — require full matrix green (not just target/label).
38+
2. A-gate Sol reviewer.
39+
3. Post merge-with-note; merge only if in approved GUI scope.
40+
41+
Terminal: DONE = merged SHA OR merge-readiness note posted.

0 commit comments

Comments
 (0)