You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Automated canary-rollout blocker. The release gate is holding pr-review and will not promote it until this clears. Filed + maintained by the Canary Rollout workflow (gate standard: .github#548); this issue is regenerated each run and auto-closes when the gate passes — do not edit the table below by hand.
field
value
agent
pr-review
transition
next->ring0
candidate
82b654f7ef11
host repo
combsmith/.github-private
cumulative failures
3 (startup_failures: 0)
triage
PRE_EXISTING
⚠️PRE_EXISTING — the failure is pre-existing/environmental (reusable byte-identical to the prior channel). Report only; the gate will not roll back or advance. Fix-forward, and the armed timer auto-promotes once clean.
Failing runs in the per-candidate window
combsmith/.github-private — run 44020718238; failed steps: Review each PR (cascade)
combsmith/.github-private — run 43974314418; failed steps: Review each PR (cascade)
combsmith/.github-private — run 43873329908; failed steps: Review each PR (cascade)
Whole-fleet status is in the Canary Rollout workflow run's job summary (Actions → Canary Rollout → latest run → Summary).
Dev-Lead: issue #1169 implementation failed — will retry
The engine failed (engine-error — e.g. a timeout or transient engine error). This was attempt 1 of 3; dev-lead will retry automatically. You can also re-apply the dev-lead label to retry now.
Warning: 256-color support not detected. Using a terminal with at least 256-color support is recommended for a better visual experience.
Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY.
Ripgrep is not available. Falling back to GrepTool.
Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY.
Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY.
Error when talking to Gemini API Full report available at: /tmp/gemini-client-error-Turn.run-sendMessageStream-2026-09-23T05-25-48-208Z.json ModelNotFoundError: This model models/gemini-2.5-pro is no longer available to new users. Please update your code to use models/gemini-3.1-pro-preview for the latest features and improvements. We recommend you to use the Interactions API (https://ai.google.dev/gemini-api/docs/get-started).
at classifyGoogleError (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:307106:12)
at retryWithBackoff (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:307883:29)
at process.processTicksAndRejections (node:internal/process/task_queues:103:5)
at async GeminiChat.makeApiCallAndProcessStream (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:331852:28)
at async GeminiChat.streamWithRetries (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:331605:29)
at async Turn.run (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:332527:24)
at async GeminiClient.processTurn (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:346188:22)
at async GeminiClient.sendMessageStream (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:346285:14)
at async file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/gemini-LUNNHKPJ.js:11868:26
at async main (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/gemini-LUNNHKPJ.js:17295:5) {
code: 404
}
{
"session_id": "ff90a430-36de-41a5-badd-09423dc76495",
"error": {
"type": "Error",
"message": "This model models/gemini-2.5-pro is no longer available to new users. Please update your code to use models/gemini-3.1-pro-preview for the latest features and improvements. We recommend you to use the Interactions API (https://ai.google.dev/gemini-api/docs/get-started).",
"code": 1
}
}
Dev-Lead: Root-cause analysis — fix belongs in .github-private, not this repo (needs-human)
Provisional / in-progress note. No source change is being made in petry-projects/.github by this run, because the defect is not in this repo's tracked source (details below). Nothing is pushed; this is analysis, not a delivered result.
Root cause (confirmed from both failing runs)
The pr-review cascade fails at tier 1 (triage) only after the engine chain has fallen all the way through to the Copilot fallback:
[claude] model claude-fable-5 throttled → claude-opus-4-8 throttled → claude-opus-4-7 throttled
Gemini rate limit hit — switching to Copilot engine for remaining PRs
[copilot] calling gh copilot (model=openai/o4-mini, timeout=300s, flags=--deny-tool *)
Invalid --deny-tool value. Error: Invalid rule format: *
##[error]cascade failed at tier 1 (triage process exit 1)
Identical failure in both windowed runs — 35821116812 (PR #1913) and 35821080436 (PR #1912).
The bug: the Copilot triage branch invokes gh copilot ... --deny-tool "*". The Copilot CLI's --deny-tool/--allow-tool take a rule (e.g. shell(git)), so * is rejected as an invalid rule format and gh copilot exits non-zero, failing the triage tier. The Copilot CLI is non-agentic by default (it only enables tools when --allow-tool/--yolo is passed), so the "deny all tools in triage" intent is already the default — the --deny-tool "*" argument is both invalid and unnecessary.
This is why the canary triage is PRE_EXISTING / environmental: the reusable is byte-identical to the prior channel, and the fault only surfaces when Claude and Gemini are simultaneously throttled and the cascade reaches the Copilot fallback. Candidate eee2898e72dc did not introduce it; it is a latent bug on the Copilot fallback path.
Exact fix (fix-forward)
In petry-projects/.github-private → scripts/engine.sh, function run_triage(), the copilot) branch (currently ~lines 1043 & 1045):
Remove the --deny-tool "*" argument from both copilot_chat "$prompt_file" "$TRIAGE_TIMEOUT_SEC" --deny-tool "*" invocations (Copilot denies all tools by default in non---yolo mode), or, if an explicit deny is desired, pass a valid rule (e.g. --deny-tool 'shell' / --deny-tool 'write') rather than *.
Add a bats case for the copilot triage invocation asserting no invalid flag is passed to gh copilot (mirror the existing engine tests), per the org "decision logic in a pure, tested script" standard.
Why this run cannot deliver the fix (scope mismatch → needs-human)
This dev-lead run is scoped to petry-projects/.github; its automation commits the working tree here and opens a PR against petry-projects/.github. But:
scripts/engine.sh (and the whole pr-review agent) is not tracked in petry-projects/.github — git ls-files | grep engine.sh → none; git grep 'deny-tool' → none.
Per AGENTS.md ("What lives where"), agent scripts/reusables are owned by .github-private; .github must not become the source of truth for agent code. Adding engine.sh here would be drift.
The only copy present locally is the gitignored .dev-lead/ working clone of .github-private (git check-ignore .dev-lead → ignored; nested .git → origin = petry-projects/.github-private). Edits there are not committed by this run and would not reach .github-private.
I am therefore not fabricating a change in .github to make this run look complete (that would be the #1445 / "Fail Loud, Never Fake" anti-pattern) and not opening a cross-repo PR (out of this run's sanctioned git scope).
Action needed: route the one-line fix + test above to petry-projects/.github-private (a .github-private dev-lead run or a human PR). Once that lands and a clean per-candidate window accrues, the armed timer auto-promotes pr-review next->ring0 and this blocker auto-closes.
Dev-Lead: Diagnostic (no .github-side change warranted)
Provisional / in-progress note — nothing pushed. This is a root-cause finding, not a completion claim.
I investigated the two failing runs in the per-candidate window. Both fail with an identical, candidate-independent signature, which confirms the automated PRE_EXISTING triage.
Root cause
All Claude models were rate-limited (claude-fable-5, claude-opus-4-8, claude-opus-4-7 throttled).
The Gemini fallback was unavailable — "prepayment credits depleted" (environmental).
The cascade fell through to the Copilot engine, whose triage branch invokes gh copilot … --deny-tool "*".
→ triage exits 1 → cascade failed at tier 1 (triage process exit 1).
Evidence (both runs identical): run 35821116812 (PR #1913) and run 35821080436 (PR #1912).
Why this is PRE_EXISTING (and why the candidate is not at fault)
The pr-review reusable (pr-review.yml) is byte-identical to the prior channel (differs=0). Candidate eee2898e72dc introduced neither the --deny-tool "*" invocation nor the depleted-credits/rate-limit environment.
The failure only manifests on the Copilot fallthrough path, which is reached only because the primary (Claude) + secondary (Gemini) engines were both unavailable — an environmental condition, exactly as the gate's PRE_EXISTING note describes.
Fix-forward location — NOT this repo
The defective invocation lives in petry-projects/.github-private/scripts/engine.sh (the copilot triage branch passes --deny-tool "*", which the gh copilot CLI does not accept). That is a different repository; none of the failing code is tracked in petry-projects/.github. Per AGENTS.md multi-repo isolation ("work only on its assigned repo — do not modify files in other repos"), I am not making a change here to code that lives elsewhere, and I will not fabricate an unrelated .github edit to force a green PR.
Recommended fix-forward (to be applied in .github-private): replace the invalid --deny-tool "*" with a rule-format the gh copilot CLI accepts for "deny all tools" (or achieve the same triage restriction via the engine's supported deny-all mechanism), and add a bats case around the copilot triage invocation so the argument is validated pre-merge (AGENTS.md "Decision Logic Lives in a Pure, Tested Script").
Disposition
No source change is appropriate in petry-projects/.github.
The gate note states this PRE_EXISTING blocker is report-only and the armed timer auto-promotes once clean; it will clear once the environmental engine availability recovers and/or the .github-private fix-forward lands.
Dev-Lead: Triage report — no in-repo code change is warranted (PRE_EXISTING)
I investigated issue #1169 as the dev-lead for petry-projects/.github. This is an
auto-generated canary-rollout blocker with triage PRE_EXISTING, not a feature or
bug request. My conclusion is that there is no source change in this repository that
would resolve it, and fabricating one would be incorrect. Details below.
Triage PRE_EXISTING — per the gate core (scripts/lib/canary-rollout.sh → classify_failure), this fires when the candidate reusable is byte-identical to the
prior channel (differs=0, so it cannot be candidate-caused) or the failure is an
environmental class (comment-cap | rate-limit | infra | data). The gate's own directive
is: "report, do NOT rollback, do NOT advance."
Root cause of the two failing runs (verified from logs)
Both runs (35821116812, 35821080436, workflow "PR Review Agent — Trigger", the review / review cascade job) failed on API rate-limit / credit-depletion, not a code
defect:
Gemini fallback unavailable — prepayment credits depleted — replenish via Google AI Studio
Claude rate limit hit but Gemini fallback unavailable — falling through to Copilot
This is precisely the rate-limit environmental class that maps to PRE_EXISTING, and it is
consistent with the candidate reusable being byte-identical to the prior channel.
Why no code change here
The failing workflow and the pr-review agent/reusable (pr-review.yml, pr-review/v1-next → eee2898e72dc) live in petry-projects/.github-private, not in
this repo. Per AGENTS.md, .github-private owns agents/skills/reusables; this repo owns
standards. I cannot (and per my constraints must not) modify .github-private.
The failure is operational (depleted Gemini prepayment credits + Claude/Codex rate-limit
windows), not a logic bug. No edit to petry-projects/.github changes that condition.
The issue banner explicitly says this issue is regenerated each run, auto-closes when the
gate passes, and the table must not be hand-edited. It is Report only.
Fix-forward (operational, outside this repo)
Replenish Gemini prepayment credits (Google AI Studio billing) and/or let the Claude /
Codex rate-limit windows reset.
The armed timer auto-promotes pr-review next->ring0 once a clean run occurs, and this
issue auto-closes.
Disposition
No PR will be opened from this repo for #1169 — there is nothing here to implement. I am
deliberately not manufacturing code or tests to force a "done" state (AGENTS.md
"Fail Loud, Never Fake" / YAGNI). Recording this as the triage report the PRE_EXISTING
disposition calls for.
Dev-Lead: Root-cause analysis — no fix-forward lands in petry-projects/.github
I investigated this canary blocker. Summary up front: the failing code is not in this repository, so there is nothing for a dev-lead PR against petry-projects/.github to change. This note records the root cause so the fix-forward can be applied where it belongs.
Root cause
Both failing runs fail at the same point — the pr-reviewtriage cascade (tier 1), in the copilot engine-fallback path:
Primary advisory-gate engine is rate-limited: [advisory-gate] ✗ chatgpt-codex-connector → RATE_LIMITED → [approve] rate limit detected — exiting with code 2 for engine fallback (environmental/transient).
Triage falls back to the copilot engine and invokes: [copilot] calling gh copilot (model=openai/o4-mini, timeout=300s, flags=--deny-tool *)
The CLI rejects the argument: Invalid --deny-tool value. Error: Invalid rule format: *
triage exited with code 1 → cascade failed at tier 1 (triage process exit 1).
So the surface failure ("Review each PR (cascade)") is a latent bug in the copilot fallback invocation: --deny-tool * is not a valid gh copilot rule format. It is only reached when the primary engine is rate-limited, which is why triage classified it PRE_EXISTING (reusable byte-identical to the prior channel; the bug predates candidate eee2898e72dc).
Why there is no change to make in this repo
The cascade/triage engine and the --deny-tool invocation live in petry-projects/.github-private/.github/workflows/pr-review.yml (candidate eee2898e72dc, pr-review/v1-next). This repo (petry-projects/.github) contains only the mention→dispatch trigger (pr-review-mention-reusable.yml) and the canary-rollout gate that filed this issue — neither contains any --deny-tool/cascade/triage logic (grep for deny-tool, cascade failed, advisory-gate returns nothing here). Per AGENTS.md, agent reusable workflows are scoped to .github-private, not this repo.
Consistent with this issue's own PRE_EXISTING triage ("Report only; the gate will not roll back or advance. Fix-forward, and the armed timer auto-promotes once clean"), I am not fabricating a change in this repo.
Recommended fix-forward (in petry-projects/.github-private)
In the copilot engine invocation inside pr-review.yml, replace the invalid --deny-tool * argument with a form gh copilot accepts (e.g. a concrete tool/rule pattern, or drop the bare * glob). Landing that in pr-review/v1-next and letting a clean run through the per-candidate window will clear this gate; the armed timer then auto-promotes.
No file changes were made in this workspace; no PR should be opened against petry-projects/.github for this issue.
Dev-Lead: Triage — no implementable change in this repo (report-only blocker)
This is an auto-generated PRE_EXISTING canary blocker. Per the issue itself the gate is report-only ("the gate will not roll back or advance … auto-promotes once clean") and the reusable is byte-identical to the prior channel. I investigated the failing runs rather than manufacture a TDD change, and the evidence confirms there is nothing to implement in petry-projects/.github.
What the failing runs actually show
All failing steps execute inside the pr-review.yml reusable + cascade scripts in petry-projects/.github-private (candidate eee2898e72dc, tag pr-review/v1-next). None of that code lives in this repo — .github owns the canary-rollout engine and standards (and other agents' reusables like pr-auto-review / pr-review-mention), not pr-review.
1. "Verify auth scopes" — run 35916202618
The bot GH_TOKEN is a fine-grained PAT (or a classic PAT missing read:org). The step is designed to fail loud here. → secret/token configuration in the calling repo/org, not a code defect.
2. "Review each PR (cascade)" — run 35822489297 and the rest of the window
The engine cascade exhausts every tier:
Claude → API 404: "the selected model (claude-sonnet-5-0) … may not exist or you may not have access to it" → model entitlement/config.
Copilot fallback → Invalid --deny-tool value. Error: Invalid rule format: * → a bad gh copilot invocation in .github-private's cascade (--deny-tool * is rejected by the CLI).
Why no code change here
The reusable is byte-identical to the prior channel, so the candidate didn't introduce these failures — they are pre-existing/environmental (auth token, model access, billing) plus one CLI-arg bug that lives in .github-private.
--deny-tool does not appear anywhere in petry-projects/.github.
Editing anything here (or hand-editing the blocker table) would be drift/fabrication and is explicitly out of scope for this report-only gate.
Where the fixes belong (fix-forward, outside this repo)
Replace the bot PAT with a classic PAT carrying repo, workflow, read:org (calling repo/org secret).
Restore Claude model access for claude-sonnet-5-0 (or repoint the engine to an entitled model).
Replenish Gemini prepayment credits, or accept Copilot-only fallback.
Fix the gh copilot … --deny-tool * argument in .github-private's pr-review cascade (the * rule format is invalid).
No PR will be opened from this repo for #1169. The gate auto-closes this issue once the pr-review runs go green.
combsmith-release-manager commented on Sep 24, 2026
Dev-Lead: Triage (investigation only — no fix-forward belongs in this repo)
This is an automated PRE_EXISTING / report-only canary blocker. Per the gate note, the release gate "will not roll back or advance" and "the armed timer auto-promotes once clean." I investigated and confirmed there is no source change to make in petry-projects/.github. I have not edited the auto-maintained table, overridden the gate, or weakened any threshold.
1. The candidate did not touch the pr-review engine reusable
Candidate 1d6b38b52387 changed only docs, scripts/lib/* helpers, and tests in .github-private. It does not modify .github/workflows/pr-review.yml (the reusable the canary gate byte-diffs). That is exactly why the gate triaged this PRE_EXISTING (reusable byte-identical to the prior channel → differs=0) — the candidate cannot have introduced this failure.
2. Root cause is environmental, not a regression
Failing run 36000423865 failed at Review each PR (cascade) due to:
deep review did not produce valid JSON at /tmp/cascade/deep.json (transient deep-tier output)
This is a fallback-chain/billing/environmental failure, consistent with the PRE_EXISTING triage.
3. The failing code lives in .github-private, not here
This repo (.github) owns only the canary registry (standards/canary-rings.json) and the rollout scripts. The pr-review.yml engine and its cascade/fallback logic live in petry-projects/.github-private (registry _note confirms this). Any hardening fix-forward (e.g. more graceful handling when the deep tier emits invalid JSON) belongs there, not in this repo.
Recommended remediation (ops, not code-in-this-repo)
Replenish the Gemini prepayment credits (Google AI Studio billing) so the Claude→Gemini→Copilot cascade has a working fallback; the transient invalid-JSON clears on its own.
No --override and no gate-weakening (benign_failure_classes / allow_pre_existing) — the org standard forbids bypassing CI gates, and this issue auto-closes when a clean run lands.
Investigation note only — nothing pushed; no completion claim is being made.
Dev-Lead: Triage — no code change (report-only, PRE_EXISTING)
I investigated this blocker. It is a PRE_EXISTING canary-rollout blocker, and per the gate standard and this repo's docs/initiatives/canary-rollout-adr.md, that classification means there is no candidate-introduced defect to fix — so I am deliberately not landing a code change.
What the failing run actually shows
Run 36000423865 failed at Review each PR (cascade) while reviewing petry-projects/.github-private#1906 (a docs PR).
The annotation cascade failed at tier 2 maps to scripts/review-one-pr.sh:1440: the deep-review engine did not emit valid JSON for that one PR. The code path explicitly rules out a rate-limit (exit 2) and a CLI-format error before reaching this line, so this is a transient engine-execution failure, not a code defect.
Why no fix is warranted
Triage is PRE_EXISTING, which the gate assigns only when the candidate's reusable blob is byte-identical to the prior channel (differs=0) — i.e. not a candidate regression (docs/initiatives/canary-rollout-adr.md; candidate 1d6b38b52387 is HEAD of the reusable, unchanged from the prior good release).
The issue body itself states this is report-only: "the gate will not roll back or advance … the armed timer auto-promotes once clean."
Resolution paths per the ADR are (a) the environmental condition clears and the armed timer auto-promotes, or (b) a human runs promote --allow-pre-existing. Neither is a source change in this repo.
Fabricating a code change here would break the byte-identical invariant the gate relies on and would violate the org "Fail Loud, Never Fake" standard. No files changed; nothing to push. This issue auto-closes when the gate next passes; if it persists, a maintainer can advance it with promote --allow-pre-existing.
Activity
don-petry commentedon Sep 23, 2026
Dev-Lead: issue #1169 implementation failed — will retry
The engine failed (
engine-error— e.g. a timeout or transient engine error). This was attempt 1 of 3; dev-lead will retry automatically. You can also re-apply thedev-leadlabel to retry now.engine-errorLast 40 lines of engine output (redacted)
Warning: 256-color support not detected. Using a terminal with at least 256-color support is recommended for a better visual experience. Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY. Ripgrep is not available. Falling back to GrepTool. Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY. Both GOOGLE_API_KEY and GEMINI_API_KEY are set. Using GOOGLE_API_KEY. Error when talking to Gemini API Full report available at: /tmp/gemini-client-error-Turn.run-sendMessageStream-2026-09-23T05-25-48-208Z.json ModelNotFoundError: This model models/gemini-2.5-pro is no longer available to new users. Please update your code to use models/gemini-3.1-pro-preview for the latest features and improvements. We recommend you to use the Interactions API (https://ai.google.dev/gemini-api/docs/get-started). at classifyGoogleError (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:307106:12) at retryWithBackoff (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:307883:29) at process.processTicksAndRejections (node:internal/process/task_queues:103:5) at async GeminiChat.makeApiCallAndProcessStream (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:331852:28) at async GeminiChat.streamWithRetries (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:331605:29) at async Turn.run (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:332527:24) at async GeminiClient.processTurn (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:346188:22) at async GeminiClient.sendMessageStream (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/chunk-M6NSK26M.js:346285:14) at async file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/gemini-LUNNHKPJ.js:11868:26 at async main (file:///home/runner/.npm-global/lib/node_modules/@google/gemini-cli/bundle/gemini-LUNNHKPJ.js:17295:5) { code: 404 } { "session_id": "ff90a430-36de-41a5-badd-09423dc76495", "error": { "type": "Error", "message": "This model models/gemini-2.5-pro is no longer available to new users. Please update your code to use models/gemini-3.1-pro-preview for the latest features and improvements. We recommend you to use the Interactions API (https://ai.google.dev/gemini-api/docs/get-started).", "code": 1 } }don-petry commentedon Sep 23, 2026
Dev-Lead: Root-cause analysis — fix belongs in
.github-private, not this repo (needs-human)Root cause (confirmed from both failing runs)
The
pr-reviewcascade fails at tier 1 (triage) only after the engine chain has fallen all the way through to the Copilot fallback:Identical failure in both windowed runs — 35821116812 (PR #1913) and 35821080436 (PR #1912).
The bug: the Copilot triage branch invokes
gh copilot ... --deny-tool "*". The Copilot CLI's--deny-tool/--allow-tooltake a rule (e.g.shell(git)), so*is rejected as an invalid rule format andgh copilotexits non-zero, failing the triage tier. The Copilot CLI is non-agentic by default (it only enables tools when--allow-tool/--yolois passed), so the "deny all tools in triage" intent is already the default — the--deny-tool "*"argument is both invalid and unnecessary.This is why the canary triage is PRE_EXISTING / environmental: the reusable is byte-identical to the prior channel, and the fault only surfaces when Claude and Gemini are simultaneously throttled and the cascade reaches the Copilot fallback. Candidate
eee2898e72dcdid not introduce it; it is a latent bug on the Copilot fallback path.Exact fix (fix-forward)
In
petry-projects/.github-private→scripts/engine.sh, functionrun_triage(), thecopilot)branch (currently ~lines 1043 & 1045):--deny-tool "*"argument from bothcopilot_chat "$prompt_file" "$TRIAGE_TIMEOUT_SEC" --deny-tool "*"invocations (Copilot denies all tools by default in non---yolomode), or, if an explicit deny is desired, pass a valid rule (e.g.--deny-tool 'shell'/--deny-tool 'write') rather than*.batscase for the copilot triage invocation asserting no invalid flag is passed togh copilot(mirror the existing engine tests), per the org "decision logic in a pure, tested script" standard.Why this run cannot deliver the fix (scope mismatch → needs-human)
This dev-lead run is scoped to
petry-projects/.github; its automation commits the working tree here and opens a PR againstpetry-projects/.github. But:scripts/engine.sh(and the wholepr-reviewagent) is not tracked inpetry-projects/.github—git ls-files | grep engine.sh→ none;git grep 'deny-tool'→ none..github-private;.githubmust not become the source of truth for agent code. Addingengine.shhere would be drift..dev-lead/working clone of.github-private(git check-ignore .dev-lead→ ignored; nested.git→origin = petry-projects/.github-private). Edits there are not committed by this run and would not reach.github-private.I am therefore not fabricating a change in
.githubto make this run look complete (that would be the #1445 / "Fail Loud, Never Fake" anti-pattern) and not opening a cross-repo PR (out of this run's sanctioned git scope).Action needed: route the one-line fix + test above to
petry-projects/.github-private(a.github-privatedev-lead run or a human PR). Once that lands and a clean per-candidate window accrues, the armed timer auto-promotespr-review next->ring0and this blocker auto-closes.don-petry commentedon Sep 23, 2026
Dev-Lead: Diagnostic (no
.github-side change warranted)I investigated the two failing runs in the per-candidate window. Both fail with an identical, candidate-independent signature, which confirms the automated PRE_EXISTING triage.
Root cause
claude-fable-5,claude-opus-4-8,claude-opus-4-7throttled).gh copilot … --deny-tool "*".gh copilotrejects that argument:cascade failed at tier 1 (triage process exit 1).Evidence (both runs identical): run 35821116812 (PR #1913) and run 35821080436 (PR #1912).
Why this is
PRE_EXISTING(and why the candidate is not at fault)pr-reviewreusable (pr-review.yml) is byte-identical to the prior channel (differs=0). Candidateeee2898e72dcintroduced neither the--deny-tool "*"invocation nor the depleted-credits/rate-limit environment.PRE_EXISTINGnote describes.Fix-forward location — NOT this repo
The defective invocation lives in
petry-projects/.github-private/scripts/engine.sh(thecopilottriage branch passes--deny-tool "*", which thegh copilotCLI does not accept). That is a different repository; none of the failing code is tracked inpetry-projects/.github. Per AGENTS.md multi-repo isolation ("work only on its assigned repo — do not modify files in other repos"), I am not making a change here to code that lives elsewhere, and I will not fabricate an unrelated.githubedit to force a green PR.Recommended fix-forward (to be applied in
.github-private): replace the invalid--deny-tool "*"with a rule-format thegh copilotCLI accepts for "deny all tools" (or achieve the same triage restriction via the engine's supported deny-all mechanism), and add a bats case around the copilot triage invocation so the argument is validated pre-merge (AGENTS.md "Decision Logic Lives in a Pure, Tested Script").Disposition
petry-projects/.github.PRE_EXISTINGblocker is report-only and the armed timer auto-promotes once clean; it will clear once the environmental engine availability recovers and/or the.github-privatefix-forward lands.don-petry commentedon Sep 23, 2026
Dev-Lead: Triage report — no in-repo code change is warranted (PRE_EXISTING)
I investigated issue #1169 as the dev-lead for
petry-projects/.github. This is anauto-generated canary-rollout blocker with triage PRE_EXISTING, not a feature or
bug request. My conclusion is that there is no source change in this repository that
would resolve it, and fabricating one would be incorrect. Details below.
What the blocker is
pr-review next->ring0, candidateeee2898e72dc, host repopetry-projects/.github-private, cumulative failures 2 (startup_failures 0).scripts/lib/canary-rollout.sh→classify_failure), this fires when the candidate reusable is byte-identical to theprior channel (
differs=0, so it cannot be candidate-caused) or the failure is anenvironmental class (
comment-cap | rate-limit | infra | data). The gate's own directiveis: "report, do NOT rollback, do NOT advance."
Root cause of the two failing runs (verified from logs)
Both runs (
35821116812,35821080436, workflow "PR Review Agent — Trigger", thereview / reviewcascade job) failed on API rate-limit / credit-depletion, not a codedefect:
Gemini fallback unavailable — prepayment credits depleted — replenish via Google AI Studio[advisory-gate] ✗ chatgpt-codex-connector → RATE_LIMITEDClaude rate limit hit but Gemini fallback unavailable — falling through to CopilotThis is precisely the
rate-limitenvironmental class that maps to PRE_EXISTING, and it isconsistent with the candidate reusable being byte-identical to the prior channel.
Why no code change here
pr-reviewagent/reusable (pr-review.yml,pr-review/v1-next→eee2898e72dc) live inpetry-projects/.github-private, not inthis repo. Per AGENTS.md,
.github-privateowns agents/skills/reusables; this repo ownsstandards. I cannot (and per my constraints must not) modify
.github-private.windows), not a logic bug. No edit to
petry-projects/.githubchanges that condition.gate passes, and the table must not be hand-edited. It is Report only.
Fix-forward (operational, outside this repo)
Codex rate-limit windows reset.
pr-review next->ring0once a clean run occurs, and thisissue auto-closes.
Disposition
No PR will be opened from this repo for #1169 — there is nothing here to implement. I am
deliberately not manufacturing code or tests to force a "done" state (AGENTS.md
"Fail Loud, Never Fake" / YAGNI). Recording this as the triage report the PRE_EXISTING
disposition calls for.
don-petry commentedon Sep 23, 2026
Dev-Lead: Root-cause analysis — no fix-forward lands in
petry-projects/.githubI investigated this canary blocker. Summary up front: the failing code is not in this repository, so there is nothing for a dev-lead PR against
petry-projects/.githubto change. This note records the root cause so the fix-forward can be applied where it belongs.Root cause
Both failing runs fail at the same point — the
pr-reviewtriage cascade (tier 1), in the copilot engine-fallback path:The sequence in each run:
[advisory-gate] ✗ chatgpt-codex-connector → RATE_LIMITED→[approve] rate limit detected — exiting with code 2 for engine fallback(environmental/transient).[copilot] calling gh copilot (model=openai/o4-mini, timeout=300s, flags=--deny-tool *)Invalid --deny-tool value. Error: Invalid rule format: *triage exited with code 1→cascade failed at tier 1 (triage process exit 1).So the surface failure ("Review each PR (cascade)") is a latent bug in the copilot fallback invocation:
--deny-tool *is not a validgh copilotrule format. It is only reached when the primary engine is rate-limited, which is why triage classified it PRE_EXISTING (reusable byte-identical to the prior channel; the bug predates candidateeee2898e72dc).Why there is no change to make in this repo
The cascade/triage engine and the
--deny-toolinvocation live inpetry-projects/.github-private/.github/workflows/pr-review.yml(candidateeee2898e72dc,pr-review/v1-next). This repo (petry-projects/.github) contains only the mention→dispatch trigger (pr-review-mention-reusable.yml) and the canary-rollout gate that filed this issue — neither contains any--deny-tool/cascade/triage logic (grepfordeny-tool,cascade failed,advisory-gatereturns nothing here). Per AGENTS.md, agent reusable workflows are scoped to.github-private, not this repo.Consistent with this issue's own PRE_EXISTING triage ("Report only; the gate will not roll back or advance. Fix-forward, and the armed timer auto-promotes once clean"), I am not fabricating a change in this repo.
Recommended fix-forward (in
petry-projects/.github-private)In the copilot engine invocation inside
pr-review.yml, replace the invalid--deny-tool *argument with a formgh copilotaccepts (e.g. a concrete tool/rule pattern, or drop the bare*glob). Landing that inpr-review/v1-nextand letting a clean run through the per-candidate window will clear this gate; the armed timer then auto-promotes.No file changes were made in this workspace; no PR should be opened against
petry-projects/.githubfor this issue.[-]Canary blocker: pr-review next->ring0 (cum_fail=2, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=27, PRE_EXISTING)[/+]don-petry commentedon Sep 24, 2026
Dev-Lead: Triage — no implementable change in this repo (report-only blocker)
This is an auto-generated PRE_EXISTING canary blocker. Per the issue itself the gate is report-only ("the gate will not roll back or advance … auto-promotes once clean") and the reusable is byte-identical to the prior channel. I investigated the failing runs rather than manufacture a TDD change, and the evidence confirms there is nothing to implement in
petry-projects/.github.What the failing runs actually show
All failing steps execute inside the
pr-review.ymlreusable + cascade scripts inpetry-projects/.github-private(candidateeee2898e72dc, tagpr-review/v1-next). None of that code lives in this repo —.githubowns the canary-rollout engine and standards (and other agents' reusables likepr-auto-review/pr-review-mention), notpr-review.1. "Verify auth scopes" — run 35916202618
The bot
GH_TOKENis a fine-grained PAT (or a classic PAT missingread:org). The step is designed to fail loud here. → secret/token configuration in the calling repo/org, not a code defect.2. "Review each PR (cascade)" — run 35822489297 and the rest of the window
The engine cascade exhausts every tier:
404: "the selected model (claude-sonnet-5-0) … may not exist or you may not have access to it" → model entitlement/config.Invalid --deny-tool value. Error: Invalid rule format: *→ a badgh copilotinvocation in.github-private's cascade (--deny-tool *is rejected by the CLI).Why no code change here
.github-private.--deny-tooldoes not appear anywhere inpetry-projects/.github.Where the fixes belong (fix-forward, outside this repo)
repo,workflow,read:org(calling repo/org secret).claude-sonnet-5-0(or repoint the engine to an entitled model).gh copilot … --deny-tool *argument in.github-private's pr-review cascade (the*rule format is invalid).No PR will be opened from this repo for #1169. The gate auto-closes this issue once the pr-review runs go green.
combsmith-release-manager commentedon Sep 24, 2026
✅ Gate cleared —
pr-reviewis nowSOAKING. Closed automatically by canary-rollout.[-]Canary blocker: pr-review next->ring0 (cum_fail=27, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=1, PRE_EXISTING)[/+]don-petry commentedon Sep 24, 2026
Dev-Lead: Triage (investigation only — no fix-forward belongs in this repo)
1. The candidate did not touch the pr-review engine reusable
Candidate
1d6b38b52387changed only docs,scripts/lib/*helpers, and tests in.github-private. It does not modify.github/workflows/pr-review.yml(the reusable the canary gate byte-diffs). That is exactly why the gate triaged thisPRE_EXISTING(reusable byte-identical to the prior channel →differs=0) — the candidate cannot have introduced this failure.2. Root cause is environmental, not a regression
Failing run 36000423865 failed at
Review each PR (cascade)due to:deep review did not produce valid JSON at /tmp/cascade/deep.json(transient deep-tier output)Gemini fallback unavailable — prepayment credits depletedcascade failed at tier 2for PR #1906 → exit 1This is a fallback-chain/billing/environmental failure, consistent with the
PRE_EXISTINGtriage.3. The failing code lives in
.github-private, not hereThis repo (
.github) owns only the canary registry (standards/canary-rings.json) and the rollout scripts. Thepr-review.ymlengine and its cascade/fallback logic live inpetry-projects/.github-private(registry_noteconfirms this). Any hardening fix-forward (e.g. more graceful handling when the deep tier emits invalid JSON) belongs there, not in this repo.Recommended remediation (ops, not code-in-this-repo)
--overrideand no gate-weakening (benign_failure_classes/allow_pre_existing) — the org standard forbids bypassing CI gates, and this issue auto-closes when a clean run lands.Investigation note only — nothing pushed; no completion claim is being made.
don-petry commentedon Sep 24, 2026
Dev-Lead: Triage — no code change (report-only, PRE_EXISTING)
I investigated this blocker. It is a
PRE_EXISTINGcanary-rollout blocker, and per the gate standard and this repo'sdocs/initiatives/canary-rollout-adr.md, that classification means there is no candidate-introduced defect to fix — so I am deliberately not landing a code change.What the failing run actually shows
petry-projects/.github-private#1906(a docs PR).cascade failed at tier 2maps toscripts/review-one-pr.sh:1440: the deep-review engine did not emit valid JSON for that one PR. The code path explicitly rules out a rate-limit (exit 2) and a CLI-format error before reaching this line, so this is a transient engine-execution failure, not a code defect.Why no fix is warranted
differs=0) — i.e. not a candidate regression (docs/initiatives/canary-rollout-adr.md; candidate1d6b38b52387is HEAD of the reusable, unchanged from the prior good release).promote --allow-pre-existing. Neither is a source change in this repo.PRE_EXISTING) was auto-closed by the canary-rollout automation with no dev-lead code change once the gate cleared.Fabricating a code change here would break the byte-identical invariant the gate relies on and would violate the org "Fail Loud, Never Fake" standard. No files changed; nothing to push. This issue auto-closes when the gate next passes; if it persists, a maintainer can advance it with
promote --allow-pre-existing.87 remaining items
[-]Canary blocker: pr-review next->ring0 (cum_fail=4, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=5, PRE_EXISTING)[/+][-]Canary blocker: pr-review next->ring0 (cum_fail=5, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=7, PRE_EXISTING)[/+]combsmith-release-manager commentedon Nov 13, 2026
✅ Gate cleared —
pr-reviewis no longer BLOCKED. Closed automatically by canary-rollout.[-]Canary blocker: pr-review next->ring0 (cum_fail=7, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=2, PRE_EXISTING)[/+]combsmith-release-manager commentedon Nov 14, 2026
✅ Gate cleared —
pr-reviewis no longer BLOCKED. Closed automatically by canary-rollout.[-]Canary blocker: pr-review next->ring0 (cum_fail=2, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=1, PRE_EXISTING)[/+]combsmith-release-manager commentedon Nov 15, 2026
✅ Gate cleared —
pr-reviewis no longer BLOCKED. Closed automatically by canary-rollout.[-]Canary blocker: pr-review next->ring0 (cum_fail=1, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=2, PRE_EXISTING)[/+][-]Canary blocker: pr-review next->ring0 (cum_fail=2, PRE_EXISTING)[/-][+]Canary blocker: pr-review next->ring0 (cum_fail=3, PRE_EXISTING)[/+]