[aw-failures] [aw] Failure Investigator Report — 2026-08-02 (6h)

[github-actions]

Executive summary

Fix the tool-permission scoping regression first — it independently crashed Daily Testify Uber Super Expert and Linter Miner today via the same read($GITHUB_WORKSPACE) denial, burning ~25 minutes and ~14 minutes of compute respectively before giving up. Everything else this cycle is either already tracked or too low-severity to act on.

5 failures analyzed (of 8 failed runs in the last 6h — 3 lacked detailed logs in the pre-fetch payload and were not deep-dived). 1 new P0 sub-issue filed. 0 issues closed — no fresh evidence that any of the 12 currently-open agentic-workflows issues are fixed or stale.

Failure cluster table

#	Severity	Workflow(s)	Run(s)	Signature	Tracking
1	P0	Daily Testify Uber Super Expert, Linter Miner	§30760639664, §30759473080	Read-tool permission denied on the workflow's own checkout root → tool-denial threshold exceeded → session aborts	New — filed as sub-issue below
2	P1	Design Decision Gate 🏗️	§30754639273	Inconclusive: 21 turns completed, then exit 1; chroot ownership-transfer WARN on the safeoutputs temp dir precedes it but causality unconfirmed	Not filed — evidence too thin, see roadmap
3	P1 (recurring)	PR Code Quality Reviewer	§30754639269	Matches known background task-tool sub-agent failure pattern	Already tracked: #49022
4	P2	PR Sous Chef	§30755002904	safe_outputs job "Process Safe Outputs" step failure, isolated single occurrence	Not filed — too low-severity

3 additional failed run IDs from the lookback window without detailed log payloads

30753821917, 30751969165, 30751208732 — present in failed_run_ids but not in the detailed failures array; not deep-dived this cycle per the pre-fetch-first policy. Flag for next cycle if they recur.

Evidence

Cluster 1 (P0) — audit-diff and cross-workflow correlation

audit-diff comparing Daily Testify's failed run against its own successful baseline (§30570103985): posture regressed write_capable → read_only, turns dropped 1 → 0, GitHub Core API points consumed dropped 77% (252 → 59) — the agent barely started before being denied. Zero firewall/domain anomalies, which rules out a network/egress explanation.

audit on Linter Miner's failed run shows the identical degradation class against its own baseline: posture write_capable → read_only, turns 3 → 0. Two unrelated scheduled workflows, same signature, ~30 minutes apart — this is a systemic regression, not a workflow-specific config error. Full detail is in the filed sub-issue.

Cluster 2 — Design Decision Gate, inconclusive

Unlike cluster 1, this run completed 21 agent turns (reading PR metadata, ADR pre-fetch summary, etc.) before failing with a bare Process completed with exit code 1. The only anomaly in the log tail is [entrypoint][WARN] Failed to transfer /host/.../safeoutputs ownership to chroot user, emitted at container entrypoint — but the 50-line tail captured doesn't reach the actual failing step, so this is a lead, not a confirmed cause. Does not match either existing Design Decision Gate issue (#49144: pre-fetch step fails before agent starts; #49096: LLM invocation cap from shell-expanded paths) — this run got well past both of those failure points.

Existing issue correlation

• [aw-failures] Fleet-wide: Copilot CLI subagent model allocation fails with 'No model available' (hit PR Code Quality Reviewer, L [Content truncated due to length] #49022 (PR Code Quality Reviewer background task-tool failures) — still active, cluster 3 is a fresh recurrence, high-confidence match, no new issue needed.
• [aw-failures] Design Decision Gate: 'Pre-fetch ADR gate PR context' step fails before agent starts #49144, [aw-failures] Design Decision Gate exhausts LLM invocation cap after permission-denied retries on shell-expanded file paths #49096 (Design Decision Gate) — checked against cluster 2, do not match; this run failed later in its lifecycle than either tracked scenario.
• [aw-failures] dispatch_workflow safe-output fails the whole Smoke Copilot run when the ephemeral trigger branch is gone #49583, [aw-failures] Fix: unguarded arithmetic on github.event.pull_request.stack in applyPullRequestStackFilter #49446, [aw-failures] Conclusion job's setup_globals.cjs MODULE_NOT_FOUND cascade has regressed (repeat of #28149) #49023, [aw-failures] MCP Gateway github SSE reconnect hang + Ollama BYOK backend 503 fast-fail #49246, [aw] No-Op Runs #48838, [evals] Daily Evals Feature Report - 2026-08-02 #49735 and the three prior daily report issues ([aw-failures] [aw] Failure Investigator Report — 2026-07-31 (6h) #49245, [aw-failures] [aw] Failure Investigator Report — 2026-07-30 (6h) #49095, [aw-failures] [aw] Failure Investigator Report — 2026-07-29 (6h) #48897) — no overlap with this cycle's clusters.
• No open issue is contradicted by fresh evidence, so none closed this cycle.

Fix roadmap

• P0 — Fix the read-tool permission scoping regression. It's cross-workflow, wastes real compute (up to 25 min/run before failing), and has zero existing tracking. See the linked sub-issue for root cause and remediation steps.
• P1 — Get full raw logs for Design Decision Gate run §30754639273. The pre-fetched 50-line tail doesn't reach the actual failure; pull the complete agent job log before filing a fix, since the current evidence (a WARN, not an ERROR) isn't strong enough to act on.
• P2 — No action on PR Sous Chef's safe_outputs failure. Single occurrence, isolated to one run; revisit only if it recurs in a future cycle.

Sub-issues created

1. Read-tool permission scoping regression crashes read-only workflows on their own checkout root (P0) — linked below.

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 129.2 AIC · ⌖ 36.9 AIC · ⊞ 5.3K · ◷

• expires on Aug 9, 2026, 11:21 AM UTC-08:00

[github-actions]

🎉 Automatically closed by Sub-Issue Closer

All sub-issues have been completed. This parent issue is now closed automatically.

Sub-issues status: 2/2 closed (100%)

Generated by 🔧 Sub-Issue Closer · auto · 33.3 AIC · ⌖ 3.34 AIC · ⊞ 7.4K · ◷