[aw-failures] [aw] Failure Investigator Report — 2026-08-18 19:00 UTC (6h)

[github-actions]

Executive summary

Fix the malformed-JSON crash in PR Sous Chef's safe_outputs job first — it kills the job on every single trigger and accounts for 18 of the 24 failed runs (75%) in this window. Root cause is pinned to one line of generated YAML; no existing issue tracks it. Everything else this window is a single-occurrence, lower-severity blip already covered by prior reports.

24 failed runs in the last 6h → 1 confirmed new P0 (18 runs, one root cause), 4 singleton P2/likely-tracked blips, 0 issues closed (no fresh disconfirming evidence found for any of the 9 open agentic-workflows issues checked).

Failure cluster table

Cluster	Severity	Workflow	Runs	Representative	Comparator	Signature	Status
A	P0	PR Sous Chef	18/24	§32171918501	§32170093369 (identical crash, prior run)	safe_outputs job crashes at startup: ERR_VALIDATION: Handler manager failed: ERR_PARSE: Failed to parse GH_AW_SAFE_OUTPUTS_HANDLER_CONFIG	New gap — no tracking, fix below
B	P1	Scout	2/24	§32153071114	§32152602055	Start MCP Gateway step fails	Likely matches #53191 (MCP Gateway digest-pin) — not re-verified this window
C	P1	Architecture Guardian, Linter Miner	2/24	§32147417635	§32166116673	Execute GitHub Copilot CLI step fails	Likely matches #52253 (Copilot proxy port 10002 refusals) — not re-verified this window
D	P2	Daily Fact	1/24	§32147275242	—	Check KVM availability for docker-sbx fails	Single occurrence, no action taken
E	P2	Super Linter Report	1/24	§32146482329	—	Run super-linter step fails	Single occurrence, no action taken

Evidence

Cluster A — malformed JSON crashes PR Sous Chef's safe_outputs job on every run (P0)

What happens: every recent PR Sous Chef run completes its agent job successfully, then the safe_outputs job dies immediately with:

##[error]ERR_VALIDATION: Handler manager failed: ERR_PARSE: Failed to parse GH_AW_SAFE_OUTPUTS_HANDLER_CONFIG: Expected ',' or '}' after property value in JSON at position 200 (line 1 column 201)

Confirmed identically in 5 sampled runs spanning the full 6h window (§32171918501, §32170093369, §32168627250, §32167555737, §32166311424); near-certain for the remaining 13 PR Sous Chef failures since the bug is a deterministic code-generation defect, not a flaky condition.

Root cause, pinned: .github/workflows/pr-sous-chef.lock.yml:2566 embeds the allow-list as a quoted JSON string:

\"allowed_pull_requests\":\"${{ needs.approval_allowlist.outputs.eligible_pull_request_numbers }}\"

eligible_pull_request_numbers (lock.yml:1515) is produced by jq -c '[.prs[]?.number | tostring]', i.e. it already renders as a JSON array string like ["53698","53725"]. GitHub Actions splices that text in verbatim with no re-escaping, so the surrounding quotes collide with the array's own quotes and the JSON breaks exactly at allowed_pull_requests (verified locally: parsing the captured config text reproduces the same "expected ',' or '}'" error).

Compiler-side, the culprit is AddTemplatableStringSlice in pkg/workflow/compiler_safe_outputs_builder.go:63-75: when a field's value is a single-element slice matching ${{ ... }} (isExpression, pkg/workflow/expression_patterns.go:95), it stores the raw expression as a scalar string, which json.Marshal then wraps in quotes — safe only when the expression's runtime value is a plain scalar. It silently breaks for any expression whose output is itself JSON (arrays, objects), which is exactly what eligible_pull_request_numbers is. approve_workflow_run.cjs's parseAllowedPullRequests (actions/setup/js/approve_workflow_run.cjs:36-49) already expects a real array or bracket-string, so this is a pure emission bug, not a consumer bug.

Blast radius: because the crash happens before create_issue's own close_older_issues/group_by_day dedup logic ever runs, PR Sous Chef's separate "Failed jobs" auto-notifier has created 30+ open, undeduplicated [aw] Failed jobs: PR Sous Chef issues (#53245 → #53802) since ~2026-08-17. These will keep multiplying every ~15–20 min until the emission bug is fixed. Not closing them here — they self-expire (each carries a 7-day expiry checkbox) and closing them individually would blow the update_issue budget for no lasting benefit; fixing the root cause stops the flood at the source.

Existing issue correlation

• Searched is:issue label:agentic-workflows state:open (9 issues) plus a targeted ERR_PARSE / allowed_pull_requests / SAFE_OUTPUTS_HANDLER_CONFIG search — no open issue covers this crash. [aw-failures] safe_outputs job hard-fails entire batch on one non-retryable error #53263 ("safe_outputs job hard-fails entire batch on one non-retryable error") is the closest by keyword but describes a different failure mode (a valid per-item API error propagating as a job failure, not a JSON parse crash before any item is processed) — not a match, left open.
• [aw-failures] [P1] PR Sous Chef: Start DIFC Proxy step fails with no captured root cause #52502 ("PR Sous Chef: Start DIFC Proxy step fails") is a different step, different (unrelated, single-occurrence) failure from 2026-08-13 — left open, no fresh evidence either way.
• No issues closed this round: none of the 9 open agentic-workflows issues had disconfirming evidence surface in this window's data.

Fix roadmap

• P0 — do now: stop quoting single-expression fields whose runtime value is itself JSON. Either (a) emit "allowed_pull_requests":${{ needs.approval_allowlist.outputs.eligible_pull_request_numbers }} unquoted so the array splices in as real JSON, or (b) give AddTemplatableStringSlice a way to mark an expression as "array-valued" and skip the scalar-string branch for it. Option (a) is the smaller diff — sub-issue below.
• P1 — verify, don't fix yet: re-run audit on the latest Scout ([aw-failures] [P1] Two new job-failure root causes: AWF binary curl reset in Daily SPDD evals + MCP Gateway serena digest-pin sc [Content truncated due to length] #53191-adjacent) and Copilot-proxy ([aw-failures] [P1] Copilot engine proxy (port 10002) refuses connections despite passing key validation — Daily Assign Issue To [Content truncated due to length] #52253-adjacent) failures next window to confirm they're still the same tracked signature before any new issue is opened.
• P2 — monitor: Daily Fact's KVM check and Super Linter Report are single occurrences; no action unless they recur.

Sub-issues created

• #aw_fixp0 — Fix malformed JSON emission for approve_workflow_run.allowed_pull_requests in the safe-outputs config builder (linked as sub-issue of this report).

References:

• https://github.com/github/gh-aw/actions/runs/32171918501
• https://github.com/github/gh-aw/actions/runs/32170093369
• https://github.com/github/gh-aw/actions/runs/32168627250

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 165.4 AIC · ⌖ 12.6 AIC · ⊞ 5.9K · ◷

• expires on Aug 25, 2026, 11:15 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: 1/1 closed (100%)

Generated by 🔧 Sub-Issue Closer · auto · 34.7 AIC · ⌖ 2.18 AIC · ⊞ 7.9K · ◷