[aw-failures] pull_request max-stack filter crashes non-PR workflow runs — 29 workflows hit startup_failure

[github-actions]

Executive Summary

Fix the max-stack PR filter now — it silently breaks every workflow with a pull_request trigger. Commit f9181ad ("Add stacked PR max-stack trigger protection") regenerated 30 .lock.yml files with a new activation if: clause that performs arithmetic on github.event.pull_request.stack.position/.size — fields that don't exist outside a real stacked-PR event. On any non-pull_request trigger (push, workflow_dispatch, etc.) this crashes GitHub Actions' expression evaluator, and the entire run fails at startup_failure before a single job executes. This is a repo-wide P0 regression, not an isolated flake.

Failure Cluster Table

Cluster	Severity	Workflows affected	Representative run	Comparator run	Signature
C1 — Stack-filter arithmetic crash	P0	~29 of the 32 failures in this window: all smoke-*.lock.yml (Pi, Antigravity, agent-public-approved, agent-public-none, agent-all-merged, agent-all-none, agent-scoped-approved, call-workflow, ci, claude, codex, copilot-arm, create-cross-repo-pr, gemini, multi-pr, opencode, project, temporary-id, test-tools, update-cross-repo-pr), plus ai-moderator, changeset, design-decision-gate, firewall-escape, impeccable-skills-reviewer, mattpocock-skills-reviewer, pr-code-quality-reviewer, pr-description-caveman, refiner, test-quality-sentinel, visual-regression-checker	30668964442 (Smoke Pi)	30576616018 (last green Smoke Pi)	startup_failure: 0 jobs run, 0s duration, created_at == updated_at
C2 — Detection continue-on-error still broken	P1 (tracked)	Detection Analysis Report	30672938234	—	GH_AW_DETECTION_CONTINUE_ON_ERROR=true set but job still marked failure
C3 — Ollama BYOK backend failure	P2 (tracked)	Daily BYOK Ollama Test	30670580818	—	Failure in "Execute GitHub Copilot CLI" step, ollama orphan process terminated in cleanup

Evidence

Fix the null/undefined .stack dereference in applyPullRequestStackFilter — it's the confirmed single point of failure. pkg/workflow/filters.go (added in f9181ad) generates, for every workflow whose on: includes pull_request (default max-stack: 1, no opt-in required):

github.event_name != 'pull_request' || github.event.pull_request.stack == null || github.event.pull_request.stack.position + 1 > github.event.pull_request.stack.size

This is spliced into the job-level activation if: via BuildConditionTree/RenderCondition (see diff to smoke-pi.lock.yml, activation and pre_activation jobs). github.event.pull_request.stack is not a real GitHub webhook field for ordinary PRs — on any event where github.event.pull_request itself is absent (push, workflow_dispatch, schedule...), GitHub Actions' expression evaluator hits an unsupported arithmetic operation (null + 1) while evaluating the job condition, which invalidates the whole run before job scheduling.

Verified via audit-diff between the failing run and the last successful Smoke Pi run:

• Failing run (30668964442): 0 GitHub API calls, 0 token usage — the run never reached the agent/activation logic.
• Comparator (30576616018): 12 API calls, normal core-quota consumption — a normal execution.
• run_metrics_diff.github_rate_limit_details.api_calls_change: "+∞" confirms the failing run made no calls at all, consistent with a pre-job startup_failure, not an in-job error.

This is not new as of this window: an earlier Smoke Pi run (30606865041, 2026-07-31T05:27:45Z, ~17h before this window) shows the identical zero-duration/zero-job signature — the underlying filter bug predates this 6h lookback; commit f9181ad expanded its blast radius by regenerating 30 additional lock files with the same broken clause.

Existing Issue Correlation

• C1 has no existing tracking issue — none of the 13 open agentic-workflows issues reference max-stack, stack filter, or this signature. This is why a new parent is being opened (per policy: new parent only when a P0 cluster has zero coverage).
• C2 matches [aw-failures] Detection job reports failure conclusion despite GH_AW_DETECTION_CONTINUE_ON_ERROR=true #49351 ("Detection job reports failure conclusion despite GH_AW_DETECTION_CONTINUE_ON_ERROR=true") — still open, still recurring (this run is 2h newer than the issue's last update). No new issue needed; that issue remains valid and unresolved.
• C3 matches [aw-failures] MCP Gateway github SSE reconnect hang + Ollama BYOK backend 503 fast-fail #49246 ("MCP Gateway github SSE reconnect hang + Ollama BYOK backend 503 fast-fail") — same workflow, consistent failure mode. No new issue needed.

Fix Roadmap

• P0 — Ship now: Guard applyPullRequestStackFilter's generated condition so it never dereferences .stack.position/.size unless github.event_name == 'pull_request' is already true (short-circuit via nested &&/fromJSON default, not a flat || chain that GitHub's evaluator still type-checks). See sub-issue below.
• P1 — Same PR: Add a compiler-level regression test that renders the activation/pre_activation if: for a workflow with on: [push, pull_request] and asserts the expression evaluates without error under a synthetic push event context.
• P2 — Follow-up: Re-run gh-aw compile across the repo once fixed and confirm all 30 previously-regenerated lock files no longer contain the broken clause; re-enable/verify the smoke-test matrix goes green.

Sub-Issues Created

• Sub-issue opened against pkg/workflow/filters.go's applyPullRequestStackFilter for the concrete arithmetic-on-null fix (linked below).

Generated by 🔍 [aw] Failure Investigator (6h) · agent · 102.6 AIC · ⌖ 32.8 AIC · ⊞ 6.8K · ◷

• expires on Aug 7, 2026, 5:27 PM UTC-08:00