[workflow-analysis] Weekly Workflow Analysis (Oct 5–12): Two Persistent CI Failures Found

[github-actions[bot]]

Overview

Reviewed GitHub Actions activity for github/gh-aw (337 tracked agentic workflows: 332 active, 5 disabled). Broad log collection across all 337 workflows for the full week timed out (repo's automation volume is very high — ~50 runs/hour observed in sampling), so this report combines a repo-wide recent-run sample with deep root-cause audits on the two workflows found to be failing consistently.

Key Issues

1. Daily Go Test Parallelizer — recent regression, 10/10 runs failing
Every run since §38147689674 (run #673, 2026-10-11 14:33 UTC) through the latest §38226989695 (run #682) has failed. Root cause: the Codex CLI agent (gpt-5.6-terra) exits after ~33s with errorCodes=["invalid_request_body"] — an API-level rejection of the model request, not a Go test or workflow YAML bug. This began abruptly within the last day, suggesting a model/engine-side change (version bump, payload-shape incompatibility) rather than a code regression in this repo.

2. Windows Grower — longstanding startup_failure, 10/10 runs failing
Every daily run since §37112823450 (run #36, 2026-10-03) through the latest §38231300010 (run #45, 2026-10-12) has ended in startup_failure with zero jobs ever dispatched (list_workflow_jobs returns total_count: 0). This is a pre-dispatch validation failure, not an in-job error — GitHub is rejecting the workflow (which calls the reusable windows.lock.yml via uses: with secrets: inherit) before any job starts. The exact "Invalid workflow file" annotation text could not be retrieved via available tooling (no job/log artifacts exist to query). This has been broken for 9+ days, independent of recent commits.

3. Frequent action_required conclusions on interactive workflows
In a 1-hour sample window, Agentic Commands, Content Moderation, AI Moderator, and Squad all showed multiple action_required runs. This is likely expected approval-gating (e.g. environment protection rules) rather than a defect, but the frequency is worth periodic monitoring for contributor friction.

Reliability Metrics (sampled)

Workflow	Recent runs checked	Failures	Pattern
Daily Go Test Parallelizer	10	10/10 failure	New regression (~18h old)
Windows Grower	10	10/10 startup_failure	Longstanding (9+ days)
Repo-wide (1h sample, 50 runs)	50	2 non-success (excl. action_required/skipped)	Normal background noise

View Details

Daily Go Test Parallelizer — job log excerpt (run #682):

[113] agent/agent session.result sourceType=error
[114] agent/agent session.result status=failed sourceType=turn.failed
[119] execution/agent agent.execution categories=[] errorCodes=["invalid_request_body"] errorTypes=[] exitCode=1

The agent took one successful apply_patch action, then the turn failed outright with an API-level invalid_request_body. The downstream "Handle agent failure" step in the conclusion job then surfaces a generic "Agent execution exited with code 1" message, obscuring the real cause.

Windows Grower — zero-dispatch signature:
get_workflow_run shows conclusion: startup_failure with no jobs; list_workflow_jobs returns total_count: 0; get_workflow_run_logs_url 404s (no logs exist because no job ever ran). This points to GitHub's workflow-file validation rejecting the run before scheduling — most likely in how the reusable windows.lock.yml is invoked or in its secrets: inherit / permissions wiring.

Next Actions

• Daily Go Test Parallelizer: Check whether the Codex CLI / gpt-5.6-terra engine version was recently bumped; invalid_request_body suggests a payload-shape mismatch between the harness and the current model API. Compare the last-known-good run (before Test Issue for Commands #673) against the first failure for a harness/engine diff.
• Windows Grower: Pull the "Invalid workflow file" annotation directly from the Actions UI for run Fix #664 - track added actions and only emit ai-reaction action if feature is used #45 (not retrievable via current tooling) to get the exact validation error on windows-grower.lock.yml / windows.lock.yml. Given 9+ days of 100% failure with zero job dispatch, consider pausing the schedule until fixed to stop accumulating silent failures.
• Consider adding alerting (e.g. a lightweight daily check) for workflows with sustained 100% failure streaks, since both issues here persisted for a day and 9 days respectively before being surfaced by this review.

References:

• §38226989695 — Daily Go Test Parallelizer, latest failure
• §38231300010 — Windows Grower, latest startup_failure
• §37112823450 — Windows Grower, first observed startup_failure (run #36)

Generated by 🔍 Weekly Workflow Analysis · claude · agent · 255.8 AIC · ⊞ 6.3K · ◷

• expires on Oct 13, 2026, 2:03 AM UTC-08:00