DeepReport Intelligence Briefing - 2026-07-23

[github-actions[bot]]

🔍 Executive Summary

The gh-aw fleet is healthy: 96.2% raw / 96.5% excl-intentional workflow success (07-22 audit, 314 runs; healthiest window in weeks) with all failures being 0-turn driver / safe-output-job fails and zero missing-tools/data/MCP/agent-logic errors. The top finding is a governance gap, not a health gap: issue #47346 — the pr_review_buffer.cjs:554 review-comment 422 fix — was closed without the fix being applied, and the exact failure recurred in production twice in 4 days (07-20, 07-23). Most urgent watch: two unmerged PRs (#44254 CLI-hang, #43527 Q-workflow gate) are silently pinning multiple agents at 100% activation-refused and are dragging the weekly effectiveness score down (36%→27% success since Jul 8).

🚨 Top 5 Findings

• Closed-but-unfixed review 422 (HIGH). [deep-report] Broaden PR-review 422 fallback to match "Path could not be resolved" (pr_review_buffer.cjs:554) #47346 was closed yet the pr_review_buffer.cjs:554 Path-variant 422 hard-failed in prod on 07-20 AND 07-23 — 2nd hit in 4 days, unvalidated for 55 audits. Re-filed with new evidence this run.
• Safe-output observability blind spot (HIGH). The Process Safe Outputs step log is not pre-bundled and the logs CLI times out at 300s, so every safe-output job failure (07-20 pair + 07-23 pair) can only be family-attributed, never root-caused. This blocks confirming the two fixes above.
• Merge-blocked reviewers dragging the fleet (HIGH). PR fix: reduce post-completion idle watchdog and add cleanup timeouts to prevent Copilot CLI hang on exit #44254 (CLI hang-on-exit) keeps PR Code Quality Reviewer, Test Quality Sentinel & Impeccable at 100% AR; PR Add shared prompt quality gate for plateaued agent-review workflows #43527 leaves Q Workflow 100%-skipped 15+ days. Weekly quality 63→60, effectiveness 64→62. Both are DO-NOT-RE-FILE (already tracked) — they need a merge, not a new ticket.
• API quota concentrated in 4 reviewers (MEDIUM). 106,414 REST calls/day; the four skills/quality reviewers alone are ~54% (~58k), each approaching the 15k/hr ceiling. 30-day API trend is +32.6%. Matt Pocock (top burner, 100% AR, redundant) is a clean deprecation win — filed.
• Positive: recoveries + strong core. Avenger recovered (4/4 pass, was 10/10 fail), token-reporting gap resolved, Copilot SWE Agent merged 30 PRs in 24h, and the fix-task prompt core converts at 81% (firewall fixes 91.7%). Backlog is clean: 124 open / 376 closed, 0 open >7d, 4 unlabeled (auto-triage-pending).

✅ Actionable Agentic Tasks

Seven data-driven quick-win issues were filed this run (each verified non-duplicate against the open backlog; #44254, #43527, #44241, reporting-otlp/guard-otlp/smoke-engine-base #47456/47458/47459, io.Writer #47438, refactor #47520/47577, logs-CLI timeout #47491, Smoke Codex #47525 were deliberately NOT re-filed):

1. Add recover() to WASM compile goroutine (cmd/gh-aw-wasm/main.go:48) — Fast; fixes the only CRITICAL goroutine gap (browser Promise hang on panic).
2. Harden HTTP server timeouts (bootstrap GitHub App + MCP server) — Quick; closes Slowloris-style slow-client resource exhaustion.
3. Apply the pr_review_buffer.cjs:554 Path-variant 422 fallback + test — Quick; [deep-report] Broaden PR-review 422 fallback to match "Path could not be resolved" (pr_review_buffer.cjs:554) #47346 was closed unfixed, recurred in prod twice in 4 days.
4. Pre-bundle Process Safe Outputs step stdout/stderr on failure — Medium; unblocks root-causing every safe-output job failure.
5. Soft-skip add_comment/update_pull_request on context-less scheduled runs — Medium; fixes PR Sous Chef hard-fail, matches review-comment handler behavior.
6. Evaluate Matt Pocock Skills Reviewer for deprecation — Quick; 100% AR + ~15k REST/day, redundant with 3 other reviewers.
7. Join CheckForUpdatesAsync goroutine + drop time.After leak (update_check.go:249) — Fast; aligns with existing in-repo goroutine conventions.

View supporting data

Metric	Value
Runs (07-22 audit)	314 (302 ✅ / 12 ❌) = 96.2% raw · 96.5% excl-intentional · all fails 0-turn
Prod-main success	94.2% raw · 94.9% excl-intentional
Per-engine	copilot 96.1% · claude 94.7% · pi 97.6% · codex 7/7
Fleet tokens (07-22)	10.84M (reporting gap resolved)
Safe-output in-scope success (07-23)	99.24% (262/264; 2 fail, 1 gated-skip)
REST API calls (07-23, partial)	106,414 · 30d +32.6% · 7d -3.4% · top-4 reviewers ~54%
Weekly issues (500/7d)	124 open / 376 closed · 0 open >7d · 4 unlabeled
Agent perf (wk)	quality ~60 (↓63) · effectiveness ~62 (↓64) · success 27% (↓36%) · AR 42%
Prompt clustering	1,000 PRs · 76.3% merge (flat) · fix 81% · docs-site 23.8% (worst) · release chores 54.8%
Goroutine hygiene	7 bare prod goroutines · 2 no-recover (WASM critical) · 1/2 http servers no-timeout
Workflows	261

Sources analyzed: Agentic Workflow Audit (#47409), Safe Output Health (#47473), API Consumption (#47556), Agent Performance (#47575), Auto-Triage (#47570), Repository Quality (#47586), Prompt Clustering (#47549), Workflow Skill Extractor (#47460), weekly issues (500), prior DeepReport (#47350).

Continuity: prior DeepReport was 2026-07-22 (#47350). Repo-memory refreshed (patterns, trends, flagged items); a 21 KB stale legacy memory/default/ set (Dec-2025) was removed to satisfy the 10 KB memory patch limit. Already-tracked items intentionally not re-filed: #44254 (reviewer CLI hang), #43527 (Q gate), #44241 (AI Moderator engine), and the reporting-otlp / io.Writer / refactor / logs-CLI-timeout code-quality tickets. #47346 was re-filed because it was closed without the fix landing and recurred in production.

References:

• §30009821130
• §30008301535
• §30002849032

Generated by 🔬 Deep Report · age00 · 242.8 AIC · ⌖ 11.4 AIC · ⊞ 10.3K · ◷

• expires on Jul 30, 2026, 7:32 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Deep Report.

A newer discussion is available at Discussion #47817.