DeepReport Intelligence Briefing - 2026-07-13

[github-actions[bot]]

🔍 Executive Summary

The gh-aw agent fleet is healthy but stuck in config-drift fatigue: success holds in its stable ~85% band (85.1% over 148 runs, 85.6% over 201), safe-output jobs are 100% clean (62/62), and every non-copilot engine (claude, pi, codex, gemini, antigravity) ran 100% reliable. The top finding is that the entire failure surface is copilot-engine 0-turn driver failures for the 7th+ consecutive audit, while the antigravity engine-catalog drift flagged since 07-09 remains unfixed across code, schema, and docs. Urgent action: land the small, concrete engine-catalog/schema fixes (they keep getting re-filed without merge) and repair PR Sous Chef's pre-agent PR-queue fetch, which fails 50% of runs before the agent ever starts.

Continuity: last full DeepReport was #44775 (2026-07-10), which filed 7 issues (detection batch, dispatch-repository kebab rename, engine docs, MCP list_pull_requests bounding, get_me→context identity, Smoke CI API footprint). This run avoids duplicating those and pruned a stale 21 KB orphan memory/default/ folder that had reappeared in repo memory.

🚨 Top 5 Findings

• Copilot is the entire failure surface (7th+ audit) — copilot-engine runs 74.4% vs 100% for claude/pi/codex/gemini/antigravity; all real failures are 0-turn driver failures at the Execute GitHub Copilot CLI step. Smoke CI 100% failure (5/5) is the loudest instance and runs with detection explicitly disabled.
• Engine-catalog drift (antigravity) unfixed 4+ days — [Schema Consistency] 🔍 Schema Consistency Check - 2026-07-13 #45205 reconfirms: antigravity missing from main_workflow_schema.json and engines.md; the NewEngineCatalog comment in engine_definition.go still omits pi + antigravity. Re-filed repeatedly without merge = churn.
• PR Sous Chef fails 50% before the agent starts — 3/6 runs (16.7% in the detection window) die at the pre-agent shell step Fetch open non-draft PR queue with AIC=0. A fixable gh/shell data-fetch bug, not a model problem.
• 0% write-actuation across the fleet — no write-capable safe outputs emitted across 148–201 runs (every SafeItemsCount=0), yet the API report separately counts 254 safe-output writes. Either intentional staged operation or a silent actuation regression — worth reconciling.
• API consumption concentrated in the PR review cluster — 79.2K REST calls/day; top-5 workflows = ~61% of quota, led by PR Code Quality Reviewer (15.1%). No hourly-limit breach, but caching contents/pulls reads is the clear optimisation target as PR volume grows. (MCP list_pull_requests remains the worst token bloat at 5,520 tok/PR.)

✅ Actionable Agentic Tasks

Seven GitHub issues were filed from this analysis (all new; none duplicate the 07-10 #44775 batch):

1. Fix PR Sous Chef pre-agent Fetch open non-draft PR queue step (50% fail, AIC=0) — Quick — from [workflow-analysis] Weekly Workflow Analysis — 148 recent runs: 85% success, Smoke CI 100% failure, 0% actuation #45250/[detection-analysis] Detection Analysis Report — 2026-07-13 (last 24h) #45269.
2. Add gh-aw-detection: true to 5 flagged audit workflows (Typist, Daily AIC Usage Audit, MCP Structural Analysis, Prompt Clustering, Weekly Workflow Analysis) — Quick — from [detection-analysis] Detection Analysis Report — 2026-07-13 (last 24h) #45269.
3. Fix stale engine-catalog comment in engine_definition.go (add pi, antigravity) — Fast — from [Schema Consistency] 🔍 Schema Consistency Check - 2026-07-13 #45205.
4. Add antigravity to main_workflow_schema.json + engines.md — Quick — from [Schema Consistency] 🔍 Schema Consistency Check - 2026-07-13 #45205.
5. Document engine.driver in frontmatter.md reference — Fast — from [Schema Consistency] 🔍 Schema Consistency Check - 2026-07-13 #45205.
6. Confirm whether 0% safe-output actuation is intentional (reconcile vs 254-writes API figure) — Medium — from [workflow-analysis] Weekly Workflow Analysis — 148 recent runs: 85% success, Smoke CI 100% failure, 0% actuation #45250/[safe-output-health] Safe Output Health Report - 2026-07-13 #45199/[api-consumption] 📊 GitHub API Consumption Report — 2026-07-13 #45255.
7. Reduce PR review-agent cluster API consumption via caching (61% of quota, PR Code Quality Reviewer 15.1%) — Medium — from [api-consumption] 📊 GitHub API Consumption Report — 2026-07-13 #45255.

View Full Details — metrics snapshot

Metric	Value (2026-07-13)
Runs/day	~500+; success 85.1% (148-run slice), 85.6% (201-run day)
Engine failures	100% copilot; claude/pi/codex/gemini/antigravity all 100% success
Safe-output jobs	62/62 success (100%); actuation 0 write-capable outputs (24h)
REST API	79.2K calls/day; top-5 = 61%; PR Code Quality Reviewer 15.1%
Detection	enabled vs regular both 82.9%; detection ~1.6× tokens
Issues (7d)	116 open / 384 closed; 0 open >7d; 1 unlabeled (#45244)
Discussions (7d)	39 agentic reports
MCP	10,161 tok/day, avg usefulness 3.36/5; list_pull_requests 5,520 tok/PR

References:

• §29249325390 — Typist run flagged for missing detection
• §29247097459 — GitHub API consumption run
• §29242348400 — Weekly Workflow Analysis run

Generated by 🔬 Deep Report · 164.8 AIC · ⌖ 17.1 AIC · ⊞ 10K · ◷

• expires on Jul 20, 2026, 7:44 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Deep Report.

A newer discussion is available at Discussion #45514.