DeepReport Intelligence Briefing - 2026-05-20

[github-actions[bot]]

🔍 Executive Summary

The gh-aw fleet's mixed-health pattern from yesterday has tipped slightly worse: the open [aw] failed cohort grew again (21 → 22), the code-quality score dropped sharply from 89.6 → 72.6/100 in one day (churn-driven), and — most urgently — today's Safe Output Health Monitor report (#33469) shipped with a body that is literally the string PLACEHOLDER, burning 30.4M tokens to produce zero analysis. Meanwhile, the documented microsoft/apm integration path is now broken end-to-end on every released version and on main (#33572). The Agentic Maintenance orchestrator (P1, #33555) has been down for two consecutive days, and the 18-day quality/effectiveness plateau (Q:74 / E:71) is unchanged.

🚨 Top 5 Findings

1. Safe Output Health Monitor produced literal PLACEHOLDER body — the canonical daily safe-output health report is unusable today; 30.4M tokens spent for no analysis. Prompt regression or stripped templating. (#33469, §26143483059)
2. microsoft/apm shared workflow is uncompilable — the documented integration path (gh aw add microsoft/apm/.github/workflows/shared/apm.md) fails on v0.74.4, v0.74.7, and main. Two schema errors: import-schema/apps rejects properties, and jobs/apm/strategy requires matrix to be an object. External users cannot install APM. (#33572)
3. safe-outputs base_branch derivation produces multi-MB phantom patches — in multi-repo workflows, git rev-parse --abbrev-ref HEAD runs in the target repo's feature-branch checkout. Patch-generation Strategy 3 then computes a merge-base against an ancient commit and emits a 5.7 MB / 122k-line patch containing the entire repo history — for a real change of 8 files. PR creation rejects with E003 or falls silently to an issue. (#33545)
4. [aw] failed cohort still growing — 22 open today vs 21 yesterday vs 11 two days ago. Smoke-test failures (Smoke Pi, Codex, Copilot, Claude, Gemini, Agent Container) dominate; Auto-Triage Issues silently noops due to sandbox correctly denying mkdir/python3 while the prompt asks the agent to write a Python script (#33560).
5. Code-quality score regressed 17 points in 24h — 89.6 → 72.6/100 (#33383). Driven by churn (1,247 source files changed in 7 days) and 335 large files. Test ratio still healthy at 2.076, but the trend is the sharpest single-day drop in 30 days.

✅ Actionable Agentic Tasks

Seven concrete, low-effort improvements were filed as issues during this run. Each carries a clear success criterion and is sized for a single agent session.

#	Task	Effort	Driver
1	Fix Safe Output Health Monitor PLACEHOLDER regression	Small (<2h)	Top finding above; #33469
2	Emit deprecation warning when frontmatter uses infer	Small (<2h)	Schema Consistency audit (#33486) finding #1
3	Add applyTo / inputs to main_workflow_schema.json	Small (<2h)	Schema-parser drift; #33486 finding #2
4	Add generic x-deprecation-message walker to compiler	Medium (2-4h)	Class-of-failure fix; subsumes #2 above and prevents future repeats
5	Strip trailing periods from 34 CLI command short descriptions	Small (<1h)	CLI Consistency #33565
6	Add report-formatting guidelines to 8 non-compliant workflow prompts	Small (<2h)	Workflow Style #33554
7	Implement OTel partial_success taxonomy in gh-aw.run.status	Small (<2h)	Reliability gap #33518

Issues filed under deep-report + ai-generated labels — see the issue list for full descriptions.

📊 Notable Metrics

• Tokens (24h): 13.6M across 32 authoritative LLM calls in 19 workflows. Top 3 consumers = ~50% (Matt Pocock Skills Reviewer 2.42M, Test Quality Sentinel 2.31M, Glossary Maintainer 2.06M). Input:output ≈ 35:1.
• PR merge gap: copilot-swe-agent 80.8% vs app/github-actions 32.8% (43/64 PRs closed without merge in 30d) — P2 #33556.
• Sentry observability: 100% null on span.status, release, service.version, gen_ai.response.finish_reasons, gh-aw.agent.conclusion — OTLP mapping broken, blocks all standard error queries (#33544 P1).
• Lint debt: 2,456 findings — 2,409 function-length violations dominate (#33446).
• Lockfiles: 231 workflows, 22.17 MB total, 0 malformed, all use scoped permissions (#33389).

References

• §26143483059 — Safe Output Health Monitor (PLACEHOLDER run)
• §26165726464 — Agent Performance Analyzer (latest)
• §26165337753 — Auto-Triage Issues (silent noop fallback)

Generated by 🔬 DeepReport - Intelligence Gathering Agent · ● 11.6M · ◷

• expires on May 27, 2026, 4:04 PM UTC