DeepReport Intelligence Briefing - 2026-05-19

[github-actions[bot]]

🔍 Executive Summary

The gh-aw fleet is in mixed health: two of yesterday's highest-severity issues (OTLP 401 storm, silent 422 PR-review failures) are confirmed resolved, but the open [aw] failed cohort nearly doubled overnight (11 → 21 issues) — driven primarily by smoke-test failures and a new Copilot engine-termination pattern (copilot engine terminated before producing output). The top urgent finding is a fresh schema-vs-runtime mismatch on stdio_mcp_tool.network: the schema explicitly tells users the field is ignored, but the parser at pkg/parser/mcp.go:625-637 actively reads network.proxy-args and shapes compilation output.

🚨 Top 5 Findings

1. stdio_mcp_tool.network schema actively lies — the schema declares deprecated: true and "This field is ignored"; the parser reads proxy-args and uses it. Critical schema-vs-runtime mismatch. (#33233)
2. [aw] failed cohort doubled in 24h: 21 open today vs 11 yesterday. Heavy smoke-test failures (Smoke CI, Pi, Claude, Copilot, Gemini, OTEL), and 5 daily Copilot workflows now failing with a new engine-termination-before-output signature.
3. tools.grep deprecation message is dead documentation — schema carries an x-deprecation-message, but no compiler code emits it. Same class of failure as findings on tools.github.toolset (deprecated but never read by parser). Fix: add a generic x-deprecation-message walker so future deprecations honor themselves. (#33233)
4. safe-outputs body-only PR-review fallback drops inline-comment bodies on the floor — confirmed today on §26077668174: 6 buffered comments (~4900 chars of agent analysis) discarded when the handler fell back. Fallback succeeds, work doesn't survive. (#33222)
5. 18-day agent performance plateau (Q:74, E:71, H:63) with Agentic Maintenance down day 2, CGO/CJS failing 90+ days, and Codex sandbox-blocked for 12 workflows — these three structural blockers compound the plateau. (#33302)

✅ Actionable Agentic Tasks (Filed Today)

Each task was filed as a GitHub issue with concrete file references, expected impact, and a suggested agent.

1. Reconcile stdio_mcp_tool.network schema vs parser — schema says deprecated, parser reads proxy-args. Pick one direction. (Critical mismatch — single-commit fix.)
2. Add generic x-deprecation-message walker; emit warning for tools.grep — eliminate the whole class of "schema says deprecated, parser accepts silently" with one parser-init pass.
3. Decide fate of tools.github.toolset singular — deprecated in schema but parser never reads it. Either add a singular reader (with migration warning) or drop the schema entry.
4. safe-outputs PR review fallback: carry dropped inline-comment bodies into the body-only fallback — preserve agent analysis when GitHub 422s on unresolvable line refs.
5. Add defer recover() to 2 goroutines in pkg/cli/mcp_inspect_inspector.go — restore the codebase's panic-isolation convention at lines 139 and 201. ~12 line diff.
6. Treat zero-byte rpc-messages.jsonl as MCP telemetry failure, not silent warning — observed in §26060881909. Closes a recurring false-green gap.
7. Audit and restrict allowed: ["*"] on 8 high-privilege MCP servers — azure, jupyter, semgrep need explicit tool lists; read-only public servers may keep wildcards with documentation.

📌 Methodology

• Data sources: 100 most recent discussions (50 created in past 7d), 500 weekly issues (64 OPEN / 436 CLOSED), 10-run sample from agentic-workflows logs MCP tool, prior repo-memory snapshot from 2026-04-03 (46d stale — full refresh executed).
• De-duplication: candidate findings cross-checked against open issues from past 7d before filing. Items already tracked (errstringmatch, tools.serena, engine.bare, README engine mention, Agentic Maintenance compile, AWF chroot Node) were not re-filed and are tracked in repo-memory flagged_items instead.
• Memory updates: /tmp/gh-aw/repo-memory/default/memory/deep-report/ refreshed with current trends, resolved items, and 7 new flagged items.

🔭 Watching Next Briefing

• Whether the smoke-test failure cluster (6 distinct smoke workflows down) resolves within 24h or persists.
• Whether the 5 daily Copilot workflows shift back to chroot-Node failure mode or remain on engine-termination-before-output.
• DIFC pressure on Issue Triage Agent (82% of DIFC events 5/18).
• Test-to-source LOC ratio dip (-1.78% in 7d) — continue if persists 2+ weeks.

References:

• #33222 — Safe Output Health (resolved 2 high-severity clusters)
• #33233 — Schema Consistency Check (source for 3 of 7 tasks)
• #33302 — Agent Performance plateau
• #33212 — Sergo Run 13 (errstringmatch + goroutine convention)
• #33135 — MCP Inspector inventory
• #33174 — Observability Coverage

Generated by 🔬 DeepReport - Intelligence Gathering Agent · ● 23.5M · ◷

• expires on May 26, 2026, 4:04 PM UTC

[github-actions[bot]]

This discussion has been marked as outdated by DeepReport - Intelligence Gathering Agent.

A newer discussion is available at Discussion #33585.