DeepReport Intelligence Briefing - 2026-07-17

[github-actions[bot]]

🔍 Executive Summary

The gh-aw fleet is healthy on baseline and — for the first time in five days — writing again: the 4-consecutive read-only streak (07-13 → 07-16) broke today with a write-rich, 100%-clean safe-outputs day (59 items across 18 write-runs, 0 failed actuation). Non-smoke production success sits at 91.7% (66/72) and there were zero agent-logic, missing-tool, or MCP failures fleet-wide. The top finding is unchanged: Smoke CI's copilot-CLI 0-turn driver remains the entire failure surface and the single largest API burner (23.6% of REST quota) — the agent-performance report now pins the fix on merging PR #44276 (EACCES mkdir), so no new issue is filed for it. Urgent watch: the logs MCP 120s timeout still forces partial datasets on three audits daily, and two silent-failure bug reports (#46184, #46259) sit unlabeled.

Continuity: prior full DeepReport was 2026-07-16 (#46029, 7 issues filed). This run does not re-file open items. The fleet self-files most code-quality quick wins same-day (logs_download options structs → #46109/#46111/#46249; copilot/npm bool-param structs → #46112/#46113), and the logs-timeout fix (#45510, closed) plus token-total gap (#45511, closed) are deliberately not re-filed (re-filing-loop trap) — kept as watch items. Also pruned the stale memory/default/ orphan (reappeared again; it pushed repo memory to 24 KB over the 12 KB cap) so memory validates at 4 KB.

🚨 Top 5 Findings

• Read-only streak BROKE (positive). After 4 zero-actuation days, 07-17 actuated 59 safe outputs across 18 write-runs with 100% job success — create_pull_request_review_comment ×14, add_comment ×14, create_issue ×9, add_labels ×7, plus PRs/discussions. The write path is confirmed healthy end-to-end.
• Smoke CI is still the whole failure story — and the fix is now named. 0/18 runs succeed; 23.6% of REST quota (top-3 workflows = 42.0%). Non-smoke production is 91.7%; the raw 60.6% fleet rate is a Smoke-canary artifact. Agent-perf: blocked on merging PR fix: reclaim root-owned sandbox firewall dirs to prevent EACCES on reused runners #44276 (EACCES mkdir) — "no new issue needed."
• logs MCP 120s timeout keeps degrading fleet metrics. Safe Output Health (82/131 conclusions), API Consumption (109 run dirs, morning only), and the workflow audit (109/322) all ran on partial datasets today — daily health/cost numbers are computed on incomplete data. Fix [deep-report] Fix recurring 60s timeout in gh-aw logs MCP tool degrading audit datasets #45510 was closed but the symptom persists.
• Two high-severity silent-failure bugs are unlabeled. safeoutputs gateway registers 0 tools → every safe-output call fails "unknown tool" → silent green no-op (reproduces on v0.82.9, after #43057) #46184 (safeoutputs gateway registers 0 tools → every call fails 'unknown tool' → silent green no-op, v0.82.9 after Fail fast on empty safeoutputs schema and add Copilot soft-time convergence guard #43057) and Copilot engine: <mcp-clis> omits CLI-mounted github/azure-devops while <github-mcp-tools> still assumes native tools → intermittent GitHub access / silent no-ops #46259 (Copilot <mcp-clis> omits CLI-mounted github/azure-devops → intermittent access/silent no-ops). Both mask real breakage as green; neither is labeled — Auto-Triage coverage gap.
• Error-chain transparency debt is large but tractable. 788 non-wrapping fmt.Errorf (37%) discard upstream cause, concentrated in pkg/workflow/cli/parser; zero errgroup despite 8 fire-and-forget goroutines. The [repository-quality] report scopes 5 fixes — 4 filed below.

✅ Actionable Agentic Tasks

Seven new GitHub issues filed this run, each dedup-checked against open + closed issues (none duplicate the fleet's same-day self-filed batch or yesterday's items):

1. Introduce sentinel errors for recurring string matches — promote rematched error strings to typed sentinels; remove 2 errstringmatch nolint suppressions — Quick — from [repository-quality] [repository-quality] Repository Quality Improvement Report — Error Chain Transparency & Non-Wrapping fmt.Errorf Gap #46253.
2. Audit + wrap upstream YAML errors in the parser import pipeline (%w) — stop swallowing structured causes in user-visible import errors — Medium — from [repository-quality].
3. Extend a linter to flag fmt.Errorf with an error arg but no %w — enforce the invariant across the 788-call surface — Medium — from [repository-quality].
4. Adopt errgroup for structured concurrency, starting with mcp_inspect_inspector.go — Medium — from [repository-quality].
5. Populate SafeItemsCount in run_summary.json so audits stop reporting 0 safe-output writes and dropping to fallbacks — Medium — from [api-consumption] [api-consumption] 📊 GitHub API Consumption Report — 2026-07-17 #46226.
6. Triage + label 2 unlabeled open bug reports (safeoutputs gateway registers 0 tools → every safe-output call fails "unknown tool" → silent green no-op (reproduces on v0.82.9, after #43057) #46184, Copilot engine: <mcp-clis> omits CLI-mounted github/azure-devops while <github-mcp-tools> still assumes native tools → intermittent GitHub access / silent no-ops #46259) — close the Auto-Triage gap on two silent-failure bugs — Fast — from weekly-issues analysis.
7. Add gh-aw-detection: true to Duplicate Code Detector + Daily Agentic Workflow AIC Usage Audit — clear 2 recurring detection-misconfig entries — Fast — from [detection-analysis] [detection-analysis] Detection Analysis Report — 2026-07-17 (last 24h) #46254.

View Full Details — metrics snapshot & watch items (deliberately NOT re-filed)

Metric	Value (2026-07-17)
Safe-output	59 items / 18 write-runs / 100% clean — breaks 4-day read-only streak
Workflow audit (07-16 partial ~5h, 109 runs)	non-smoke prod 91.7% (66/72); raw 60.6% (Smoke canary); 0 agent-logic/tool/MCP fails
REST API (partial, 109 runs, morning)	91,806 calls; Smoke CI 23.6% (was 42.3% on 07-16); top-3 42.0%
Detection (24h, 157 runs)	28.7% detection-enabled; 91.1% vs 79.5% success; ~2x tokens (147K vs 71K); 8 misconfigs
Per-engine (audit window)	pi 100% (20), codex 100% (1), claude 75% (0-turn driver), copilot 45.8% raw (Smoke-skewed)
Agent performance (week)	avg quality 61/100 (↓2), effectiveness 62/100; top Copilot SWE Agent 92; bottom Smoke CI/CWI/CGO/CJS
Issues (7d, 500)	125 open / 375 closed; 4 unlabeled (2 open bugs); 0 open >7d; 473/500 bot-authored
Discussions (7d)	45 (all github-actions); 32 Audits, 8 General, 4 Announcements, 1 Daily News

Watch items (already tracked / upstream — deliberately NOT re-filed):

• Smoke CI copilot-CLI 0-turn ([aw] Smoke CI failed #45429/[aw] Smoke CI failed #45931) — blocked on PR fix: reclaim root-owned sandbox firewall dirs to prevent EACCES on reused runners #44276 merge; rejig docs #1 failure + quota driver. No new issue per agent-perf.
• logs MCP 120s timeout — [deep-report] Fix recurring 60s timeout in gh-aw logs MCP tool degrading audit datasets #45510 closed but symptom recurs on 3 audits daily. Re-filing trap; watch.
• Token reporting: metrics.TokenUsage empty fleet-wide ([deep-report] Surface fleet-wide token total in audit telemetry (token-usage-reporting-gap, recur #4) #45511 closed, recur Add workflow: githubnext/agentics/weekly-research #4) — recovered from token_usage_summary. Watch.
• CWI regression (back to 100% agent-required, 8/8) + Content Moderation regressed — agent-perf watch.
• 6 remaining detection misconfigs (4 detection-enabled-but-failed runs: Avenger, daily-experiment-report, Copilot PR NLP, Daily Max AI Credits) — transient run failures, watch not file.

References:

• §29575919558 — GitHub API Consumption (Smoke CI 23.6%, logs timeout)
• §29510196605 — prior DeepReport (DeepReport Intelligence Briefing - 2026-07-16 #46029)
• §29590371002 — this DeepReport run

Generated by 🔬 Deep Report · 208.7 AIC · ⌖ 11.4 AIC · ⊞ 10K · ◷

• expires on Jul 24, 2026, 7:16 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Deep Report.

A newer discussion is available at Discussion #46873.