DeepReport Intelligence Briefing - 2026-07-16

[github-actions[bot]]

🔍 Executive Summary

The gh-aw fleet is healthy but read-only, and one workflow now dominates both failure and cost: safe-output jobs are 100% clean for a 4th consecutive day (14/14 today, 0 write actuations), yet Smoke CI's copilot-CLI 0-turn failure has grown from 26.6% → 42.3% of the 24h REST-API quota (36,567 calls across 25 runs, all non-success) while remaining 0% success and unfixed since 06-30 (#45429/#45931). The top finding: one driver root-cause retires the entire failure surface (fleet is 81.8% → 90.0% success ex-Smoke CI) and the #1 quota burner simultaneously. Bright spot: the 6-week Copilot "provenance inversion" broke today — CI-gate workflow-types recorded real completions for the first time (38%, a 9-day high). Urgent action: root-cause the copilot-CLI startup driver; the logs MCP tool 120s timeout continues to force partial datasets on three audits daily.

Continuity: prior full DeepReport was 2026-07-15 (7 issues filed). This run deliberately does not re-file open items and confirms the fleet self-files most quick wins same-day — Typist clusters 1/2/3/6 → #46000/#46001/#46002, reporting-otlp bundle → #45954/#45895, smoke-base → #45898, eslint consolidations → #45952/#45955, antigravity schema/docs → #45808 (the re-filing-loop trap; NOT re-filed). Also pruned the stale memory/default/ orphan (again — it had reappeared) so repo memory validates under the 12 KB limit.

🚨 Top 5 Findings

• Smoke CI is the whole story — and its cost share is accelerating. 25/25 runs non-success today; 42.3% of REST quota (up sharply from 26.6% on 07-15), top-3 workflows = 55.7%. Unfixed since 06-30 ([aw] Smoke CI failed #45429). A single driver fix lifts fleet success from 81.8% → 90.0% and recovers the largest quota slice.
• Fleet is fully read-only for a 4th straight day. 14/14 safe-output jobs clean but 0 write messages actuated (07-13 62/62, 07-14 35/35, 07-15 51/51, 07-16 14/14). Four days of zero actuation is a pattern needing an intentional-vs-regression reconciliation — and it hides latent write-path bugs (see the Changeset Generator signature below).
• logs MCP tool 120s timeout keeps degrading fleet metrics. API Consumption (102 of ~270 run dirs), the workflow audit (100 of 225), and prior windows all fall back to partial datasets — daily health numbers are computed on incomplete data.
• Highest-priority latent write bug has no tracking issue. Changeset Generator push_to_pull_request_branch hard-fails with dubious ownership (bridge runs as a different user/HOME than the container, so git config safe.directory never reaches it). Latent since 06-26 only because the fleet is read-only — filed a tracking issue this run.
• POSITIVE — provenance inversion broke. For the first time in 6 tracked weeks, recurring CI-gate Copilot workflow-types recorded real successes (completion 38%, 9-day high, distributed across 7 branches). Loop detection 0%, orphaned branches 0%. Genuine reliability improvement worth watching to confirm it holds.

✅ Actionable Agentic Tasks

Seven new GitHub issues were filed — each dedup-checked against open + closed issues; none duplicate the fleet's same-day self-filed batch or yesterday's items:

1. Extract downloadArtifactsOptions struct + delete 2 //nolint:excessivefuncparams 9-param suppressions (logs_run_processor.go:104, update_actions.go:606, 4 siblings in logs_download.go) — Quick — from [repository-quality].
2. Replace consecutive boolean positional params with flags structs in buildCopilotStepEnv (3 bools) and GenerateNpmInstallStepsWithScope (4 bools) — Quick — from [repository-quality].
3. Unify the token-usage metrics quartet + fix camelCase/snake_case json-tag inconsistency across copilot_events_jsonl.go / token_usage.go (Typist Cluster 8) — Medium.
4. Extract shared base for AccessLogSummary/FirewallLogSummary in logs_report_firewall.go (Typist Cluster 4) — Medium.
5. Extract shared analysis base for DomainAnalysis/FirewallAnalysis (access_log.go/firewall_log.go, Typist Cluster 5) — Medium.
6. Pre-bundle the Process Safe Outputs step stdout on failure so safe-output hardfails are diagnosable from artifacts — Medium — from [safe-output-health] (its top recurring observability ask).
7. Tracking issue: Changeset Generator push_to_pull_request_branch dubious-ownership hardfail (bridge safe.directory / user+HOME alignment) — Medium — from [safe-output-health] watch-list item 1.

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

Metric	Value (2026-07-16)
Detection window (24h, 101 runs)	regular 78.9% / detection 90.0%; avg tokens 18.8K vs 28.9K
REST API (partial, ~102 runs, mid-day)	86.4K calls; Smoke CI 42.3% (36,567 / 25 runs, all fail); top-3 55.7%
Workflow audit (07-15 partial ~5.6h)	81.8% (90.0% ex-Smoke CI); 18 fails (17 copilot + 1 codex), all known CLI-driver, 0 novel
Per-engine	pi 100% (18), claude 100% (16), copilot 72.6%, codex 66.7%
Safe-output jobs	14/14 = 100%; 4th consecutive read-only day (0 actuations)
Copilot sessions	completion 38% (9-day high); provenance inversion BROKE; loop 0%, orphaned 0%
Issues (7d, 500)	102 open / 398 closed; 3 unlabeled (#46012/#45919/#45918); 0 open >7d; 475/500 bot-authored
Custom linters / eslint-factory rules	45 (was 43) / 18 (was 12)

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

• Smoke CI copilot-CLI 0-turn ([aw] Smoke CI failed #45429/[aw] Smoke CI failed #45931) — root-cause the driver; rejig docs #1 API burner + whole failure surface.
• copilot-sdk-longrun (known issue, recurrence Logs revamp. #25 since 06-02) — burns 40–61K tokens before 0-turn drop; most wasteful mode.
• codex-gh-aw-binary-not-found-for-mcp (recurrence add cli flag to guard dropping a agentic workflow instructinos file #6 since 06-15) — ~100% fail.
• antigravity schema/docs ([Code Quality] Add antigravity engine to docs/reference/engines.md and JSON schema #45808 OPEN) — the re-filing-loop trap; do NOT re-file.
• logs MCP 120s timeout — degrading 3+ audits daily via partial datasets.
• Read-only fleet — 4th straight zero-actuation day; reconcile intentional vs regression.
• Detection misconfigs (5 workflows, incl. Daily Agentic Workflow AIC Usage Audit missing gh-aw-detection) — relates to CLOSED [deep-report] Add gh-aw-detection: true to 5 flagged audit/analysis workflows #45284; watched, not re-filed.
• actions_list MCP tool ignores perPage (returns 30 of 330) — page defensively; likely upstream.

References:

• §29493910926 — GitHub API Consumption (Smoke CI 42.3%, logs timeout)
• §29450871518 — Smoke CI copilot-CLI 0-turn failure example
• §29510196605 — this DeepReport run

Generated by 🔬 Deep Report · 247.5 AIC · ⌖ 12.4 AIC · ⊞ 10K · ◷

• expires on Jul 23, 2026, 7:30 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Deep Report.

A newer discussion is available at Discussion #46271.