DeepReport Intelligence Briefing — 2026-05-21

[github-actions[bot]]

Overview

This briefing aggregates intelligence from 100 recent discussions (51 in the past 7 days), 500 weekly issues (82 OPEN / 418 CLOSED), and cross-referenced workflow audit data. The fleet is operationally degraded but functional: workflow success rate slid from 92.8% → 88.9% → 85.9% across the last three audits, three P0/P1 critical issues are open against agent families, and the REST-API budget hit an all-time high of 87.5K calls/day. Counterbalancing those signals, safe-output health is strong (97.9%), Copilot PRs merge at 80.5% across 1,052 tasks in 30 days, and DIFC integrity enforcement is working as designed.

Seven concrete quick wins have been filed as [deep-report] issues — see the Actionable Tasks section below.

Critical / P0–P1 Status

#	Item	Owner	Notes
#33765	CGO/CJS Regression Agent stuck 90+ days at 0%	P0	Recommend disable, not fix
#33766	Codex AI Moderator dead — model 404 + missing OPENAI_API_KEY	P1	All Codex agents blocked
#33767	app/github-actions 70% PR rejection rate (33→10 merged in 30d)	P1	Pre-flight validation absent

Top Patterns Across Reports

Cross-cutting infrastructure regressions

• MCP rpc-messages.jsonl zero-bytes — #33621 closed but recurring on Auto-Triage runs (2/2 failed 5-20). Possibly a fix that didn't fully land, or a new variant.
• OTLP collector 404 — was 401 on 2026-05-17 (auth fix landed), now 404 (endpoint drift). Repeated across nearly every Claude/Copilot conclusion span. Silently undermines Grafana GenAI / Sentry dashboards.
• push_repo_memory 60 KB cap — Daily Code Metrics agent burned $4.20 / 57 turns then lost its commit when 64 KB > 60 KB. Recurring class of failure for high-fidelity agents.
• gh ... | head SIGPIPE — #33769 fixes Matt Pocock alone; the pattern likely lurks in other lock files.
• PR-review firewall block rate 40–59 % — Test Quality Sentinel, Matt Pocock Skills Reviewer, PR Code Quality Reviewer routinely block ~half their network. Either expand allowlist or stop probing.

Code-quality drift

• console.FormatErrorMessage is used 70× vs. fmt.Errorf 1,844× — 3.8 % adoption of the project's own error-formatting standard (#33768).
• 64 fmt.Println calls in pkg/cli/ emit diagnostics to stdout (Unix convention violation; breaks pipeable workflows).
• ~20 bare return err calls in the compile pipeline lose context.
• Schema/parser drift continues: applyTo/inputs accepted by parser but missing from main_workflow_schema.json (#33580); infer deprecation flag in schema not surfaced by parser (#33579).

MCP tool wrapper gaps

• get_me returns 403 in the workflow environment — 2+ consecutive days of Resource not accessible by integration.
• list_workflows ignores per_page=1 — returns all 30 workflows (~3,500 tokens) regardless.
• list_label has no per_page parameter — returns first 100 of 559 labels.

Source: MCP Structural Analysis 2026-05-21.

Workflow failure recurrence (last 7 days)

Workflow	Failures	Notes
Smoke Pi	8	Auto-files & auto-closes 5×; no root cause investigation
Design Decision Gate	7	Mostly transient; sample audit showed 9/9 success
Smoke Codex	6	Tied to Codex P1 issue
PR Sous Chef	6	Mixed
Daily Issues Report Generator	6	Same fingerprint as April's 11-day failure streak
Step Name Alignment	5	Unclear
Chaos PR Bundle Fuzzer	5	allowed-files config gaps

Trends & Metrics (snapshot vs prior briefing 2026-04-03)

Metric	2026-04-03	2026-05-21	Direction
Workflow success rate (24h)	n/a	85.9 % (was 92.8% 5-17)	⬇
Cost (24h)	n/a	$26.72	–
Tokens (24h, effective)	99.5M	336.9M	⬆ 3.4×
REST API calls (24h)	n/a	87,513 (all-time high)	⬆
Open P0/P1 critical	3	3	=
Copilot PR merge rate	78.4 %	80.5 % (n=1052)	⬆
Safe-output success	100 %	97.9 %	⬇ slightly

Positive Signals

• Safe-output health 97.9 % — only 1 of 48 dispatched jobs failed in the last 24h, and that failure was a single PR-review atomic-submit rejection (discussion #33705).
• Copilot PR merge rate up to 80.5 % across 1,052 tasks in the last 30 days; the lowest-performing cluster (awf / firewall / config) is the focused improvement target.
• Compiler code quality holding: latest sample of 3 files averaged 81/100 with test-to-source ratios up to 31.5:1.
• DIFC enforcement working: 10 integrity-filtered events in 24h, all on none:all / unapproved:all resources — expected behavior.

Actionable Tasks Filed Today

Seven [deep-report] quick-win issues have been filed (auto-prefix + auto-labels automation, improvement, quick-win, cookie):

1. Sweep gh ... | head -n N SIGPIPE pattern across all workflow lock files (sibling of #33769).
2. Fix OTLP traces collector 404 endpoint drift (distinct from #33740 INPUT_JOB_NAME fix).
3. Raise or trim push_repo_memory max-patch-size to prevent Daily Code Metrics truncation.
4. Auto-start docs server in docs-noob-tester workflow setup.
5. Migrate 64 CLI diagnostic fmt.Println calls to stderr via console.FormatErrorMessage.
6. Add per_page support to MCP list_workflows / list_label wrappers.
7. Triage recurring Daily Issues Report Generator failures (6 of last 7 days; matches April's 11-day streak).

Each is sized at < 4 hours of agent time and targets a concrete, well-evidenced gap.

Cited Sources

• Daily Audit 2026-05-20 — 5 failures + Codex/AI Moderator broken, OTLP shifts to 404 (#33643)
• Safe Output Health Report 2026-05-21 (#33705)
• Repository Quality Improvement Report — Error Message Quality (#33768)
• GitHub API Consumption Report 2026-05-21 (#33747)
• Copilot Agent Prompt Clustering 2026-05-21 (#33741)
• MCP Structural Analysis 2026-05-21 (#33752)
• Docs Noob Test Report 2026-05-21 (#33700)
• Agent Performance Report — Week of May 21 (#33764)
• Token Consumption Report 2026-05-21 (#33751)

References: §26236916951

Generated by 🔬 DeepReport - Intelligence Gathering Agent · ● 17.9M · ◷

• expires on May 28, 2026, 4:03 PM UTC

[github-actions[bot]]

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

A newer discussion is available at Discussion #34057.