DeepReport Intelligence Briefing — April 27, 2026

[github-actions[bot]]

🔍 Executive Summary

The gh-aw agent ecosystem on April 27, 2026 spans 68 active workflows across 138 daily runs, with an overall agent success rate of 92% (Quality: 74/100, Effectiveness: 71/100). Token consumption is on a sustained downward trend — the 7-day rolling average of 76.4M tokens/day is 23% below the April 3 baseline of 99.5M, driven by recent optimization investments in Test Quality Sentinel, Smoke CI, and Architecture Guardian. Safe outputs maintained 100% execution success.

Two persistent infrastructure failures continue to drag on the ecosystem: Smoke CI at 100% error rate (22/22 runs this week — trigger misconfiguration unresolved) and Daily News + Daily Issues Report Generator with node-not-found in chroot mode (20+ consecutive days). The most expensive single-run workflow in the fleet is Schema Consistency Checker at 8.1M tokens and 138 turns per run. On the security side, the daily firewall shows a 15% block rate (313/2,085 requests), with 91% of those blocks attributed to (unknown) traffic that cannot be traced to specific domains — a visibility gap requiring investigation.

This briefing covers 7 days of workflow telemetry, 45 recent discussions (35 audits + 8 announcements + 2 general), and 500 weekly issues (71 open, 429 closed). Seven GitHub issues were filed for high-impact quick-win tasks.

---

📊 Pattern Analysis

Positive Patterns ✅

Token consumption declining sustainably — Weekly total for Apr 21–27 was 67.7M tokens (down from 76.6M on Apr 22), continuing a multi-week trend. No single workflow dominates: the top consumer (Contribution Check) holds only 7.9% of the weekly total. The optimization investments from previous DeepReport cycles are measurably compounding.

Safe output reliability stable at 100% — All safe-output jobs executed cleanly. The Safe Output Health Monitor confirmed 100% execution success with no error clusters detected. The ecosystem's self-monitoring layer is healthy.

Agent quality tooling cluster self-maintaining — Safe Outputs Conformance Checker, Safe Output Tool Optimizer, and Safe Output Integrator all ran and succeeded in the latest window — a virtuous loop where agents monitor and improve other agents.

Zero escalation-eligible episodes in 30-day window — The Agentic Observability Kit analyzed 138 runs and found 0 escalation-eligible episodes and 0 MCP failure counts. Risk is concentrated exclusively in two known-bad workflows (Smoke CI, Visual Regression Checker), not in agentic control behavior.

Stale lock files resolved — The April 3 concern about 19 stale lock files has been fully resolved. 200 lock files are now standardized on schema v3, firewall v0.25.28, with 100% concurrency controls.

Concerning Patterns ⚠️

Smoke CI 100% error rate persisting — 22 errors across 22 runs this week (22/22 in the optimization kit's data). Root cause is a known trigger misconfiguration: no paths: filter on the push trigger causes every push-to-main to cascade through the queue. This was documented in the April 24 DeepReport's weekly workflow analysis and still hasn't been fixed. No patch complexity — a 3-line YAML change would restore the workflow entirely.

Model pin drift pattern recurring — GitHub Remote MCP Authentication Test has been failing for 5+ consecutive days because it hardcodes gpt-5.1-codex-mini, a model no longer supported on the subscription tier. The April 24 DeepReport filed a fix issue; the workflow is still failing. This is the third distinct model-pin drift incident in the past 30 days (after Daily Community Attribution and Auto-Triage). No systematic solution (supported-model registry, CI lint) exists yet.

Resource-heavy workflows not yet optimized — Three workflows persistently appear in the top-token-consuming tier with resource_heavy_for_domain: HIGH classifications and agentic_fraction=0.50 (half turns are data-gathering, replaceable with deterministic pre-steps):

• Daily Syntax Error Quality Check: 3.4M tokens, 56 turns
• Package Specification Extractor: 3.0M tokens, 51 turns
• Documentation Unbloat: 4.8M tokens, 58 turns (also a new failure today at 34.8 minutes)

Firewall unknown traffic blind spot — 284 of 313 blocked requests (91%) are attributed to (unknown) and span 12 workflows. This is structurally blind: the audit cannot determine whether these are legitimate services needing allowlisting or unwanted outbound traffic. Workflows with block rates above 40% include: Copilot PR Conversation NLP Analysis (40.7%), Agentic Observability Kit (49.5%), Daily CLI Tools Exploratory Tester (60%), Architecture Diagram Generator (54.8%).

Daily Community Attribution duplicate issues — The workflow creates a new failure issue on every run that fails, producing issue pile-ups (#28025, #28235, etc.) with identical titles. No deduplication check before create_issue is in the prompt.

Emerging Patterns 🔎

Agent quality score plateauing at 74/100 — Quality has been stable at 74 for 3+ days (Apr 25–27), Effectiveness at 70–71. The plateau corresponds exactly to the P1 issues remaining open (Remote MCP Auth Test #27965, node chroot failures). Resolving these would push the score back toward 80+.

29-workflow shared component gap widening — The Workflow Skill Extractor identified 29 workflows that should adopt shared/daily-audit-base.md but haven't. Today's Copilot Opt agent independently filed #28707 flagging prompt drift across reporting workflows — confirming the gap is already causing quality degradation.

---

📈 Trend Intelligence

Metric	Apr 3	Apr 24	Apr 27	Trend
Tokens/day (7d avg)	99.5M	~77M	76.4M	↓ improving
Weekly tokens total	—	—	67.7M	↓ week-over-week
Agent quality score	—	72	74	↑ slow progress
Agent success rate	—	93%	92%	→ stable
Safe output success	100%	100%	100%	✅ stable
Open issues (7d)	56	63	71	↑ slight increase
Smoke CI error rate	—	100%	100%	→ unresolved
WHM score	72	72	72	→ flat
Compiled workflows	198	201	201+	→ slow growth
Firewall block rate	—	15.8%	15.0%	↓ slightly improving

The token reduction trend is the clearest positive signal. The open issue count increasing slightly (63 → 71) reflects new agent-generated issues rather than a backlog growth crisis — 429 of 500 weekly issues were closed.

---

🚨 Notable Findings

🔴 Smoke CI: 100% error rate for 22 consecutive runs — No agent turns start, no tokens consumed. The trigger misconfiguration causes immediate queue cascade. At 22 wasted GitHub Actions runs per week, this is pure infrastructure overhead with zero agent value. A 3-line YAML fix would restore it.

🔴 Schema Consistency Checker: 8.1M tokens, 138 turns — fleet's most expensive single run — The 14.9% cache efficiency compounds the cost: most context is re-read rather than cached. This is the highest-cost single-run workflow observed in this analysis period. The Weekly Workflow Analysis, Observability Kit, and Token Audit all independently flag it as the top optimization target.

⚠️ Documentation Unbloat: new failure today, 34.8 minutes — A previously-stable workflow failed for the first time on April 27 at 58 turns and 4.8M tokens over 34.8 minutes. The Agent Performance Report notes this could be related to the AWF binary CDN 502 issue (#28529) or a new regression. No max_turns guard exists.

⚠️ Visual Regression Checker: 4–7 errors across 5 runs — Consistent docs-build failure on the add-shadow-ops branch. The docs build pre-step exits non-zero and the agent never starts, wasting 4 consecutive runs. This is branch-scoped and will self-heal when the PR closes.

✅ Token optimization investments compounding — The 23% reduction in the 7-day token average (99.5M → 76.4M) since the April 3 DeepReport is the most meaningful positive metric in this analysis. It demonstrates that the optimization recommendations from previous briefings are being acted on and are measurably effective.

✅ Hippo Memory operational at 500+ memories — The memory consolidation layer is working at scale with verified institutional recall (highest-confidence memory correctly captures the stale lock file pattern at confidence score 1.007). This is a significant capability expansion since the April 3 baseline.

---

🔮 Predictions and Recommendations

Smoke CI will continue failing until a human or agent applies the paths filter fix. The April 24 Weekly Workflow Analysis documented the exact fix; GitHub issue filed in this briefing. Recommend assigning to Daily Workflow Updater for immediate execution.

Schema Consistency Checker will become the fleet's #1 cost driver if left unoptimized. At 8.1M tokens/run with bi-weekly+ cadence, it will account for 15–20% of monthly token budget. Recommend the Agentic Optimization Kit generate and apply the right-sizing plan as its next target.

Model pin drift will recur until a systematic solution is deployed. Three incidents in 30 days establishes a pattern. Recommend adding a CI lint step that validates model pins against a supported-models manifest before any workflow is compiled.

Token usage volatility (day-to-day variance from 51M to 142M) will continue unless per-workflow max_tokens or scheduling caps are introduced for high-cost one-shot workflows. Consider converting Daily Community Attribution, Package Specification Extractor, and Documentation Noob Tester from daily to weekly cadence.

---

✅ Actionable Agentic Tasks (Quick Wins)

Seven GitHub issues were created as part of this briefing. Each is ready for immediate agent assignment:

1. Optimize Daily Syntax Error Quality Check — Cut 56→35 turns via bash pre-step, saving ~1.36M tokens/run (~9.5M tokens/week). Evidence: run §24991120216. Quick (< 1h)

2. Fix Smoke CI trigger misconfiguration — Add paths filter + concurrency group to eliminate 100% error rate (22/22 runs). 3-line YAML fix. Fast (< 30 min)

3. Right-size Schema Consistency Checker — Add deterministic schema-diff pre-step + max-turns: 60 to reduce 8.1M tokens, 138 turns → ~3M token target. Medium (1–4h)

4. Consolidate Daily DIFC Analyzer + Daily Firewall Reporter — Merge into single Daily Security Observability workflow. Same domain, daily, 0.70 overlap score. Saves ~2.49M tokens/week. Medium (1–4h)

5. Add deduplication to Daily Community Attribution Updater — Search for existing open issue before creating new failure notification. Eliminates issue queue pollution. Quick (< 1h)

6. Migrate 29 workflows to shared/daily-audit-base.md — Systematic migration to end prompt drift across reporting workflows, identified by Workflow Skill Extractor. Medium (1–4h)

7. Add max_turns + pre-steps to Documentation Unbloat + Package Specification Extractor — Prevent 4.8M-token failed runs (Unbloat) and reduce 3M-token overkill runs (Pkg Extractor). Medium (1–4h)

---

📚 Source Attribution

Discussions analyzed (past 7 days):

• #28742 — Weekly Agentic Optimization Report (Apr 21–27)
• #28722 — Safe Output Health Report (Apr 27)
• #28687 — Weekly Workflow Analysis (Apr 20–27)
• #28714 — Daily Firewall Report (Apr 27)
• #28713 — Daily Token Usage Audit (Apr 27)
• #28682 — Agentic Observability Report (Apr 27)
• #28662 — Agent Performance Report — Week of Apr 27
• #28303 — DeepReport Intelligence Briefing — April 24 (previous)

Issues analyzed: 500 issues (7-day window): 71 open, 429 closed. Top labels: automation (27), agentic-workflows (24), cookie (21).

Repo memory: Previous analysis from 2026-04-03 (24-day gap to prior stored memory); April 24 briefing used as primary prior context.

Time range: 2026-04-20 to 2026-04-27 (7-day window)

References:

• §25003521428 — This DeepReport run
• §24997761991 — Agentic Optimization Kit (weekly)
• §24988530318 — Weekly Workflow Analysis

Generated by DeepReport - Intelligence Gathering Agent · ● 473.9K · ◷

• expires on May 4, 2026, 3:35 PM UTC

[github-actions[bot]]

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

A newer discussion is available at Discussion #29150.