[ambient-context] Daily Ambient Context Optimizer - 2026-08-03

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows (issue-monster excluded as self; MCP Inspector Agent excluded — failed with no usable request artifact).
• Median first-request size ≈17,075 chars, P95 ≈19,843 chars.
• ⚠️ Auto-pause active: PR close-rate for this optimizer is 40% (2 closed / 5 settled in 7d, ≥3 threshold met) — recommendation count capped at 1 per auto-pause rule.
• Highest-confidence finding: aw-failure-investigator.md (aic 224.7, highest in sample) carries a ~3.6KB "MCP Response Size Limits" pagination boilerplate block (~22% of its 16.5KB first request) that appears in only 1 of 4 sampled workflows, suggesting it is non-core, generically-injected content rather than mission-specific instruction.

Note on data provenance: canonical API-proxy event-logs.jsonl / session-state events.jsonl artifacts were not present for these sampled runs in the current agenticworkflows logs download; prompt.txt was used as the request-text source (secondary/compilation artifact per instructions), so sizes reflect compiled prompt text rather than the literal first DLLM wire payload.

Highest-Leverage Changes

1. [workflow-md, high impact] Trim or conditionally-gate the "MCP Response Size Limits" pagination guidance block in aw-failure-investigator.md's effective prompt — it is generic gh-proxy/MCP usage guidance, not task-specific to failure investigation, and inflates the first request by ~3.6KB (22%) without adding mission value for this workflow's audit/logs tool calls (already capped to 2 audit calls with artifacts scoping per its own cluster-evidence-extractor agent instructions).

(Only 1 recommendation included due to auto-pause rule; additional candidates — inline-agent count review for aw-failure-investigator.md's 3 sub-agents, and tone-variant experiment cleanup — were considered but withheld to respect the cap.)

CI-Validation Checklist for Implementing Agents

Any agent implementing workflow-file recommendations must complete every item below before opening a PR:

• Run make recompile for every modified .github/workflows/*.md file — zero compilation errors required
• Confirm no target workflow file is reduced by more than 40% of its current character count
• Run make agent-report-progress before the final commit and confirm it passes
• Check the blocked_files list in /tmp/gh-aw/ambient-context/closed-pr-targets.json (written by Step 4) — do not re-attempt changes to any file that appears in a closed ambient-context optimization PR from the last 14 days
• Include a smoke-test result or dry-run output in the PR description for every changed workflow spec
• Document all generated .lock.yml changes in the PR body
• Do not submit as a draft PR

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	17,075
P95 chars	19,843
Largest sampled request	20,235 chars (Daily Code Metrics and Trend Tracking Agent)
Merged optimizer PRs (7d)	3
Closed optimizer PRs (7d)	2
Optimizer PR close-rate (7d)	40% (auto-pause triggered)

Per-Run First-Request Metrics

Run	Workflow	Chars	Headings	Inline Agents	Dup Line Ratio	AIC
30843574517	Daily Code Metrics and Trend Tracking Agent	20,235	28	0	0.044	201.0
30848881820	Matt Pocock Skills Reviewer	17,623	21	0	0.0	47.9
30844681199	[aw] Failure Investigator (6h)	16,526	42	3 (rendered as text, not ## agent: markers post-compile)	0.0	224.7
30844497827	Daily Safe Output Integrator	13,284	13	0	0.0	78.4

Repeated Ambient Context Signals

• The "MCP Response Size Limits" pagination section (headings: Common Scenarios, Pagination Best Practices, Tool-Specific Guidance, Error Messages to Watch For, Performance Tips, Example Workflow Pattern) is present verbatim in aw-failure-investigator.md's compiled prompt but absent from the other 3 sampled workflows — it is the single largest contiguous non-mission block found (~3.6KB / 122 lines).
• No inline ## agent:/## linter:/## skill: markers survived into any compiled prompt.txt sample (they are resolved/stripped at compile time), so duplicate-line/paragraph ratios across samples were low (0.0–0.044) — bloat in this sample set is concentrated in generic tool-usage documentation rather than repeated prose.
• aw-failure-investigator.md already has gh-proxy and cli-proxy: true enabled, so its size is not attributable to missing proxy configuration.

Deterministic Analysis Output

analyze_requests.py (stdlib-only) computed per-run char/word/heading/list/table/dup-ratio metrics and aggregate median/P95/top-workflow-by-size stats, written to /tmp/gh-aw/ambient-context/request-analysis.{json,md}. Key aggregate: median 17,075 chars, P95 19,843 chars, largest sampled request 20,235 chars (Daily Code Metrics and Trend Tracking Agent). No API-proxy request_input_tokens were available in this run's logs download, so char-to-token ratios could not be computed for this cycle.

Recommendations by Category

Workflow Markdown

• Move the generic "MCP Response Size Limits" / pagination-guidance text out of aw-failure-investigator.md's always-loaded prompt body and into an on-demand ## skill: reference, since it duplicates general gh-proxy/MCP usage guidance not specific to failure-investigation logic, and only 1 of 4 sampled workflows needed it in-context. (Needs manual review: confirm no other workflow relies on this text being always-present before extracting to a shared skill.)

Skills

• None recommended this cycle (capped by auto-pause rule; withheld pending close-rate recovery).

Agents

• None recommended this cycle (capped by auto-pause rule; withheld pending close-rate recovery).

References

• §30844681199
• §30843574517
• §30848881820

Generated by 🌫️ Daily Ambient Context Optimizer · auto · 71.3 AIC · ⌖ 6.72 AIC · ⊞ 12.5K · ◷

• expires on Aug 10, 2026, 12:27 PM UTC-08:00