[ambient-context] Daily Ambient Context Optimizer - 2026-09-10

[github-actions]

Executive Summary

• 4 runs sampled (all completed), covering 4 distinct workflows: PR Sous Chef, Matt Pocock Skills Reviewer, Impeccable Skills Reviewer, Issue Monster.
• Median first-request proxy size (prompt.txt, since sandbox firewall/session event logs were not readable under this run's restricted permissions): ~22.2K chars; P95: 24.8K chars (largest sampled: Issue Monster at 37.5K chars).
• Highest-leverage finding: PR Sous Chef shows a Working-Set Rebuild Factor (WSRF) of 12.17x — its ~45K peak-token context is being resent almost every turn across 15 invocations (550K cumulative vs 45K peak), the clearest sign of avoidable re-sent ambient context in the sample.
• Two PR-review workflows (Matt Pocock, Impeccable) each unconditionally bundle multiple full skill sets even though only 1–2 are used per run.

Highest-Leverage Changes

1. Reduce PR Sous Chef's per-turn context rebuild (WSRF 12.17x) by trimming what the pr-processor sub-agent resends each invocation and caching static PR context instead of re-declaring it every turn.
2. Trim Matt Pocock Skills Reviewer's skills: list (5 full skill packages imported unconditionally) to lazy/on-demand loading based on detected change type, matching the "Select Skills" step it already documents.
3. Move Impeccable Skills Reviewer's inline mode table + fallback guidance into the imported impeccable skill so it's loaded once rather than duplicated inline in every run's first request.
4. Simplify Issue Monster's oversized steps: block (rate-limit GraphQL probing logic) — it is deterministic bookkeeping that inflates the compiled prompt/workflow body without needing agent reasoning.
5. Configure explicit GitHub MCP toolsets scoping (already partially done) and confirm cli-proxy: true is set consistently across all 4 sampled workflows to avoid raw gh-CLI-oriented wording bleeding into prompts.

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	22,208
P95 chars	24,824
Largest sampled request	Issue Monster — 37,473 chars
Merged optimizer PRs (7d)	0
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	n/a (insufficient settled PRs, <3)
WSRF (audited runs)	PR Sous Chef 12.17x; Matt Pocock 1.07x

Per-Run First-Request Metrics

Run	Workflow	Conclusion	Request chars (proxy)	Input tokens	char/token	WSRF	AIC
§34518723997	Matt Pocock Skills Reviewer	success	19,593	7,308	2.68	1.07	100.4
§34520309155	PR Sous Chef	success	24,824	8,882	2.79	12.17	27.4
34518724065	Impeccable Skills Reviewer	success	14,021	5,332	2.63	1.01	70.0
§34521691967	Issue Monster	failure	37,473	n/a	n/a	n/a	n/a

Note: request_chars uses prompt.txt size as a fallback proxy — the canonical sandbox/firewall/logs/api-proxy-logs and sandbox/agent/logs/copilot-session-state paths for other runs were not readable from this sandboxed session (read-only/permission-denied on /tmp/gh-aw/aw-mcp/logs/run-*), so this report is best-effort from audit MCP metrics only.

Repeated Ambient Context Signals

• Matt Pocock Skills Reviewer imports 5 full skill packages (diagnosing-bugs, tdd, improve-codebase-architecture, grill-with-docs, codebase-design) unconditionally in front-matter, but its own "Step 3: Identify Change Type and Select Skills" logic implies only 1–2 are typically relevant per PR.
• Impeccable Skills Reviewer inlines a 6-row review-mode selection table plus fallback instructions directly in the workflow body rather than in the imported impeccable skill, duplicating content that's also available via the installed SKILL.md.
• Issue Monster's steps: block contains ~90 lines of deterministic GraphQL rate-limit-probing JS that doesn't need to be in the agent-facing prompt path — it's pure pre-processing.
• PR Sous Chef's pr-processor sub-agent has a WSRF of 12.17 (550K cumulative vs 45K peak input tokens across only 15 invocations) — strongly suggests full context (PR diff/candidate list) is being redeclared each turn instead of referenced incrementally.

Deterministic Analysis Output

analyze_requests.py (stdlib only) computed per-run and aggregate metrics from the 4 sampled run-*.json metadata files (derived from audit MCP prompt_analysis, metrics.ambient_context, and metrics.working_set, since raw request text was inaccessible this run):

• Median request size 22,208 chars; P95 24,824 chars.
• Char-to-token ratio consistent (~2.6–2.8) across the 3 successful runs with token data — no anomalous inflation there.
• Only 1 of 4 sampled runs (PR Sous Chef) shows WSRF > 2x; it is the standout signal in this sample and the primary driver of aggregate cost (552K total tokens vs the other runs' 10–47K).

Recommendations by Category

Workflow Markdown

• PR Sous Chef (.github/workflows/pr-sous-chef.md, 32,526 chars): reduce per-turn context resend in the pr-processor sub-agent — pass only the compact per-PR JSON already written to /tmp/gh-aw/agent/pr-sous-chef-candidates-compact.json instead of re-including full candidate/eligibility context on each of the (up to 4) sub-agent invocations. Evidence: WSRF 12.17x, 550K cumulative vs 45K peak input tokens, 15 invocations. Expected impact: high. Needs manual review (behavior-sensitive sub-agent loop).
• Issue Monster (.github/workflows/issue-monster.md, 41,569 chars): the steps: rate-limit-detection GraphQL/JS block (~90 lines) is fully deterministic and could be trimmed/simplified without touching agent-facing prompt content — it doesn't reduce the first request directly but keeps the compiled workflow lean for future changes. Expected impact: low. Safe immediately (no agent-facing wording changes).

Skills

• Matt Pocock Skills Reviewer (.github/workflows/mattpocock-skills-reviewer.md): 5 full skill packages are imported unconditionally via skills: even though "Step 3" already selects 1–2 based on PR change type. Recommend keeping the skill directory available on disk (as today) but confirming no per-skill inline content is duplicated into the prompt beyond the already-present short bullet summaries. Evidence: ambient_context.input_tokens 7,308 vs 5,332 for the single-skill Impeccable Reviewer. Expected impact: medium. Needs manual review (skill-selection logic must stay correct).
• Impeccable Skills Reviewer (.github/workflows/impeccable-skills-reviewer.md): move the 6-row mode-selection table and fallback guidance (lines ~113–131) into the impeccable skill's SKILL.md so it's loaded on demand from the installed skill rather than duplicated inline in every compiled prompt. Evidence: workflow inlines the same table structurally available in the installed skill. Expected impact: medium. Safe immediately if the skill file is updated in the same PR.

Agents

• No inline-agent removal recommended this cycle — the sampled workflows' sub-agents (pr-processor, pr-triage) appear justified by their scoped, single-purpose roles. Focus instead on reducing what's passed into pr-processor per invocation (see PR Sous Chef recommendation above).

References

• §34518723997
• §34520309155
• §34521691967

Generated by 🌫️ Daily Ambient Context Optimizer · copilot · auto · 145.4 AIC · ⌖ 7.67 AIC · ⊞ 12.4K · ◷

• expires on Sep 17, 2026, 12:05 PM UTC-08:00