[ambient-context] Daily Ambient Context Optimizer - 2026-06-06

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows (Matt Pocock Skills Reviewer, Contribution Check, Test Quality Sentinel, Documentation Unbloat)
• Median first-request size: 17,077 chars · P95: 18,909 chars · Outlier: Test Quality Sentinel at 28,417 chars (66% above P95)
• All 4 workflows have cli-proxy: true; Matt Pocock Skills Reviewer is missing tools.github.mode: gh-proxy
• Highest-leverage wins: extract 2 inline agent defs from Contribution Check (~1,500 chars saved), collapse duplicated per-skill guidance in Matt Pocock (~2,700 chars), and trim Test Quality Sentinel's embedded report template (~3,000 chars)

Highest-Leverage Changes

1. [agents] Extract report-formatter and comment-dispatcher inline agent definitions from contribution-check.md into separate .agent.md files — removes ~1,500 chars from the orchestrator first-request (high, needs manual review)
2. [workflow-md] Collapse ## Available Matt Pocock Skills inline descriptions + Step 4 per-skill guidance blocks in mattpocock-skills-reviewer.md into single-line SKILL.md references — saves ~2,700 chars from the most expensive run by AIC (high, needs manual review)
3. [workflow-md] Trim Test Quality Sentinel's Step 7 embedded full report-comment template (~100 lines of placeholder prose) to a compact skeleton — the largest prompt at 28,417 chars with 10.1% duplicate-line ratio (medium, needs manual review)
4. [workflow-md] Consolidate 13–19 scattered safe-output instruction mentions per workflow (across Matt Pocock, Contribution Check, Test Quality Sentinel) into one canonical block near the top of each workflow — reduces per-step repetition (medium, partially safe to automate)
5. [workflow-md] Delete the ## Important section in contribution-check.md (1,109 chars) — it re-states rules already present in the numbered orchestrator steps (medium, safe to automate)
6. [workflow-md] Add tools.github.mode: gh-proxy to mattpocock-skills-reviewer.md — the highest-AIC run (772.9 AIC, 34 turns) currently uses only cli-proxy: true, missing proper MCP tool routing for GitHub reads (low, safe to automate)

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	17,077
P95 chars	18,909
Largest sampled request	Test Quality Sentinel — 28,417 chars

Per-Run First-Request Metrics

Run	Workflow	Chars	AIC	Tokens	Turns	Dup-Line Ratio	Source
§27069914142	Test Quality Sentinel	28,417	403.3	1,096,917	31	10.1%	prompt.txt
§27068039186	Documentation Unbloat	18,909	385.8	1,835,970	34	6.8%	prompt.txt
§27069819359	Matt Pocock Skills Reviewer	15,245	772.9	2,209,814	34	3.2%	prompt.txt
§27068700952	Contribution Check	12,388	488.2	1,379,626	45	5.3%	prompt.txt

Repeated Ambient Context Signals

• Safe-output instructions repeated per step: Matt Pocock (13 mentions), Contribution Check (19 mentions), Test Quality Sentinel (13 mentions) — all three contain the same safe-output guardrail language in multiple steps rather than once at the top
• Inline skill summaries duplicated by external SKILL.md: Matt Pocock Skills Reviewer embeds a ~200-char description for each of 6 skills inside the prompt, while skill_md_count=1 confirms those same skills are already available from loaded SKILL.md files at runtime
• Per-skill guidance blocks in Step 4 (/diagnose, /tdd, /zoom-out, /improve-codebase-architecture, /grill-with-docs) — each ~200 chars, totalling the largest section (1,560 chars) in the Matt Pocock prompt
• ## Important section in Contribution Check re-states orchestrator identity, pre-computed step guidance, and safe-output noop rules already covered in the numbered workflow steps
• Embedded report-comment template in Test Quality Sentinel Step 7: a ~100-line markdown template with placeholder syntax is fully inlined in the prompt, contributing to the 28,417-char first request and 10.1% duplicate-line ratio
• Two inline agent definitions (report-formatter, comment-dispatcher) embedded in Contribution Check — loaded on every orchestrator turn rather than fetched lazily by the sub-agent runtime

Deterministic Analysis Output

Script: /tmp/gh-aw/ambient-context/analyze_requests.py — stdlib only, no third-party packages.

Key per-run findings:

• Test Quality Sentinel: 51 headings, 26 code fences, 12 <details> blocks, 19 table rows, dup_line_ratio=10.1%, dup_para_ratio=2.1% — highest structural complexity in the sample; cat /tmp/gh-aw/agent/diff-numstat.txt appears twice (duplicated shell snippet across Step 5)
• Contribution Check: safe_output_mentions=19 (highest in sample); two inline ## agent: blocks add ~1,500 chars to the orchestrator first request; dup_line_ratio=5.3%
• Matt Pocock Skills Reviewer: skill_md_count=1 (external SKILL.md loaded) yet also contains full inline skill summaries for all 6 skills; largest section by chars is Step 4 (1,560 chars); tools.github.mode not configured
• Documentation Unbloat: 34 headings, 20 code fences, dup_line_ratio=6.8%, 4 <details> blocks — second-largest prompt with no skill references

Aggregate: most common repeated fragments are shell snippets (cat /tmp/gh-aw/agent/diff-numstat.txt, JSON event literals) reused across steps.

Recommendations by Category

Workflow Markdown

• Remove ## Important section in contribution-check.md (1,109 chars) — duplicates orchestrator identity and noop rules from numbered steps. Safe to remove immediately.
• Consolidate safe-output instructions in mattpocock-skills-reviewer.md, contribution-check.md, and test-quality-sentinel.md: replace per-step repetition with a single canonical block. Needs light manual review to verify no step-specific nuances are lost.
• Shrink Step 7 report template in test-quality-sentinel.md: replace the full ~100-line comment-format example with a 10-line skeleton plus a link to a shared template. Estimated savings: ~3,000 chars from the 28,417-char prompt.
• Add tools.github.mode: gh-proxy to mattpocock-skills-reviewer.md (only cli-proxy: true is set). Safe to add immediately; enables gh-proxy routing for GitHub reads.

Skills

• Collapse inline skill summaries in mattpocock-skills-reviewer.md: the ## Available Matt Pocock Skills section (1,125 chars) and per-skill Step 4 guidance blocks (1,560 chars combined) duplicate content already in the loaded SKILL.md. Replace both with: Apply each selected skill according to its SKILL.md definition. Saves ~2,700 chars; needs manual verification that no guidance is lost.

Agents

• Extract report-formatter and comment-dispatcher from contribution-check.md into separate .github/agents/*.agent.md files. Their full prompts (~1,500 chars) are loaded into every orchestrator request even though they are only invoked at the end. Referencing them by name instead lets the runtime load them lazily. Needs manual review to ensure sub-agent routing still resolves correctly.

References

• §27069819359 — Matt Pocock Skills Reviewer (highest AIC: 772.9)
• §27068700952 — Contribution Check (45 turns, 488.2 AIC)
• §27069914142 — Test Quality Sentinel (largest prompt: 28,417 chars)

Generated by 🌫️ Daily Ambient Context Optimizer · 723.1 AIC · ⌖ 23.2 AIC · ⊞ 21K ambient context · ◷

• expires on Jun 13, 2026, 8:13 PM UTC