[ambient-context] Daily Ambient Context Optimizer - 2026-08-22

[github-actions]

Executive Summary

• Runs sampled: 4 completed runs across 4 distinct workflows (last ~20 min of the 24h window; broader pagination was rate-limited by the logs MCP tool at count=20/60).
• Workflows covered: Avenger (CI fixer), Matt Pocock Skills Reviewer, Ponytail Reviewer, Impeccable Skills Reviewer.
• Median first-request size: 17,160 chars · P95: 19,514 chars.
• Conclusion: The three PR-review workflows share a near-identical inline instruction skeleton (## Process, numbered Step N sections, safe-output JSON examples, <details> templates) duplicated per-workflow instead of factored into a shared skill/import. Matt Pocock Reviewer's pr-triage inline agent and full skill-guidance-by-skill text block are the single largest structural contributor (~19.5K chars). All three reviewers already use gh-proxy/cli-proxy; Ponytail Reviewer has no tools: proxy config at all despite similar diff-fetch needs.

Highest-Leverage Changes

1. Extract the shared review-workflow skeleton (Steps 1–7 headings, JSON examples, progressive-disclosure <details> template) from mattpocock-skills-reviewer.md and impeccable-skills-reviewer.md into a common shared/skill-review-process.md import.
2. Move the per-skill "Focus areas by skill" guidance block (5 skill writeups, ~2.3K chars) in mattpocock-skills-reviewer.md into a loaded ## skill: block instead of always-inline text.
3. Enable tools.github.mode: gh-proxy and cli-proxy: true on ponytail-reviewer.md, matching its siblings, and confirm reliance on any raw gh shell calls is minimized.
4. Review whether the inline pr-triage sub-agent in Matt Pocock Reviewer needs its full JSON schema/skill-mapping table restated in the parent prompt, or if that can be trimmed since the agent definition itself carries it.

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	17,160
P95 chars	19,514
Largest sampled request	19,514 chars (Matt Pocock Skills Reviewer, run 32594519979)
Merged optimizer PRs (7d)	0
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	n/a (insufficient settled data)
WSRF (audited runs)	1.001–1.141 (Avenger 1.001, Ponytail 1.005, Matt Pocock 1.099, Impeccable 1.141)

Per-Run First-Request Metrics

Run	Workflow	Chars	Lines	Headings	Details blocks	Inline agents	WSRF
32593997838	Avenger	17,160	259	—	—	0	1.001
32594519979	Matt Pocock Skills Reviewer	19,514	293	14	1	1 (pr-triage)	1.099
32594519982	Ponytail Reviewer	12,059	138	—	—	0	1.005
32594520028	Impeccable Skills Reviewer	14,011	205	—	—	0	1.141

Note: request text source was aw-prompts/prompt.txt for all four runs — no event-logs.jsonl/events.jsonl artifact was present in the cached run data for this sample, so this is used per the fallback rule with the compiler-debug caveat acknowledged.

Repeated Ambient Context Signals

• All three PR-review workflows (Matt Pocock, Impeccable, Ponytail) repeat the same shape: ## Process → numbered Step N: sections → JSON/markdown examples for inline review comments → progressive-disclosure <details> guidance — duplicated per workflow rather than shared.
• Matt Pocock Reviewer inlines a full 5-skill "Focus areas by skill" reference (~2.3K chars) unconditionally, even though only 1–2 skills are typically selected by pr-triage per PR.
• The inline pr-triage agent definition restates its full JSON schema and skill-mapping table inside the same file already described in the parent prompt's Step 3, adding redundant explanatory text.
• Ponytail Reviewer has no tools: block (gh-proxy/cli-proxy unset) while its two-sibling workflows already use both, despite similar diff/PR-data needs — a config-parity gap.

Deterministic Analysis Output

• Aggregate script output (/tmp/gh-aw/ambient-context/request-analysis.json): 4 sampled runs, 4 distinct workflows, median 17,160 chars, P95 19,514 chars.
• Repeated markdown code-fence markers (``` / ```bash) appeared in all 4 samples — expected baseline noise, not a targeted signal.
• Largest recurring headings across samples (by section length in lines): "## Process" (61 lines), "### Step 4: Review Using Selected Skills" (39 lines), "### Step 5: Post Inline Review Comments" (20 lines), "## Step 2: Recompile workflows" (17 lines, Avenger), "### Step 1: Load Pre-fetched PR Data" (16 lines) — these are the highest-leverage candidates for shared-import extraction.

Recommendations by Category

Workflow Markdown

1. Extract shared review skeleton — mattpocock-skills-reviewer.md, impeccable-skills-reviewer.md (and optionally ponytail-reviewer.md) share near-identical Step 1/4/5/6 structure. Move the common process scaffolding into a shared/skill-review-process.md import. Evidence: "## Process" and "Step N" headings are the top 3 largest repeated sections across all 3 reviewer samples. Impact: medium (est. 20–30% reduction in reviewer prompt duplication). Safe to implement with manual review (must preserve per-workflow skill lists and safe-output config).
2. Enable gh-proxy/cli-proxy on Ponytail Reviewer — ponytail-reviewer.md has no tools: block while its siblings both set cli-proxy: true and github.mode: gh-proxy. Evidence: config diff against mattpocock-skills-reviewer.md/impeccable-skills-reviewer.md. Impact: low-medium (consistency + reduces raw gh shell calls). Safe to implement immediately.

Skills

3. Load per-skill guidance on demand — Matt Pocock Reviewer's "Focus areas by skill" block inlines full guidance for all 5 skills (~2.3K chars) even though pr-triage typically narrows to 1–2 relevant skills. Evidence: dup_paragraph_ratio and section-size metrics for run 32594519979 (largest sampled request, 19,514 chars). Convert to a ## skill: reference loaded conditionally based on pr-triage output. Impact: medium. Needs manual review (must not break the "Only read a skill file when inline guidance is insufficient" fallback logic).

Agents

4. Trim inline pr-triage agent definition — the agent restates its full JSON schema/skill-mapping table separately from the parent prompt's Step 3 description, adding redundant text within the same first request. Evidence: inline_agents: 1 + 19,514-char total for run 32594519979, the largest sample. Impact: low-medium. Safe to implement with manual review (verify JSON schema is still enforced correctly for fallback parsing).

References

• §32594519979
• §32593997838
• §32594520028

Generated by 🌫️ Daily Ambient Context Optimizer · auto · 68 AIC · ⌖ 15.1 AIC · ⊞ 13.1K · ◷

• expires on Aug 29, 2026, 11:59 AM UTC-08:00

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #55156

View newer issue

---

This action was performed automatically by the Daily Ambient Context Optimizer workflow.