[ambient-context] Daily Ambient Context Optimizer - 2026-07-12

[github-actions]

Executive Summary

• Runs sampled: 4 (Daily Code Metrics, Failure Investigator (6h), Smoke Copilot, Smoke Claude)
• Distinct workflows: 4
• Median first-request size: 17,010 chars
• P95 first-request size: 20,794 chars
• Key finding: All 4 sampled runs carry an identical 3,481-char system-injected safe-outputs block (~17-28% of total prompt); the top two by AIC have additional bloat from inline agent definitions and redundant structured examples.

Highest-Leverage Changes

1. [agents] Extract 3 inline agent definitions in aw-failure-investigator.md to shared imports — saves ~2,050 chars per invocation
2. [workflow-md] Compress verbose Required Investigation Steps in aw-failure-investigator.md — saves ~900 chars
3. [workflow-md] Remove inline bash pruning script and redundant JSON schema block in daily-code-metrics.md — saves ~1,050 chars
4. [workflow-md] Trim experiments: frontmatter metadata fields in aw-failure-investigator.md — saves ~500 chars once runtime non-consumption confirmed
5. [workflow-md] Compress verbose Test Requirements lists in smoke-claude.md (16 items) and smoke-copilot.md (21 items) — saves ~600-900 chars each

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 — do not re-attempt changes to any file 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	17,010
P95 chars	20,794
Largest sampled request	20,794 (Smoke Claude)
Merged optimizer PRs (7d)	1
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	N/A (< 3 settled PRs)

Per-Run First-Request Metrics

Run ID	Workflow	Chars	Lines	Headings	CodeFences	DupLineRatio	AIC
29204294047	Daily Code Metrics	19,540	336	19	12	0.007	330.2
29205172839	Failure Investigator (6h)	12,335	268	23	8	0.015	228.9
29205557277	Smoke Copilot	14,480	271	18	14	0.012	100.1
29205138142	Smoke Claude	20,794	383	22	18	0.010	64.5

Source: prompt.txt (API proxy event logs absent from all sampled run log directories).

Repeated Ambient Context Signals

• 3,481-char safe-outputs block is identical across all 4 runs (system-injected)
• Formatting rules (Use ### headers only, move long detail into details blocks, Structure reports as: overview → key metrics → collapsible detail) appear verbatim in ≥2 runs — shared import duplication
• No SKILL.md references found in any sampled prompt (no over-broad skill loading)

Deterministic Analysis Output

• aw-failure-investigator.md: 20,106 chars — 3 inline agents (~2,050 chars), multi-field experiments: frontmatter (~500 chars), verbose investigation steps (~900 chars of prose rationale)
• daily-code-metrics.md: 11,409 chars — 620-char bash pruning snippet replaceable by one instruction sentence; 570-char JSON schema already re-stated in prose
• smoke-claude.md: 12,100 chars — 16-item Test Requirements section (~3,179 chars); smoke-copilot.md: 10,161 chars — 21-item section (~2,556 chars); both contain explanatory sub-bullets that can be compressed to single-line imperatives
• Audit: ambient_context.input_tokens 8,395 (Daily Code Metrics) and 5,186 (Failure Investigator); cached tokens 3.2M and 2.0M indicate stable prompts

Recommendations by Category

Workflow Markdown

R1 — Remove redundant Data Storage schema + bash pruning script (daily-code-metrics.md) — medium, safe immediately

• JSON schema block (~570 chars) duplicates prose below it; bash pruning snippet (~620 chars) can be one sentence.
• Savings: ~1,050 chars. File: 11,409 chars → ~10,350 chars (9% reduction, well within 40% guard).

R2 — Compress Required Investigation Steps (aw-failure-investigator.md) — medium, needs review

• ~900 chars of prose rationale (clustering definitions, cap explanations) can move into inline agent descriptions.
• File: 20,106 chars → ~19,200 chars.

R3 — Strip experiments: frontmatter metadata (aw-failure-investigator.md) — medium, needs review

• Fields hypothesis, analysis_type, tags, secondary_metrics, guardrail_metrics are human metadata; verify runtime non-consumption then remove.
• Savings: ~500 chars.

R4 — Compress Test Requirements lists (smoke-claude.md, smoke-copilot.md) — medium, safe

• Reduce 16 and 21 numbered test items to compact single-line imperatives by removing explanatory sub-bullets.
• Savings: ~600-900 chars per file.

Skills

No SKILL.md references found in any sampled prompt. No over-broad skill loading detected.

Agents

R5 — Extract 3 inline agents in aw-failure-investigator.md to shared imports — high, needs review

• failure-classifier, issue-matcher, cluster-evidence-extractor embed ~2,050 chars per run. Move to shared/ imports for reuse and independent updates. Requires verification that import syntax supports inline-agent extraction.

References

• §29204294047 — Daily Code Metrics (AIC 330.2, 19,540 chars)
• §29205172839 — Failure Investigator (AIC 228.9, 12,335 chars)
• §29205138142 — Smoke Claude (AIC 64.5, 20,794 chars)

Generated by 🌫️ Daily Ambient Context Optimizer · 112.5 AIC · ⌖ 12.7 AIC · ⊞ 7.7K · ◷

• expires on Jul 19, 2026, 12:13 PM UTC-08:00