[ambient-context] Daily Ambient Context Optimizer

Executive Summary

Sampled 4 completed runs across 4 distinct workflows (last 24h): Ponytail Reviewer, Daily Go Test Parallelizer, Issue Monster, AI Moderator.
Median first-request size: 15,980 chars; P95: ~29,721 chars.
Both audited runs (Issue Monster, AI Moderator) show high Working-Set Rebuild Factor (WSRF) — 4.40 and 5.14 respectively — meaning large static context is resent almost from scratch on every turn instead of growing incrementally.
Highest conclusion: Issue Monster's first request (32,122 chars, 18 code fences, 8.2% duplicate-line ratio) is the single largest and most duplicative sample; consolidating its inline skill blocks and large report templates is the top lever.

Highest-Leverage Changes

Move Issue Monster's large inline report-formatting template (## skill: issue-monster-report-formatting) and token-budget block fully into on-demand skill files rather than always-inlined sections.
Enable tools.cli-proxy: true on AI Moderator (currently false) to route CLI calls through the proxy instead of raw shell instructions, shrinking prompt boilerplate.
Reduce duplicated step-numbering scaffolding in Issue Monster's Step-by-Step Process (7 subsections, 288 lines) by tightening repeated guardrail phrasing flagged by the 8.2% duplicate-line ratio.
Given high WSRF on both audited runs, prioritize converting static/unchanging sections (guidelines, success criteria, error handling) into deterministic steps: or cached context rather than resending each turn.

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	15,980
P95 chars	29,721
Largest sampled request	32,122 chars (Issue Monster, run 33006498846)
Merged optimizer PRs (7d)	2
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	insufficient data (<3 settled PRs)
WSRF (audited runs)	Issue Monster: 4.40, AI Moderator: 5.14

Per-Run First-Request Metrics

Run	Workflow	Chars	Headings	Code fences	Dup line ratio	WSRF
32998381104	Ponytail Reviewer	12,003	7	6	0.040	n/a
33003613538	Daily Go Test Parallelizer	15,845	7	6	0.030	n/a
33006498846	Issue Monster	32,122	17	18	0.082	4.40
33007869619	AI Moderator	16,115	13	2	0.000	5.14

Note: no API-proxy (event-logs.jsonl) or session-state artifacts were present in the downloaded log bundle for any sampled run; prompt.txt was used as the fallback request-text source per the compilation-debug cross-check rule.

Repeated Ambient Context Signals

Issue Monster's request contains the largest concentration of repeated headings/step-scaffolding: "1. Review Pre-Searched and Prioritized Issue List" (114 lines), "2. Select Up to Three Issues to Work On" (34 lines), plus 5 more sub-steps — collectively a large fraction of the 32k-char request.
A safeoutputs create_pull_request MCP usage example block appears as the single largest heading-associated span (146 lines) across the sample set — likely shared boilerplate reused verbatim across multiple workflows' safe-output guidance.
Issue Monster shows the highest duplicate-line ratio (8.2%) of the sample, consistent with its high WSRF (4.40) — static guardrail phrasing is likely being re-sent on each turn.
AI Moderator has zero duplicate lines within its single request but still shows the highest WSRF (5.14), indicating the repetition cost is turn-over-turn (context rebuild) rather than within-request duplication.

Deterministic Analysis Output

Script: /tmp/gh-aw/ambient-context/analyze_requests.py (stdlib only) processed 4 sampled run-*.txt/run-*.json pairs.
Aggregate: 4 distinct workflows, median 15,980 chars, p95 ~29,721 chars.
Full metrics: /tmp/gh-aw/ambient-context/request-analysis.json.
Issue Monster source file: 816 lines / 42,352 bytes on disk, with two large ## skill: blocks (issue-monster-token-budget, issue-monster-report-formatting) that are already skill-scoped but their surrounding step-by-step guardrail prose is not.
AI Moderator source file: 288 lines / 12,718 bytes, tools.cli-proxy: false (not enabled), while Issue Monster already has cli-proxy: true and github.mode: gh-proxy.

Recommendations by Category

Workflow Markdown

Issue Monster (.github/workflows/issue-monster.md, 42,352 chars currently — safe to trim, well above 1,000-char floor and any single change should stay under 40% reduction): Consolidate the 7 step-by-step sub-sections' repeated guardrail phrasing (8.2% duplicate-line ratio) into a single shared reference instead of restating similar constraints per step. Evidence: 32,122-char first request, 18 code fences, highest dup-line ratio in sample. Expected impact: high. Needs manual review (step semantics must be preserved).
AI Moderator (.github/workflows/ai-moderator.md, 12,718 chars): Enable tools.cli-proxy: true (currently false) and route any raw gh/shell CLI wording through explicit agentic-workflows MCP-tool calls. Evidence: cli-proxy: false confirmed in frontmatter; Issue Monster's equivalent config (cli-proxy: true, github.mode: gh-proxy) is smaller relative to its complexity. Expected impact: medium. Likely safe immediately (config-only change), verify via make recompile.

Skills

Issue Monster: The issue-monster-report-formatting skill block already exists but a large recommended-report-structure template is duplicated inline near it — move any inline example templates fully into the skill file so they load on demand rather than unconditionally inflating every run's first request. Evidence: 18 code fences concentrated late in the 32k-char request; ## skill: markers present twice, but explanatory prose duplicated around them. Expected impact: medium. Safe immediately if scoped to prose-only relocation.
Cross-workflow: The 146-line safeoutputs create_pull_request usage example (with heredoc/jq guidance) recurs as the largest single section across sampled workflows' MCP instructions — consider extracting it into one shared ## skill: reference used by all workflows that need it, rather than inlining the full example per workflow. Expected impact: medium. Safe immediately (no semantic change, pure extraction).

Agents

No inline ## agent: blocks were detected in the 4 sampled first requests (inline_agent_count: 0 across all samples), so no agent-pruning recommendation is warranted this cycle. Continue monitoring for inline-agent growth in future daily samples.

References

Generated by 🌫️ Daily Ambient Context Optimizer · copilot · auto · 53.6 AIC · ⌖ 7.52 AIC · ⊞ 11.8K · ◷

expires on Sep 2, 2026, 12:32 PM UTC-08:00