[ambient-context] Daily Ambient Context Optimizer - 2026-08-29

[github-actions]

Executive Summary

• Sampled 4 completed runs across 4 distinct workflows in the last ~1h window (count-limited by MCP pagination; date range narrower than 24h).
• Median first-request size: 15,985 chars; P95: 16,124 chars.
• Largest sampled first request: Design Decision Gate (27,549 chars, 44 headings).
• No canonical event-logs.jsonl API-proxy artifact or session-state events were available for the downloaded run bundles; all 4 samples used prompt.txt as a documented fallback source (compilation-debug artifact, flagged per source-of-truth rules).
• Two highest-cost runs audited: Daily Go Test Parallelizer (WSRF 5.78, 78 invocations, 549,927 rebuild-excess tokens) and Design Decision Gate (WSRF 1.74, 15 invocations, 20,979 rebuild-excess tokens). The Parallelizer's high WSRF despite a small (4,677 char) source file indicates static context (imports + inline agent) is being resent near-peak on nearly every turn rather than growing incrementally.

Highest-Leverage Changes

1. Daily Go Test Parallelizer — reduce turn-over-turn context rebuild (WSRF 5.78) by trimming repeated static rule text sent to the parallel-safety-batch-checker sub-agent call and offloading the safety-rule list to a referenced skill instead of inline repetition across turns.
2. Design Decision Gate — largest sampled first request (44 headings, 4 report templates embedded inline: ADR-required, ADR-verified, ADR-diverges, plus skill block). Move the 3 large report templates (currently under ## skill: adr-report-templates, lines 378-441) fully out of the main prompt body into an on-demand skill file if not already lazy-loaded, since they inflate every invocation whether or not the corresponding branch is taken.
3. AI Moderator — tools.cli-proxy: false and github.mode: local (not gh-proxy). Enable cli-proxy: true and github.mode: gh-proxy to move deterministic GitHub reads off the main prompt path, consistent with the pattern already used by design-decision-gate.md and pr-description-caveman.md.
4. PR Description Updater — lowest WSRF (1.26) and smallest sample (12,780 chars); no immediate action needed — used here as a size baseline.
5. Repository-wide: none of the 4 sampled first requests contained ## skill:-imported SKILL.md references (skill_md_reference_count: 0 in all samples) — output templates and static rule blocks are being kept inline in workflow markdown rather than deferred to skills, which is the main lever available across this sample.

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	15,985
P95 chars	16,124
Largest sampled request	27,549 chars (Design Decision Gate)
Merged optimizer PRs (7d)	4
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	0%
WSRF (audited runs)	Daily Go Test Parallelizer: 5.78; Design Decision Gate: 1.74

Per-Run First-Request Metrics

Run	Workflow	Chars	Headings	Skill refs	Inline agents	Dup line ratio	WSRF
33268456857	Daily Go Test Parallelizer	15,845	7	0	0	0.0	5.78
33269199660	AI Moderator	16,124	13	0	0	0.0	6.21
33269954213	Design Decision Gate 🏗️	27,549	44	0	0	0.0	1.74
33269962979	PR Description Updater	12,780	13	0	0	0.0	1.26

Repeated Ambient Context Signals

• No cross-run duplicate line/paragraph fragments detected (each workflow's static context is workflow-specific, not shared boilerplate) — so cross-workflow dedup is low-value here; the leverage is within-run turn repetition (WSRF) and inline template bloat, not shared text across workflows.
• All 4 samples show inline_agent_count: 0 and inline_skill_count: 0 at the top level of the extracted first request (agent/skill blocks appear later in the source file, e.g. ## agent: in Daily Go Test Parallelizer and PR Description Updater, ## skill: adr-report-templates in Design Decision Gate) — these get compiled into the prompt but were not present in the first-request preamble slice analyzed.
• Design Decision Gate embeds 3 full report templates (ADR-required / ADR-verified / diverges) inline under one skill heading rather than as 3 separately-loadable skills — all 3 load unconditionally even though only one applies per run.

Deterministic Analysis Output

• Script: /tmp/gh-aw/ambient-context/analyze_requests.py (stdlib only), outputs request-analysis.json / .md.
• Confirms 0% duplicate-line ratio across all 4 samples — bloat is structural (heading/template count), not literal repetition.
• Design Decision Gate has 44 headings vs. 7-13 for the other 3 workflows — a 3-4x heading density outlier.

Recommendations by Category

Workflow Markdown

• daily-go-test-parallelizer.md: safe immediately — tighten the sub-agent call contract in "Batched analysis agent" section to avoid restating the full safety-rule list on repeated invocations across turns (targets the WSRF 5.78 finding); impact: high.
• ai-moderator.md: safe immediately — flip tools.cli-proxy: false → true and tools.github.mode: local → gh-proxy to match the proxy pattern already adopted elsewhere in the repo; impact: medium.
• design-decision-gate.md: needs manual review — do not shrink below 1,000 chars or by >40% (currently 20,946 chars); relocate the 3 embedded report templates into separate on-demand skill files loaded only on the matching branch; impact: high.

Skills

• Extract Design Decision Gate's 3 report templates (### 🏗️ ADR Required, ### ✅ ADR Verified, ### ⚠️ ADR Diverges) into 3 discrete ## skill: blocks or files so only the applicable one is loaded per run instead of all three unconditionally; impact: medium, needs manual review (verify report-format compatibility with downstream consumers first).

Agents

• No inline-agent removal recommended this cycle — parallel-safety-batch-checker (Daily Go Test Parallelizer) and chunk-analyzer/pr-description-synthesizer (PR Description Updater) are each single-purpose and appropriately scoped; impact: low/none.

References

• §33268456857
• §33269954213
• §33269199660

Generated by 🌫️ Daily Ambient Context Optimizer · copilot · auto · 66.1 AIC · ⌖ 7.15 AIC · ⊞ 11.8K · ◷

• expires on Sep 5, 2026, 12:00 PM UTC-08:00