[ambient-context] Daily Ambient Context Optimizer - 2026-08-24

[github-actions]

Executive Summary

• Sampled 4 completed runs across 4 distinct workflows (last ~1h due to MCP logs count-limit continuation window; broader 24h sampling was not reachable within the tool timeout).
• No sandbox/firewall/logs/api-proxy-logs/event-logs.jsonl or sandbox/agent/logs/copilot-session-state/*/events.jsonl artifacts were present in the downloaded log bundles for any sampled run — analysis fell back to prompt.txt as the first-request source (documented as a fallback, not the preferred source).
• Median first-request size: 16,891 chars; P95: 28,259 chars (Design Decision Gate).
• Conclusion: the largest driver of first-request size across samples is a large shared preamble (~6.3–8.2K chars per run) plus a repeated safeoutputs create_pull_request CLI usage block (~2.5K chars) that appears verbatim in every sampled request — both are boilerplate injected once per run, not workflow-specific content, so they are the highest-leverage place to trim tokens without touching task logic.

Highest-Leverage Changes

1. Move the large inline "ADR Required / ADR Verified / Diverges" comment templates in design-decision-gate.md fully out of the prompt body into the already-present ## skill: adr-report-templates block's on-demand loading path (currently the templates are still large and always loaded).
2. Reduce heading/section sprawl in design-decision-gate.md (55 headings, 28K chars) — consolidate near-duplicate step/condition subsections (3a/3b/3c, Step 4a/4b) into fewer top-level steps.
3. Investigate high Working-Set Rebuild Factor (WSRF 3.62) on mattpocock-skills-reviewer — static context is being re-sent at ~3.6x its peak size across 36 invocations; move static skill/agent instructions to content that isn't resent every turn.
4. Trim the 3 inline ## agent: blocks in aw-failure-investigator.md (failure-classifier, issue-matcher, cluster-evidence-extractor) — evaluate whether all three need to be inline vs. converted to on-demand skills, given the workflow's already-large 493-line source.
5. impeccable-skills-reviewer.md shows 2 imported skill refs pulling full SKILL.md content into a lean 156-line workflow — verify only the necessary skill sections are being fused rather than entire skill files.

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	16,891.5
P95 chars	28,259
Largest sampled request	Design Decision Gate — 28,259 chars
Merged optimizer PRs (7d)	1
Closed optimizer PRs (7d)	0
Optimizer PR close-rate (7d)	insufficient data (<3 settled)
WSRF (audited runs)	Failure Investigator: 1.74; Matt Pocock Skills Reviewer: 3.62

Per-Run First-Request Metrics

Run	Workflow	Status	Chars	Headings	Inline agents/skills	Dup line ratio	WSRF
§32766584101	Design Decision Gate	failure	28,259	55	0 agents / 1 skill	0.0	n/a
§32766584217	Matt Pocock Skills Reviewer	success	19,477	21	1 agent / 0 skills	0.0	3.62
§32765579386	[aw] Failure Investigator (6h)	success	14,306	12	3 agents / 0 skills	0.0	1.74
32766584175	Impeccable Skills Reviewer	success	13,974	11	0 agents / 2 imported skill refs	0.0	n/a

Repeated Ambient Context Signals

• A ~2.5K-char safeoutputs create_pull_request CLI-usage instruction block (from the shared <mcp-clis> section) appears verbatim in all 4 sampled requests — static, unconditional content resent every run.
• The immutable <system> security-policy preamble (sandbox/firewall/injection rules) is ~6.3–8.2K chars in every sample — necessary for safety but a candidate for compaction (e.g., shorter phrasing) rather than removal.
• No cross-run textual duplication beyond the shared boilerplate was detected (duplicate_line_ratio = 0.0 for all samples), indicating each workflow's task-specific content is largely unique — good, but means further savings must come from shrinking boilerplate/skill fusion, not de-duplicating task text.

Deterministic Analysis Output

• Script: /tmp/gh-aw/ambient-context/analyze_requests.py (stdlib only) computed byte/char/line/word counts, heading/list/code-fence/details/table counts, inline agent/linter/skill counts, imported SKILL.md reference counts, duplicate line/paragraph ratios, and longest sections per run.
• Longest sections by heading were dominated by (preamble) (6.3K–8.2K chars) and the recurring safeoutputs create_pull_request CLI block (~2.5K chars) in every run — confirming these are the top two byte contributors independent of workflow identity.
• design-decision-gate stood out with 55 headings and 22 code fences in 28K chars — the highest heading density of the sample, consistent with its long, template-heavy ADR report format still embedded near the skill boundary.

Recommendations by Category

Workflow Markdown

• design-decision-gate.md (evidence: 55 headings, 28,259 chars, largest sampled request): consolidate the 3a/3b/3c ADR-detection subsections and Step 4a/4b branches into fewer top-level steps; the report templates are already partially under ## skill: adr-report-templates — ensure remaining large <details> blocks (5 template variants) stay skill-scoped rather than duplicated in the main body. Expected impact: high. Needs manual review (do not reduce >40% or below 1,000 chars in one pass).
• aw-failure-investigator.md (evidence: 3 inline ## agent: blocks in a 493-line file, 14,306-char first request): assess whether failure-classifier, issue-matcher, and cluster-evidence-extractor are all needed as always-loaded inline agents vs. converting low-frequency-path agents to skills loaded on demand. Expected impact: medium. Needs manual review.
• Both audited workflows already have tools.cli-proxy: true and tools.github.mode: gh-proxy configured — no proxy-enablement rewrite is needed for this sample.

Skills

• impeccable-skills-reviewer.md (evidence: 2 imported SKILL.md references fused into a lean 156-line workflow, 13,974-char first request): confirm skill fusion only pulls the specific sections needed per review mode rather than entire skill files, to avoid unconditional inflation as more skills are added. Expected impact: medium. Likely safe immediately (verification-only, no content removal).
• design-decision-gate.md / adr-report-templates skill: fully move the ADR Required/Verified/Diverges comment templates (currently ~2.5K+ chars of markdown examples in the workflow body) into the skill file so they load only when a comment is actually being posted, not on every gate evaluation. Expected impact: high. Needs manual review.

Agents

• mattpocock-skills-reviewer.md (evidence: WSRF 3.62 across 36 invocations, cumulative input tokens 19,916 vs. peak 5,503 — static context resent ~3.6x its peak size): investigate why the single inline ## agent: pr-triage and shared instructions are being rebuilt near-peak-size repeatedly rather than growing incrementally; this is the strongest per-run WSRF signal in the sample and suggests context is not being cached/reused across turns. Expected impact: high. Needs manual review.
• aw-failure-investigator.md: given only 1.74 WSRF (healthier) but 3 inline agents, no urgent agent-count change is needed here beyond the workflow-md consolidation above. Expected impact: low. Safe immediately (monitoring only).

References

• §32766584101
• §32766584217
• §32765579386

Generated by 🌫️ Daily Ambient Context Optimizer · copilot · auto · 61.3 AIC · ⌖ 17.2 AIC · ⊞ 11.8K · ◷

• expires on Aug 31, 2026, 12:05 PM UTC-08:00