Executive Summary
- Runs listed: 60 (partial, 17:43–19:45 UTC); audited 1 (
37508923389). Reduced-data report: the run artifact directory /tmp/gh-aw/aw-mcp/logs was permission-denied in this sandbox, so no first-request text (event-logs.jsonl) could be extracted and the deterministic Python analysis was not run. No char/token metrics are claimed.
- Evidence used: run metadata (aic/tokens), one audit, and current workflow source sizes.
- Conclusion: the largest controllable ambient context is in large workflow markdown bodies (
issue-monster.md 41.5 KB, deep-report.md 18.8 KB) with inline step/report templates.
Highest-Leverage Changes
issue-monster.md (workflow-md, impact medium, manual review): ~41.5 KB; its prompt body spans ~230 lines of step-by-step process (lines 565–800) after ~500 lines of frontmatter/pre-steps. Move deterministic issue scoring/filtering fully into steps: and keep only decision rules in the prompt; the existing ## skill: blocks (token-budget, report-formatting) show the pattern. Run had aic 32.5 and 632,912 tokens. Keep any reduction ≤40%.deep-report.md (workflow-md + agents, impact medium, manual review): 18.8 KB, highest aic in sample (421.8, 46,368 tokens, failed). Report structure template (lines ~292–375, 8 headed sections) is inline; move it into a ## skill: block loaded on demand. Re-evaluate whether the inline agent: issues-analyst (line 380) justifies being in every first request.impeccable-skills-reviewer.md (skills, impact low): aic 175 / 74,490 tokens in a 1-turn run despite a 6 KB source — suggests skill/ambient content outside the workflow body dominates; capture first-request text next run to confirm before editing. Safe to investigate only.- Tooling readiness:
issue-monster has cli-proxy: true but only toolsets: [issues]; confirm needed toolsets and replace any raw gh aw shell wording with agentic-workflows MCP tool calls. Safe immediately after review.
- Not recommended (blocked by closed PR
#65568 in last 14 days): mattpocock-skills-reviewer.md, pr-code-quality-reviewer.md (both high-cost in sample: aic 133 and 46, tokens up to 715K).
CI-Validation Checklist for Implementing Agents
Key Metrics
| Metric |
Value |
| Sampled runs |
0 with first-request text (60 listed, 1 audited) |
| Distinct workflows |
n/a |
| Median chars |
n/a (artifacts inaccessible) |
| P95 chars |
n/a |
| Largest sampled request |
n/a |
| Merged optimizer PRs (7d) |
0 |
| Closed optimizer PRs (7d) |
1 |
| Optimizer PR close-rate (7d) |
null (<3 settled PRs; auto_pause false) |
| WSRF (audited runs) |
1.02 (run 37508923389; peak 5,418 vs cumulative 5,546 input tokens, 65 invocations) |
Per-Run First-Request Metrics
| Run |
Workflow |
aic |
tokens |
turns |
| 37511664816 |
Deep Report (failure) |
421.8 |
46,368 |
n/a |
| 37508923389 |
Impeccable Skills Reviewer |
175.0 |
74,490 |
1 |
| 37508897038 |
Matt Pocock Skills Reviewer |
132.8 |
41,971 |
2 |
| 37508923161 |
PR Code Quality Reviewer |
46.0 |
715,214 |
2 |
Request char/line metrics unavailable.
Repeated Ambient Context Signals / Deterministic Output
Not computed: first-request artifacts were not readable. WSRF of 1.02 for the audited run indicates context is not being rebuilt across turns, so per-turn resend is not the problem there; first-request size is.
Recommendations by Category
Workflow Markdown
Items 1, 2, 4 above.
Skills
Item 3 above.
Agents
Item 2 (inline issues-analyst agent in deep-report.md).
References
Generated by 🌫️ Daily Ambient Context Optimizer · copilot · auto · 23.7 AIC · ⌖ 14.5 AIC · ⊞ 12K · ◷
Executive Summary
37508923389). Reduced-data report: the run artifact directory/tmp/gh-aw/aw-mcp/logswas permission-denied in this sandbox, so no first-request text (event-logs.jsonl) could be extracted and the deterministic Python analysis was not run. No char/token metrics are claimed.issue-monster.md41.5 KB,deep-report.md18.8 KB) with inline step/report templates.Highest-Leverage Changes
issue-monster.md(workflow-md, impact medium, manual review): ~41.5 KB; its prompt body spans ~230 lines of step-by-step process (lines 565–800) after ~500 lines of frontmatter/pre-steps. Move deterministic issue scoring/filtering fully intosteps:and keep only decision rules in the prompt; the existing## skill:blocks (token-budget, report-formatting) show the pattern. Run had aic 32.5 and 632,912 tokens. Keep any reduction ≤40%.deep-report.md(workflow-md + agents, impact medium, manual review): 18.8 KB, highest aic in sample (421.8, 46,368 tokens, failed). Report structure template (lines ~292–375, 8 headed sections) is inline; move it into a## skill:block loaded on demand. Re-evaluate whether the inlineagent: issues-analyst(line 380) justifies being in every first request.impeccable-skills-reviewer.md(skills, impact low): aic 175 / 74,490 tokens in a 1-turn run despite a 6 KB source — suggests skill/ambient content outside the workflow body dominates; capture first-request text next run to confirm before editing. Safe to investigate only.issue-monsterhascli-proxy: truebut onlytoolsets: [issues]; confirm needed toolsets and replace any rawgh awshell wording withagentic-workflowsMCP tool calls. Safe immediately after review.#65568in last 14 days):mattpocock-skills-reviewer.md,pr-code-quality-reviewer.md(both high-cost in sample: aic 133 and 46, tokens up to 715K).CI-Validation Checklist for Implementing Agents
make recompilefor every modified.github/workflows/*.mdfile — zero compilation errors requiredmake agent-report-progressbefore the final commit and confirm it passesblocked_fileslist in/tmp/gh-aw/ambient-context/closed-pr-targets.json— do not re-attempt changes to any file that appears in a closed ambient-context optimization PR from the last 14 days.lock.ymlchanges in the PR bodyKey Metrics
Per-Run First-Request Metrics
Request char/line metrics unavailable.
Repeated Ambient Context Signals / Deterministic Output
Not computed: first-request artifacts were not readable. WSRF of 1.02 for the audited run indicates context is not being rebuilt across turns, so per-turn resend is not the problem there; first-request size is.
Recommendations by Category
Workflow Markdown
Items 1, 2, 4 above.
Skills
Item 3 above.
Agents
Item 2 (inline
issues-analystagent indeep-report.md).References