[ambient-context] Daily Ambient Context Optimizer - 2026-06-04

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows in the last 24h
• Workflows: Failure Investigator (6h), Daily Code Metrics, Smoke Copilot, Copilot Agent PR Analysis
• Median first-request: 31,680 chars; P95: 36,827 chars
• Highest-cost: Failure Investigator — 10.3M effective tokens (34 turns); every duplicated byte costs ~34×
• Top drivers: repeated Python dataviz boilerplate (12.9% dup line ratio); preambles at 22–46% of prompt; gh-proxy missing on 2/4 workflows; redundant safe-output prose in bodies; over-structured headings (up to 129/975 lines)

Highest-Leverage Changes

1. Dedup shared/python-dataviz.md boilerplate — 4× repeated matplotlib/seaborn setup drives 12.9% dup lines in Daily Code Metrics; one canonical setup block replaces all copies [workflow-md, high, safe]
2. Add tools.github.mode: gh-proxy to daily-code-metrics.md and copilot-agent-analysis.md — removes pre-auth CLI narrative from the injected preamble on every turn [workflow-md, high, safe]
3. Strip redundant safe-output prose from aw-failure-investigator.md body — 12 safe-output keyword hits; preamble is already authoritative; 34-turn multiplier amplifies the cost [workflow-md, high, safe]
4. Defer 3,495-char report template in daily-code-metrics.md to a late step — conditional experiment variant content is dead weight on every first request [workflow-md, high, needs review]
5. Flatten heading density in daily-code-metrics.md (129h/975l) and copilot-agent-analysis.md (75h/748l) — merge stub sections into parents [workflow-md, medium, safe]
6. Fix duplicate Serena context in smoke-copilot.md — Workspace/Memory/Project Activation lines appear 2× from a shared import collision [workflow-md, medium, safe]

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	31,680
P95 chars	36,827
Largest request	Daily Code Metrics (36,827 chars, 129 headings, 29 code fences)

Per-Run First-Request Metrics

Run	Workflow	Chars	Lines	Headings	Code Fences	Dup%	Eff.Tokens	gh-proxy
§26974650815	Failure Investigator (6h)	12,479	217	21	1	1.3%	10.3M	✓
§26973556771	Daily Code Metrics	36,827	975	129	29	12.9%	8.9M	✗
§26975370978	Smoke Copilot (failed)	16,833	291	22	4	5.9%	8.1M	✓
§26973141611	Copilot Agent PR Analysis	31,680	748	75	19	9.0%	6.8M	✗

All 4 prompts sourced from prompt.txt. Proxy columns reflect source frontmatter config.

Repeated Ambient Context Signals

Preamble dominance: auto-injected system preamble is the #1 largest section in all 4 prompts (5,691–8,245 chars; 22–46% of prompt total).

Daily Code Metrics repeated lines (from shared/python-dataviz.md multi-example blocks):

• import matplotlib.pyplot as plt — 4×
• import seaborn as sns / sns.set_style("whitegrid") — 4× each
• sns.set_palette("husl") / fig, ax = plt.subplots(...) — 3× each

Copilot Agent PR Analysis: jq 'length' and gh pr list patterns appear 2× (Alternative Approaches section duplicates primary).

Smoke Copilot: Workspace, Memory, ### Project Activation all appear 2× from shared import collision.

Deterministic Analysis Output

Script: analyze_requests.py (stdlib only) → request-analysis.json + request-analysis.md

Ambient context only (preamble, from audit): 5,220 tokens (Failure Investigator), 15,629 tokens (Daily Code Metrics). The 10.3M vs 1.8M actual-token gap in Failure Investigator implies ~5.5× per-turn context-window multiplier.

Keyword density (lines mentioning topic):

Workflow	tools	agents	safe-out	workflow
Failure Investigator	18	6	12	11
Daily Code Metrics	26	4	17	22
Smoke Copilot	42	6	18	13
Copilot Agent PR Analysis	29	36	13	18

Recommendations by Category

Workflow Markdown

1. shared/python-dataviz.md — replace 4 per-example setup stubs with one canonical ## Chart Setup block [high, safe — affects all importers]
2. daily-code-metrics.md + copilot-agent-analysis.md — add github: { mode: gh-proxy } to tools: frontmatter [high, safe]
3. aw-failure-investigator.md — audit body for safe-output re-documentation (12 hits) and remove; preamble is authoritative [high, safe]
4. daily-code-metrics.md — move 3,495-char conditional report template to late-step skill or <!-- include --> [high, needs review — verify turn sequencing]
5. daily-code-metrics.md + copilot-agent-analysis.md — merge #### stub sections (<4 body lines) into parents to cut heading count [medium, safe]
6. smoke-copilot.md — remove duplicate shared/mcp/serena-go.md import or deduplicate activation block [medium, safe]

Skills

SKILL.md refs = 0 across all 4 runs. Skills load lazily and did not contribute to first-request bloat in this sample. No changes needed.

Agents

Inline agent count = 0 across all 4 runs. Sub-agent dispatch is prose-described; no inline agent definitions found. No agent-definition changes recommended.

References

• §26974650815 — Failure Investigator (6h), 10.3M eff. tokens, 34 turns
• §26973556771 — Daily Code Metrics, 36,827 char prompt, 12.9% dup
• §26973141611 — Copilot Agent PR Analysis, 31,680 chars, 75 headings

Generated by 🌫️ Daily Ambient Context Optimizer · sonnet46 6.5M · ◷

• expires on Jun 11, 2026, 8:38 PM UTC