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

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows from the last 24 hours
• Workflows covered: Go Logger Enhancement, Sergo - Serena Go Expert, PR Sous Chef, Smoke Copilot
• Median first-request size: 17,140 chars | P95: 20,216 chars
• Key finding: The highest-cost run (Go Logger, 26.9M effective tokens / 93 turns) has a compact 6,583-char prompt — the problem is exploratory turn proliferation, not prompt bloat. Sergo (20,216 chars, 90 headings, 13 code fences) is the largest prompt and has structural over-engineering. Smoke Copilot (3.02% duplicate line ratio, 4,805-char Test Requirements section) has addressable prompt redundancy.

Highest-Leverage Changes

1. [Go Logger] Promote cache-check to a pre-flight gate step — currently buried mid-workflow; moving it to a dedicated no-LLM step eliminates all 93 exploration turns on unchanged days (HIGH)
2. [Go Logger] Replace iterative file-scan turns with a deterministic bash step — emit a structured manifest before any LLM invocation; ~50% of 93 turns are pure data-gathering (HIGH)
3. [Sergo] Collapse 90-heading structure to ≤15 headings — 1 heading per 7 lines inflates effective token weight; merge template/decorative headings into flat prose (HIGH)
4. [Sergo] Extract 13 embedded code fences into a conditional skill — static Go snippets and format examples add ~3–5k chars on every daily run; move to a sergo-examples skill loaded only when formatting guidance is needed (HIGH)
5. [Smoke Copilot] Trim ## Test Requirements section (4,805 chars) — largest single section; split into a shared import or collapse verbose sub-requirements (MEDIUM)
6. [Sergo] Cap lazy-loaded strategy memory to 3 most recent entries — ### 2.1 Load Previous Strategies loads full repo-memory history unconditionally; cap to recent entries (MEDIUM)
7. [Go Logger] Add structured safe-output schema for logger diffs — reduces per-turn token volume by replacing prose reasoning with compact patch objects (MEDIUM)

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	17,140
P95 chars	20,216
Largest sampled request	Sergo – 20,216 chars (646 lines, 90 headings, 13 code fences)
Highest effective tokens	Go Logger – 26,941,267 (93 turns, 6,583-char prompt)

Per-Run First-Request Metrics

Run	Workflow	Chars	Lines	Headings	Code Fences	Dup Line Ratio	Eff Tokens
26930135373	PR Sous Chef	14,315	198	8	0	0.00%	7,762,110
26931276933	Go Logger Enhancement	6,583	201	17	2	0.00%	26,941,267
26931476596	Smoke Copilot	19,965	304	22	4	3.02%	7,693,470
26932367410	Sergo - Serena Go Expert	20,216	646	90	13	0.20%	10,173,346

• All 4 runs use gh-proxy and cli-proxy — no MCP-to-CLI rewrite needed
• Request source: workflow .md source (claude runs) and agent process log (copilot runs)

Repeated Ambient Context Signals

• Smoke Copilot: 3.02% duplicate line ratio — safe-output boilerplate and tool-access rules repeated across sections
• Sergo: 90 headings create fragmented micro-sections; template placeholders like ### Task [N]: [Short Title] appear multiple times as copy-paste scaffolding
• PR Sous Chef: ## Token efficiency rules (mandatory) (963 chars) and ## Required skip rules per PR (749 chars) are verbose inline guardrails that could be condensed to a short shared import
• Go Logger: ## Efficiency First: Check Cache (922 chars) is effective guidance but placed too late — it runs after LLM context is already loaded

Deterministic Analysis Output

• Script: /tmp/gh-aw/ambient-context/analyze_requests.py (stdlib only)
• Output: request-analysis.json, request-analysis.md
• Key findings:
  • Sergo has the highest heading density: 90 / 646 lines = 0.14 headings/line
  • Smoke Copilot ## Test Requirements is the single largest section at 4,805 chars (24% of total prompt)
  • PR Sous Chef ## Required nudges for eligible PRs is 1,908 chars (13% of prompt) with dense bullet lists
  • Go Logger ambient_context effective_tokens=5,239 — prompt is efficient; token cost comes entirely from 93 LLM turns
  • Sergo ambient_context effective_tokens=12,560 — moderate; driven by 90-heading structure and 13 code fences

Recommendations by Category

Workflow Markdown

1. [Go Logger] Pre-flight cache gate (go-logger.md) — Move the cache check to a steps: entry before any LLM turn. On days with no new Go files, the workflow exits immediately with zero LLM calls. Expected impact: HIGH. Safe immediately: yes.
2. [Go Logger] Deterministic manifest step (go-logger.md) — Add a bash step that scans Go files and writes a JSON manifest (files needing logger, missing imports, call sites) to /tmp. The LLM prompt then reads this file instead of exploring iteratively. Eliminates ~46 data-gathering turns. Expected impact: HIGH. Safe immediately: yes.
3. [Sergo] Flatten heading structure (sergo.md) — Target ≤15 headings by merging the ~75 decorative/template sub-headings into numbered prose. Reduces effective token multiplier from heading-anchored context. Expected impact: HIGH. Safe immediately: yes (structural only).
4. [Sergo] Cap repo-memory load (sergo.md) — Replace unconditional Load Previous Strategies with a conditional load capped to the 3 most recent entries. Prevents unbounded memory growth inflating every future run. Expected impact: MEDIUM. Needs review.
5. [Smoke Copilot] Split ## Test Requirements (smoke-copilot.md) — At 4,805 chars (24% of prompt), this section is the largest. Move stable sub-requirements to a shared import and keep only the dynamic per-run discriminators inline. Expected impact: MEDIUM. Safe immediately: yes.

Skills

6. [Sergo] Create sergo-examples skill — Extract the 13 embedded code fences (Go snippets, output format templates) into .github/skills/sergo-examples/SKILL.md. Load conditionally in the workflow only when the agent needs output formatting guidance. Reduces base prompt by ~3–5k chars on every daily run. Expected impact: HIGH. Needs review.
7. [Go Logger] Structured safe-output patch schema — Define a compact safe-output type for logger diffs (file path, import line, call-site list) to replace multi-turn prose reasoning with a single structured emission. Expected impact: MEDIUM. Needs design review.

Agents

• No inline agent (## agent:) or skill (## skill:) definitions found in any sampled run. All 4 workflows use gh-proxy and cli-proxy correctly. No agent restructuring recommended at this time.

References

• §26931276933 — Go Logger Enhancement (26.9M eff tokens, 93 turns)
• §26932367410 — Sergo - Serena Go Expert (10.2M eff tokens, 90 headings)
• §26931476596 — Smoke Copilot (7.7M eff tokens, 3.02% dup lines)

Generated by 🌫️ Daily Ambient Context Optimizer · sonnet46 5M · ◷

• expires on Jun 11, 2026, 5:48 AM UTC