[ambient-context] Daily Ambient Context Optimizer - 2026-06-07

[github-actions]

Executive Summary

• 4 runs sampled from 4 distinct workflows over the last 24 hours
• Workflows covered: Design Decision Gate, Test Quality Sentinel, Documentation Unbloat, Linter Miner
• Median first-request size: 23,656 chars · P95: 28,404 chars
• Highest-cost run: Documentation Unbloat at AIC 527 / 68 turns — prompt size is modest (18.9K chars) but runaway turn count drives cost
• Largest prompt: Design Decision Gate at 34,771 chars, where the imported adr-writer agent definition accounts for 43% of workflow content
• All 4 workflows already have gh-proxy and cli-proxy enabled ✓

Highest-Leverage Changes

1. [agents] Slim adr-writer.agent.md in Design Decision Gate — 11,713 chars (43%) is ADR template boilerplate loaded on every PR gate check; extract the normative spec/template to a lazy-loaded skill or strip to decision-only instructions
2. [workflow-md] Fix Documentation Unbloat Step 2 — agent re-runs search()+find when candidate-files.txt is already pre-computed; eliminating this saves an estimated 10–20 turns per run (primary driver of AIC 527)
3. [workflow-md] Scope serena-go.md import in Linter Miner to the code-pattern-scanner sub-agent — saves 3,432 chars (21% of first request) since the main agent never calls Serena tools directly
4. [agents] Write Linter Miner sub-agent outputs to /tmp/gh-aw/agent/candidates.json instead of returning inline — likely driver of 669K effective tokens in a single-turn run
5. [workflow-md] Remove duplicated exclusion guards from Documentation Unbloat Steps 2+3 — pre-flight bash already enforces all exclusions; ~900 chars of redundant prose
6. [skills] Drop shared/otlp.md from Documentation Unbloat — 635 chars with no functional use in a doc-editing workflow
7. [skills] Extract inline linter-writer agent body in Linter Miner to a reusable skill — ~60 lines of file layout and compilation commands duplicated from the existing go-linters skill

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	23,656
P95 chars	28,404
Largest sampled request	Design Decision Gate (34,771 chars)
Highest AIC run	Documentation Unbloat (526.96 AIC, 68 turns)
Avg duplicate line ratio	7.7%

Per-Run First-Request Metrics

Run ID	Workflow	Chars	AIC	Effective Tokens	Turns	Dup Line %	Code Fences	Headings
§27098774407	Documentation Unbloat	18,909	526.96	1,756,540	68	6.8%	20	34
§27100128932	Linter Miner	16,161	132.66	669,577	1	4.9%	4	18
§27100525416	Design Decision Gate	34,771	100.04	333,459	—	8.9%	28	90
§27100525434	Test Quality Sentinel	28,404	115.80	386,007	—	10.1%	26	51

Repeated Ambient Context Signals

• shared/reporting.md (2,231 chars) appears identically in 3 of 4 sampled prompts; loaded by 53 workflows total — expected baseline but confirms universal overhead
• ADR template boilerplate (# ADR-{NNNN}: ..., **Date**: {YYYY-MM-DD}, ## Examples of ADR-Worthy Decisions) appears twice within the Design Decision Gate prompt — narrative template duplicated in the normative spec section
• cat /tmp/gh-aw/agent/diff-numstat.txt appears twice in Test Quality Sentinel (Steps 1 and a later verification step)
• "event": "REQUEST_CHANGES" JSON fragment appears twice in Test Quality Sentinel — duplicated across two verdict-path examples
• Exclusion rules (blog/, frontmatter-full.md, disable-agentic-editing) appear 13 times in Documentation Unbloat Steps 2+3 despite pre-flight bash already enforcing them
• The (preamble) section (system block + imports) is the largest single section in all 4 runs, ranging from 69–89 lines

Deterministic Analysis Output

Script: analyze_requests.py (stdlib only), reading 4 × run-*.txt / run-*.json samples.

Key computed metrics:

• System block (runtime-injected): 5,667–7,702 chars depending on safe-output configuration
• Workflow-specific content after </system>: 8,459–27,134 chars
• adr-writer.agent.md inlined: 11,713 chars = 43.2% of Design Decision Gate workflow content
• serena-go.md + transitive serena.md = 3,432 chars in Linter Miner preamble; main agent never calls Serena tools
• shared/reporting.md = 2,231 chars; shared/otlp.md = 635 chars (loaded by Documentation Unbloat but functionally unused)
• Code fence density highest in Design Decision Gate (28) and Test Quality Sentinel (26) — dense inline bash examples per step
• Duplicate line ratios: Test Quality Sentinel 10.1%, Design Decision Gate 8.9%, Documentation Unbloat 6.8%, Linter Miner 4.9%

Recommendations by Category

Workflow Markdown

R1 — Documentation Unbloat: replace Step 2 file-discovery with cat candidate-files.txt · Impact: high · Safe to automate

• Step 2 instructs the agent to call search() + find even though pre-agent-steps already writes /tmp/gh-aw/agent/candidate-files.txt with all exclusions applied
• Step 0 acknowledges the file exists but Step 2 ignores it — forces ≥2 extra turns before file selection
• With 68 turns × ~25K avg tokens/turn, each eliminated turn pair saves ~50K tokens
• Fix: collapse Step 2 to a single cat /tmp/gh-aw/agent/candidate-files.txt instruction

R2 — Documentation Unbloat: deduplicate exclusion guards · Impact: medium · Safe to automate

• Exclusion rules for blog/, frontmatter-full.md, disable-agentic-editing appear 13 times across Steps 2–3; pre-flight already enforces them
• Removing inline NEVER/IMPORTANT warnings saves ~900 chars with no behavioral change

R3 — Linter Miner: scope serena-go.md import to code-pattern-scanner sub-agent · Impact: high · Safe to automate

• serena-go.md + transitive serena.md = 3,432 chars injected in the main-agent preamble
• Main agent never calls activate_project, find_symbol, search_for_pattern, or find_referencing_code_snippets — only code-pattern-scanner does
• Move the import from top-level imports: to the code-pattern-scanner agent block
• Saves 3,432 chars (21% of 16,161 first-request chars)

Skills

R4 — Documentation Unbloat: remove shared/otlp.md import · Impact: medium · Safe to automate

• shared/otlp.md (635 chars, 26 lines) covers OTLP tracing configuration; a doc-editing workflow never emits traces
• Removal cuts ~3.4% from the 18,909-char first request at zero functional cost

R5 — Linter Miner: extract linter-writer agent body to a reusable skill · Impact: medium · Needs manual review

• Inline ## agent: linter-writer block duplicates file layout, package structure, and compilation commands partly covered by .github/skills/go-linters/SKILL.md
• Extracting to .github/skills/linter-writer/SKILL.md makes the definition shareable across linter-creating workflows

Agents

R6 — Design Decision Gate: trim adr-writer agent to gate-check scope · Impact: high · Needs manual review

• adr-writer.agent.md (270 lines, 9,787 bytes) fully inlined: 11,713 chars = 43% of workflow content on every PR gate check
• Most runs only need the ## Procedure: Writing a New ADR section; the full RFC 2119 normative spec, dual-part template, and writing-quality standards are only needed when actually drafting
• Options: (a) split into slim gate-check agent + separate ADR-drafting skill loaded lazily, or (b) move normative spec and template body to a skill referenced only when drafting

R7 — Linter Miner: write sub-agent candidate outputs to temp file · Impact: high · Needs manual review

• discussion-miner and code-pattern-scanner return candidate lists inline into the main agent context, likely driving 669K effective tokens for a single-turn run
• Pattern: instruct sub-agents to write output to /tmp/gh-aw/agent/candidates.json (consistent with existing prior-linters.json pattern); main agent reads the file in Step 3
• Keeps sub-agent verbosity outside the main-agent context window

References

• §27098774407 — Documentation Unbloat (AIC 527, 68 turns)
• §27100128932 — Linter Miner (AIC 133, 669K eff. tokens, 1 turn)
• §27100525416 — Design Decision Gate (AIC 100, 34.8K chars)

Generated by 🌫️ Daily Ambient Context Optimizer · 937.1 AIC · ⌖ 35.6 AIC · ⊞ 21K · ◷

• expires on Jun 14, 2026, 12:17 PM UTC-08:00

[pelikhan]

/q to review how the analysis was done for this report, ensure it is based on the API proxy event log. That reflects the text actually sent to the LLM, rather than the prompt.

Inline agents and inline linters should be removed from the prompt and automatically refactored out. If they are not, that is a bug that needs to be fixed.

[github-actions]

🎩 Mission equipment ready! Q has optimized your workflow. Use wisely, 007! 🔫