[ambient-context] Daily Ambient Context Optimizer - 2026-06-12

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows (Daily Safe Output Integrator, Test Quality Sentinel, Matt Pocock Skills Reviewer, Failure Investigator)
• Median first-request: 21,450 chars | P95: 30,318 chars
• Largest bloat: DSOI at 30,318 chars with 13 near-identical inline sub-agent stubs (~15,000 chars, 49% of prompt)
• Critical hidden bloat: MPSR prompt.txt shows 14,861 chars but actual first-request = 27,826 tokens (char/tok=0.53), because the runtime injects all 11 .github/agents/ files (~140KB) regardless of workflow scope

Note: No event-logs.jsonl found; sizes are from aw-prompts/prompt.txt. Token counts are from token-usage.jsonl first-entry.

---

Highest-Leverage Changes

1. DSOI: Parameterize 13 type-specific inline sub-agents into a template table — saves ~13,000 chars (43%)
2. All Copilot workflows: Filter irrelevant .github/agents/ injection — create-safe-output-type.agent.md (23,730 bytes) + custom-engine-implementation.agent.md (27,203 bytes) are irrelevant to PR-review workflows
3. TQS: Move binary-penalty scoring algorithm to a steps: script — 10,118-char embedded formula can be pre-computed
4. DSOI: Replace Phase 5 static Go field-mapping table with a source-derive instruction — 1,530 chars of static data already derivable from safe_outputs_validation_config.go
5. shared/reporting.md: Remove design-philosophy and example prose — 1,200 chars of "Airbnb-inspired" copy across all importing workflows
6. MPSR: Compress Steps 4 & 5 to tight bullets — 2,686 chars → ~1,486 chars; high-AIC workflow (372, 22 turns)

---

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	21,450
P95 chars	30,318
Largest request	30,318 chars (DSOI)
Avg char/token ratio	1.33
Lowest char/tok	0.53 — Matt Pocock Skills Reviewer

Per-Run First-Request Metrics

Run	Workflow	Chars	Lines	Hdgs	Fences	Dup%	Input Tokens	Char/Tok	AIC
§27438027894	Failure Investigator	12,212	207	21	2	0.0%	5,117	2.39	399.95
§27439111642	Matt Pocock Skills Reviewer	14,861	254	20	10	0.0%	27,826	0.53	372.03
§27438448655	Test Quality Sentinel	28,038	512	51	26	0.7%	27,245	1.03	221.50
§27437733590	Daily Safe Output Integrator	30,318	831	68	50	0.6%	22,245	1.36	217.48

Repeated Ambient Context Signals

• shared/reporting.md: 1,934 chars of ## Report Structure Guidelines injected identically into TQS and DSOI
• 13 DSOI type sub-agents: # Test Copilot Update Pull Request etc. repeat same ## Task / ## Steps / ## Output scaffold 13× (only type name differs)
• TQS scoring verbosity: ### Calibration (1,519) + ### Red Flags to Detect (1,863) + ### Verdict (1,127) = 4,509 chars that could be compressed
• <system> boilerplate: 5,273–6,828 chars of runtime security/safe-output prose across all 4 runs
• Most repeated fragments: for sotype in safe_output_types: (×2), cat /tmp/gh-aw/agent/diff-numstat.txt (×2)

Deterministic Analysis Output

Longest sections per workflow:

• DSOI: ## Phase 5: Add Go Compiler Tests = 2,745 chars
• TQS: ### Red Flags to Detect = 1,863 chars; embedded scoring formula block = 10,118 chars
• MPSR: ### Step 4: Review Using Selected Skills = 1,602 chars
• FI (Claude Code): ## Tone Variant Instructions = 801 chars; only 5,117 tokens — most efficient engine by far

Key structural insight: FI uses Claude Code engine (char/tok=2.39, 5,117 tokens); all Copilot-engine workflows are 4–5× more expensive at first request. Agent injection is the primary driver for MPSR's 0.53 char/tok ratio.

---

Recommendations by Category

Workflow Markdown

1. DSOI — Parameterize 13 type-specific sub-agents [HIGH] · saves ~13,000 chars

• Replace 13 # Test Copilot <TypeName> blocks with one compact markdown table (type | perm-scope | task | tool-key) + existing template
• Safe to apply immediately

2. TQS — Extract scoring algorithm to steps: script [HIGH] · saves ~8,000–10,000 chars

• Move # Binary penalty: deduct all 10 points... (10,118 chars) to a pre-agent shell step; emit result to /tmp/gh-aw/agent/inflation-score.txt
• Needs functional testing before applying

3. DSOI — Replace static Phase 5 field-mapping table with source-derive instruction [MEDIUM] · saves ~1,530 chars

• 13-entry lookup table is fully derivable from pkg/workflow/safe_outputs_validation_config.go; replace with grep instruction
• Safe to apply immediately

4. shared/reporting.md — Remove design-philosophy and example prose [MEDIUM] · saves ~1,200 chars per importer

• Delete ### Design Principles (Airbnb-Inspired) and ### Example Report Structure blocks
• Safe to apply; benefits every workflow importing reporting.md

Skills

5. MPSR — Compress Steps 4 & 5 to action bullets [MEDIUM] · saves ~1,200 chars

• Collapse 1,602-char Step 4 and 1,084-char Step 5 to ≤3 bullets each
• Directly reduces AIC across 22 turns per run

Agents

6. All Copilot workflows — Scope-filter .github/agents/ injection [HIGH]

• Add applyTo filtering or an explicit agents: allowlist in workflow frontmatter
• create-safe-output-type.agent.md (23,730 bytes) + custom-engine-implementation.agent.md (27,203 bytes) irrelevant to PR-review workflows = 50,933 chars of avoidable injection
• Needs runtime support verification

---

References

• §27439111642 — MPSR (aic=372, char/tok=0.53)
• §27438448655 — TQS (aic=221, 28,038 chars)
• §27437733590 — DSOI (aic=217, 30,318 chars, failed)

Generated by 🌫️ Daily Ambient Context Optimizer · 1.9K AIC · ⌖ 26.5 AIC · ⊞ 21.8K · ◷

• expires on Jun 19, 2026, 12:44 PM UTC-08:00