[ambient-context] Daily Ambient Context Optimizer - 2026-06-14

[github-actions]

Executive Summary

• 4 runs sampled across 4 distinct workflows (Smoke Copilot, Test Quality Sentinel, Linter Miner, Contribution Check)
• Median first-request: 14,265 chars · P95: 14,746 chars · Largest: 26,348 chars (Test Quality Sentinel)
• All 4 workflows already use gh-proxy and cli-proxy — no proxy-enablement gaps
• Highest-leverage finding: Test Quality Sentinel embeds a 13,770-char verbatim HTML report template in its first request; removing it saves >50% of that workflow's prompt chars
• Cross-workflow finding: 15 distinct boilerplate lines from the <system> block appear in every run (framework-injected); <safe-outputs> alone contributes ~2,607 chars per request and is the single largest repeated section
• Low char/token ratios (0.49–0.98) confirm that runtime-injected tool definitions add large invisible token overhead beyond what prompt.txt captures

Highest-Leverage Changes

1. [TQS] Move 13,770-char report template out of main prompt — extract to a dedicated skill or step and inject only when writing the PR comment (saves ~52% of TQS first-request chars)
2. [smoke-copilot] Delete ## Tool Access Overview — 1,179-char section duplicates the framework-injected <mcp-clis> block already present in the preamble
3. [smoke-copilot + TQS] Extract ## Report Structure Guidelines + ## Workflow Run References into a shared skill — 497 chars identical across both workflows, currently inlined in both
4. [smoke-copilot + linter-miner] Extract ### Analysis Constraints + ## Serena Code Analysis into a shared skill — ~974 chars repeated verbatim across both workflows
5. [TQS] Compress illustrative <details> blocks 2 & 3 — 1,710 chars of supplementary examples that restate instruction already in the main steps; inline as short prose
6. [TQS] Flatten heading depth from 43 → ≤10 — 43 headings for 8 main steps fragments instructions into small named sections, adding token overhead without clarity gain

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	14,265
P95 chars	14,746
Largest sampled request	26,348 chars (Test Quality Sentinel)
Framework <safe-outputs> overhead	2,607 chars (present in all 4 runs)
Shared boilerplate lines (all 4 runs)	15 distinct repeated lines

Per-Run First-Request Metrics

Run	Workflow	Chars	Tokens	AIC	char/tok	dup_line%	Conclusion
§27507898911	Smoke Copilot	13,784	19,721	1015.5	0.70	0.9%	failure
§27507998075	Test Quality Sentinel	26,348	26,843	932.5	0.98	0.9%	success
§27507182843	Linter Miner	14,746	17,504	502.9	0.84	0.0%	success
§27506199838	Contribution Check	11,994	24,738	503.3	0.49	0.0%	success

Note on char/token < 1: prompt.txt captures only the workflow text. The full DLLM request also includes runtime-injected tool schema definitions (not in prompt.txt), which explains token counts exceeding prompt chars for Contribution Check.

Repeated Ambient Context Signals

Framework-injected (in all 4 runs, not changeable via workflow .md):

• <safe-outputs> block: 2,607 chars per run — immutable security + guidance boilerplate
• Security policy lines (4 sentences on injection/escape): ~600 chars repeated in every preamble
• <temporary-files>, <environment-limitations>, <markdown-generation>, <mcp-clis>, <gh-cli>: minor but consistent

Cross-workflow duplicates (fixable in workflow .md files):

• ## Report Structure Guidelines (248 chars): identical in smoke-copilot + test-quality-sentinel
• ## Workflow Run References (249 chars): identical in smoke-copilot + test-quality-sentinel
• ### Analysis Constraints (300 chars): identical in smoke-copilot + linter-miner
• ## Serena Code Analysis (263–674 chars): similar across smoke-copilot + linter-miner

Single-workflow oversizing:

• test-quality-sentinel <details> block 1: 13,770 chars — full HTML report template inlined
• smoke-copilot ## Tool Access Overview: 1,179 chars — duplicates framework <mcp-clis> block

Deterministic Analysis Output

Script: /tmp/gh-aw/ambient-context/analyze_requests.py (stdlib only)
Output: /tmp/gh-aw/ambient-context/request-analysis.json

Key script-derived findings:

• Test Quality Sentinel: heading_count=43, code_fence_count=24, list_item_count=85, html_details_count=9, table_rows=19 — substantially more structural markup than any other sampled workflow
• Contribution Check: char/tok=0.485 — lowest ratio, suggesting high tool-definition injection overhead at runtime (prompt.txt is only 11,994 chars but 24,738 tokens actually sent)
• 0 inline skills / skill SKILL.md refs across all 4 runs — no workflows in this sample use dynamic skill loading; all context is inlined in prompt.txt
• 0 inline linters across all 4 runs
• safe_output_lines per run: Contribution Check=19, Smoke Copilot=18, Linter Miner=16, TQS=13 — safe-output guidance dominates the <system> block across all runs
• Most common fragment across runs (4×): entire <safe-outputs> content paragraphs (2,607 chars total per run)
• Duplicate line ratio ≤ 0.9% in all runs — per-run duplication is low; the issue is cross-run repetition and oversized single sections

Recommendations by Category

Workflow Markdown

1. [HIGH] Extract Test Quality Sentinel's verbatim report template (test-quality-sentinel.md)

• Evidence: The first <details> block (13,770 chars, 52% of prompt) is a static HTML/Markdown rendering template with placeholder variables like {TOTAL}, {DESIGN_COUNT}. It has no instructional value at first-request time.
• Action: Move this template to a skill (e.g., tqs-report-template) or an external file. Reference it only in Step 7 when writing the PR comment.
• Impact: HIGH — saves ~13,570 chars and reduces TQS first-request tokens by ~50%
• Safe immediately: Yes — the template is purely a rendering guide; moving it to a skill does not change behavior

2. [HIGH] Remove ## Tool Access Overview from smoke-copilot.md

• Evidence: 1,179 chars (8.5% of Smoke Copilot prompt). The framework already injects a <mcp-clis> block listing all CLI proxied servers with usage examples. This section is a verbatim restatement of that injected content.
• Action: Delete the section entirely, or replace with a one-liner: "See <mcp-clis> above for CLI-proxy server usage."
• Impact: HIGH — saves ~1,050 chars per Smoke Copilot run
• Safe immediately: Yes — agents should rely on <mcp-clis> as the canonical source

3. [MEDIUM] Compress illustrative <details> blocks 2 & 3 in test-quality-sentinel.md

• Evidence: Block 2 (928 chars, example flagged test) and Block 3 (782 chars, classification reference) total 1,710 chars of supplementary scaffolding that restates instruction already in the main workflow steps.
• Action: Inline one compact example (~100 chars) directly under the relevant rule; fold the key classification distinction (design vs. implementation) into a single sentence under its parent heading. Delete both standalone <details> blocks.
• Impact: MEDIUM — saves ~1,380 chars
• Safe immediately: Yes

Skills

4. [MEDIUM] Extract cross-workflow boilerplate into a shared skill

• Evidence: ## Report Structure Guidelines (248 chars) and ## Workflow Run References (249 chars) are byte-for-byte identical in smoke-copilot.md and test-quality-sentinel.md. ### Analysis Constraints (300 chars) is byte-for-byte identical in smoke-copilot.md and linter-miner.md.
• Action: Create one shared skill (e.g., reporting-conventions) containing all three sections. Replace inline text with a skill import in each affected workflow.
• Impact: MEDIUM — saves ~717 chars per affected workflow; simplifies cross-workflow maintenance
• Safe immediately: Yes

5. [MEDIUM] Consolidate Serena Code Analysis into a shared skill

• Evidence: ## Serena Code Analysis appears in both linter-miner.md (263 chars) and smoke-copilot.md (674 chars) with nearly identical instructions for activating the Serena MCP tool on the go workspace.
• Action: Extract to a serena-go-analysis shared skill. Keep only workflow-specific overrides inline (~50 chars residual per workflow).
• Impact: MEDIUM — saves ~520–870 chars combined; ensures Serena usage guidance stays consistent across all Go-analysis workflows
• Safe immediately: Yes

Agents

6. [MEDIUM] Flatten test-quality-sentinel.md heading structure from 43 to ≤10 top-level headings

• Evidence: heading_count=43, code_fence_count=24 for 8 logical steps. The deep nesting (e.g., ### 3a. Go — Test* functions, ### Red Flags to Detect, ### Calibration) fragments continuous instruction into small named sections without improving the agent's performance.
• Action: Convert deeply nested sub-headings into bullet lists or inline prose under parent steps. Target ≤10 headings total.
• Impact: MEDIUM — saves ~660 chars; reduces first-request structural overhead
• Safe immediately: Needs manual review — verify the agent references headings programmatically before removing them

References

• §27507898911 — Smoke Copilot (AIC 1015, failure)
• §27507998075 — Test Quality Sentinel (AIC 932, success)
• §27507182843 — Linter Miner (AIC 503, success)

Generated by 🌫️ Daily Ambient Context Optimizer · 1.2K AIC · ⌖ 14.5 AIC · ⊞ 21.9K · ◷

• expires on Jun 21, 2026, 12:24 PM UTC-08:00

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #39453

View newer issue

---

This action was performed automatically by the Daily Ambient Context Optimizer workflow.