[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: Smoke Codex, Smoke Claude, [aw] Failure Investigator (6h), AI Moderator
• Median first-request size: 19,020 chars | P95: 32,686 chars
• Highest-level finding: The two smoke test workflows (Smoke Claude, Smoke Codex) include large shared skills—particularly shared/github-queries-mcp-script.md (~16.5 KB)—that are largely unused for smoke-test purposes. The Failure Investigator embeds MCP tool parameter docs inline that could be fetched at runtime, and ~50% of its 20 turns are data-gathering that should move to deterministic steps. Security policy boilerplate (4 lines × 4 runs) and safe-output noop examples (5× each run) are repeated verbatim across every prompt.

---

Highest-Leverage Changes

1. [skills] Drop shared/github-queries-mcp-script.md from smoke tests — it contributes ~16.5 KB (~50% of Smoke Claude's 32 KB prompt) but only repos and pull_requests toolsets are configured.
2. [workflow-md] Remove inline MCP tool parameter docs from aw-failure-investigator.md — the top two sections (audit + CLI usage) add >3 KB that can be fetched via --help at runtime.
3. [agents] Move failure run ID discovery to a deterministic pre-step in aw-failure-investigator.md — audit evidence shows ~50% of 20 turns are data-gathering, inflating effective tokens to 9.3 M.
4. [workflow-md] Deduplicate repeated boilerplate in smoke workflows — dup_line_ratio of 0.111 (Smoke Claude) and 0.090 (Smoke Codex) means 9–11% of lines are exact duplicates (workspace paths, memory paths, "Verify the tool call succeeds" repeated 5×).
5. [skills] Enable gh-proxy and cli-proxy on smoke workflows — none of the three largest-prompt workflows (Smoke Claude, Smoke Codex, Failure Investigator) have tools.github.mode: gh-proxy or tools.cli-proxy: true. Only AI Moderator has cli-proxy: true. Enabling these allows MCP-tool-oriented wording to be replaced with compact CLI-form commands.
6. [skills] Extract reporting structure guidelines from aw-failure-investigator.md into a referenced shared skill — currently ~2 KB of formatting rules (heading levels, progressive disclosure, report structure pattern) are inlined.
7. [workflow-md] Trim the safe-output noop example JSON (appears 5× across all 4 runs) — a one-line reference to the schema is sufficient.

---

Key Metrics

Metric	Value
Sampled runs	4
Distinct workflows	4
Median chars	19,020
P95 chars	32,686
Largest sampled request	Smoke Claude (32,686 chars)

Per-Run First-Request Metrics

Run ID	Workflow	Chars	Lines	Headings	Code Fences	Dup Line Ratio	Token Usage
26922082621	Smoke Claude	32,686	681	56	19	0.111	1,392,018
26922082608	Smoke Codex	20,451	411	23	9	0.090	1,589,240
26924689485	[aw] Failure Investigator (6h)	17,589	356	31	11	0.053	1,541,840
26924934999	AI Moderator	13,803	258	14	5	0.011	90,124

Repeated Ambient Context Signals

These lines/fragments appear verbatim across 4 or more sampled run prompts:

• [5×] {"noop": {"message": "No action needed: [brief explanation...]"}} — safe-output noop example repeated in every run
• [5×] - Verify the tool call succeeds — repeated within Smoke Claude alone
• [4×] Immutable security policy block (3 lines, ~400 chars) — injected identically into all 4 prompts; runtime injection; no action needed but confirms system overhead
• [4×] <instruction>When you need to create temporary files...</instruction> — system injection, expected
• [4×] Do NOT attempt to edit files outside these directories — system injection, expected

Largest unique skill sections driving Smoke Claude size:

• shared/github-queries-mcp-script.md — ~16.5 KB inline (~50% of total prompt)
• #### audit — Inspect a specific run — 6,693 chars
• # Works with any tool — just match the parameter names — 3,699 chars (appears in Smoke Codex and Failure Investigator too)

Deterministic Analysis Output

Script: /tmp/gh-aw/ambient-context/analyze_requests.py
Output: /tmp/gh-aw/ambient-context/request-analysis.json + .md

Key computed metrics:

• Smoke Claude: 56 headings, 186 list items, 19 code fences, 104 "tools" keyword hits, 43 "safe_outputs" keyword hits
• Smoke Codex: 23 headings, 104 list items, 9 code fences, 58 "tools" keyword hits
• Failure Investigator: 31 headings, 93 list items, eff. tokens = 9,346,226 (20 turns × ~467K avg)
• AI Moderator: 14 headings, lowest dup ratio (0.011), lowest prompt size (13,803 chars) — closest to well-scoped baseline

Audit for Failure Investigator: agentic_fraction=0.50 — half of 20 turns are pure data gathering.

---

Recommendations by Category

Workflow Markdown

• [HIGH] smoke-claude.md / smoke-codex.md: Remove shared/github-queries-mcp-script.md from skills: list. This file is 16.5 KB and provides GitHub API query scripts for issues, discussions, labels — none of which are exercised by a smoke test that only configures repos and pull_requests toolsets. Estimated saving: ~16 KB from Smoke Claude prompt.
• [HIGH] aw-failure-investigator.md: Replace the inline MCP tool parameter documentation block (audit/logs/status sections, ~3 KB) with a single bootstrapping call: agenticworkflows --help or agenticworkflows logs --help. The agent uses bash: ["*"] and can fetch this at runtime. Needs manual review to confirm all referenced parameters are discoverable via --help.
• [MEDIUM] smoke-claude.md / smoke-codex.md: Deduplicate repeated workspace/memory path lines and "Verify the tool call succeeds" boilerplate. dup_line_ratio of 0.09–0.11 means ~10% of lines are redundant. Consolidate into a single definitions block.
• [MEDIUM] smoke-claude.md / smoke-codex.md / aw-failure-investigator.md: Enable tools.github.mode: gh-proxy and tools.cli-proxy: true. Currently only ai-moderator.md has cli-proxy: true. Enabling proxies allows replacing verbose MCP-tool-oriented step wording with compact gh CLI commands, reducing both prompt size and turn count.

Skills

• [HIGH] Smoke tests: shared/github-queries-mcp-script.md should be removed from or not included by smoke test workflows. Consider gating skill inclusion on active toolset configuration (i.e., only include when issues/discussions toolsets are configured).
• [MEDIUM] aw-failure-investigator.md: Extract the reporting structure guidelines (~2 KB of heading-level rules, progressive disclosure patterns, Airbnb-inspired design principles) into a shared shared/reporting.md skill so other workflows can reference it without inlining it. Check if shared/reporting.md already exists and whether it covers this content.

Agents

• [HIGH] aw-failure-investigator.md: Move failure run ID pre-fetch to a deterministic GitHub Actions run: step before agent activation. The audit confirms ~50% of 20 turns are data-gathering (effective tokens: 9.3 M). Injecting structured run data as a cache: artifact (the workflow already has a cache block) would cut agentic turns by ~10 and save millions of effective tokens per run. Needs manual review to define the pre-fetch query scope.

---

References

• §26922082621 — Smoke Claude (32,686 chars, 1.39 M tokens)
• §26924689485 — Failure Investigator (17,589 chars, 9.35 M eff. tokens, 20 turns)
• §26922082608 — Smoke Codex (20,451 chars, 1.59 M tokens)

Generated by 🌫️ Daily Ambient Context Optimizer · sonnet46 3M · ◷

• expires on Jun 11, 2026, 2:20 AM UTC