菜单
Use when auditing cartographer recon-brief injections in the audit log. Reports (repo x day) inject uniqueness, missing-atlas repos, stale-at-inject rate, and excerpt-size distribution. Triggers on "audit recon", "recon audit", "recon injections".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use when designing a goal-oriented agent loop or reviewing one for runaway risk - build gate, machine-decidable goal, anti-Goodhart boundaries, five failure modes. Triggers on "design a loop", "review this loop", "will this loop run away".
Use when running OpenAI Codex via the codex CLI (copilot CLI fallback) for code analysis, refactoring, or editing. Triggers on "run codex", "codex analyze", "ask codex", "copilot with codex".
Use when running Google Gemini via the native Gemini CLI for code analysis, refactoring, or editing. Triggers on "run gemini", "gemini analyze", "ask gemini".
Use when running a local Qwen model (via llama.cpp and the pi agent) for code analysis, refactoring, or editing - local inference, no API cost. Triggers on "run qwen", "qwen analyze", "ask qwen", "local model task".
Use when running Anthropic Claude Sonnet via the native claude CLI (headless) for code analysis, refactoring, or editing. Triggers on "run sonnet", "sonnet analyze", "ask sonnet".
Use when converting a plan, design doc, brainstorming output, or discovery doc into a structured PRD saved to the backlog. Triggers on "create PRD", "create PRD from plan", "make PRD", "convert to PRD", "save plan as PRD".
| name | audit-recon |
| description | Use when auditing cartographer recon-brief injections in the audit log. Reports (repo x day) inject uniqueness, missing-atlas repos, stale-at-inject rate, and excerpt-size distribution. Triggers on "audit recon", "recon audit", "recon injections". |
Report on the recon-brief hook's injection uniqueness, atlas coverage,
staleness, and excerpt size. The aggregation is deterministic, so a script
computes it (PRD 00046) — never parse audit.jsonl in-model.
python3 ~/.claude/skills/audit-recon/scripts/report.py
It reads ~/.claude/cartographer/audit.jsonl (phase == "recon" events) and
prints four sections: Inject uniqueness (repo x day), Missing-atlas repos, Stale-at-inject rate, Excerpt-size distribution, plus header
totals and a malformed: line count. An empty or missing log prints zero
counts — that itself is the LOW: no recon events recorded yet finding.
THROTTLE BROKEN (>=3).
A count of exactly 2 is the documented concurrent-session race; tolerate it.Missing-atlas repos (no atlas; run /survey
there), and a non-zero stale_pct (listed repos injected a stale atlas;
run /survey --refresh).Synthesize a ## Findings section (CRITICAL/HIGH/MEDIUM/LOW) followed by a
## Summary block quoting the script's counts verbatim: recon totals
(inject / atlas-missing), double-inject groups, repos needing /survey,
stale-at-inject rate, and excerpt min/median/max.
~/.claude/hooks/cartographer-recon-brief.py (UserPromptSubmit;
injects once per repo per UTC day). Suppressed prompts are intentionally not
logged, so absence of an event after a day's first inject is expected.ts (ISO-8601 UTC), session, phase: "recon", decision
(inject | atlas-missing), repo_hash, atlas_excerpt_bytes (0 on
atlas-missing), stale (bool).