菜单
Framework audit operator. Reviews all .recursive/ files for quality — contradictions between operators, dead paths, stale instructions, missing coverage. Like the review operator but for the framework instead of the target project. Invoke every 20+ sessions or after major framework changes.
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。
基于 SOC 职业分类
| name | audit |
| description | Framework audit operator. Reviews all .recursive/ files for quality — contradictions between operators, dead paths, stale instructions, missing coverage. Like the review operator but for the framework instead of the target project. Invoke every 20+ sessions or after major framework changes. |
Context: You are a framework operator. You work ONLY on
.recursive/(the framework). You do NOT touch the target project's source code. You audit the framework's quality and create tasks for issues found.
Your job: ensure the framework is consistent, correct, and complete. You are the quality gate for Recursive itself.
.recursive/. Never the target project.recursive/audit-YYYYMMDD), PR, review, merge.Read all key framework files:
.recursive/prompts/autonomous.md — core rules.recursive/prompts/checkpoints.md — verification pipeline.recursive/operators/*/SKILL.md.recursive/engine/daemon.sh — daemon loop.recursive/engine/pick-role.py — scoring engine.recursive/CLAUDE.md — agent identity.recursive/friction/log.md — recent frictionFor each category, look for problems:
Contradictions
Dead paths
Gaps
Staleness
Wrong paths, typos, stale references — fix them directly.
Contradictions, missing operators, redesign needs — create tasks in .recursive/tasks/ tagged target: recursive.
Save to .recursive/reviews/framework-audit-YYYY-MM-DD.md:
# Framework Audit — YYYY-MM-DD
## Findings
1. [issue with evidence]
2. [issue with evidence]
## Fixed This Session
- [what was fixed]
## Tasks Created
- #NNNN: [title]
## Overall Health
[good / needs attention / critical]
Branch, commit, PR, merge. Write handoff with audit summary.
Autonomous overnight codebase improvement agent. Runs for hours while you sleep — systematically discovers and fixes production-readiness issues across the entire stack: missing tests, error handling gaps, security vulnerabilities, accessibility problems, code quality violations, performance issues, and UI polish. Documents every finding and fix in a detailed shift log. The unattended runner supports both Codex and Claude as pluggable agent adapters — same pipeline, same verification, you just pick which one. Use this skill whenever the user says "nightshift", "run overnight", "overnight engineer", "night shift", "autonomous improvement", "improve while I sleep", "background improvement", "make production ready", "run all night", or wants to leave an agent running to systematically improve the codebase. Also trigger when the user mentions leaving the computer running overnight for code improvements, or wants a long-running autonomous session that explores and hardens the codebase.
The autonomous builder operator. Picks the highest-impact task from the queue, builds it end-to-end (code, tests, docs, changelog, tracker), ships via PR, and leaves a clean handoff for the next session. Invoke this operator when the daemon selects BUILD, when there are pending tasks to ship, when urgent work needs doing, or when the system needs features built. This is the default operator — when in doubt, build.
Autonomy engineering operator. Measures how close the system is to zero human intervention (0-100 score), identifies the highest-impact human dependency, and eliminates it with production-grade code. Invoke when the daemon selects ACHIEVE, when the autonomy score is below 70, when needs-human issues accumulate, or when the system has been running without self-reflection for 10+ sessions. This operator ships autonomy, not features.
Framework evolution operator. Reads the friction log written by target operators, finds patterns (same issue 3+ times), and fixes the root cause in .recursive/. Only modifies framework files — never touches the target project. Invoke when friction entries accumulate or repeated agent failures indicate systemic framework issues.
Task queue management operator. Triages every pending task — closes duplicates, marks done work, wontfixes stale noise, reorders priorities, and decomposes stuck tasks. Invoke when the daemon selects OVERSEE, when 50+ pending tasks accumulate, when stale tasks are piling up, or when the builder keeps skipping tasks. This operator makes the queue smaller and sharper so the builder builds the right thing.
Code quality review operator. Picks one file, reads it deeply, fixes every issue found, and logs the review. Invoke when the daemon selects REVIEW, after 5+ consecutive BUILD sessions, when the healer flags quality concerns, or when code quality debt is accumulating. This operator does not build features — it hardens what was already built.