docs-drift

name	docs-drift
description	USER-INVOKED ONLY. Use only when the user explicitly invokes $docs-drift or asks to audit Codex guidance drift. Audits AGENTS.md, AGENTS.override.md, .agents/skills, .codex/config.toml, project skill guidance, and other Codex instructions against recent code changes. Inspect real files first; use official OpenAI Codex docs or live web only for current Codex behavior; report findings before edits.

Docs Drift

Audit whether Codex-facing instructions still match the repo. Report first; edit only after explicit approval.

Scope

• AGENTS.md and AGENTS.override.md files along the repo path
• repo skills under .agents/skills/ or */SKILL.md
• project-local .codex/config.toml, hooks.json, MCP/app/plugin config
• repo docs that are explicitly imported or referenced by those guidance files

Skip README/API docs unless Codex instructions depend on them.

Workflow

1. Run:

   python <skill-dir>/scripts/codex_docs_drift.py
   

   Use --since <rev> for an explicit baseline; --json for machines.
2. Read the report first: baseline, drift window, changed zones, candidate guidance files.
3. Inspect real files. Compare guidance to changed source, tests, scripts, config, generated-source boundaries, and install/runtime ownership.
4. Check official Codex docs only when current Codex behavior matters. Prefer OpenAI docs MCP; otherwise live web search restricted to official OpenAI Codex docs. See references/codex-guidance.md.
5. Report findings before patches:

   ## Findings
   - HIGH|MEDIUM|LOW: <guidance file> drifts from <source evidence>
     Evidence: <file:line or commit/source inspection>
     Fix: <precise change>
   
   ## No-Drift Areas
   - <guidance file>: current because <evidence>
   
   ## Proposed Patch
   <only if edits were requested or the user asks for the patch>
   

6. On approval, patch only relevant source guidance. Prefer stable constraints, ownership, validation, and source-of-truth rules; avoid volatile procedures.

Rules

• Spend context on source evidence, not copied docs.
• Prefer one deep source-backed pass to shallow findings.
• Mark uncertainty explicitly when source evidence is partial.
• Treat stale guidance as a contract bug only when it can mislead future agents.
• Do not add old-Codex compatibility notes unless the repo still supports them.

Drift Heuristics

Flag drift when commits:

• add, rename, or remove modules mentioned by guidance;
• move source-of-truth ownership across runtime config, generated files, repo source, live install, or external systems;
• change validation, package managers, hooks, sandbox assumptions, or install paths;
• introduce a mature local pattern that guidance still tells agents to recreate;
• touch skill metadata, SKILL.md, scripts, or agent config without updating corresponding trigger and validation guidance.

Ignore pure prose polish, generated files whose source matches guidance, and implementation details the code already makes obvious.

Resources

• scripts/codex_docs_drift.py: drift window and guidance candidate report.
• references/codex-guidance.md: official Codex documentation checkpoints and refresh triggers.