agent-docs

name	agent-docs
license	MIT
description	Create or converge agent documentation (AGENTS.md hierarchy) for a codebase — bootstrap mode generates docs from scratch, refresh mode detects and fixes drift
inputs	[{"name":"scope","required":false,"description":"Directory path relative to repo root (default: '.' for the entire repo)."},{"name":"mode","required":false,"description":"'bootstrap' (generate from scratch), 'refresh' (fix drift in existing docs), or omit to auto-detect: no agent docs in scope → bootstrap, otherwise refresh."},{"name":"depth","required":false,"description":"[bootstrap] Doc hierarchy depth: 'flat', 'hierarchical', or 'deep'. Default: auto-detect based on repo size."},{"name":"deep","required":false,"description":"[refresh] Also perform semantic verification of behavioral claims against source. Expensive (5-15 min, 200-500k tokens)."},{"name":"cross-repo","required":false,"description":"[bootstrap] Also generate integration-surface documentation describing how this codebase connects to other documented repos."}]

Agent Docs

One workflow with two modes, both converging docs and code onto the same target state: an accurate, AGENTS.md-standard doc hierarchy that an agent can trust cold.

• bootstrap — no docs exist yet. Generate them from scratch: reconnaissance → gotcha mining → hierarchy design → generation.
• refresh — docs exist but may have drifted. Detect → triage → surgical fixes → (optionally) deep semantic verification.

The modes are the same convergence loop entered at different states: a refresh that finds a severely drifted doc regenerates it with bootstrap's machinery; a refresh that finds an uncovered package bootstraps a doc for it; a bootstrap validates its output with refresh's drift detector. Both modes finish through the same shared tail below.

All docs produced in either mode MUST conform to the Doc Authoring Guidelines — read that file before generating or editing any doc.

---

Phase 0 — Resolve scope and mode

Use the user-provided scope, or default to . (the entire repo). Then run the drift detector once:

// turbo

bash scripts/check-agent-docs-freshness.sh <scope>

(The script lives in this skill's scripts/ directory; run it from the repo root. It writes a structured report to /tmp/agent-docs-freshness.json.)

Mode resolution:

User said	Detector found	Mode
nothing	docsScanned: 0	bootstrap
nothing	docs exist	refresh
bootstrap	docs exist	Confirm intent. Then: fill gaps only — generate docs for uncovered areas via bootstrap; existing docs are never clobbered, they're refresh candidates.
refresh	docsScanned: 0	Report "nothing to refresh" and suggest bootstrap. Stop.

Phase 1 — Execute the mode core

• bootstrap → follow modes/bootstrap.md (reconnaissance, gotcha & convention discovery, hierarchy design, generation, optional --cross-repo integration surface).
• refresh → follow modes/refresh.md (triage the detector report, fix critical/moderate issues, fill coverage gaps, rewrite severely drifted docs, optional --deep semantic verification). The report from Phase 0 is the input — don't re-run the detector to start.

Both mode files end by returning here.

Phase 2 — Shared tail

Every run, regardless of mode, finishes with these steps in order.

2.1 — Validate

Re-run the detector to confirm the docs now match the code:

// turbo

bash scripts/check-agent-docs-freshness.sh <scope>

• Critical issues = 0: proceed. (A fresh bootstrap MUST hit 0 on the first validation — anything else means something was generated from assumption rather than verification. Fix immediately.)
• Critical issues remain after a refresh: iterate (max 2 additional attempts), then report remaining issues to the user.

2.2 — Consistency sweep

Re-read every doc this run created or edited, plus its parent AGENTS.md, checking for rules that now contradict each other (build commands that differ, conventions stated one way at the root and another way in a nested doc, gotchas that a fix made obsolete). Contradicting instructions cause agents to pick one arbitrarily — resolve per the no-contradictions rule in the Doc Authoring Guidelines.

2.3 — CLAUDE.md bridge

Claude Code reads CLAUDE.md, not AGENTS.md — without a bridge, everything this workflow maintains is invisible to it. Unless the user opts out or the repo already has a CLAUDE.md, create one at the repo root containing:

@AGENTS.md

plus (only if needed) Claude Code-specific notes below the import: settings, permissions, hook expectations. Keep it to the import line if there is nothing Claude-specific to add — the point is one source of truth in AGENTS.md, shared across harnesses.

2.4 — Quality checklist

Before committing, verify:

• Every command in "Build & Test" actually works when copy-pasted
• Critical Gotchas are genuinely non-obvious (not just "follow the style guide")
• No section is generic filler — every line earns its place
• File paths are verified against the filesystem (use ls / find)
• Line counts use actual wc -l values with proper rounding
• Docs read as if written by a knowledgeable engineer, not a template
• An agent reading the docs cold could start productive work within 1 task

2.5 — Commit

Review git status and stage only the doc files this run touched (exclude agent-harness dirs like .windsurf/, .claude/, .cursor/ unless the bridge file was created):

git add <the doc files you changed>
git commit -m "docs: <bootstrap|refresh> agent docs — <summary>

Scope: <scope>
<bootstrap: docs created | refresh: Critical: X fixed | Moderate: Y fixed | Minor: Z fixed>
"

---

Optional CI Integration

The structural detector is fast, deterministic, and exit-code-gated — suitable as a PR check. Drift research (DOCER, EMSE 2023) found broken doc references in ~29% of the top-1000 GitHub repos, typically unnoticed for years, precisely because checking was manual; the same work ships a PR-triggered action because catching drift at change time beats periodic sweeps. Recommended split:

• In CI (structural only): run check-agent-docs-freshness.sh <scope> on PRs that touch source or docs; fail (or warn) on critical issues. No LLM involved — cheap and deterministic.
• Manual (this workflow): the LLM-driven generation, fixing, and --deep semantic verification stay human-triggered. Run refresh when you suspect drift (after refactors, after merging large PRs, periodically).

What This Workflow Does NOT Do

• Generate API documentation — this produces agent orientation docs, not user-facing API docs
• Convention drift detection — if the team adopts new patterns, that requires human input to update "Code Conventions" sections
• Cross-repo knowledge synthesis — data flows between services need a different mechanism (--cross-repo documents this repo's surface, not the other repos)
• Make architectural decisions — it documents what IS, not what SHOULD BE
• Replace human review — generated docs should be reviewed by a human who knows the codebase before being treated as authoritative
• Automatic doc rewriting in CI — only the structural detector belongs in CI (see above)