Gather agent docs

Gather Agent Docs

Build a small, durable, agent-readable doc set for the project's key topics. Navigation map first, captured detail on request — future agents read the local map to land on the right upstream URL in one hop, then fetch upstream when they need depth.

Output layout

Use kebab-case slugs for topic directories (gemini-cli, cloud-run, firebase-app-hosting):

.agents/docs/
  INDEX.md            # one-line entry per topic, pointing to its DOC.md
  <topic>/
    DOC.md            # entry point: summary, key concepts, documentation map, optional patterns
    <SUBTOPIC>.md     # optional deep dive (rare — see below)

Frontmatter

DOC.md and every <SUBTOPIC>.md open with YAML frontmatter the other skills key off:

---
last_verified: 2026-05-08 # required, ISO date
upstream_commit: <sha> # optional; commit at verification — enables smart refresh
sources: # required, at least one URL; richer set = more refresh short-circuits
  docs: https://... # main documentation page
  repo: https://github.com/... # source repository
  changelog: https://... # changelog / release-notes — enables smart refresh
  release: https://github.com/.../releases
---

INDEX.md carries last_verified only — set to the oldest date among topic files.

What each DOC.md contains

1. Summary — one paragraph: what the topic is and where it fits.
2. Key concepts — the primitives an agent needs, one terse line each. No config snippets, no flag tables.
3. Documentation map — structured list of upstream pages, grouped by upstream nav section. Each entry: [Page title](url) — one-line hook on what's there.
4. Patterns (optional, deepen-only) — paste-ready snippets for surfaces agents reach for repeatedly. Capped at ~5. Each block ends with an inline _Source: …_ link to the upstream anchor that backs it.
5. Subtopics (optional) — list sibling files with a one-line hook each.

Soft cap: ~100 lines for DOC.md, ~150 for a subtopic deep-dive.

Capture rules

• T1 — link map. Page title + URL + one-line hook. Default tier for every entry in the upstream nav.
• T2 — snippet. Small, paste-ready block tied to a single upstream anchor. Captured only when deepening, capped at ~5 per topic, each with an inline _Source: …_.
• T3 — volatile. Model IDs, pricing, preview/beta flags, version numbers, deprecation timers. Linked only, never captured.
• Verbatim only. Captured details must come from upstream just fetched in this session, or a command run in this session. Never paraphrase, infer, or auto-complete from training memory.
• Stay reusable. Avoid project-specific framing; pages should travel between repos.

Topic discovery

Pick topics from what the project actually declares as a dependency, not what it might plausibly use:

• Package manager manifests (package.json, pyproject.toml, go.mod, Cargo.toml, ...).
• CI / build configs (.github/workflows/, cloudbuild.yaml, ...).
• Infrastructure / runtime manifests (Dockerfile, terraform/*.tf, firebase.json, apphosting.yaml, ...).
• Tool-version pins (mise.toml, .tool-versions, .nvmrc, ...).
• Skip: the project's own code, generic standards (Markdown, JSON), anything not wired into the build/runtime. Confirm the shortlist with the user when non-obvious.

Building the documentation map

For each topic, fetch one upstream nav source — try in order until one resolves:

1. Docs sidebar / table-of-contents page (e.g. a /docs/ index, a _sidebar file).
2. The repo's README.md "Documentation" section, or the docs/ folder structure on the default branch.
3. The docs home page when it links to every section.
4. sitemap.xml as a last resort (raw, needs grouping).

Extract section headings, page titles, and URLs; preserve the upstream grouping so the map mirrors how upstream organises itself. If a heading has no children, list the section as a single entry.

Optional subtopic files

Split a subsystem into a sibling file (HOOKS.md, EXTENSIONS.md) only when all of:

• It has its own canonical upstream page distinct from the topic's overview.
• The deepened version (with captured snippets) would push past the soft cap.
• Agents reach for that depth repeatedly in this project, not just hypothetically.

The link map alone rarely justifies a split. Subtopic files share DOC.md's shape and drift on their own clock; surface them from DOC.md's "Subtopics" section. Fold back into DOC.md if a subtopic shrinks below its own weight.

INDEX.md shape

INDEX.md carries last_verified frontmatter (the oldest date among all topic files), a brief intro line, then one entry per topic: - [YYYY-MM-DD] [topic-slug](topic-slug/DOC.md): one-line hook.. The leading date is the oldest last_verified among that topic's files. Don't enumerate subtopic files; DOC.md's Subtopics section handles that. If INDEX.md already exists, extend it; only add entries for missing topics.

Workflow

Default (no args, or "gather docs"):

1. Identify topics using the discovery signals; confirm shortlist when non-obvious.
2. Build each link map. Fetch one nav source per topic; extract title + URL + one-line hook per page, preserving upstream grouping.
3. Write each DOC.md — frontmatter, summary, key concepts, documentation map. No Patterns section.
4. Write INDEX.md — frontmatter + dated topic entries.
5. Hand off — don't commit; let the user review.

Deepen (args like "deepen gemini-cli", "expand all"): for each topic in scope, capture up to ~5 paste-ready snippets (T2) — fetch the linked upstream page, copy a small block, add an inline _Source: …_. Add or extend a Patterns section. Bump last_verified. Hand off.

Pairing

• use-agent-docs — reads these files before web research; can backfill T2 snippets when fetching upstream for depth.
• refresh-agent-docs — re-verifies against upstream; smart path uses upstream_commit / sources.changelog for cheap no-op, falls through to a link-rot scan.