antigravity-docs-index-builder

name	antigravity-docs-index-builder
description	Rebuild the docs-manifest.json that the antigravity-docs skill relies on, by extracting the DOCS_STRUCTURE map from the live Antigravity web app JS bundle and enriching it with titles/descriptions from llms.txt. Use when Antigravity has shipped new/renamed/removed doc pages, when antigravity-docs reports a stale or missing page, or when you simply want to refresh the manifest. Manually invoked (e.g. /antigravity-docs-index-builder).
argument-hint	["--force"]
allowed-tools	Bash, Read, Write

Antigravity Docs Index Builder

Regenerates .claude/skills/coding-agent-docs/skills/antigravity-docs/references/docs-manifest.json — the static index that the antigravity-docs skill (published as part of the coding-agent-docs plugin) reads to answer questions. This is a maintenance skill you run by hand whenever the Antigravity docs may have changed; antigravity-docs itself never runs it.

Why this skill has to exist

Unlike Claude Code and Codex — whose doc pages have raw .md twins you can fetch directly — antigravity.google/docs/* is a client-rendered SPA. A plain fetch of a doc page returns an empty shell, and there is no .md twin at the page URL and no llms-full.txt. So the doc list can't be scraped by fetching pages.

But the app ships the map and the content as static resources, if you know where to look:

1. The JS bundle (main-<hash>.js) embeds a DOCS_STRUCTURE array of {section, path, slug, filename} records — the authoritative list of every doc page.
2. The real Markdown for each page lives at https://antigravity.google/assets/docs/<path>/<filename>.md (pristine, with frontmatter — this is what antigravity-docs fetches).
3. llms.txt lists - [Title](/docs/<slug>): description — human-readable titles and one-line descriptions, joinable to DOCS_STRUCTURE on slug (verified: a 1:1 join).

This builder stitches (1) + (3) into the manifest and records the content URLs from (2).

Critical: the bundle must be fetched as raw bytes (curl/urllib), never through a markdown-converting fetcher like WebFetch — that strips <script> tags and the DOCS_STRUCTURE data with them. That's why this skill uses Bash, not WebFetch.

Procedure

1. Run the builder script

python3 .claude/skills/antigravity-docs-index-builder/scripts/build_manifest.py

• Set today's date so the manifest is stamped: ANTIGRAVITY_BUILD_DATE=<YYYY-MM-DD> python3 .../build_manifest.py
• Pass --force (or $ARGUMENTS) to rebuild even when the bundle hash is unchanged.

What it does: discovers the current main-<hash>.js from https://antigravity.google/docs/home, and short-circuits if that bundle name matches the one already recorded in the manifest (docs structure unchanged → nothing to do). Otherwise it downloads the bundle, parses DOCS_STRUCTURE, fetches and joins llms.txt, and writes the manifest — printing the added/removed page slugs versus the previous version.

2. Report the result

Relay to the user: the bundle name, page count, and the added/removed diff. If pages were added or removed, that's the signal that antigravity-docs now covers new material (or dropped stale slugs).

3. If the script errors

The script fails loudly (rather than writing a bad manifest) in two cases, both meaning Antigravity changed something structural:

• "could not find main-.js" — the app shell markup changed. curl -s https://antigravity.google/docs/home | grep -oE 'src="[^"]*"' to find the new bundle-reference pattern, then update find_bundle_name.
• "DOCS_STRUCTURE not found" — the record shape changed. Download the bundle, grep -oE '.{40}filename:"[^"]*".{20}' bundle.js to see the new object shape, then update parse_structure.

Fix the script, re-run, and note what changed in this skill's CHANGELOG.md.

Rules

• Never hand-edit docs-manifest.json. It is generated. Any manual fix will be silently overwritten on the next run — fix the builder instead.
• Raw fetch only for the bundle. WebFetch cannot see the DOCS_STRUCTURE data. Use curl/Python.
• Verify the join. If the script reports pages with no llms.txt description, that's tolerable (it falls back to a filename-derived title) but worth a glance — a large unmatched count usually means the slug scheme drifted between the bundle and llms.txt.
• Don't widen scope. This builder only writes the manifest for antigravity-docs. It does not fetch or cache page content — that's antigravity-docs' job at answer time.