manage-sources

name	manage-sources
description	Use when user asks to "add a source", "add docs", "ingest a site", "add this repo to the corpus", "remove a source", "disable a source", or gives a documentation URL to include in the Solana MCP index. Edits ingestion/sources.yaml, validates with pnpm gen:sources, and opens a PR.
user-invocable	true

Manage ingestion sources

Add, remove, or disable entries in ingestion/sources.yaml, the single source of truth for what the Databricks ingestion job crawls and what list_sections exposes. lib/sources.generated.ts is gitignored and rebuilt at build time, so a PR only ever touches the yaml.

Operating procedure

1. Dedup check (always first)

Search ingestion/sources.yaml for the requested URL's domain and any plausible id. A source counts as already present when its primary_url shares the domain/path or an existing entry clearly covers the same docs. If found: report the existing entry (id, enabled state, sections) and stop. If it exists but enabled: false, ask whether to re-enable instead of adding.

For removals: find the entry by id or url. If nothing matches, list the closest candidates and stop.

2. Author the entry (add) or delete it (remove)

Read the header comment of ingestion/sources.yaml before writing — it defines the schema, the closed section taxonomy, and the tagging guidance. Key decisions:

• kind: github for a GitHub repo (docs live in .md/.mdx), web for a docs site, openapi for a JSON spec.
• id: kebab-case, unique. GitHub sources use the gh- prefix (gh-anchor), web sources typically <project>-docs.
• sections: 1–3 tags from the closed taxonomy in the file header. Do not invent tags; the server boot check rejects unknown ones.
• use_cases: comma-separated routing keywords, strongest hooks first. Written for an agent deciding whether this source answers a query.
• Fetch the URL to confirm it resolves and to inform use_cases/sections. Prefer sitemaps when the site publishes one, otherwise start_urls. Add url_include filters for multi-chain doc sites so only Solana pages are kept.

Insert the entry into the matching # --- ... --- group in the file, not at the end. Follow the exact field order and style of neighboring entries.

Entry shapes to copy:

  - id: example-docs
    name: Example Docs
    kind: web
    enabled: true
    sections: [defi]
    use_cases: "Example AMM, swap API, liquidity pools"
    primary_url: https://docs.example.xyz
    start_urls:
      - https://docs.example.xyz

  - id: gh-example
    name: GitHub example-org/example
    kind: github
    enabled: true
    sections: [clients]
    use_cases: "Example SDK source, TypeScript client"
    primary_url: https://github.com/example-org/example
    github: { owner: example-org, repo: example, include_readmes: true, include_source_code: true }

  - id: example-api
    name: Example API (OpenAPI)
    kind: openapi
    enabled: true
    sections: [data]
    use_cases: "Example REST API spec, historical data"
    primary_url: https://api.example.xyz/docs
    spec_url: https://api.example.xyz/openapi.json

To take a source out of rotation without losing its config, prefer enabled: false over deletion; delete only when the source is gone for good or was added in error.

3. Validate

pnpm install
pnpm gen:sources
pnpm typecheck

gen:sources enforces structure (unique ids, kinds, non-empty sections/use_cases); the typecheck's freezeSources boot check rejects unknown section tags. All three must pass before committing.

4. PR

• Branch from main: feat/sources-<id> (or chore/sources-remove-<id>).
• Conventional Commit, e.g. feat(sources): add Example docs / chore(sources): remove example-docs.
• .claude/* is commonly globally gitignored; when editing this skill itself use git add -f.
• Open the PR with a body noting: source id, kind, sections, why it belongs in the corpus, and the validation commands run.
• Do not merge; a human reviews.

After merge (mention in the PR body)

The Databricks copy of sources.yaml refreshes on the next just deploy (asset bundle). The daily 09:00 UTC crawl only picks the source up after that deploy. The Cloud Run server side needs nothing extra: push to main redeploys and regenerates the catalogue.