refresh-provider-docs

Refresh Provider Documentation

Verify what agentconfig.org publishes against what providers actually document, and cite a primary source for every claim.

The site's credibility rests on being right about file paths, lifecycle events, and precedence rules that change without notice. This skill exists so that no provider panel, comparison row, or tutorial is expanded on memory.

The split this skill enforces

Reading documentation is a judgment task. Deciding whether the site is wrong is not.

Step	Owner	Why
Retrieve registered sources	Script	Reproducible, cited, and offline afterwards
Read the retrieved documents	Agent	Prose is not parseable by regex
Write normalized claims	Agent	Extraction is judgment, so it must be explicit and citable
Validate, compare, and report	Script	A verdict must be deterministic and reviewable

Never skip the middle two steps by guessing, and never hand-write a finding. If a claim cannot be traced to a retrieved snapshot, it does not get recorded.

When to Use

Use this skill when:

• A provider's configuration location, hook events, permissions, or precedence rules need checking against current documentation.
• A new provider is being considered for the compatibility index.
• The scheduled refresh runs and needs an evidence report.
• Someone asks which primary source backs a published claim.

Do not use this skill to edit content. It produces evidence and findings; the edits belong to add-primitive, add-provider, and generate-llms.

Workflow

1. Retrieve the sources

bun .github/skills/refresh-provider-docs/scripts/provider-docs.ts fetch --allow-network
bun .github/skills/refresh-provider-docs/scripts/provider-docs.ts fetch --allow-network --provider claude

Retrieval requires --allow-network every time. Nothing in this skill runs during an ordinary site build, and the site never depends on network access to render.

Snapshots and a manifest.json are written to a fresh run-specific directory outside the repository, printed at the end of the run. Each run gets its own directory so a partial or provider-scoped retrieval can never leave older snapshots beside the new manifest, where they would be read back as current evidence. Each snapshot carries its source id, canonical URL, requested URL, final URL, content type, authority, and retrieval timestamp in a header comment. Responses are limited to 512 KB, non-document content types fail closed, and unexpected redirect targets are rejected. An unknown --provider value fails closed rather than reporting an empty run as a success. Scheduled runs add --tracked-only so candidate providers cannot silently enter the compatibility refresh. If any source fails, the command exits non-zero: fix or retire the registry entry rather than publishing a claim you could not retrieve.