GitHub via the official OpenAPI 3 spec — gives the LLM full request and response types so it doesn't have to guess pagination shape, field names, or wrapper-vs-array.

Untyped GitHub tooling forces the LLM to shape-probe on every call. Common failures observed in production traces:

• .forEach on the listIssues result (it's {issues, totalCount, pageInfo}, not a bare array)
• i.user.login vs i.author.login field-name guessing
• q vs query argument-name confusion

This pack reads the canonical GitHub OpenAPI 3 spec from github/rest-api-description and surfaces it via the assistant's existing OpenAPI learner. The LLM gets full typed interfaces.ts including pagination wrappers and response field names.

Methods land under gateway.gh. E.g.:

• gateway.gh.issues_list_for_repo({owner, repo, ...})
• gateway.gh.users_get_by_username({username})
• gateway.gh.search_issues({q})
• gateway.gh.issues_create({owner, repo, title, body})

See prompts/examples.md for usage patterns.

Set GITHUB_PERSONAL_ACCESS_TOKEN in the credential store or environment. Fine-grained PATs and classic PATs both work (Bearer auth).

The full spec is 600+ operations. This pack filters to five high-value tags: issues, pulls, repos, users, search. Edit apis/apis.yml to add more (e.g. actions, git, code-scanning).