GitHub Radar

A searchable archive of repositories starred by @AndyTheFactory.

Capture: star a repository on GitHub, then close the tab. A twice-daily GitHub Action imports it into this repository.

Browse: start with the human-readable catalog, or inspect CATALOG.md and the raw catalog/ entries.

Search: use GitHub Code Search with repo:AndyTheFactory/github-radar plus terms such as agent orchestration, evaluation, or OCR. Indexing new files may take some time.

Deterministic browse pages

python scripts/build_browse.py generates browse/README.md, subject pages under browse/domains/, types, and an alphabetical directory. These pages are rebuilt after each sync and enrichment; they require no model, network calls, or extra dependencies. Repositories awaiting classification appear in Unclassified with their original descriptions. Do not edit generated browse pages manually.

To regenerate them without importing or invoking Copilot, run Build readable catalog manually from GitHub Actions.

How it works

• Each starred repository is written once to catalog/<owner>/<repo>.md.
• Entries contain metadata and a short, searchable excerpt of the upstream README where available.
• Existing entries are never overwritten: manual edits and notes are safe.
• Removing a star does not delete an archived entry.
• Repositories owned by AndyTheFactory are excluded; any previously imported entries under that owner are removed on the next sync.
• Each run imports up to 50 new repositories, newest first. Repeated runs gradually backfill older stars.
• Sync checks all star pages, so it does not rely on a fragile timestamp checkpoint.
• No LLM key, database, or third-party dependencies are necessary.

Running it

From Actions → Sync starred repositories, choose Run workflow to begin the initial import. Subsequent imports run at 00:00 and 14:00 fixed EET (UTC+2; GitHub may delay scheduled jobs).

Or run locally (Python 3.11+):

export GH_TOKEN="$(gh auth token)"
python scripts/sync.py --user AndyTheFactory --max-new 50
python -m unittest discover -s tests -v

Local sync writes files but does not commit them. To backfill faster, repeat with a larger --max-new.

Optional GitHub Copilot enrichment

The regular sync imports stars without requiring Copilot. After import, Actions checks for entries without a generated summary. If there are none, the Copilot SDK is not installed or invoked.

To enable enrichment:

1. Create a GitHub personal access token with Copilot Requests permission, associated with an account that can use Copilot.
2. Set repository Actions secret COPILOT_TOKEN under Settings → Secrets and variables → Actions.
3. Set the optional Actions repository variable COPILOT_MODEL to one of:
   • cheapest (default): query the Copilot SDK model API and pick the lowest known premium-request multiplier; where multipliers are unavailable, compare estimated AI-credit token prices. If pricing cannot be queried, fail without a costly fallback.
   • auto: let Copilot select the model according to your plan and policies; it is not necessarily cheapest.
   • An explicit model ID (for example claude-haiku-4.5): use that model only; fail if unavailable.

Each scheduled run (00:00 and 14:00 fixed EET, UTC+02:00) processes at most 10 unenriched entries. A manual Run workflow exposes an enrichment_limit input (default 100, allowed 1–100) for backfilling; it never processes more than 100 enrichments in one run. Importing remains functional without the Copilot secret. Individual failed entries stay pending, and successful entries are retained. Nothing triggers Copilot when all records are already enriched. Expect usage of your Copilot entitlement and possible AI-credit charges.

You can preview the pending queue without a token:

python scripts/enrich.py --dry-run

The enrichment output is stored in an explicitly machine-managed block at the bottom of each Markdown entry. Reprocessing an enriched entry is not automatic; manual notes are preserved. The taxonomy is specified in taxonomy.yaml and CLASSIFICATION.md.

Classification quality checks

The classifier follows the artifact-first rules in CLASSIFICATION.md: capability tags must describe what a repository itself delivers, not concepts discussed in a reading list or functionality inherited from dependencies.

Run python scripts/audit_catalog.py to print suspicious classification combinations without calling an LLM. This advisory check also runs during sync and does not block importing or publishing entries. Corrected catalog classifications retain their original descriptions, README excerpts, and manual notes. Generated category and browse indexes are rebuilt on the next workflow run.

Data and privacy

This is a public catalog of public starred repositories. It may reflect your interests and approximate discovery dates. Don't put internal client details or private notes into public entries.

The catalog is useful without AI-generated summaries. Copilot enrichment is optional, and requires the repository secret described above.