GitHub watch — repos, profiles, and fork trees

In big projects, important changes show up as PR/issue/fork activity well before the headline release. Watching repos, the people behind them, and their fork trees surfaces that build-up early.

How to fetch (feeds first, no auth, no rate-limit worries)

Prefer GitHub's public Atom feeds and plain pages — the same RSS pattern as the lab sweep, readable via tvly with no API key and no rate limit:

• Releases: https://github.com/<o>/<r>/releases.atom
• Tags: https://github.com/<o>/<r>/tags.atom
• Commits on default branch: https://github.com/<o>/<r>/commits.atom
• A user's public activity: https://github.com/<u>.atom
• PR list / issues / a fork's page: extract the HTML via tvly (no feed exists).

The REST API (api.github.com) is OPTIONAL — use it only if it is reachable and authenticated in this environment (it may not be: the network often routes through Tavily only, and unauthenticated API is 60 req/hr). When available it gives cleaner structured data (exact ahead_by, merged_at); when not, the feeds and page extraction above are sufficient. Never let the watch block on the API.

Check activity since the last scan (previous date in logs/source_rotation.md). Log the watch every run, even when quiet ("github-watch: no movement").

1. Watched repositories

For each repo in SOURCES.md → "Watched repositories":

• Releases / tags (releases.atom / tags.atom) — a new release is a citable

artifact (evidence-eligible against a trend).

• Merged PRs — extract the PRs page (/<o>/<r>/pulls?q=is:pr+is:merged);

a notable merge (feature, format, kernel, model support) is a citable artifact.

• Hot issues / discussions — extract /<o>/<r>/issues?q=is:open+sort:comments;

a suddenly high-activity thread is a QUEUE signal until it lands.

2. Watched profiles / users

For each user in SOURCES.md → "Watched profiles/users", read their activity feed (https://github.com/<u>.atom) to see where they're heading: brand-new repos, a repo getting heavy pushes, a new release on any of their repos. New repos/releases are citable; early push activity on an unreleased repo is a queue signal ("watch <user>/<repo>, pre-release").

3. Fork-tree analysis (depth 3) + Fork Notability Score

For each project in SOURCES.md → "Fork-tree analysis", walk the fork tree up to depth 3 (the repo's forks page / /network/members, then forks of the active ones, etc.). Prune aggressively: only descend into a fork that is itself active (updated in the last ~30 days) — dead forks have no live children worth chasing.

Compute the Fork Notability Score (FNS) per fork. Read the values from the fork's page ("This branch is N commits ahead", star count, Releases tab) — or from the REST API if it's available:

gate:   updated within last 30 days          (else skip — stale)
FNS  =  commits_ahead_of_upstream            (the "N commits ahead" figure)
      + 2 × (fork's own stars)
      + 20 if the fork has its own releases/tags
      + 15 if the fork has an open PR back to upstream

Surface a fork under the project's trend when FNS ≥ 50 AND commits_ahead ≥ 20.

• A fork that clears the bar AND ships a release/artifact → citable evidence

under that project's trend (cite the fork repo/release).

• A fork that clears the bar but is still pre-release (just diverging fast) →

observation_queue signal ("fork <o>/<r> diverging, N commits ahead").

• Name the fork in the trend's notes/evidence ("fork X carries feature Y not yet

upstream"). Tune the threshold over time and log changes in the report.

Rationale: in this ecosystem the real innovation often lives in an active, far-ahead fork (performance/quant variants) months before it merges upstream — FNS is meant to catch exactly those, not the thousands of bookmark forks.

Self-maintenance (agent owns this)

Add repos/users/projects that become central to a tracked trend; drop ones that go silent or off-scope (one-line reason in the report). Adjust the FNS weights and threshold as you learn what they surface — record any change. Keep all lists in SOURCES.md.