Two fixes surfaced by the live feed.
modelwatch: some repos report a bogus tiny `safetensors.total` (e.g.
deepreinforce-ai/Ornith-1.0-35B came through as "1M params"), which both
misrendered and — worse — let an oversized model slip past the size cap. Add
`params_from_name` (parse the first <num>B|M token from the repo id, which by
convention is the total size) and `effective_params` (trust safetensors.total
only when it's >= 100M, else fall back to the name). Use it for the size filter
and the summary. New unit tests cover the name parser, the fallback, and that a
70B reporting a 1M total is still rejected as too large. (Forward-looking: the
idempotent ingest won't re-render items already in the feed.)
fetch: feed requests now send browser-shaped `Accept` / `Accept-Language`
headers. reqwest sends no Accept by default, which some anti-bot filters 403.
NB this is hygiene, not a Reddit fix — Reddit fingerprints the TLS/HTTP client
(rustls) and hard-403s it where curl only gets rate-limited (429); headers don't
change the fingerprint. The control path (github releases.atom, etc.) is
unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016fKZzDpvjiJ9eYbPGgJvUP
Add newsfeed-modelwatch: a one-shot, systemd-timer-driven producer that watches
the Hugging Face Hub for open-weight releases and pushes filtered candidates to
the ingest endpoint. Answers "how do I automate the HF firehose into my feed" —
the RSS half already works on the pull rail; this is the JSON/API half.
How it works: each run polls the HF models API (a watchlist of orgs +
trendingScore), applies an admission predicate (license allowlist, total-params
cap, safetensors present, not gated, pipeline_tag), and POSTs the survivors to
POST /v1/ingest/candidates under a bearer token, tagged source="huggingface".
Two properties of the existing ingest rail shape the design:
- Ingest is idempotent on (user, external_id). Using the repo id as external_id
makes the producer STATELESS — no dedup table; re-runs/overlapping timers just
re-submit and the server drops repeats.
- HF `tags` are copied onto the candidate, so the user's per-interest weights do
the ranking. The predicate is only an ADMISSION filter (what's worth surfacing
at all) and stays subordinate to explicit weights, per the house rule.
Layout: newsfeed-modelwatch is a client of the API, not an internal component —
it holds no core/data deps. Pure logic (HF types, predicate, mapping to
CandidateSubmission, param formatting) lives in the lib with unit tests; the bin
does the HTTP polling/posting. CandidateSubmission gains Serialize so the
in-workspace producer can build and post one.
Ops:
- asset/systemd/newsfeed-modelwatch.{service,timer}: oneshot + 3-hourly timer.
The ingest token is a secret, kept in /etc/newsfeed/modelwatch.env
(NEWSFEED_TOKEN, mapped to `token` by figment) so a redeploy of the config
never clobbers it.
- asset/config/modelwatch.toml.tmpl: watchlist + predicate, no secret.
- infra-setup.sh installs the units, creates the env placeholder (never
overwritten), enables the timer, and prints the token step.
- deploy.yml builds + ships the binary and the (secret-free) config each deploy.
Verified: unit tests (gated union, license precedence, admit accept/reject,
mapping); dry-run against live HF; full loop against a local API — minted token,
submitted a real release, confirmed it landed in the feed with the huggingface
source linked and HF tags carried through. fmt/clippy/test all green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016fKZzDpvjiJ9eYbPGgJvUP
