Files
newsfeed/CLAUDE.md
rob thijssen 3db1f586a8
Some checks failed
deploy / build-web (push) Successful in 1m30s
deploy / deploy-web (push) Successful in 4s
deploy / build-api (push) Successful in 6m35s
deploy / deploy-api (push) Failing after 10s
feat(modelwatch): push producer for open-weight model releases
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
2026-07-08 14:29:32 +03:00

81 lines
4.6 KiB
Markdown

# CLAUDE.md — newsfeed
Agent-facing notes for working in this repo. Read the workspace-wide conventions in
`~/git/architecture/*.md` first; this file records only what's specific to newsfeed,
especially the deliberate deviations.
## What this is
A self-hosted, user-controlled news feed. The core idea: the user owns the ranking. Never
introduce an implicit/opaque signal that overrides the user's explicit weights. Signals
(view/click/save/dismiss) may *inform* future ranking but must remain subordinate to the
per-source and per-interest weights the user sets.
## Layout & boundaries (strict)
- `newsfeed-entities` — types/DTOs only, no I/O. Wire DTOs derive `ts_rs::TS` with
`#[ts(export)]`; `cargo test -p newsfeed-entities` regenerates `web/src/api/bindings/`
(path set by `.cargo/config.toml`'s `TS_RS_EXPORT_DIR`). Don't hand-edit bindings.
- `newsfeed-core` — business logic + data-access **ports** (traits). Pure where possible
(ranking, auth hashing). No DB or network calls.
- `newsfeed-data` — SQLite **adapters** implementing the core ports.
- `newsfeed-fetch` — the outbound-feed adapter (RSS fetch, feed discovery, OPML parse);
implements the `FeedProbe` port. The one crate that talks HTTP to the wider web for
feeds — used by the api (discovery on add) and the worker (polling). Kept separate from
`data` (SQLite-only) on purpose.
- `newsfeed-api` / `newsfeed-worker` — thin binaries; wire config/logging/signals and the
concrete store. No business logic that could live in a library crate.
- `newsfeed-modelwatch` — a **push producer** (lib + one-shot bin): watches the Hugging
Face Hub for open-weight releases, applies an admission predicate, and POSTs candidates
to `/v1/ingest/candidates` under a token. A *client* of the API, not an internal
component — it holds no DB/core deps; its pure logic (predicate, mapping) is in the lib.
New types → entities. New logic → core. New feed I/O → `fetch`. New DB I/O → `data`. Add a
port to `core::ports`, implement it in the matching adapter crate.
## Deliberate deviations from `generic.md`
1. **SQLite instead of Postgres (§5).** Explicit project choice. Because SQLite is
single-file/single-host, **api and worker co-locate** and share `/var/lib/newsfeed/newsfeed.db`.
There is no central DB cluster, no mTLS DB auth, no `pg_ident` mapping. The worker does
in-process scheduling — the `FOR UPDATE SKIP LOCKED` guidance does not apply.
2. **Runtime sqlx queries, not `query!` macros (§5).** SQLite's dynamic typing makes
compile-time checking low-value and forces a live DB or fiddly offline cache into CI.
We use `sqlx::query`/`query_as` with `FromRow` row structs (see `data/src/rows.rs`) and
map to entities explicitly. No `.sqlx/` cache; CI builds need no database.
Both are documented at their site (`data/src/lib.rs` header, `readme.md`) — keep them in
sync if you change the approach.
## Datastore conventions
TEXT UUIDs (hyphenated), RFC3339 timestamps, INTEGER booleans, JSON-in-TEXT for small
collections (`tags`, `media`). Migrations are immutable once committed (§5): add a new
`crates/newsfeed-data/migrations/NNNN_*.sql`, never edit a landed one. `RETURNING` and
`ON CONFLICT` are used, so the target needs SQLite ≥ 3.35.
## Auth model
- Humans: Argon2id password hash + opaque session cookie (`nf_session`), hash stored.
- Producers: per-user bearer API tokens (`nf_<prefix>_<random>`), SHA-256 hashed at rest,
shown once. Ingest is attributed to the token's owner.
Crypto/token logic lives in `core::auth`; the flow orchestration in `core::service`.
## Deploy topology
api + worker (+ the `newsfeed-modelwatch` systemd timer, 3-hourly, whose ingest token lives
in `/etc/newsfeed/modelwatch.env` — a secret kept out of the deployable config so redeploys
never clobber it) → `slartibartfast.kosherinata.internal`; SPA → `oolon.kosherinata.internal`
(nginx serves + proxies `/v1` to the API over the mesh; TLS terminates at oolon). API
port 22672 (registered in architecture `port-allocations.md`; deliberately not the crowded
`8081`), plain HTTP behind firewalld. See `.gitea/workflows/deploy.yml` (infra truth)
and `script/infra-setup.sh` (one-time host prep). `Type=exec`, not `notify`, because axum
doesn't `sd_notify`.
## Before you commit
- `cargo fmt --all` · `cargo clippy --workspace --all-targets -- -D warnings` · `cargo test --workspace`
- `pnpm --dir web build` (tsc typecheck + vite) · `pnpm --dir web lint`
- If you touched an entity DTO, regenerate bindings and include them in the commit.
- Conventional Commits; scope = crate/area (`feat(api):`, `fix(core):`, …).