CLAUDE.md@tools/data-seed git:20260603.4cd87f7A

Immutable. This exact content never changes and is served at /api/v1/blob/a999ce7b6f55d306.

# CLAUDE.md — tools/data-seed/

Standalone uv project. Mirrors `openlatch-platform/tools/data-seed/`. The CLI talks to the SaferSkills API only — never imports from `services/api/`.

## Hard rules

1. **Purge is loopback-only + runtime-discovered.** `purge run --apply` exits 2 unless the **DB host** resolves to loopback (`HOST_ALLOWLIST` = `localhost`/`127.0.0.1`/`::1`). It `TRUNCATE`s every public table *discovered at runtime* from `pg_tables` **except `alembic_version`** (which pins the schema version) — there is **no hardcoded table list to drift against the schema** (new tables like `scan_runs`/`scan_events`/`upload_files`/`artifact_blobs`/`item_sources` are covered automatically, no FK orphans). Widening the host allowlist requires editing `saferskills_data_seed/domains/purge/app.py` + a review.
2. **Paced publish, loopback-exempt.** `catalog publish` paces at ≤2 scans/sec to avoid hammering the API (the 50-item seed takes ~25-30s). It does **not** rely on staying under the public 10/day/IP cap — that cap is a daily *count*, not a rate, so pacing can't keep a 50-item corpus under it. Instead the API exempts loopback callers from the per-IP scan-submit limit (`services/api/app/routers/scans.py::_is_loopback`), so seeding from the operator's own machine is unlimited while public submissions stay capped. Seed against a remote host and the cap applies again.
3. **No imports from `services/api/`.** The seed tool talks to the API through HTTP only — **except `purge`**, which connects to Postgres directly via psycopg (there is no admin bulk-delete endpoint by design — deletion is vendor-appeals / operator-runbook only, see `security.md`). Even so, `purge` imports **no** `services/api` Python — it only takes a libpq DSN (default mirrors `config.py::database_url`, sans the `+asyncpg` suffix). Sharing Python modules makes the seed brittle to backend refactors.
4. **Fixture corpus is canonical.** `saferskills_data_seed/domains/catalog/files/catalog.yaml` is the demo dataset. ~50 entries spanning all 5 PRD categories with the score distribution from D-FE-14. A1 ships with 8 representative entries; the full 50 lands in a follow-up.

## Architecture

```
saferskills_data_seed/
├── __main__.py          # python -m saferskills_data_seed
├── cli.py               # Typer root + global flags
├── shared/
│   ├── context.py       # GlobalContext dataclass
│   └── http_client.py   # httpx.Client factory
└── domains/
    ├── catalog/         # list / describe / publish
    ├── scans/           # list / run
    ├── vendors/         # verify-issue / verify-redeem / seed
    ├── doctor/          # preflight
    └── purge/           # describe / run — DB-direct TRUNCATE (host allowlist + --apply + --yes)
```

Every domain is a `typer.Typer()` subapp registered in `cli.py`. Shared context is attached in the root callback (`attach_context`) and read by subcommands (`get_context`).

## When endpoints don't exist yet

A1 ships before Phase B (scan backend) and Phase C (vendor right-of-reply). Endpoints that don't exist yet return 404. Each subcommand handles 404 explicitly — prints a yellow info message and exits 0 (not an error; the surface just isn't shipped yet). The structure exists so each phase can wire the corresponding endpoint without touching the CLI.

## CI

No dedicated CI lane — this CLI is dev tooling, not production code, so it
runs locally on demand and isn't gated. Linting + smoke happen as part of the
normal local workflow:

```bash
cd tools/data-seed
uv sync
uv run ruff check .
uv run ruff format --check .
uv run saferskills-data-seed --help
uv run saferskills-data-seed catalog list
```

If a regression slips through, the next time the founder reaches for the CLI
they'll see it. That tradeoff is intentional — paying for a CI lane on a
tool that runs maybe once a week wasn't worth the overhead.