chore(activity-log): post-merge cleanup + graduate context to CLAUDE.md

- remove plans/activity-log/ (feature merged; learnings in CLAUDE.md + git history)
- server/CLAUDE.md: Activity/Audit Log architecture + extension points (recorder, fire_entity_event hook, sanitize_display, events allowlist, retention, API auth posture)
This commit is contained in:
2026-06-10 18:42:15 +03:00
parent b43f821046
commit e584235676
9 changed files with 12 additions and 1157 deletions
+12
View File
@@ -28,6 +28,18 @@ API key authentication via Bearer token in the `Authorization` header (`Authoriz
- When `api_keys` is **set**: a valid Bearer token is required from every client (loopback included).
- `require_authenticated()` rejects even loopback-anonymous callers on sensitive endpoints (e.g. backup download, secret reveal).
## Activity / Audit Log
Persistent, queryable audit log of meaningful actions (auth, device, entity CRUD, capture, system), surfaced in the WebUI (Activity tab + Dashboard widget + Settings retention panel).
- **Storage is NOT a `BaseSqliteStore`.** `storage/activity_log_repository.py` is a purpose-built repository over a dedicated indexed `activity_log` table (migration `002_add_activity_log`) — query-on-demand with **keyset pagination** (`seq` cursor), never load-all-into-memory. Don't route it through the entity-store pattern.
- **Recording.** `core/activity_log/recorder.py` (`ActivityRecorder`) is best-effort (never raises into the audited action) and **thread-safe** (inline on the event loop; `loop.call_soon_threadsafe` from non-loop threads, e.g. zeroconf discovery). It persists the entry **and** fires an `activity_logged` realtime event. Actor comes from the `current_actor` `ContextVar` (set in `verify_api_key`), default `"system"`.
- **Entity CRUD is auto-audited** via the `fire_entity_event()` choke point in `api/dependencies.py` — every create/update/delete already calls it. **Delete handlers must pass `entity_name`** (the entity is gone by record time). Non-entity events use explicit `recorder.record(...)` (get it via `get_activity_recorder()` DI or `get_module_recorder()` for engine/thread sites).
- **Never log secrets.** API-key tokens are never recorded. Wrap any untrusted/attacker-controllable string (mDNS names, headers, user-authored names) with `sanitize_display()` (`core/activity_log/sanitize.py`) before it enters a `message`/`metadata` field. Per-IP throttle bounds auth-failure audit writes.
- **Adding a new audited event:** pick a dotted `action` (e.g. `"thing.created"`), call the recorder; for it to render localized in the UI, add `activity_log.msg.<action>` to all three `static/locales/*.json` (the frontend `localizeMessage()` maps action→template; falls back to the server `message`). Entity-type labels live under `activity_log.entity_type.<type>`.
- **Adding a new realtime event type** (`pm.fire_event({"type": ...})`): add it to `_ALLOWED_SERVER_EVENT_TYPES` in `static/js/core/events-ws.ts` AND keep `tests/test_events_ws_parity.py` green.
- **Retention + API.** `core/activity_log/retention.py` prunes by `max_days` + `max_entries` (settings persisted via `db.set_setting("activity_log")`); the recorder's `enabled` flag is rehydrated from those settings on startup. REST in `api/routes/activity_log.py`: `GET /activity-log` (list, `AuthRequired`), `GET /export` (CSV/JSON stream — `require_authenticated`; chunked keyset so it never holds the DB lock across the stream; CSV formula-injection guarded), `GET|PUT /settings` (PUT is `require_authenticated`), `DELETE` (clear — `require_authenticated`, self-audited). The table is covered by the existing whole-DB backup (no `STORE_MAP` change needed).
## Common Tasks
### Adding a new API endpoint