Third family of the v0.5.7 NATS-and-categories campaign. Fire is the heaviest of the campaign -- four distinct fixes plus a category audit. Two of the four were broken in production: FIRMS subscribed to a syntactically invalid pattern, and WFIGS tombstones were silently dropped.
FIX 1 -- FIRMS NATS pattern (the canonical bug). Pre-v0.5.7-fire `_subjects_for("firms","us.id")` returned `["central.fire.hotspot.>.us.id"]`, which is INVALID NATS (the `>` multi-level wildcard is only legal at the tail token). It also wouldn't have matched anything Central publishes: per the Central v0.10.0 consumer integration guide §firms, the actual published pattern is `central.fire.hotspot.<satellite>.<confidence>` (5 tokens, no us.<state> suffix). The two slots after "hotspot" are satellite name and confidence band -- NOT tile coordinates or region tokens.
Note on prompt vs. guide discrepancy: the v0.5.7-fire task spec described a tile-coord/state pattern `central.fire.hotspot.*.*.us.id` (7 tokens with us.<state> tail). That's neither what Central v0.10.0 publishes nor what its guide documents. We follow the guide. Subscribing to the prompt's 7-token pattern would silently match zero messages in production (token-count mismatch). State filtering for FIRMS happens client-side via data.latitude / data.longitude against the configured region bbox.
New subscription: `central.fire.hotspot.>` -- tail-only `>`, NATS-legal, matches all <satellite>.<confidence> combinations.
FIX 2 -- WFIGS tombstone subjects. Per guide §wfigs_incidents and §wfigs_perimeters, WFIGS publishes:
active: central.fire.incident.<state>.<county> (Convention A, depth-3 state)
active: central.fire.perimeter.<state>.<county>
tombstone: central.fire.incident.removed.<state> (5 tokens, "removed" at depth-3)
tombstone: central.fire.perimeter.removed.<state>
Pre-v0.5.7-fire `_subjects_for("fires","us.id")` subscribed only to the active subjects (`central.fire.incident.id.>` and `central.fire.perimeter.id.>`). The tombstone subjects have "removed" at depth-3 instead of the state token, so the active-subject `>` filters silently dropped EVERY tombstone. Fall-off signals never reached meshai's inhibitor, so old incidents stayed "live" in the pipeline indefinitely.
Added the two tombstone subjects to the subscription list. Both are 5-token literals with no wildcards -- trivially NATS-legal.
FIX 3 -- WFIGS tombstone dedup. Per guide §wfigs_incidents removal semantics, the tombstone env_id has the shape `<IrwinID>:removed:<iso_now>` -- the `:removed:` is sandwiched in the middle, with a timestamp tail. Pre-v0.5.7-fire the consumer.py group_key recovery was `re.sub(r":removed$", "", group_key)` -- a literal trailing `:removed` match -- which DID NOT FIRE on the WFIGS form (the regex required `:removed` at the very end of the string, but the WFIGS form has `:<iso>` after it).
Consequence: WFIGS tombstones' group_key was the full `<IrwinID>:removed:<iso>` string instead of the bare `<IrwinID>`. The pipeline grouper/inhibitor never matched tombstones to their original incidents, so the lapse signal was lost.
Fixed by switching the regex to `re.sub(r":removed(:.*)?$", "", group_key)` -- handles both the WFIGS `<IrwinID>:removed:<iso>` form AND the legacy GDACS `<id>:removed` form. The `is_tombstone` detection also gained an explicit `":removed:" in env_id` check for the WFIGS shape.
Per the guide: "the same incident can have one or more removal tombstones over its lifecycle" (it can re-enter and re-fall-off). To preserve per-tombstone distinctness for downstream lifecycle accounting, the full env_id is stashed on `Event.data["_central_tombstone_id"]` (the group_key collapses to the IrwinID by design, but the original env_id with the :<iso> tail survives on data).
FIX 4 -- ALERT_CATEGORIES fire-family audit + removed parametric entries. Per Matt's direct feedback ("fire near mesh has its own set of parameters that I don't even know what they could be. like how far is near mesh? I don't know I can't set that."), the parametric `fire_proximity` and the duplicate-named `wildfire_proximity` (both labeled "Fire Near Mesh" with parametric radius-based descriptions) were unselectable in the new Advanced Rules UI. Removed both.
Cross-referenced what FIRMS and WFIGS actually emit (per the guide and the native adapter code) and audited the registry:
Native emit:
firms.py -> new_ignition (when adapter flags new_ignition)
or wildfire_hotspot (otherwise) [v0.5.7-fire: was wildfire_proximity]
fires.py -> wildfire_incident
Central path emit (via map_category):
fire.hotspot.* -> wildfire_hotspot
fire.incident.* -> wildfire_incident
fire.perimeter.* -> wildfire_incident (perimeters merge to the incident)
fire.<other> -> wildfire_incident (catchall)
Registry after v0.5.7-fire:
{new_ignition, wildfire_hotspot, wildfire_incident}
Parity confirmed. No orphans, no missing.
Aligning firms.py to emit `wildfire_hotspot` (matching the central FIRMS map) means native + central FIRMS produce identical categories regardless of which feed path is enabled.
Composer (`_CATEGORY_EMOJI`, `_CATEGORY_LABEL`) and router (three source-attribution tables) updated to drop the removed categories and add the new ones.
Deferred to v0.5.8: distance_max_km field on rules for actual proximity filtering. Replaces the parametric fire_proximity registry entry with a parameterized rule field that the user CAN configure ("alert me about wildfire_incident within 30 km" instead of an opaque "Fire Near Mesh" toggle).
Tests
-----
PYTHONPATH=. pytest -q: 380 passed (was 366; +14 net).
- tests/test_fire_v057.py (new): FIRMS subject is tail-only `>` with no mid-subject placement; WFIGS subjects cover active + four tombstones; WFIGS tombstone strips `:removed(:.*)?$` for group_key; two same-IrwinID tombstones both propagate through _handle and share group_key, with the original env_id preserved on data["_central_tombstone_id"]; legacy GDACS `:removed` shape still strips cleanly; fire_proximity / wildfire_proximity absent from ALERT_CATEGORIES; no "Fire Near Mesh" name duplicates; fire-family parity (native + central emit == registry); required-fields check on the three fire entries.
- tests/test_central_region_routing.py: updated FIRMS test (tail-only `>`) and WFIGS test (includes tombstone subjects).
- tests/test_pipeline_toggle_filter.py, tests/test_adapter_firms.py, tests/test_v052_dispatcher.py, tests/test_pipeline_digest.py: bulk-migrated obsolete category references (wildfire_proximity -> wildfire_hotspot, fire_proximity -> wildfire_incident) so the existing test suites continue to exercise the same routing/digest/dispatch paths with the new category names.
Safe-mode preserved (master off, all family toggles off, all adapters native, central disabled). No live toggle flipped. Not tagging yet -- v0.5.7 tag waits until all families ship.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
build_pipeline previously constructed its own LLMBackend from
config.llm, which:
- duplicated main.py's already-running backend instance
- failed to inherit env-loaded LLM_API_KEY when called from
short-lived scripts (eyeball checks, tests), forcing fallback
- prevented pipeline components from sharing the live backend
build_pipeline and build_pipeline_components now require an
llm_backend parameter. main.py passes the same instance it
constructed for its primary responder. Tests pass mocks. The
digest accumulator now uses the live, authenticated backend.
Added test_build_pipeline_uses_provided_backend to lock in the
injection contract.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Remove severity-based fork; tee pattern sends all events to both dispatcher and accumulator
- Add ToggleFilter before tee; drops events for disabled toggles
- Rework DigestAccumulator: event log instead of active/resolved tracking
- render_digest now async, calls LLM once per toggle with severity-ordered events
- Fallback to count-based summary when LLM unavailable
- Add TogglesConfig to config.py for master toggle settings
- Update scheduler to await async render_digest
- 75 tests passing
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
One-time renormalization pass under the .gitattributes added in the
previous commit. Every tracked text file now uses LF. No semantic
changes — verified via git diff --cached --ignore-all-space showing
zero real differences. Future diffs will only show real content
changes.
This commit will appear huge in git log --stat but represents zero
behavior change. Use git log --follow --ignore-all-space or
git blame -w when archaeologically tracing through this commit.
Adds DigestAccumulator tracking ACTIVE NOW and SINCE LAST DIGEST
state per toggle. Replaces StubDigestQueue in build_pipeline; the
stub class is kept for Phase 2.1 backward-compat tests.
- enqueue(): adds new events, updates in place by id, detects
resolutions (expires past, or title contains cleared/reopened/
ended/resolved/back online/recovered/lifted)
- tick(now): rolls expired actives into since_last
- render_digest(now): produces a Digest with mesh_compact (<=200
chars) and full multi-line forms; clears since_last after
- Toggle ordering and labels match the v0.3 design
- Phase 2.3b will add real scheduling on top of this
