The Archive page's GET /calls/search failed its CORS preflight (OPTIONS -> 405, no Access-Control-* headers). Allow the app origin(s) explicitly for the standard methods and the authorization/content-type headers.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Tbknwttzou4s46PAykmtix
Board minutes #62 Decision 2 (server-26#64), due 2026-08-31. CTO draft #60
finding 1 and CISO draft #61 finding 3 reached this independently.
GET/PUT /admin/features accepted only a Firebase admin token, so the unattended
runbook had no headless path and SSHed into the c2-core container to write
config/ai_features with the admin SDK. Moving a platform-wide AI cost switch
required a full container shell, and set_flags() wrote no audit entry either
way, so a flag flip was unattributable however it happened.
- New agent_service_key (AGENT_SERVICE_KEY), deliberately separate from the
Discord bot's service_key. Sharing one key would collapse two principals into
a single unattributable identity in every log line, and the bot has no
business flipping AI flags regardless.
- require_agent_key_or_admin accepts the agent key or a Firebase admin, and
rejects the Discord key. The "key is configured" guard is load-bearing:
compare_digest("", "") is a match, so a deployment that never set the key
would otherwise accept an empty credential.
- set_flags() writes an audit_log entry with before/after values and the actor,
wrapped so an audit failure cannot lose the flag write or 500 the route.
- Cascade helper sets the global doc and every system carrying an ai_flags
override in one call. A global False already beats everything, but a system
False beats a global True, so turning AI *on* could half-apply and leave a
radio system hot after shutoff. It scans for the override rather than
hardcoding the two known system IDs, so a new system cannot silently defeat
it.
- cascade defaults to False. PUT /systems/{id}/ai-flags and the AiFlagsPanel
toggle mean a per-system override is deliberate operator intent; cascading by
default would erase it on any unrelated global flip. The runbook opts in.
Issue items 5 and 6 (retiring the SSH path from drb-worksession.md) are NOT
done here and the runbook is untouched. The credential does not exist in
production yet, so the SSH path is still the only one that works; retiring it
now would break the next unattended run. Owner activation is recorded on #64.
Tests 273 -> 289.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Alert dispatch attached a 200-character raw transcript snippet to the
alert_events document and POSTed the same text to the org's Discord
webhook, with no redaction of any kind. Board minutes #42 ratified that
person names are suppressed on every surface until E&O is bound, and a
Discord channel is the least recoverable surface there is: once the text
lands we do not own it, cannot unsend it, and cannot audit who read it.
Raw transcript text now requires two independent gates, both closed by
default:
1. alert_transcript_snippet_enabled -- an operator switch in config,
set from the environment.
2. alert_snippet_opt_in on the org document -- the customer's own
explicit consent.
Gate 1 is not redundant. The frontend reads and writes Firestore directly
from the browser, so the org flag alone would let an org owner opt
themselves into receiving person names lifted from live public-safety
traffic. Capability is the operator's to grant; consent is the org's.
The gate fails closed on a Firestore error and on a call with no org
(a pre-tenancy node that has not been backfilled) -- a less informative
alert is cheap, an unrecallable disclosure is not. Alerting itself is
unchanged: the webhook still fires and still names the rule, the
talkgroup and the matched keywords.
This does not wait on the Gate B3 redactor (#43, 2026-09-30). The
snippet was a convenience field and needed no redactor to withhold.
Tests assert the person name in a sample transcript does not appear in
either the outbound payload or the Firestore write, in every combination
of the two gates.
Closes server-26#85. Refs #42, #43, #48.
#36 — the correction pass shipped in 58efdbd was right, its reference-data
shape was not. One shape now, at both scopes, every field nullable:
area_context: { municipality?, county?, state?,
center?, radius_km?, resolved_from?, resolved_at?,
local_knowledge?: [{term, meaning}] }
`state` closes the ambiguity that made "Ossining" a national guess.
`local_knowledge` replaces roads[]/landmarks[], which could not hold
intersections, schools or nicknames and carried no meanings — `11-X-ray` is
useless alone, `11-X-ray — MTA PD patrol unit` is what a corrector can act on.
Pre-#36 roads[]/landmarks[] are read forward as bare terms so nothing an
operator already entered is lost.
Nullability is the mechanism: which scope gets filled is the operator's
declaration of how homogeneous the system is. One town — fill it once at system
level. Statewide — leave it blank and fill each talkgroup.
The backend owns the derived anchor. PUT /systems/{id} merges config.talkgroups[]
against what is stored instead of writing the client's blob verbatim, which
would have erased the anchor and the pending queue — the same defect as the
ten_codes wipe.
#37 — Maps as a verifier, not as prompt stuffing. The corrector emits its
location nouns; each is geocoded against the talkgroup's anchor, and on a miss
we look for a sound-alike that does resolve there, correct to it, and propose
{term, meaning} to that talkgroup. Cost scales with location nouns, not calls.
No anchor means SKIP. An area too wide to discriminate stores no anchor at all,
because a statewide radius would confirm anything inside it — verification that
passes everything is worse than none, since it reads as a check in the data.
Also re-anchors _geocode_location, which rejected results >40km from the NODE
(server-26#6). An antenna is not a jurisdiction; distance-from-node was always
a stand-in for the anchor and is now only the fallback.
The induction loop proposes at talkgroup level and never promotes. Blast
radius: a wrong term on a channel misleads that channel, the same term
system-wide misleads one 400km away on a statewide system.
38 new tests; 240 pass. Frontend typechecks clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Correction existed, but as a line in intelligence.py's EXTRACTION_PROMPT --
which put it in the wrong place twice over. The same model call that extracted
units, location and severity emitted the correction afterwards, so extraction
reasoned over text already known to be wrong; and it sat behind
correlation_enabled, so during a cost-controlled STT-only window nothing was
ever corrected at all. That is the normal state during development.
internal/transcript_correction.py is now its own pass, between the degenerate
filter and the Firestore write. It receives an already-produced transcript plus
a reference list, so unlike a Whisper prompt it has no series to extend -- the
distinction that keeps vocabulary out of the recogniser's prompt, where an
enumerated ten-code list once made it hallucinate ten-code runs.
Reference data is merged from the talkgroup and the system, TALKGROUP FIRST. A
system spanning several counties can have a talkgroup covering one
municipality, and that municipality's streets must not be buried under a
county-wide list. A single-municipality system is the degenerate case: populate
the system level and every talkgroup inherits it. Area context is now SET --
municipality, county, roads, landmarks, on both scopes -- rather than guessed
from talkgroup names, which is what vocabulary_learner did and which is close
to useless across multiple counties.
Segments are corrected too, not just the joined text. extract_scenes builds its
prompt from numbered segments whenever there is more than one, so a correction
that only fixed the transcript would have been discarded on exactly the
multi-transmission calls carrying the most content. Alignment is enforced: an
array of the wrong length or type is dropped whole, because scenes map back to
transmissions by index and a shifted array would misattribute audio silently.
Whisper is also retried once on degenerate output. Call e49ea32c produced a
56-word ten-code counting run on one attempt and ordinary speech on the next --
same clip, same temperature=0 -- so a hallucination is a coin-flip, and
discarding on the first bad roll threw away a recoverable transcript.
Two things found on the way:
PUT /systems/{id} wiped ten_codes on every save. The systems form sends only
{name, type, config}, and model_dump() wrote every omitted field as its default
over the top. Now exclude_unset. area_context would have been the next victim,
which is why it gets its own route alongside ten-codes rather than a field on
that payload.
Closes server-26#36.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two unrelated-looking problems with the same shape: a dangerous state that
looked fine from the outside.
DEPLOY (server-26#21). The Deploy job failed on fifteen consecutive pushes
between 2026-08-18 and 08-20 and nobody noticed for two days, because the
build job was green and a red run is only visible to someone who opens Gitea.
Production served 08-18 code the whole time -- including the entire frontend
redesign, chunks 2 through 8. Three changes:
* The health check now asserts WHICH build answered, not just that something
did. CI bakes the commit into the image (Dockerfile ARG/ENV GIT_SHA) and
/health reports it, so a deploy that "succeeds" while the previous
container keeps running now fails. Liveness alone could never have caught
this.
* The image pull retries once after a prune. The actual failure was
containerd unable to extract a layer -- "failed to Lchown ... no such file
or directory" -- a corrupted entry in the snapshot store, which a prune
clears. A second failure after pruning is a real problem (check the VM's
disk) and still stops the deploy.
* A notify-failure job POSTs to DEPLOY_ALERT_WEBHOOK when anything in the
workflow fails. Unset means skip quietly, not fail.
CORS (server-26#20). allow_origins=["*"] with allow_credentials=True is not
the permissive-but-harmless setting it reads as. Starlette does not reject the
pair -- it reflects the caller's Origin back and still sends
Access-Control-Allow-Credentials: true, so the effective policy is "any
origin, WITH credentials", the opposite of what a wildcard normally means.
Rather than trust every deployment to remember CORS_ORIGINS, the pair is now
unrepresentable: a wildcard forces allow_credentials off and logs an ERROR
naming the variable to set. Correctly configured deployments that name their
origins are unaffected and keep credentialed requests.
Severity honestly: low today. c2-core is bearer-auth, and browsers do not
attach bearer tokens cross-origin the way they attach cookies. This is a
misconfiguration waiting for the day something starts trusting a cookie.
Also adds firebase_admin.auth.UserRecord and the list/update/create/delete_user
names to the conftest stub. routers/users.py annotates with UserRecord at
import time, so without it importing app.main failed at collection -- which is
why nothing had ever tested anything wired at app level, CORS included.
Tests: 5 new in test_cors_policy.py, covering the pure policy function, the
middleware actually mounted on the app (so re-hardcoding allow_credentials=True
fails here), and the presence of the build stamp.
Closeslogan/server-26#20Closeslogan/server-26#21
The 2026-08-20 production dump had 4 of 6 sampled incidents as junk chains,
the worst being f5190670: 68 calls over 4h09m, 44 units, 12 tags, at least
13 genuinely distinct events. 58 of 133 linked calls took the fast/thin
path, which is the one path that attaches a call with no fit test at all.
Three defects combined to produce that, and all three are fixed here.
1. What counted as thin was wrong.
is_thin_call was "not units and not vehicles and not coords". A real
dispatch qualified as thin whenever no unit ID parsed and the geocode
failed - six of them did in that dump, including "All units head over to
the powerhouse, 55 Hyman Hills Road ... she's 87 years old", a brand new
job that attached to the four-hour chain and then overwrote its location
and its title. A call is now substantive if it carries tags, a location
string, a severity above routine, or is a reassignment; only genuinely
content-free housekeeping ("10-4", "Copy") stays thin. Those calls now go
through _call_fits_incident like everything else, which on a dispatch
backbone with no positive signal means they open their own incident or
orphan rather than merging.
The reassignment clause closes a self-defeating guard: upload.py blanks
units when dispatch pulls a unit onto a NEW job, specifically to stop
unit-overlap chaining - and blanking units made the call thin, routing it
to the only path with no fit check. The guard produced the merge it
existed to prevent.
2. The thin path was bounded on dispatch channels only.
Every other talkgroup fell through to "thin_pool = tg_recent": any
incident idle up to tg_fast_path_idle_minutes (90), no single-candidate
requirement, no fit test. The 30-second tier-1 / single-candidate tier-2
structure now applies to all channels. Non-dispatch gets its own window,
TG_THIN_IDLE_MINUTES=15, rather than sharing the dispatch value: a
tactical channel really is dedicated to one scene so it earns longer, but
15 sits inside the 20-minute tactical-default window already used in
_call_fits_incident, so the no-evidence path is never more permissive than
the fit-tested path on the same channel.
Recency gates now compare the magnitude of the idle, not the signed value.
The re-correlation sweep anchors "now" to the call's own started_at, so
idle goes negative routinely - incident 9d376ffe recorded
corr_incident_idle_min: -4.1 - and every "idle <= window" test in this
module reads True for a negative number. Those gates had silently stopped
bounding anything for exactly the calls the sweep re-examines.
3. Nothing capped an incident's total size.
Every fit test in the correlator is pairwise: does this call belong with
that incident. Each of f5190670's 68 links was individually arguable; the
mistake was the accumulated shape, which no pairwise rule can see. Two
hard caps now remove an incident from the candidate pool entirely, before
any path can choose it - including the LLM tier, which reads the same
ctx lists.
INCIDENT_MAX_DURATION_MINUTES=120. The one incident in that dump that was
genuinely a single event ran 63 minutes (06:15 wrong-way driver to 07:18
closeout), so the cap has to clear an hour with real headroom. The four
junk chains ran 3h41m, 3h43m, 4h05m and 4h09m, so it has to sit well under
three hours. 120 also equals correlation_window_hours: the location and
slow paths already refuse a candidate older than that, and the fast path
was the only one exempt, so this removes an inconsistency rather than
inventing a number.
INCIDENT_MAX_CALLS=40. A backstop for a burst that fills up inside the
duration cap, not the primary bound. The worst chain averaged ~16
calls/hour while absorbing an entire dispatch backbone, so 40 calls in
under two hours means one incident is eating most of the channel. Set
deliberately above any plausible single-incident call volume (a
multi-alarm fire on its own tactical channel) so this cap errs toward
keeping real incidents whole and lets the duration cap do the cutting.
Capping is not truncation: the incident keeps every call it has and still
auto-resolves on the normal idle sweep. It just stops being a candidate.
Every ambiguous call here was resolved toward a separate incident rather
than a merge. A wrongly-separate incident is visibly wrong and can be
merged later; a wrongly-merged one silently corrupts every unit, tag,
severity and map pin on the incident it joined, and poisons the AI
summary written from them. The cost is some acknowledgements orphaning
instead of riding along on an incident, which is a small, visible loss.
Deliberately NOT changed, since both push toward more merging while the
current failure mode is entirely over-merging (every incident in the dump
has exactly one "new" call; there is no over-splitting left to trade
against):
- unit-overlap positive feedback on shared dispatch channels, which is
now bounded by the caps rather than fixed at its root
- the sweep retry budget expiring before the target incident exists
Tests: 31 new cases in tests/test_correlator_merge_caps.py, including a
replay of the f5190670 night - 13 unrelated jobs at their real offsets,
plus roster unit traffic and acknowledgements every two minutes. Without
the caps that traffic still builds a 125-call incident spanning 244
minutes; with the old thinness test on top, 153 calls over 247 minutes in
3 incidents. With this commit it is 13 incidents, largest 40 calls over 80
minutes. Each new case was checked to fail when the behaviour it covers is
reverted. Suite: 138 passed.
No AI feature flag was touched; correlation stays off in production.
Closeslogan/server-26#22
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three AI dependency failures in one night (retired Gemini model IDs,
depleted Gemini balance, unpayable OpenAI account) each surfaced only
as a single ERROR log line that nobody was watching. Add
app/internal/ai_health.py, a shared in-memory registry that
transcription.py and llm_correlator.py report into on every call
(success and failure), distinguishing permanent conditions (dead
model, dead billing) which alert immediately from transient ones
(rate limits, network blips) which only alert after they persist.
Alerts POST once per degradation episode and once on recovery to an
optional Discord webhook (AI_ALERT_WEBHOOK_URL), reusing alerter.py's
httpx pattern. State is exposed unauthenticated at GET /health/ai
alongside the existing /health.
Closeslogan/server-26#14
Both Gemini model IDs had been retired by Google. Production logs show every
correlation call 404ing -- "models/gemini-2.0-flash is no longer available" --
and gemini-1.5-pro is gone from the model list as well. Because a failed LLM
call falls back to the rules decision by design, nothing surfaced: the pipeline
kept producing incidents, so the LLM tier and the consensus tiebreak were dead
in production for an unknown number of days while correlation was being tuned.
Some of what recent tuning was reacting to was rules-only behaviour that was
never meant to run alone.
Cheap model becomes gemini-3.6-flash, which is the migration target named in
Google's own 404. Smart model becomes gemini-2.5-pro, the only stable Pro-tier
text model left; the tiebreak fires rarely and its value comes from being a
different, stronger model than the first pass, so a second Flash was not worth
the consensus it would give up. Model list checked against
https://ai.google.dev/gemini-api/docs/models on 2026-08-18.
The more important half is the logging. A per-call WARNING was the only signal,
and it is indistinguishable from an ordinary API hiccup, so a permanent
misconfiguration read as noise. Failures that look like a missing model (404,
"not found", "no longer available") now log once per model at ERROR, name the
config keys to change, and say plainly that correlation is running rules-only.
Transient errors keep the old per-call WARNING. Once per model, not once per
call, so the alert stays readable at radio traffic volume.
Gemini is used nowhere else in c2-core -- extraction, embeddings and summaries
all run on OpenAI -- so the blast radius was exactly the correlation LLM tier.
38 correlator tests still pass. No new environment variables: both model IDs
are config.py defaults and are not templated into any .env, so CI deploys this
without an ansible run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
With over-creation fixed, the incidents that remain are readable enough to
judge, and the ones that still do not make sense all fail the same way. A
content-free call attaches to the single active incident on its talkgroup if
that incident has been idle under tg_dispatch_thin_idle_minutes, and at 10
minutes that is long enough for the channel to have moved on to something
else. In the 00:30Z dump a "72 at Holland Station" incident absorbed a Grand
Central train-crew meet 9.6 minutes later, and a status check absorbed a
records lookup at 9.7.
Being the only candidate is not evidence. It means the channel was quiet,
which is exactly when guessing is weakest -- the single-candidate rule was
meant to avoid picking wrongly among several, not to license a match no other
signal supports.
Every correct thin attach in that dump was <= 3.4 minutes idle and every wrong
one was >= 8.2, so 5 separates them with room on both sides. Real
back-and-forth is unaffected: it runs through the 30-second tier-1 path, and
the observed conversational replies sit near zero. Tests pin both sides of the
new boundary at 4.9 and 5.1 minutes so a later change to this number has to be
deliberate. 23 pass.
Also corrects a DEFERRED.md entry written earlier today. It claimed nothing
ever closes an incident that goes quiet; summarizer.py has run a stale sweep
at incident_auto_resolve_minutes (90) the whole time. The 37 open incidents
were caused by over-creation, not by a missing sweeper, and 90 minutes may be
fine now -- worth rechecking on a fully post-fix dump before changing it.
No new environment variables: tg_dispatch_thin_idle_minutes is a config.py
default and is not templated into any .env, so CI deploys this without an
ansible run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two independent sources of garbage in the AI pipeline, both visible in the
2026-08-16 correlation dump.
1. Hallucinated transcripts. The Whisper prompt opened with an enumerated run
of ten-codes: 10-4, 10-23, 10-20, 10-97 and so on. Whisper treats prompt
text as preceding transcript, so on noisy or silent audio it continued the
series, emitting transcripts that count upward from 10-4 to 10-99. The
existing no_speech_prob filter could not catch these: the model is highly
confident in text it invented by continuing a pattern.
The prompt no longer contains a series to extend, and _is_degenerate()
rejects the three shapes this failure takes: ascending ten-code runs, one
phrase looping, and near-identical segments across a whole recording.
Verified against 13 transcripts from production: all four known
hallucinations rejected, all nine real ones kept, including terse traffic
containing legitimate codes.
2. Duplicate recordings. node-002 and node-PI-2 both cover TG 9048 and both
uploaded the same transmissions, ~1.1s apart. Nine pairs appeared in one
dump. Each was transcribed, billed and correlated twice, and the resulting
incident listed two units where there was one.
Canonical selection is by earliest started_at, tie-broken on call_id, NOT
by upload order: upload order varies with encode time and network latency,
so it would make the authoritative recording non-deterministic. Call
documents are created from MQTT call_start before uploads arrive, so both
nodes independently reach the same verdict. The loser keeps its audio (it
may be the cleaner capture) but is excluded from STT, correlation, the
re-correlation sweep and the orphan debug view.
Also fixes _sync_transcribe returning a bare None when OPENAI_API_KEY is
missing, where the caller unpacks two values. A missing key surfaced as a
misleading "Transcription failed" instead of the real warning.
Adds tests/test_dedup.py (15 cases). dedup.py reaches Firestore through an
injected callable so it stays importable without firebase-admin present.
upload_audio() could only sign a URL when GCP_CREDENTIALS_PATH pointed at a
service-account key file. The deployed VM runs on Application Default
Credentials with no key file, so every upload silently took the fallback
branch and returned a bare gs:// URI. That broke two things at once:
* Browsers cannot fetch a gs:// URI, so no recording was ever playable.
* _public_url_to_gcs_uri() only matched https://storage.googleapis.com/ and
returned None for it, so `if gcs_uri:` in the upload path was always false
and transcription never ran. Nothing was logged, which is why this looked
like an OpenAI credits problem rather than a storage one.
The fallback also interpolated the client-supplied filename instead of the
call_id-derived safe name, so the URI did not even name the object written.
Calls now store only the canonical gs:// location. A short-lived playback link
is minted per read as an HMAC over (call_id, expiry) keyed by SERVICE_KEY, and
audio is served from the private bucket by the new /media route. An <audio src>
cannot carry an Authorization header, so the link has to be the credential;
that router is therefore public with the check done inline, as enrollment.py
already does. Signing GCS URLs from the VM would have needed a
serviceAccountTokenCreator grant on its own service account — this avoids the
IAM change entirely and keeps the bucket private.
gcs_uri_for_call() reconstructs the object name from call_id, so recordings
made before this fix are reachable again without a data migration.
Frontend rows come straight from Firestore via onSnapshot and never see a
server-minted field, so CallRow fetches the link lazily on expand.
Also removes the last long-lived (1 year) signed URL and the log line that
printed it.
Edge nodes are deployed to arbitrary locations by arbitrary people, so the
broker has to be reachable from the internet and secured on its own merits
rather than by a VPN.
Three defects made that impossible. The broker only had a plaintext 1883
listener; every node shared one drb-node password; and the ACL pattern used
%c, the client-supplied client id, so any holder of that shared password
could set client_id to another node and take over its namespace. The comment
claiming this cryptographically prevented cross-node access was wrong and is
gone.
Authentication now uses mosquitto 2.x's built-in dynamic-security plugin on
the stock eclipse-mosquitto image. c2-core administers it over the control
topic, creating each node's client on approval with username=<node_id> and
password=<its node_keys api_key>, attached to a role whose ACL is nodes/%u/#
against the authenticated username. One credential, one revocation point.
An HTTP-callback plugin was implemented first and rejected: that project is
archived upstream, which is not an acceptable dependency on an
internet-facing broker.
Because dynsec state is a second source of truth alongside Firestore,
approve/reissue/delete now write to the broker first and surface a 502
rather than drifting, and c2-core reconciles every approved node into dynsec
on startup.
Adds node self-enrollment (POST /nodes/enroll, GET /nodes/{id}/credentials)
so a new node can obtain its key over HTTPS without an operator handling
secrets by hand. Enrolling an already-approved node_id is refused on the
fleet token alone — otherwise a leaked token plus a guessable id would let
an attacker steal a live node's key before the real node asked for it.
Pickup secrets are stored hashed and returned once, and the endpoint is rate
limited per source IP.
Infrastructure: an 8883 TLS listener fed by Caddy's certificate via a
systemd path unit, a firewall rule for it, and Caddy now 404s /internal/*
so the api vhost cannot proxy internal routes.
Also fixes CORS, which allowed https://app.<domain> while the frontend is
served on the bare domain — every call from the portal would have failed —
and widens the vault gitignore to a glob, since ansible-vault leaves
backup siblings that the exact-name rule left committable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
auth.py
secrets.compare_digest replaces == for service key comparison (timing-safe)
Added require_service_key — bot-only endpoints (trip/event join/leave)
Added require_service_key_or_admin — node commands/config (bot via service key OR dashboard admin via Firebase)
Added _RateLimiter with three shared instances: trip_chat_limiter (20/5min per user), summarize_limiter (5/10min per incident), bootstrap_limiter (2/hr per system)
nodes.py
send_command and assign_system now require require_service_key_or_admin — the Discord bot can still call them via service key, but regular Firebase users are blocked
tokens.py
add_token, flush_tokens, set_preferred_system, delete_token all require require_admin_token
Token masking changed from token[:10] + "…" + token[-4:] to "•••" + token[-4:]
systems.py
All write endpoints (create, update, delete, ai-flags, ten-codes, vocabulary writes, bootstrap) now require require_admin_token
bootstrap_vocabulary also calls bootstrap_limiter.check(system_id)
incidents.py
POST /incidents/summarize (bulk) now requires require_admin_token
POST /incidents/{id}/summarize now calls summarize_limiter.check(incident_id)
trips.py
join_trip, leave_trip, join_event, leave_event require require_service_key — only the Discord bot can set Discord attendee identity
delete_trip, delete_event require require_service_key_or_admin
trip_chat rate-limited per caller UID, history stripped to user/assistant roles only, user message truncated to 2000 chars, Maps query strings capped at 200 chars
upload.py
Rejects files larger than settings.upload_max_bytes (default 100MB) with 413
storage.py
_safe_audio_filename() derives GCS object name from call_id + allowlisted extension, completely ignoring the client-supplied filename
config.py
Added upload_max_bytes: int = 100 * 1024 * 1024
Both Dockerfiles — python:3.14-slim → python:3.12-slim
- correlator: unit_overlap on dispatch channels now applies content
divergence check when the call has geocoded coords but the incident
doesn't; previously this gap caused unrelated calls to merge into
stale incidents (e.g. patrol officer at a second scene 70 min later)
- STT: switch default model from gpt-4o-transcribe to whisper-1, which
faithfully transcribes all exchanges in multi-PTT recordings; gpt-4o
was silently dropping utterances, starving the correlation engine
- STT: remove vocabulary from the Whisper prompt; whisper-1 echoes
prompted terms into noise/silence, skewing extracted incident data;
vocabulary context is now applied exclusively in the GPT extraction
step (build_gpt_vocab_block) where it is used as reference only
Refactor incident_correlator.py to a decision/commit split (preview_correlation
/ apply_correlation) so the rules engine and LLM can both produce decisions before
anything is written to Firestore.
Add llm_correlator.py: cheap Gemini Flash first-pass + Gemini Pro tiebreaker.
Wire _correlate_with_consensus in upload.py — rules-only fallback when key is
absent or call is thin; agreed/tiebreak consensus written to corr_debug.
- Cap unit-continuity path at 20 min idle (unit_continuity_max_idle_minutes)
- Block time_fallback and unit-continuity matching on reassignment calls
- Expand reassignment detection to cover unit-initiated self-reassignment
- Skip GPT extraction entirely for transcripts ≤5 words (prevents hallucinated tags/units)
- Reduce geocode_max_km from 75 to 40 to reject far-out-of-area results
- Include county in geocoding query for tighter jurisdiction anchoring
geocode_max_km: 25 → 75 km. The node is a physical receiver, not the system boundary;
digital repeaters extend coverage well beyond 25km (North White Plains at 35.5km from
the Yorktown node is a legitimate Westchester County location).
Query now fully qualified: "High Street" → "High Street, Yorktown, New York".
Added _get_node_state() which reverse-geocodes the node position once (cached) using
Google Maps to get the state name, appended alongside the municipality.
Generic street names (High Street, Main Street) no longer resolve to wrong-country results.
Switch geocoding from Nominatim to Google Maps Geocoding API for accurate
local place name resolution (bounds-biased, with 25km distance rejection guard).
Remove the now-unused _get_node_place reverse-geocoder and _node_place_cache.
Map page (TOC improvements):
- Weather radar tiles auto-refresh every 5 minutes via radarEpoch key cycling
- Google Maps traffic overlay added to LayersControl
- Live 24h clock overlay at bottom-left for situational awareness
- Incident sidebar cards now show age (time since dispatch) and unit count
Nominatim's viewbox is advisory (bounded=0), so ambiguous place names like
"Pinebrook" can resolve to locations 30-40km away in the wrong town. Added
a post-geocode distance gate: results farther than geocode_max_km (default
25km) from the node are discarded with a warning log rather than written to
the incident.
Also logs distance on successful geocodes for easier audit.
New config setting: geocode_max_km (float, default 25.0)
Correlator
- Raise fast-path idle gate 30 → 90 min (tg_fast_path_idle_minutes)
- Fix disambiguate always-commits bug: run _call_fits_incident on winner
before committing; fall through to new-incident creation if it fails
- Add unit-continuity path (path 1.5): matches all_active by shared unit
IDs with a reassignment guard, bridges calls past the idle gate
- Add tag-based incident_type inference (_TAG_TYPE_HINTS) as GPT fallback,
rescuing tagged calls that would have been dropped (616 observed orphans)
- Add master/child incident model: _create_master_incident, _demote_to_child,
_add_child_to_master; new incidents stamped incident_type="master"
- Add cross-system parent detection (_find_cross_system_parent): two-signal
scoring (road overlap=0.4, embedding≥0.78=0.3, proximity=0.3, threshold=0.5)
wired into create-if-new path; creates master shell on first cross-system match
- Add maybe_resolve_parent: auto-resolves master when all children close;
called from upload pipeline (LLM closure) and summarizer stale sweep
- Add signal-based auto-resolve via units_active/units_cleared tracking:
GPT now extracts cleared_units per scene; _update_incident moves units
between active/cleared lists and resolves the incident when active empties;
stored on call doc for re-correlation sweep reuse
- Add _create_incident initialization of units_active/units_cleared fields
Re-correlation sweep
- Add corr_sweep_count + MAX_SWEEP_ATTEMPTS=3: orphans get 3 attempts
then are tombstoned as corr_path="unlinked", ending the re-sweep loop
(previously hammering each orphan 29-31 times per shift)
Intelligence extraction
- Add cleared_units to GPT prompt schema and rules
- Extract and propagate cleared_units per scene; merge across scenes;
store on call doc for re-correlation sweep
Token management
- Fix token release bug: remove release_token call on discord_connected=False
in MQTT checkin (transient Discord drops were orphaning bots mid-shift)
- Add PUT /tokens/{id}/prefer/{system_id} endpoint: lock a bot token to a
system; pass _none as system_id to clear; stored bidirectionally on both
token and system documents
- discord_join handler resolves preferred_token_id from system doc and passes
system_name in MQTT payload
- *`correlate_call`* — added units and vehicles optional params; when provided (per-scene from intelligence extraction), they take priority over the merged call-document values, preventing multi-scene unit contamination
- *Cross-TGID correlation path (2.5)* — *new path between location and slow paths*: when a call shares 2+ unit IDs with a recent same-system, same-type incident AND embedding similarity ≥ 0.85, it links them — catches multi-talkgroup pursuits like the bicycle search that split across dispatch/tactical/geographic channels
# `app/internal/intelligence.py`
- *`reassignment` field* — added to the GPT-4o-mini prompt schema and rules; `true` when dispatch is actively pulling a unit to a new, different call (not a status update or en route acknowledgement); returned in every processed scene dict
- *Tag location rule* — added explicit instruction to the prompt: tags must describe what happened, not where; place names, road names, and talkgroup names are explicitly forbidden as tags
# `app/routers/upload.py`
- Both scene correlation call sites (`_run_extraction_pipeline` and `_run_intelligence_pipeline`) now pass `units=corr_units` where `corr_units = [] if scene.get("reassignment") else scene.get("units") `— suppresses unit overlap matching when a unit is being reassigned to a new call, preventing chaining into their previous incident
- Both sites also pass `vehicles=scene.get("vehicles")` (per-scene vehicles, from the multi-scene units fix)
# `app/config.py`
- `embedding_cross_tg_threshold: float = 0.85` — threshold for the new cross-TGID path
incident_correlator.py — full rewrite: always runs on every call, fetches all active incidents cross-type, fast path collects all talkgroup matches and disambiguates by unit/vehicle overlap → location proximity → embedding, new location proximity path, slow path requires location corroboration, "Auto:" stripped from titles, "auto-generated" tag added, units/vehicles now accumulated on update
intelligence.py — resolved field in GPT schema, returned as 5th value
upload.py — both pipelines unpack 5-tuple, always call correlate, auto-resolve on resolved=True
summarizer.py — stale sweep runs each tick, resolves incidents idle for 90+ minutes
config.py — correlation_window_hours=2, embedding_similarity_threshold=0.93, location_proximity_km=0.5, incident_auto_resolve_minutes=90
Logan CusanoandClaude Sonnet 5