ci: deploy Firestore rules with a service account, not login:ci (#51)

The deploy-firestore-rules job has failed on every push because
FIREBASE_TOKEN was never set, so rule changes (e.g. aircraft/vessels for
node-26#9) never reached prod. Switch to a dedicated least-privilege
service account whose JSON key lives in FIREBASE_SA_KEY; login:ci tokens
are deprecated and carry their minter's full access. Key is written to
RUNNER_TEMP at 0600 and removed on exit.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

.gitea/workflows/deploy.yml

@@ -307,28 +307,30 @@ jobs:
 
       - name: Deploy firestore rules and indexes
         env:
-          FIREBASE_TOKEN: ${{ secrets.FIREBASE_TOKEN }}
+          FIREBASE_SA_KEY: ${{ secrets.FIREBASE_SA_KEY }}
         run: |
           set -e
           # server-26#51: this used to run over SSH on the deploy VM, gated
           # on the VM having firebase-tools installed. It never did, so it
           # silently warned-and-skipped on every single deploy for weeks.
-          # Running it here instead means the only prerequisite is a secret
+          # Auth is a dedicated service account (drb-ci-firestore-deploy,
-          # -- FIREBASE_TOKEN, from `firebase login:ci` -- rather than
+          # roles: Firebase Rules Admin, Cloud Datastore Index Admin,
-          # something installed by hand on a machine this pipeline doesn't
+          # Service Usage Consumer), its JSON key stored as the
-          # otherwise touch. A missing token now fails this job LOUDLY
+          # FIREBASE_SA_KEY secret. Not `firebase login:ci`: those tokens are
-          # (picked up by notify-failure) instead of a buried warning line
+          # deprecated and carry the full permissions of whoever minted them.
-          # nobody reads in the app deploy's logs.
+          # A missing key fails this job LOUDLY (picked up by notify-failure).
-          if [ -z "$FIREBASE_TOKEN" ]; then
+          if [ -z "$FIREBASE_SA_KEY" ]; then
-            echo "FIREBASE_TOKEN secret is not set -- cannot deploy Firestore rules/indexes." >&2
+            echo "FIREBASE_SA_KEY secret is not set -- cannot deploy Firestore rules/indexes." >&2
-            echo "Generate one with 'firebase login:ci' and add it as a Gitea Actions secret." >&2
+            echo "Add the drb-ci-firestore-deploy service account's JSON key as a Gitea Actions secret." >&2
             exit 1
           fi
+          export GOOGLE_APPLICATION_CREDENTIALS="$RUNNER_TEMP/firebase-sa.json"
+          trap 'rm -f "$GOOGLE_APPLICATION_CREDENTIALS"' EXIT
+          ( umask 077 && printf '%s' "$FIREBASE_SA_KEY" > "$GOOGLE_APPLICATION_CREDENTIALS" )
           npm install -g firebase-tools
           cd infra/firestore
           firebase deploy --only firestore:rules,firestore:indexes \
-            --project ${{ secrets.FIREBASE_PROJECT_ID }} \
+            --project ${{ secrets.FIREBASE_PROJECT_ID }} --non-interactive
-            --token "$FIREBASE_TOKEN" --non-interactive
 
   notify-failure:
     name: Report a failed deploy
@@ -371,7 +373,7 @@ jobs:
           # failed before any deploy was attempted" text even when the app
           # deployed fine and only the Firestore rules/indexes push failed.
           if deploy_result != "failure" and rules_result == "failure":
-              detail = "App deploy succeeded; Firestore rules/indexes deploy FAILED (server-26#51). Rules may be stale — check FIREBASE_TOKEN and the job log."
+              detail = "App deploy succeeded; Firestore rules/indexes deploy FAILED (server-26#51). Rules may be stale — check the FIREBASE_SA_KEY secret and the job log."
 
           # server-26#65: the old text here unconditionally claimed
           # "production is still running the previous build" -- true only

drb-c2-core/app/internal/gemini.py

@@ -0,0 +1,94 @@
+"""
+One place every Gemini call goes through: JSON-mode generation, an explicit
+thinking level, and token accounting.
+
+Why it exists: a day of replay runs (server-26#170) cost ~$5 of Gemini for
+~7 two-hour windows — roughly $0.70 per 290 calls, which projects to several
+dollars a day per live deployment for correlation alone — and nothing in DRB
+could say where it went (server-26#45). Gemini 3.x models "think" by default
+and bill that as output; the old google-generativeai SDK these calls used
+cannot even set a thinking level. A link/new/orphan choice or a transcript
+cleanup does not need extended reasoning.
+
+Every call logs its token counts, and inside a replay run they are also added
+to the run's own usage sink (see app/internal/replay.py), so a run reports
+what it actually spent instead of an estimate.
+"""
+import json
+import threading
+from contextvars import ContextVar
+from typing import Optional
+
+from app.config import settings
+from app.internal.logger import logger
+
+_client = None
+_client_lock = threading.Lock()
+# Models that rejected a thinking level: retried without one from then on.
+_no_thinking_level: set[str] = set()
+
+_usage_sink: ContextVar[Optional[dict]] = ContextVar("drb_gemini_usage", default=None)
+
+
+def collect_usage(sink: Optional[dict]):
+    """Route token counts for the current context into `sink` (a replay run). Returns a reset token."""
+    return _usage_sink.set(sink)
+
+
+def reset_usage(token) -> None:
+    _usage_sink.reset(token)
+
+
+def _get_client():
+    global _client
+    with _client_lock:
+        if _client is None:
+            from google import genai  # lazy — only when a Gemini call is made
+            _client = genai.Client(api_key=settings.gemini_api_key)
+        return _client
+
+
+def _config(thinking_level: Optional[str]):
+    from google.genai import types
+    kwargs = {"response_mime_type": "application/json"}
+    if thinking_level:
+        kwargs["thinking_config"] = types.ThinkingConfig(thinking_level=thinking_level)
+    return types.GenerateContentConfig(**kwargs)
+
+
+def _record(purpose: str, model: str, usage) -> None:
+    prompt = getattr(usage, "prompt_token_count", None) or 0
+    output = getattr(usage, "candidates_token_count", None) or 0
+    thoughts = getattr(usage, "thoughts_token_count", None) or 0
+    logger.info(f"gemini usage {purpose} {model}: in={prompt} out={output} thinking={thoughts}")
+    sink = _usage_sink.get()
+    if sink is not None:
+        row = sink.setdefault(f"{purpose}:{model}", {"calls": 0, "in": 0, "out": 0, "thinking": 0})
+        row["calls"] += 1
+        row["in"] += prompt
+        row["out"] += output
+        row["thinking"] += thoughts
+
+
+def generate_json(model: str, prompt: str, *, purpose: str,
+                  thinking_level: Optional[str] = "minimal") -> dict:
+    """
+    Synchronous (run it via asyncio.to_thread). Returns the parsed JSON body.
+    Raises on API failure, exactly like the old per-module helpers, so callers'
+    ai_health classification (billing / dead model / transient) is unchanged.
+    """
+    client = _get_client()
+    level = None if model in _no_thinking_level else thinking_level
+    try:
+        resp = client.models.generate_content(model=model, contents=prompt, config=_config(level))
+    except Exception as e:
+        # A model that doesn't accept this thinking level answers 400 for
+        # every call; drop the setting for that model rather than lose the tier.
+        if level and "thinking" in str(e).lower():
+            logger.warning(f"gemini: {model} rejected thinking_level={level!r} ({e}); retrying without it")
+            _no_thinking_level.add(model)
+            resp = client.models.generate_content(model=model, contents=prompt, config=_config(None))
+        else:
+            raise
+    _record(purpose, model, getattr(resp, "usage_metadata", None))
+    return json.loads(resp.text)

drb-c2-core/app/internal/incident_correlator.py

@@ -343,9 +343,21 @@ def clean_location(value) -> Optional[str]:
     s = str(value).strip()
     if not s or not _LOCATION_WORD_RE.search(s):
         return None
+    if _RADIO_CODE_RE.match(s):
+        return None
     return s
 
 
+# Status/disposition codes the extractor sometimes returns as a location:
+# "96 times 5" (a disposition code read aloud) titled a 09-22 replay stop
+# "Traffic Stop at 96 times 5" (server-26#170); "10-8", "signal 99", "code 4"
+# are the same shape.
+_RADIO_CODE_RE = re.compile(
+    r"^\s*(?:\d{1,3}\s*(?:times|x)\s*\d{1,3}|10[\s-]?\d{1,3}|(?:signal|code|condition)\s+\d{1,3})\s*$",
+    re.IGNORECASE,
+)
+
+
 def location_is_unit(location, units) -> bool:
     """
     True when a location label is really one of the incident's own unit

drb-c2-core/app/internal/intelligence.py

@@ -517,12 +517,33 @@ _NO_BACKSTOP_TG = re.compile(r"\b(mta|rail|railroad|train|transit|bridges? and t
                              r"rescue|ambulance|dpw|public works)\b", re.IGNORECASE)
 
 
+# A plate read aloud — two or more phonetic letters then 3-7 digits:
+# "Frank David Boy, 4514", "Lincoln, Charlie, Robert, 7-4-0-7". On a patrol
+# channel that is a unit running a car it has stopped. Held-out replay of
+# 09-21 (server-26#170): Ossining's Post 4 stops were read out only as plates
+# and never became incidents.
+_PHONETIC = (r"(?:adam|alpha|baker|boy|bravo|charlie|charles|david|delta|eddie|edward|echo|frank|"
+             r"george|golf|henry|hotel|ida|india|john|juliet|king|kilo|lincoln|lima|larry|mary|"
+             r"michael|mike|nora|nancy|november|ocean|oscar|peter|paul|papa|queen|robert|romeo|"
+             r"sam|sierra|tom|tango|union|uniform|victor|william|whiskey|x-ray|xray|young|yankee|zebra|zulu)")
+_PLATE_READ = re.compile(rf"\b{_PHONETIC}(?:[,\s]+{_PHONETIC}){{1,3}}[,\s]+\d(?:[\s-]?\d){{2,6}}\b",
+                         re.IGNORECASE)
+
+
 def _self_initiated_backstop(
     text: str, tags: list, incident_type: Optional[str], severity: str,
     talkgroup_name: Optional[str] = None,
 ) -> tuple[list, Optional[str], str]:
     if talkgroup_name and _NO_BACKSTOP_TG.search(talkgroup_name):
         return tags, incident_type, severity
+    # A plate read only stands for a stop when extraction found no other
+    # event in the call: the plate on an MVA, a tow or a parked-car complaint
+    # belongs to that event, not to a new stop.
+    if not tags and _PLATE_READ.search(text or ""):
+        tags = ["traffic-stop"]
+        incident_type = incident_type or "police"
+        if severity == "routine":
+            severity = "minor"
     for pattern, tag in _SELF_INITIATED:
         m = pattern.search(text or "")
         if not m or _NEGATED.search(text[: m.start()]):

drb-c2-core/app/internal/llm_correlator.py

@@ -20,7 +20,6 @@ Error handling: any Gemini failure returns None from decide() and the
 rules_decision from tiebreak() so the pipeline never stalls.
 """
 import asyncio
-import json
 from datetime import datetime, timezone
 from typing import Optional
 from app.internal.logger import logger
@@ -190,15 +189,8 @@ def _build_tiebreak_prompt(rules_decision: dict, llm_decision: dict, ctx: dict)
 # ─────────────────────────────────────────────────────────────────────────────
 
 def _sync_gemini(model_name: str, prompt: str) -> dict:
-    import google.generativeai as genai  # lazy import — only when needed
+    from app.internal import gemini
-
+    return gemini.generate_json(model_name, prompt, purpose="correlation")
-    genai.configure(api_key=settings.gemini_api_key)
-    model = genai.GenerativeModel(
-        model_name,
-        generation_config={"response_mime_type": "application/json"},
-    )
-    response = model.generate_content(prompt)
-    return json.loads(response.text)
 
 
 # ─────────────────────────────────────────────────────────────────────────────

drb-c2-core/app/internal/replay.py

@@ -469,6 +469,9 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
     fl_token = force_flags(_flags_for(mode))
     ai_failures: list = []
     ai_token = ai_health.collect_sandbox_failures(ai_failures)
+    from app.internal import gemini
+    usage: dict = {}
+    usage_token = gemini.collect_usage(usage)
     try:
         sem = asyncio.Semaphore(PREFETCH)
 
@@ -561,6 +564,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
         metrics = compute_metrics(incidents, sb_calls)
         metrics["est_cost_usd"] = _running_cost(progress, metrics, mode)
         metrics["ai_failures"] = dict(Counter(f"{f['tier']}: {f['problem']}" for f in ai_failures))
+        metrics["gemini_usage"] = usage
     except Exception as e:
         status = "failed"
         errors.append(f"run: {type(e).__name__}: {e}"[:300])
@@ -568,6 +572,7 @@ async def _run(run_id: str, org_id: str, calls: list[dict], mode: str,
         logger.error(f"Replay {run_id} failed: {e}")
     finally:
         ai_health._sandbox_failures.reset(ai_token)
+        gemini.reset_usage(usage_token)
         unforce_flags(fl_token)
         fstore.exit_sandbox(sb_token)
         _cancel.discard(run_id)

drb-c2-core/app/internal/transcript_correction.py

@@ -43,7 +43,6 @@ another equally plausible word.
 """
 
 import asyncio
-import json
 import re
 from typing import Any, Optional
 
@@ -228,14 +227,11 @@ def build_context_block(context: dict, talkgroup_name: Optional[str]) -> str:
 
 
 def _sync_gemini(model_name: str, prompt: str) -> dict:
-    import google.generativeai as genai  # lazy import — only when needed
+    from app.internal import gemini
-
+    # Correction rewrites text against vocabulary; keep a little reasoning
-    genai.configure(api_key=settings.gemini_api_key)
+    # ("low") rather than the correlator's "minimal" until a replay shows
-    model = genai.GenerativeModel(
+    # minimal doesn't hurt it.
-        model_name,
+    return gemini.generate_json(model_name, prompt, purpose="correction", thinking_level="low")
-        generation_config={"response_mime_type": "application/json"},
-    )
-    return json.loads(model.generate_content(prompt).text)
 
 
 async def correct(

drb-c2-core/requirements.txt

@@ -6,6 +6,7 @@ firebase-admin
 google-cloud-storage
 openai
 google-generativeai
+google-genai
 numpy
 httpx
 python-multipart

drb-c2-core/tests/test_clearance_tracking.py

@@ -140,3 +140,24 @@ def test_short_stop_report_opens_a_scene():
             return await intelligence.extract_scenes("c1", "Adam 3 on a stop.", "Ch 1 (Patched with 155.310)")
     scenes = asyncio.run(run())
     assert len(scenes) == 1 and scenes[0]["tags"] == ["traffic-stop"] and scenes[0]["incident_type"] == "police"
+
+
+def test_plate_read_on_a_patrol_channel_is_a_stop():
+    from app.internal.intelligence import _self_initiated_backstop as b
+    ch = "Ossining - Police Dispatch"
+    for t in ("Post 4. 52-62, 3-3. Hemlock Circle. Frank David Boy, 4514. 10-8.",
+              "4, Ossining. 52-22, Ramapo, New York. Lincoln, Charlie, Robert, 7-4-0-7 on a Chevy.",
+              "New York, Mary, Charlie, Nora, 5-8-6-7."):
+        assert b(t, [], None, "routine", ch) == (["traffic-stop"], "police", "minor"), t
+    # a plate on a call that is already about something else stays with it
+    assert b("MVA, plate Mary George Sam 2740", ["mva"], "accident", "moderate", ch) == (["mva"], "accident", "moderate")
+    # not on rail/bridge channels, and not without digits
+    assert b("Frank David Boy 4514", [], None, "routine", "MTA Bridges and Tunnels - Whitestone") == ([], None, "routine")
+    assert b("Charlie, David, go ahead.", [], None, "routine", ch) == ([], None, "routine")
+
+
+def test_radio_codes_are_not_locations():
+    for junk in ("96 times 5", "96 x 1", "10-8", "Signal 99", "code 4"):
+        assert ic.clean_location(junk) is None, junk
+    for place in ("West Main Street", "Route 9", "96 Main Street", "Exit 17 southbound"):
+        assert ic.clean_location(place) == place, place

drb-c2-core/tests/test_gemini_helper.py

@@ -0,0 +1,70 @@
+"""
+app/internal/gemini.py — thinking level, fallback when a model rejects it,
+and token accounting into a replay's usage sink (server-26#170 cost finding).
+"""
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from app.internal import gemini
+
+
+class _FakeModels:
+    def __init__(self, reject_thinking=False):
+        self.reject_thinking = reject_thinking
+        self.configs = []
+
+    def generate_content(self, model, contents, config):
+        self.configs.append(config)
+        if self.reject_thinking and config.get("thinking_level"):
+            raise RuntimeError("400 INVALID_ARGUMENT: thinking_level is not supported for this model")
+        return SimpleNamespace(
+            text='{"action": "link"}',
+            usage_metadata=SimpleNamespace(prompt_token_count=1200, candidates_token_count=30,
+                                           thoughts_token_count=0),
+        )
+
+
+def _patched(models):
+    client = SimpleNamespace(models=models)
+    return (patch.object(gemini, "_get_client", return_value=client),
+            patch.object(gemini, "_config", lambda level: {"thinking_level": level}))
+
+
+def test_minimal_thinking_by_default_and_usage_lands_in_the_sink():
+    models = _FakeModels()
+    a, b = _patched(models)
+    sink = {}
+    tok = gemini.collect_usage(sink)
+    try:
+        with a, b:
+            assert gemini.generate_json("m1", "p", purpose="correlation") == {"action": "link"}
+    finally:
+        gemini.reset_usage(tok)
+    assert models.configs == [{"thinking_level": "minimal"}]
+    assert sink == {"correlation:m1": {"calls": 1, "in": 1200, "out": 30, "thinking": 0}}
+
+
+def test_model_that_rejects_thinking_level_falls_back_once():
+    models = _FakeModels(reject_thinking=True)
+    a, b = _patched(models)
+    gemini._no_thinking_level.discard("m2")
+    with a, b:
+        gemini.generate_json("m2", "p", purpose="correlation")
+        gemini.generate_json("m2", "p", purpose="correlation")
+    # first call: tried minimal, retried without; second call: straight without
+    assert models.configs == [{"thinking_level": "minimal"}, {"thinking_level": None}, {"thinking_level": None}]
+    gemini._no_thinking_level.discard("m2")
+
+
+def test_other_failures_still_raise_for_ai_health():
+    class Boom(_FakeModels):
+        def generate_content(self, **kw):
+            raise RuntimeError("429 insufficient_quota")
+    a, b = _patched(Boom())
+    with a, b:
+        try:
+            gemini.generate_json("m3", "p", purpose="correlation")
+        except RuntimeError as e:
+            assert "insufficient_quota" in str(e)
+        else:
+            raise AssertionError("should raise")

drb-frontend/components/admin/ReplayTab.tsx

@@ -361,6 +361,14 @@ function RunDetail({ run }: { run: ReplayRun }) {
           paths: {Object.entries(m.corr_path).map(([k, v]) => `${k} ${v}`).join(" · ")}
         </p>
       )}
+      {m?.gemini_usage && Object.keys(m.gemini_usage).length > 0 && (
+        <p className="text-xs font-mono text-gray-500">
+          Gemini tokens:{" "}
+          {Object.entries(m.gemini_usage)
+            .map(([k, u]) => `${k} ${u.calls} calls, in ${u.in.toLocaleString()} / out ${u.out.toLocaleString()} / thinking ${u.thinking.toLocaleString()}`)
+            .join(" · ")}
+        </p>
+      )}
       {m?.ai_failures && Object.keys(m.ai_failures).length > 0 && (
         <p className="text-xs font-mono text-amber-400">
           AI failures: {Object.entries(m.ai_failures).map(([k, v]) => `${k} ×${v}`).join(" · ")}

drb-frontend/lib/types.ts

@@ -340,6 +340,7 @@ export interface ReplayMetrics {
   llm_decisions: number;
   est_cost_usd: number;
   ai_failures?: Record<string, number>;
+  gemini_usage?: Record<string, { calls: number; in: number; out: number; thinking: number }>;
 }
 
 export interface ReplayRun {

infra/firestore/firestore.rules

@@ -8,13 +8,11 @@
 // hand-set in the Firebase console: unversioned, unreviewed, unknown. See
 // SAAS_PLAN.md B1.
 //
-// DEPLOY IS A MANUAL, OUT-OF-BAND STEP — nothing in CI or this codebase
+// DEPLOYED BY CI on every push to main (.gitea/workflows/deploy.yml, job
-// pushes these rules to Firebase:
+// deploy-firestore-rules, service-account auth via the FIREBASE_SA_KEY
-//   firebase deploy --only firestore:rules --project <project-id>
+// secret — server-26#51). That job is separate from the app deploy, so a
-// (from this directory, or point --config at infra/firestore/firebase.json
+// green app deploy does NOT mean these rules are live: check that job too.
-// from the repo root). Do this before or immediately after the code that
+// Editing rules in the Firebase console is overwritten by the next push.
-// starts stamping org_id ships — until these rules are live, the
-// console-configured rules are still what's actually enforced.
 //
 // MODEL: c2-core (firebase-admin SDK, server-side) bypasses these rules
 // entirely and is the sole writer for every collection below — that was