perf: take per-API-call token accounting off the turn thread

agent/codex_runtime.py

@@ -74,7 +74,10 @@ def _record_codex_app_server_usage(agent, turn) -> dict[str, Any]:
             try:
                 if not agent._session_db_created:
                     agent._ensure_db_session()
-                agent._session_db.update_token_counts(
+                # Enqueued for the SessionDB background writer — keeps the
+                # per-call accounting write off the turn thread (see
+                # conversation_loop's queue_token_counts call).
+                agent._session_db.queue_token_counts(
                     agent.session_id,
                     model=agent.model,
                     billing_provider=agent.provider,
@@ -154,7 +157,8 @@ def _record_codex_app_server_usage(agent, turn) -> dict[str, Any]:
         try:
             if not agent._session_db_created:
                 agent._ensure_db_session()
-            agent._session_db.update_token_counts(
+            # Enqueued for the SessionDB background writer (see above).
+            agent._session_db.queue_token_counts(
                 agent.session_id,
                 input_tokens=canonical_usage.input_tokens,
                 output_tokens=canonical_usage.output_tokens,

agent/conversation_loop.py

@@ -2998,7 +2998,12 @@ def _perform_api_call(next_api_kwargs):
                                     _cost_delta = (_cost_delta or 0.0) + float(_moa_ref_cost)
                                 except (TypeError, ValueError):  # pragma: no cover
                                     pass
-                            agent._session_db.update_token_counts(
+                            # Enqueued, not written: the background writer
+                            # applies the delta off the turn thread (a cold
+                            # state.db UPDATE here stalled the tool loop for
+                            # up to hundreds of ms per API call). Drained at
+                            # turn finalize via _persist_session.
+                            agent._session_db.queue_token_counts(
                                 agent.session_id,
                                 input_tokens=canonical_usage.input_tokens,
                                 output_tokens=canonical_usage.output_tokens,

agent/insights.py

@@ -113,6 +113,13 @@ def generate(self, days: int = 30, source: str = None) -> Dict[str, Any]:
         """
         cutoff = time.time() - (days * 86400)
 
+        # Token/cost totals may still sit on the SessionDB's async
+        # accounting queue; drain so the report reflects exact counters.
+        # (self.db may be a raw sqlite3 connection in tests — guard.)
+        flush = getattr(self.db, "flush_token_counts", None)
+        if callable(flush):
+            flush()
+
         # Gather raw data
         sessions = self._get_sessions(cutoff, source)
         tool_usage = self._get_tool_usage(cutoff, source)