fix: capture crewai token usage from CrewOutput.token_usage

CrewAI 1.6.x routes through litellm and silently discards LangChain
callbacks when converting a passed ChatAnthropic / ChatOpenAI into
CrewAI's native LLM (see crewai/agent/core.py:create_llm — it copies
only model/temperature/api_key, never callbacks). As a result the
framework's TokenUsageCallback never fired, AgentDecision.tokens_used
was always 0, and token_cost_estimate was always $0.0000.

Add BaseAgent._extract_tokens helper that reads the canonical source —
CrewOutput.token_usage.total_tokens — populated by Crew.calculate_usage_metrics
after every kickoff(). Wire it into all 6 LLM-driven agents
(CCA, Logic, Performance, Documentation, TestCoverage, RevisionProposer)
and drop the no-op callbacks=[self.token_callback] arg + dead
reset/get_token_count calls.

TokenUsageCallback itself is left in place as harmless dead code; a
follow-up commit can remove it together with the constructor wiring.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: 2 people

agents/base.py

@@ -119,6 +119,22 @@ def _get_token_count(self) -> int:
         """Get total tokens used from callback."""
         return self.token_callback.total_tokens
 
+    @staticmethod
+    def _extract_tokens(crew_result: Any) -> int:
+        """Read aggregated tokens from a CrewAI Crew.kickoff() result.
+
+        CrewAI 1.6.x routes through litellm and silently drops langchain
+        callbacks during Agent LLM conversion (see crewai/agent/core.py:
+        create_llm), so the framework's TokenUsageCallback never fires.
+        CrewOutput.token_usage (UsageMetrics) is the reliable source —
+        total_tokens aggregates input + output across every LLM call
+        inside the Crew.
+        """
+        token_usage = getattr(crew_result, "token_usage", None)
+        if token_usage is not None:
+            return int(getattr(token_usage, "total_tokens", 0))
+        return 0
+
     def _validate_findings(self, findings: List[Finding]) -> List[Finding]:
         """Validate findings meet evidence requirements."""
         valid = []

agents/change_context_analyst.py

@@ -76,21 +76,10 @@ def _build_analysis_context(self, context: PRContext, prompt_template: str) -> s
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task and extract results."""
         try:
-            self._reset_token_tracking()
-            # Pass callback to Crew to track token usage
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            # Fallback: if callback didn't capture tokens, try to get from result
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {
                 "reasoning": output[:500] or "Analysis completed",
                 "tokens": tokens,

agents/documentation_reviewer.py

@@ -79,19 +79,10 @@ def _build_analysis_context(self, context: PRContext, prompt_template: str) -> s
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task."""
         try:
-            self._reset_token_tracking()
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {"reasoning": output[:500], "tokens": tokens, "raw_output": output}
         except Exception as e:
             logger.error(f"Error executing CrewAI task: {e}")

agents/logic_reviewer.py

@@ -76,19 +76,10 @@ def _build_analysis_context(self, context: PRContext, prompt_template: str) -> s
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task and extract results."""
         try:
-            self._reset_token_tracking()
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {
                 "reasoning": output[:500] or "Analysis completed",
                 "tokens": tokens,

agents/performance_reviewer.py

@@ -164,19 +164,10 @@ def _build_analysis_context(self, context: PRContext, prompt_template: str) -> s
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task."""
         try:
-            self._reset_token_tracking()
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {"reasoning": output[:500], "tokens": tokens, "raw_output": output}
         except Exception as e:
             logger.error(f"Error executing CrewAI task: {e}")

agents/revision_proposer.py

@@ -128,19 +128,10 @@ def _generate_single_patch(
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task."""
         try:
-            self._reset_token_tracking()
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {"reasoning": output[:500], "tokens": tokens, "raw_output": output}
         except Exception as e:
             logger.error(f"Error executing CrewAI task: {e}")

agents/test_reviewer.py

@@ -210,19 +210,10 @@ def _build_analysis_context(self, context: PRContext, prompt_template: str) -> s
     def _execute_task(self, agent: Agent, task: Task) -> dict:
         """Execute CrewAI task."""
         try:
-            self._reset_token_tracking()
-            crew = Crew(
-                agents=[agent],
-                tasks=[task],
-                verbose=False,
-                callbacks=[self.token_callback]
-            )
+            crew = Crew(agents=[agent], tasks=[task], verbose=False)
             result = crew.kickoff()
             output = str(result) if result else ""
-            tokens = self._get_token_count()
-            if tokens == 0 and hasattr(result, 'usage_metadata'):
-                if hasattr(result.usage_metadata, 'total_tokens'):
-                    tokens = result.usage_metadata.total_tokens
+            tokens = self._extract_tokens(result)
             return {"reasoning": output[:500], "tokens": tokens, "raw_output": output}
         except Exception as e:
             logger.error(f"Error executing CrewAI task: {e}")