Tests: Extended Stage 7 benchmarks with latency + token metrics after tuning

benchmarks/plan_generation_benchmark.py

@@ -8,6 +8,8 @@ from __future__ import annotations
 
 from pathlib import Path
 
+import tiktoken
+
 from cleveragents.agents.graphs.plan_generation import PlanGenerationGraph
 from cleveragents.domain.models.core import (
     Context,
@@ -79,9 +81,25 @@ class PlanGenerationSuite:
         self.plan = _sample_plan()
         self.contexts = _sample_contexts()
 
+        enc = tiktoken.get_encoding("cl100k_base")
+        self.prompt_tokens = {
+            "analyze": len(enc.encode(self.graph.analyze_prompt.template)),
+            "generate": len(enc.encode(self.graph.generate_prompt.template)),
+            "validate": len(enc.encode(self.graph.validate_prompt.template)),
+        }
+
     def time_invoke(self) -> None:
         self.graph.invoke(self.project, self.plan, self.contexts, thread_id="bench")
 
     def time_stream(self) -> None:
         for _ in self.graph.stream(self.project, self.plan, self.contexts):
             pass
+
+    def track_analyze_prompt_tokens(self) -> int:
+        return self.prompt_tokens["analyze"]
+
+    def track_generate_prompt_tokens(self) -> int:
+        return self.prompt_tokens["generate"]
+
+    def track_validate_prompt_tokens(self) -> int:
+        return self.prompt_tokens["validate"]