Add context_chars to ActionStep and Step log line

- Add  field to  dataclass
  in memory.py — the natural carrier alongside token_usage and timing.

- In  (monitoring.py), compute the context char
  count once from , store it on
  , and append it to the Step log line:
    [Step N: Duration X.XX seconds| Input tokens: ... | Context: N chars]

- In  (agents.py), prefer the already-stored
   value (survives model_input_messages cleanup)
  before falling back to re-computing from messages.

Zero code duplication: one calculation site, one storage location,
two consumers (Step log + Turn summary line).

Co-Authored-By: bpsa2 <241537330+bpsa2@users.noreply.github.com>

Author: joaopauloschuler

src/smolagents/agents.py

@@ -843,8 +843,11 @@ def get_context_char_size(self) -> int:
         """Return the character count of the last context sent to the LLM."""
         from smolagents.memory import ActionStep, count_messages_chars
         for step in reversed(self.memory.steps):
-            if isinstance(step, ActionStep) and step.model_input_messages:
-                return count_messages_chars(step.model_input_messages)
+            if isinstance(step, ActionStep):
+                if step.context_chars is not None:
+                    return step.context_chars  # already computed by Monitor.update_metrics
+                if step.model_input_messages:
+                    return count_messages_chars(step.model_input_messages)
         return 0
 
     def cleanup_model_input_messages(self, keep_last: int = 2):

src/smolagents/memory.py

@@ -74,6 +74,7 @@ class ActionStep(MemoryStep):
     observations_images: list["PIL.Image.Image"] | None = None
     action_output: Any = None
     token_usage: TokenUsage | None = None
+    context_chars: int | None = None
     is_final_answer: bool = False
     actionstep_id: int | None = None
     _archived_observations: str | None = None

src/smolagents/monitoring.py

@@ -113,6 +113,14 @@ def update_metrics(self, step_log):
             console_outputs += (
                 f"| Input tokens: {self.total_input_token_count:,} | Output tokens: {self.total_output_token_count:,}"
             )
+
+        # Compute context length once and store on the step for later use
+        if hasattr(step_log, "model_input_messages") and step_log.model_input_messages:
+            from smolagents.memory import count_messages_chars  # lazy to avoid circular import
+            ctx_chars = count_messages_chars(step_log.model_input_messages)
+            step_log.context_chars = ctx_chars
+            console_outputs += f"| Context: {ctx_chars:,} chars"
+
         console_outputs += "]"
         self.logger.log(Text(console_outputs, style="dim"), level=1)