perf: cap marathon sessions at 200 messages with forced compression (closes #326)

Add max_messages limit (default 200, configurable via HERMES_MAX_MESSAGES
env var). When conversation message count exceeds the cap, force a
compression pass that creates a new session lineage.

Prevents error cascading and tool fixation in marathon sessions (200+
messages) which empirically show 45-84% error rates.

Uses existing _compress_context mechanism — agent continues seamlessly
in a fresh session with compressed history.

run_agent.py

@@ -586,6 +586,9 @@ class AIAgent:
 
         self.model = model
         self.max_iterations = max_iterations
+        # Marathon session cap: force compression when messages exceed this.
+        # Eliminates error-cascading in 200+ message sessions (#326).
+        self.max_messages = int(os.environ.get("HERMES_MAX_MESSAGES", "200"))
         # Shared iteration budget — parent creates, children inherit.
         # Consumed by every LLM turn across parent + all subagents.
         self.iteration_budget = iteration_budget or IterationBudget(max_iterations)
@@ -7328,6 +7331,24 @@ class AIAgent:
                 pass
 
         while api_call_count < self.max_iterations and self.iteration_budget.remaining > 0:
+            # Marathon session guard: if message count exceeds max_messages,
+            # force a compression pass to start a fresh session lineage.
+            # Prevents error cascading and tool fixation in long sessions. #326
+            if (
+                self.compression_enabled
+                and len(messages) > self.max_messages
+            ):
+                if not self.quiet_mode:
+                    self._safe_print(
+                        f"\n🔄 Marathon session limit hit ({len(messages)} msgs > "
+                        f"{self.max_messages} cap). Compressing and continuing..."
+                    )
+                messages, active_system_prompt = self._compress_context(
+                    messages, system_message,
+                    approx_tokens=estimate_messages_tokens_rough(messages),
+                    task_id=effective_task_id,
+                )
+            
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
             self._checkpoint_mgr.new_turn()