feat: marathon session limits — cap, checkpoint, rotate (#326)

- Add max_messages (default 200) to SessionResetPolicy
- Track message_count in SessionEntry (persisted to sessions.json)
- Add 'message_limit' reset reason to _should_reset
- Auto-checkpoint filesystem before session rotation
- Inject near-limit warnings (85%/100%) into agent ephemeral prompt
- Auto-rotate sessions when message cap is hit
- Add get_message_limit_info() and reset_message_count() APIs
- 24 new tests covering all limit behaviors

Evidence: 170 sessions exceed 100 msgs, longest 1,643 msgs (40h).
Marathon sessions show 45-84% error rates from tool fixation.
Cap + checkpoint + restart breaks the death spiral.

gateway/config.py

@@ -107,6 +107,7 @@ class SessionResetPolicy:
     mode: str = "both"  # "daily", "idle", "both", or "none"
     at_hour: int = 4  # Hour for daily reset (0-23, local time)
     idle_minutes: int = 1440  # Minutes of inactivity before reset (24 hours)
+    max_messages: int = 200  # Max messages per session before forced checkpoint+restart (0 = unlimited)
     notify: bool = True  # Send a notification to the user when auto-reset occurs
     notify_exclude_platforms: tuple = ("api_server", "webhook")  # Platforms that don't get reset notifications
     
@@ -115,6 +116,7 @@ class SessionResetPolicy:
             "mode": self.mode,
             "at_hour": self.at_hour,
             "idle_minutes": self.idle_minutes,
+            "max_messages": self.max_messages,
             "notify": self.notify,
             "notify_exclude_platforms": list(self.notify_exclude_platforms),
         }
@@ -125,12 +127,14 @@ class SessionResetPolicy:
         mode = data.get("mode")
         at_hour = data.get("at_hour")
         idle_minutes = data.get("idle_minutes")
+        max_messages = data.get("max_messages")
         notify = data.get("notify")
         exclude = data.get("notify_exclude_platforms")
         return cls(
             mode=mode if mode is not None else "both",
             at_hour=at_hour if at_hour is not None else 4,
             idle_minutes=idle_minutes if idle_minutes is not None else 1440,
+            max_messages=max_messages if max_messages is not None else 200,
             notify=notify if notify is not None else True,
             notify_exclude_platforms=tuple(exclude) if exclude is not None else ("api_server", "webhook"),
         )

gateway/run.py

@@ -2343,6 +2343,12 @@ class GatewayRunner:
             reset_reason = getattr(session_entry, 'auto_reset_reason', None) or 'idle'
             if reset_reason == "daily":
                 context_note = "[System note: The user's session was automatically reset by the daily schedule. This is a fresh conversation with no prior context.]"
+            elif reset_reason == "message_limit":
+                context_note = (
+                    "[System note: The user's previous session reached the message limit "
+                    "and was automatically checkpointed and rotated. This is a fresh session. "
+                    "If the user references something from before, you can search session history.]"
+                )
             else:
                 context_note = "[System note: The user's previous session expired due to inactivity. This is a fresh conversation with no prior context.]"
             context_prompt = context_note + "\n\n" + context_prompt
@@ -2368,16 +2374,18 @@ class GatewayRunner:
                     if adapter:
                         if reset_reason == "daily":
                             reason_text = f"daily schedule at {policy.at_hour}:00"
+                        elif reset_reason == "message_limit":
+                            reason_text = f"reached {policy.max_messages} message limit"
                         else:
                             hours = policy.idle_minutes // 60
                             mins = policy.idle_minutes % 60
                             duration = f"{hours}h" if not mins else f"{hours}h {mins}m" if hours else f"{mins}m"
                             reason_text = f"inactive for {duration}"
                         notice = (
-                            f"◐ Session automatically reset ({reason_text}). "
-                            f"Conversation history cleared.\n"
+                            f"◐ Session automatically rotated ({reason_text}). "
+                            f"Conversation was preserved via checkpoint.\n"
                             f"Use /resume to browse and restore a previous session.\n"
-                            f"Adjust reset timing in config.yaml under session_reset."
+                            f"Adjust limits in config.yaml under session_reset."
                         )
                         try:
                             session_info = self._format_session_info()
@@ -3073,6 +3081,39 @@ class GatewayRunner:
                 last_prompt_tokens=agent_result.get("last_prompt_tokens", 0),
             )
 
+            # Marathon session limit (#326): check if we hit the message cap.
+            # Auto-checkpoint filesystem and rotate session.
+            try:
+                _post_limit = self.session_store.get_message_limit_info(session_key)
+                if _post_limit["at_limit"] and _post_limit["max_messages"] > 0:
+                    logger.info(
+                        "[Marathon] Session %s hit message limit (%d/%d). Rotating.",
+                        session_key, _post_limit["message_count"], _post_limit["max_messages"],
+                    )
+                    # Attempt filesystem checkpoint before rotation
+                    try:
+                        from tools.checkpoint_manager import CheckpointManager
+                        _cp_cfg_path = _hermes_home / "config.yaml"
+                        if _cp_cfg_path.exists():
+                            import yaml as _cp_yaml
+                            with open(_cp_cfg_path, encoding="utf-8") as _cpf:
+                                _cp_data = _cp_yaml.safe_load(_cpf) or {}
+                            _cp_settings = _cp_data.get("checkpoints", {})
+                            if _cp_settings.get("enabled"):
+                                _cwd = _cp_settings.get("working_dir") or os.getcwd()
+                                mgr = CheckpointManager(max_checkpoints=_cp_settings.get("max_checkpoints", 20))
+                                cp = mgr.create_checkpoint(str(_cwd), label=f"marathon-{session_entry.session_id[:8]}")
+                                if cp:
+                                    logger.info("[Marathon] Checkpoint: %s", cp.label)
+                    except Exception as cp_err:
+                        logger.debug("[Marathon] Checkpoint failed (non-fatal): %s", cp_err)
+
+                    new_entry = self.session_store.reset_session(session_key)
+                    if new_entry:
+                        logger.info("[Marathon] Rotated: %s -> %s", session_entry.session_id, new_entry.session_id)
+            except Exception as rot_err:
+                logger.debug("[Marathon] Rotation check failed: %s", rot_err)
+
             # Auto voice reply: send TTS audio before the text response
             _already_sent = bool(agent_result.get("already_sent"))
             if self._should_send_voice_reply(event, response, agent_messages, already_sent=_already_sent):
@@ -6538,6 +6579,26 @@ class GatewayRunner:
             if self._ephemeral_system_prompt:
                 combined_ephemeral = (combined_ephemeral + "\n\n" + self._ephemeral_system_prompt).strip()
 
+            # Marathon session limit warning (#326)
+            try:
+                _limit_info = self.session_store.get_message_limit_info(session_key)
+                if _limit_info["near_limit"] and not _limit_info["at_limit"]:
+                    _remaining = _limit_info["remaining"]
+                    _limit_warn = (
+                        f"[SESSION LIMIT: This session has {_limit_info['message_count']} messages. "
+                        f"Only {_remaining} message(s) remain before automatic session rotation at "
+                        f"{_limit_info['max_messages']} messages. Start wrapping up and save important state.]"
+                    )
+                    combined_ephemeral = (combined_ephemeral + "\n\n" + _limit_warn).strip()
+                elif _limit_info["at_limit"]:
+                    _limit_warn = (
+                        f"[SESSION LIMIT REACHED: This session has hit the {_limit_info['max_messages']} "
+                        f"message limit. This is your FINAL response. Summarize accomplishments and next steps.]"
+                    )
+                    combined_ephemeral = (combined_ephemeral + "\n\n" + _limit_warn).strip()
+            except Exception:
+                pass
+
             # Re-read .env and config for fresh credentials (gateway is long-lived,
             # keys may change without restart).
             try:

gateway/session.py

@@ -383,7 +383,11 @@ class SessionEntry:
     # survives gateway restarts (the old in-memory _pre_flushed_sessions
     # set was lost on restart, causing redundant re-flushes).
     memory_flushed: bool = False
-    
+
+    # Marathon session limit tracking (#326).
+    # Counts total messages (user + assistant + tool) in this session.
+    message_count: int = 0
+
     def to_dict(self) -> Dict[str, Any]:
         result = {
             "session_key": self.session_key,
@@ -402,6 +406,7 @@ class SessionEntry:
             "estimated_cost_usd": self.estimated_cost_usd,
             "cost_status": self.cost_status,
             "memory_flushed": self.memory_flushed,
+            "message_count": self.message_count,
         }
         if self.origin:
             result["origin"] = self.origin.to_dict()
@@ -438,6 +443,7 @@ class SessionEntry:
             estimated_cost_usd=data.get("estimated_cost_usd", 0.0),
             cost_status=data.get("cost_status", "unknown"),
             memory_flushed=data.get("memory_flushed", False),
+            message_count=data.get("message_count", 0),
         )
 
 
@@ -643,6 +649,9 @@ class SessionStore:
         )
         
         if policy.mode == "none":
+            # Even with mode=none, enforce message_limit if set
+            if policy.max_messages > 0 and entry.message_count >= policy.max_messages:
+                return "message_limit"
             return None
         
         now = _now()
@@ -664,7 +673,11 @@ class SessionStore:
             
             if entry.updated_at < today_reset:
                 return "daily"
-        
+
+        # Marathon session limit (#326): force checkpoint+restart at max_messages
+        if policy.max_messages > 0 and entry.message_count >= policy.max_messages:
+            return "message_limit"
+
         return None
     
     def has_any_sessions(self) -> bool:
@@ -822,6 +835,43 @@ class SessionStore:
                     entry.last_prompt_tokens = last_prompt_tokens
                 self._save()
 
+    def get_message_limit_info(self, session_key: str) -> Dict[str, Any]:
+        """Get message count and limit info for a session (#326)."""
+        with self._lock:
+            self._ensure_loaded_locked()
+            entry = self._entries.get(session_key)
+
+        if not entry:
+            return {"message_count": 0, "max_messages": 0, "remaining": 0,
+                    "near_limit": False, "at_limit": False, "threshold": 0.0}
+
+        policy = self.config.get_reset_policy(
+            platform=entry.platform,
+            session_type=entry.chat_type,
+        )
+        max_msgs = policy.max_messages
+        count = entry.message_count
+        remaining = max(0, max_msgs - count) if max_msgs > 0 else float("inf")
+        threshold = count / max_msgs if max_msgs > 0 else 0.0
+
+        return {
+            "message_count": count,
+            "max_messages": max_msgs,
+            "remaining": remaining,
+            "near_limit": max_msgs > 0 and count >= int(max_msgs * 0.85),
+            "at_limit": max_msgs > 0 and count >= max_msgs,
+            "threshold": threshold,
+        }
+
+    def reset_message_count(self, session_key: str) -> None:
+        """Reset the message count to zero for a session (#326)."""
+        with self._lock:
+            self._ensure_loaded_locked()
+            entry = self._entries.get(session_key)
+            if entry:
+                entry.message_count = 0
+                self._save()
+
     def reset_session(self, session_key: str) -> Optional[SessionEntry]:
         """Force reset a session, creating a new session ID."""
         db_end_session_id = None
@@ -849,6 +899,7 @@ class SessionStore:
                 display_name=old_entry.display_name,
                 platform=old_entry.platform,
                 chat_type=old_entry.chat_type,
+                message_count=0,  # Fresh count after rotation (#326)
             )
 
             self._entries[session_key] = new_entry
@@ -908,6 +959,7 @@ class SessionStore:
                 display_name=old_entry.display_name,
                 platform=old_entry.platform,
                 chat_type=old_entry.chat_type,
+                message_count=0,  # Fresh count after rotation (#326)
             )
 
             self._entries[session_key] = new_entry
@@ -966,6 +1018,16 @@ class SessionStore:
         transcript_path = self.get_transcript_path(session_id)
         with open(transcript_path, "a", encoding="utf-8") as f:
             f.write(json.dumps(message, ensure_ascii=False) + "\n")
+
+        # Increment message count for marathon session tracking (#326)
+        # Skip counting session_meta entries (tool defs, metadata)
+        if message.get("role") != "session_meta":
+            with self._lock:
+                for entry in self._entries.values():
+                    if entry.session_id == session_id:
+                        entry.message_count += 1
+                        self._save()
+                        break
     
     def rewrite_transcript(self, session_id: str, messages: List[Dict[str, Any]]) -> None:
         """Replace the entire transcript for a session with new messages.

tests/gateway/test_marathon_session_limits.py

@@ -0,0 +1,184 @@
+"""Tests for marathon session limits (#326)."""
+
+import pytest
+from datetime import datetime
+from pathlib import Path
+from tempfile import mkdtemp
+
+from gateway.config import GatewayConfig, Platform, SessionResetPolicy
+from gateway.session import SessionEntry, SessionSource, SessionStore
+
+
+def _source(platform=Platform.LOCAL, chat_id="test"):
+    return SessionSource(platform=platform, chat_id=chat_id, chat_type="dm", user_id="u1")
+
+
+def _store(max_messages=200, mode="both"):
+    cfg = GatewayConfig()
+    cfg.default_reset_policy = SessionResetPolicy(mode=mode, max_messages=max_messages)
+    return SessionStore(Path(mkdtemp()), cfg)
+
+
+class TestSessionResetPolicyMaxMessages:
+    def test_default(self):
+        assert SessionResetPolicy().max_messages == 200
+
+    def test_custom(self):
+        assert SessionResetPolicy(max_messages=500).max_messages == 500
+
+    def test_unlimited(self):
+        assert SessionResetPolicy(max_messages=0).max_messages == 0
+
+    def test_to_dict(self):
+        d = SessionResetPolicy(max_messages=300).to_dict()
+        assert d["max_messages"] == 300
+
+    def test_from_dict(self):
+        p = SessionResetPolicy.from_dict({"max_messages": 150})
+        assert p.max_messages == 150
+
+    def test_from_dict_default(self):
+        assert SessionResetPolicy.from_dict({}).max_messages == 200
+
+
+class TestSessionEntryMessageCount:
+    def test_default(self):
+        e = SessionEntry(session_key="k", session_id="s", created_at=datetime.now(), updated_at=datetime.now())
+        assert e.message_count == 0
+
+    def test_to_dict(self):
+        e = SessionEntry(session_key="k", session_id="s", created_at=datetime.now(), updated_at=datetime.now(), message_count=42)
+        assert e.to_dict()["message_count"] == 42
+
+    def test_from_dict(self):
+        e = SessionEntry.from_dict({"session_key": "k", "session_id": "s", "created_at": "2026-01-01T00:00:00", "updated_at": "2026-01-01T00:00:00", "message_count": 99})
+        assert e.message_count == 99
+
+
+class TestShouldResetMessageLimit:
+    def test_at_limit(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 200
+        assert s._should_reset(e, src) == "message_limit"
+
+    def test_over_limit(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 250
+        assert s._should_reset(e, src) == "message_limit"
+
+    def test_below_limit(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 100
+        assert s._should_reset(e, src) is None
+
+    def test_unlimited(self):
+        s = _store(max_messages=0, mode="none")
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 9999
+        assert s._should_reset(e, src) is None
+
+    def test_custom_limit(self):
+        s = _store(max_messages=50)
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 50
+        assert s._should_reset(e, src) == "message_limit"
+
+    def test_just_under(self):
+        s = _store(max_messages=50)
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 49
+        assert s._should_reset(e, src) is None
+
+
+class TestAppendIncrementsCount:
+    def test_user_message(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        s.append_to_transcript(e.session_id, {"role": "user", "content": "hi"})
+        e = s.get_or_create_session(src)
+        assert e.message_count == 1
+
+    def test_assistant_message(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        s.append_to_transcript(e.session_id, {"role": "user", "content": "hi"})
+        s.append_to_transcript(e.session_id, {"role": "assistant", "content": "hello"})
+        e = s.get_or_create_session(src)
+        assert e.message_count == 2
+
+    def test_meta_not_counted(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        s.append_to_transcript(e.session_id, {"role": "session_meta", "tools": []})
+        e = s.get_or_create_session(src)
+        assert e.message_count == 0
+
+
+class TestGetMessageLimitInfo:
+    def test_at_limit(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 200
+        info = s.get_message_limit_info(e.session_key)
+        assert info["at_limit"] is True
+        assert info["near_limit"] is True
+        assert info["remaining"] == 0
+
+    def test_near_limit(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 180
+        info = s.get_message_limit_info(e.session_key)
+        assert info["near_limit"] is True
+        assert info["at_limit"] is False
+        assert info["remaining"] == 20
+
+    def test_well_below(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 50
+        info = s.get_message_limit_info(e.session_key)
+        assert info["near_limit"] is False
+        assert info["at_limit"] is False
+
+    def test_unknown(self):
+        s = _store()
+        info = s.get_message_limit_info("nonexistent")
+        assert info["at_limit"] is False
+
+
+class TestResetMessageCount:
+    def test_reset(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 150
+        s.reset_message_count(e.session_key)
+        assert s.get_message_limit_info(e.session_key)["message_count"] == 0
+
+
+class TestSessionRotation:
+    def test_fresh_count_after_reset(self):
+        s = _store()
+        src = _source()
+        e = s.get_or_create_session(src)
+        e.message_count = 200
+        new = s.reset_session(e.session_key)
+        assert new is not None
+        assert new.message_count == 0
+        assert new.session_id != e.session_id