fix(memory): archive short idle sessions for Dream

nanobot/agent/autocompact.py

@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Any, Callable, Coroutine, cast
 
 from loguru import logger
 
-from nanobot.session.manager import Session, SessionManager
+from nanobot.session.manager import MIN_COMPACTED_REPLAY_MESSAGES, Session, SessionManager
 
 if TYPE_CHECKING:
     from nanobot.agent.memory import Consolidator
@@ -16,7 +16,7 @@ if TYPE_CHECKING:
 
 
 class AutoCompact:
-    _RECENT_SUFFIX_MESSAGES = 8
+    _RECENT_SUFFIX_MESSAGES = MIN_COMPACTED_REPLAY_MESSAGES
     _INTERNAL_SESSION_PREFIXES = ("dream:",)
 
     def __init__(self, sessions: SessionManager, consolidator: Consolidator,
@@ -45,25 +45,9 @@ class AutoCompact:
             return False
         return idle_seconds >= self._ttl * 60
 
-    def _has_compactable_idle_tail(self, key: str) -> bool:
+    def _has_unarchived_messages(self, key: str) -> bool:
         session = self.sessions.get_or_create(key)
-        tail = list(session.messages[session.last_consolidated:])
-        if not tail:
-            return False
-        probe = Session(
-            key=session.key,
-            messages=tail,
-            created_at=session.created_at,
-            updated_at=session.updated_at,
-            metadata={},
-            last_consolidated=0,
-        )
-        result = probe.retain_recent_legal_suffix(
-            self._RECENT_SUFFIX_MESSAGES,
-            extend_to_user=True,
-        )
-        messages_to_remove = result.dropped[result.already_consolidated_count:]
-        return bool(messages_to_remove)
+        return session.last_consolidated < len(session.messages)
 
     @staticmethod
     def _format_summary(text: str, last_active: datetime) -> str:
@@ -88,7 +72,7 @@ class AutoCompact:
             if key in active_session_keys:
                 continue
             updated_at = info.get("updated_at")
-            if self._is_expired(updated_at, now) and self._has_compactable_idle_tail(key):
+            if self._is_expired(updated_at, now) and self._has_unarchived_messages(key):
                 session = self.sessions.get_or_create(key)
                 try:
                     runtime = resolve_runtime(session)

nanobot/agent/memory.py

@@ -21,7 +21,7 @@ from typing import TYPE_CHECKING, Any, Callable, Iterator, cast
 from loguru import logger
 
 from nanobot.runtime_context import public_history_messages
-from nanobot.session.manager import Session, SessionManager
+from nanobot.session.manager import MIN_COMPACTED_REPLAY_MESSAGES, Session, SessionManager
 from nanobot.utils.gitstore import GitStore
 from nanobot.utils.helpers import (
     content_with_media_breadcrumbs,
@@ -858,14 +858,13 @@ class Consolidator:
         return last_boundary
 
     @staticmethod
-    def _full_unconsolidated_history(
+    def _full_replay_history(
         session: Session,
     ) -> list[dict[str, Any]]:
-        """Return the whole unconsolidated tail for consolidation decisions."""
-        unconsolidated_count = len(session.messages) - session.last_consolidated
-        if unconsolidated_count <= 0:
+        """Return all messages that can reach the next model prompt."""
+        if not session.messages:
             return []
-        return session.get_history(max_messages=unconsolidated_count)
+        return session.get_history(max_messages=len(session.messages))
 
     @staticmethod
     def _replay_overflow_boundary(
@@ -948,8 +947,8 @@ class Consolidator:
         *,
         runtime: LLMRuntime,
     ) -> tuple[int, str]:
-        """Estimate prompt size from the full unconsolidated session tail."""
-        history = self._full_unconsolidated_history(session)
+        """Estimate prompt size from the full replayable session history."""
+        history = self._full_replay_history(session)
         channel = session.key.split(":", 1)[0] if ":" in session.key else None
         # Include archived summary in estimation so the budget accounts for it.
         meta = session.metadata.get("_last_summary")
@@ -1160,42 +1159,37 @@ class Consolidator:
         session_key: str,
         *,
         runtime: LLMRuntime,
-        max_suffix: int = 8,
+        max_suffix: int = MIN_COMPACTED_REPLAY_MESSAGES,
     ) -> str | None:
-        """Archive an idle prefix and hide it from replay without deleting it."""
+        """Archive the full idle tail while keeping recent messages replayable.
+
+        ``max_suffix`` remains accepted for SDK compatibility. Replay retention
+        is now derived independently from archive progress using the project-wide
+        compacted-session window.
+        """
+        if max_suffix != MIN_COMPACTED_REPLAY_MESSAGES:
+            logger.debug(
+                "Idle-session compact for {} uses the fixed replay window ({}, requested {})",
+                session_key,
+                MIN_COMPACTED_REPLAY_MESSAGES,
+                max_suffix,
+            )
         lock = self.get_lock(session_key)
         async with lock:
             self.sessions.invalidate(session_key)
             session = self.sessions.get_or_create(session_key)
 
-            messages_to_summarize = list(session.messages[session.last_consolidated:])
-            if not messages_to_summarize:
-                self.sessions.save(session)
-                return ""
-
-            probe = Session(
-                key=session.key,
-                messages=messages_to_summarize.copy(),
-                created_at=session.created_at,
-                updated_at=session.updated_at,
-                metadata={},
-                last_consolidated=0,
-            )
-            result = probe.retain_recent_legal_suffix(max_suffix, extend_to_user=True)
-            visible_suffix = probe.messages
-            messages_to_remove = result.dropped
-
-            if not messages_to_remove:
-                self.sessions.save(session)
+            archive_start = session.last_consolidated
+            messages_to_archive = list(session.messages[archive_start:])
+            if not messages_to_archive:
                 return ""
 
             last_active = session.updated_at
-            # The visible suffix informs the summary but stays out of raw fallback.
+            archive_end = archive_start + len(messages_to_archive)
             summary = await self.archive(
-                messages_to_remove,
+                messages_to_archive,
                 runtime=runtime,
                 session_key=session_key,
-                summary_messages=messages_to_summarize,
             )
 
             if summary and summary != "(nothing)":
@@ -1204,16 +1198,22 @@ class Consolidator:
                     "last_active": last_active.isoformat(),
                 }
 
-            # Preserve history and advance only the replay boundary.
-            session.last_consolidated = len(session.messages) - len(visible_suffix)
+            # A turn can append while the provider call is in flight. Advance only
+            # through the captured batch so new messages remain eligible next time.
+            session.last_consolidated = archive_end
             session.provider_state = None
             self.sessions.save(session)
 
+            visible = session.get_history(
+                max_messages=MIN_COMPACTED_REPLAY_MESSAGES,
+                extend_to_user=True,
+            )
+
             logger.info(
                 "Idle-session compact for {}: archived={}, visible={}, retained={}, summary={}",
                 session_key,
-                len(messages_to_remove),
-                len(visible_suffix),
+                len(messages_to_archive),
+                len(visible),
                 len(session.messages),
                 bool(summary),
             )

nanobot/session/manager.py

@@ -36,6 +36,7 @@ from nanobot.utils.subagent_channel_display import scrub_subagent_announce_body
 FILE_MAX_MESSAGES = 2000
 SESSION_CACHE_MAX_SIZE = 128
 MIN_REPLAY_MAX_MESSAGES = 120
+MIN_COMPACTED_REPLAY_MESSAGES = 8
 REPLAY_TOKENS_PER_MESSAGE = 100
 _MESSAGE_TIME_PREFIX_RE = re.compile(r"^\[Message Time: [^\]]+\]\n?")
 _LOCAL_IMAGE_BREADCRUMB_RE = re.compile(r"^\[image: (?:/|~)[^\]]+\]\s*$")
@@ -191,19 +192,37 @@ class Session:
         extend_to_user: bool = False,
         include_runtime_context: bool = True,
     ) -> list[dict[str, Any]]:
-        """Return unconsolidated messages for LLM input.
+        """Return recent replayable messages for LLM input.
 
         History is sliced by message count first (``max_messages``), then by
         token budget from the tail (``max_tokens``) when provided.
         """
-        unconsolidated = self.messages[self.last_consolidated:]
+        replay_start = self.last_consolidated
+        if replay_start:
+            # ``last_consolidated`` is archive progress, not a replay boundary.
+            # Keep a small raw suffix for continuity, extending back to the user
+            # that started an assistant/tool sequence when necessary.
+            recent_start = recent_message_start_index(
+                self.messages,
+                MIN_COMPACTED_REPLAY_MESSAGES,
+                extend_to_user=True,
+            )
+            replay_start = min(replay_start, recent_start)
+
+        replayable = self.messages[replay_start:]
         max_messages = max_messages if max_messages > 0 else FILE_MAX_MESSAGES
-        start_idx = recent_message_start_index(
-            unconsolidated,
-            max_messages,
-            extend_to_user=extend_to_user,
-        )
-        sliced = unconsolidated[start_idx:]
+        unarchived_count = len(self.messages) - self.last_consolidated
+        if replay_start < self.last_consolidated and unarchived_count < max_messages:
+            # The archived replay suffix can exceed the nominal count when one
+            # tool-heavy turn spans the boundary. Preserve that complete turn.
+            start_idx = 0
+        else:
+            start_idx = recent_message_start_index(
+                replayable,
+                max_messages,
+                extend_to_user=extend_to_user,
+            )
+        sliced = replayable[start_idx:]
 
         # Avoid starting mid-turn when possible, except for proactive
         # assistant deliveries that the user may be replying to.