[eric] fix: autocompact now actually trims the model's history instead of just showing a marker

backend/apps/agents/agent_manager.py

@@ -1807,7 +1807,10 @@ class AgentManager:
                 if session.needs_fork:
                     session.needs_fork = False
             elif len(session.messages) > 1:
-                history = _build_history_prefix(_get_branch_messages(session))
+                history = _build_history_prefix(
+                    _get_branch_messages(session),
+                    cutoff_msg_id=session.compacted_through_msg_id,
+                )
                 if history:
                     if isinstance(prompt_content, str):
                         prompt_content = history + "\n\n" + prompt_content

backend/apps/agents/manager/session/history_compaction.py

@@ -48,8 +48,17 @@ def _get_branch_messages(session) -> list:
     return result
 
 
-def _build_history_prefix(messages) -> str:
-    """Format branch messages into a conversation summary for context injection."""
+def _build_history_prefix(messages, cutoff_msg_id: str | None = None) -> str:
+    """Format branch messages into a conversation summary for context injection.
+
+    When `cutoff_msg_id` is provided (session.compacted_through_msg_id), drop every
+    message up to and including that id so the marker the UI shows actually matches
+    what the model sees. Missing cutoff id falls through to full history.
+    """
+    if cutoff_msg_id:
+        skip_idx = next((i for i, m in enumerate(messages) if m.id == cutoff_msg_id), -1)
+        if skip_idx >= 0:
+            messages = messages[skip_idx + 1:]
     lines = []
     for m in messages:
         if m.role not in ("user", "assistant") or getattr(m, "hidden", False):

backend/main.py

@@ -716,9 +716,10 @@ async def mcp_meta(action: str, request: Request):
 async def session_compact(session_id: str):
     """Force a compaction pass on a session (Phase 2 /compact slash cmd).
 
-    Cheap programmatic summarization (no aux LLM call), so it's safe to
-    invoke at any time. Sets needs_fork=True so the next turn rebuilds
-    options and ships the compacted prefix.
+    User explicitly clicked compact, so we accept the prompt-cache loss in exchange
+    for a real visible trim: needs_fresh_session drops the SDK convo so the next turn
+    rebuilds from history with compacted_through_msg_id actually applied (auto-compact
+    only sets the marker; the button is the user opting into the cost).
     """
     from backend.apps.agents.agent_manager import agent_manager
     from backend.apps.agents.core.ws_manager import ws_manager as _ws
@@ -726,7 +727,8 @@ async def session_compact(session_id: str):
     if not session:
         return JSONResponse({"error": "session not found"}, status_code=404)
     did_compact = agent_manager._maybe_compact(session, force=True)
-    session.needs_fork = True
+    if did_compact:
+        session.needs_fresh_session = True
     await _ws.send_to_session(session_id, "agent:context_status", {
         "session_id": session_id,
         "reason": "compacted_manual" if did_compact else "noop",