[eric] agents: drop orphaned compaction comment block (code moved in refactor) + tighten a 15-line narrow-wrapped comment to 3

backend/apps/agents/agent_manager.py

@@ -50,27 +50,6 @@ class AgentManager(SessionLifecycle, SessionPersistence, Messaging, SessionContr
         self.cancel_events: Dict[str, asyncio.Event] = {}
 
 
-
-
-    # ------------------------------------------------------------------
-    # Compaction & token guard (Phase 2)
-    #
-    # Triggered by *live* context-usage ratio, not turn count. The signal
-    # is the same `ctx_used_pct` we already broadcast to the UI on every
-    # turn: input_tokens / context_window. Three escalating thresholds:
-    #   - compact_threshold_pct (default 0.65): summarize stale tool_results
-    #     and old user/assistant pairs before the next query() call
-    #   - context_soft_cap_pct (default 0.90): pre-send hard guard. After
-    #     compaction, if still over, LRU-trim active_mcps
-    #   - >= 1.0 hits the proxy/Anthropic 200K ceiling, friendly card
-    #     surfaces from the catch-all
-    # ------------------------------------------------------------------
-
-
-
-
-
-
     @typechecked
     async def run_agent_loop(self, session_id: str, prompt: str, images: Optional[List] = None, context_paths: Optional[List] = None, forced_tools: Optional[List[str]] = None, attached_skills: Optional[List] = None, fork_session: bool = False, selected_browser_ids: Optional[List[str]] = None, selected_app_output_ids: Optional[List[str]] = None, selected_setting_ids: Optional[List[str]] = None):
         """Run the Claude Agent SDK query loop for a session."""

backend/apps/agents/manager/streaming/handle_stream_event.py

@@ -80,21 +80,9 @@ async def handle_stream_event(
             tool_msg_id = uuid4().hex
             turn.stream_tool_msg_ids_ordered.append(tool_msg_id)
             turn.stream_block_index_map[index] = tool_msg_id
-            # Stream-level tool count for the
-            # consolidated thinking pill. The
-            # AssistantMessage path (further down)
-            # ALSO increments turn.tool_count when
-            # ToolUseBlocks fully arrive, but for
-            # OpenAI/Gemini through 9Router the
-            # AssistantMessage envelope is sometimes
-            # incomplete, so this stream-level count
-            # is what guarantees the "N tools used"
-            # segment renders cross-provider. To
-            # avoid double-counting we DON'T also
-            # increment on AssistantMessage when
-            # this code path already fired, see
-            # the dedupe at the AssistantMessage
-            # block below.
+            # Stream-level tool count for the thinking pill. OpenAI/Gemini-through-9Router
+            # AssistantMessage envelopes are sometimes incomplete, so this stream count guarantees
+            # "N tools used" renders cross-provider; the AssistantMessage path dedupes against it.
             turn.tool_count += 1
             await ws_manager.send_to_session(session_id, "agent:stream_start", {
                 "session_id": session_id,