Stabilize HITL bundle UX and resume.

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/resume.py

@@ -22,7 +22,12 @@ def hitlrequest_action_count(pending_value: Any) -> int:
 
 
 def fan_out_decisions_to_match(resume_value: Any, expected_count: int) -> Any:
-    """Pad a single-decision resume to N entries so an ``action_requests=N`` bundle accepts it."""
+    """Legacy fallback: pad a 1-decision resume to N for an ``action_requests=N`` bundle.
+
+    Modern frontend submits N decisions per bundle (one per action_request) so
+    this is a no-op; kept for backwards compatibility with old in-flight
+    threads or non-bundle clients that send a single decision.
+    """
     if expected_count <= 1:
         return resume_value
     if not isinstance(resume_value, dict):

surfsense_backend/app/agents/multi_agent_chat/main_agent/graph/middleware/checkpointed_subagent_middleware/task_tool.py

@@ -139,13 +139,6 @@ def build_task_tool_with_parent_config(
             if resume_value is not None:
                 expected = hitlrequest_action_count(pending_value)
                 resume_value = fan_out_decisions_to_match(resume_value, expected)
-                logger.info(
-                    "Forwarding surfsense_resume_value into subagent %r "
-                    "(action_requests=%d, targeted_id=%s)",
-                    subagent_type,
-                    expected,
-                    pending_id is not None,
-                )
                 result = subagent.invoke(
                     build_resume_command(resume_value, pending_id),
                     config=sub_config,
@@ -211,13 +204,6 @@ def build_task_tool_with_parent_config(
             if resume_value is not None:
                 expected = hitlrequest_action_count(pending_value)
                 resume_value = fan_out_decisions_to_match(resume_value, expected)
-                logger.info(
-                    "Forwarding surfsense_resume_value into subagent %r "
-                    "(action_requests=%d, targeted_id=%s)",
-                    subagent_type,
-                    expected,
-                    pending_id is not None,
-                )
                 result = await subagent.ainvoke(
                     build_resume_command(resume_value, pending_id),
                     config=sub_config,

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/markdown/main_agent_tool_routing.md

@@ -13,4 +13,15 @@ later in this system prompt, after citation rules).
 
 Do not treat live third-party state as if it were already in the indexed knowledge
 base; reach it via **task**.
+
+Never emit more than one **task** tool call in the same turn. Bundle related work
+for the same specialist into a single **task** invocation (the subagent itself can
+call its own tools in parallel inside that one run). Parallel **task** calls would
+fan out into multiple concurrent subagent runs whose human-approval interrupts
+cannot be coordinated; one **task** at a time is required.
 </tool_routing>
+
+<!-- TODO: lift the single-task constraint once the runtime supports parallel task
+interrupts end-to-end (multi-interrupt SSE + interrupt-id-keyed Command(resume)
++ keyed surfsense_resume_value side-channel). Until then this nudge is the only
+guard; the parent graph's resume cannot address multiple pending interrupts. -->