diff --git a/backend/apps/agents/agent_manager.py b/backend/apps/agents/agent_manager.py
index 2a817d32..1979f746 100644
--- a/backend/apps/agents/agent_manager.py
+++ b/backend/apps/agents/agent_manager.py
@@ -3882,6 +3882,8 @@ class AgentManager:
                         f"<request>\n{user_prompt[:2000]}\n</request>"
                     ),
                 }],
+                # Binds this aux call to its query's free-trial run; ignored off the free lane.
+                extra_headers={"X-Openswarm-Task-Id": session_id},
             ) as stream:
                 async for text in stream.text_stream:
                     chunks.append(text)
@@ -3936,6 +3938,8 @@ class AgentManager:
                 max_tokens=1,
                 system="You are a helpful assistant. Reply with one character.",
                 messages=[{"role": "user", "content": "ping"}],
+                # Binds the cache-warm ping to its query's free-trial run; ignored off the free lane.
+                extra_headers={"X-Openswarm-Task-Id": session_id},
             )
             logger.debug(f"Cache pre-warm fired for session {session_id}")
         except Exception as e:
@@ -4008,6 +4012,8 @@ class AgentManager:
                 max_tokens=aux_max_tokens_for(aux_model, base=300),
                 system=system,
                 messages=[{"role": "user", "content": user_content}],
+                # Binds this aux call to its query's free-trial run; ignored off the free lane.
+                extra_headers={"X-Openswarm-Task-Id": session_id},
             ) as stream:
                 async for text in stream.text_stream:
                     chunks.append(text)
diff --git a/backend/apps/agents/agents.py b/backend/apps/agents/agents.py
index 1cab53f8..cec4a8c6 100644
--- a/backend/apps/agents/agents.py
+++ b/backend/apps/agents/agents.py
@@ -80,7 +80,7 @@ async def send_message(session_id: str, body: dict):
 
         async def _emit_preflight():
             try:
-                result = await run_preflight(prompt)
+                result = await run_preflight(prompt, task_id=session_id)
                 if result.get("suggestions") or result.get("is_vague"):
                     await _ws.send_to_session(session_id, "agent:mcp_suggestions", {
                         "session_id": session_id,
diff --git a/backend/apps/agents/core/mcp_preflight.py b/backend/apps/agents/core/mcp_preflight.py
index 9de5ca70..fad2c739 100644
--- a/backend/apps/agents/core/mcp_preflight.py
+++ b/backend/apps/agents/core/mcp_preflight.py
@@ -85,7 +85,7 @@ def _is_obviously_local(prompt: str) -> bool:
     return False
 
 
-async def run_preflight(prompt: str, timeout_s: float = 2.0) -> dict:
+async def run_preflight(prompt: str, timeout_s: float = 2.0, task_id: str | None = None) -> dict:
     """Classify the prompt and return {is_vague, suggestions}; never raises."""
     default: dict[str, Any] = {"is_vague": False, "suggestions": []}
 
@@ -100,7 +100,7 @@ async def run_preflight(prompt: str, timeout_s: float = 2.0) -> dict:
         available = _build_available_shortlist(settings)
 
         result = await asyncio.wait_for(
-            _call_classifier(settings, prompt, available),
+            _call_classifier(settings, prompt, available, task_id),
             timeout=timeout_s,
         )
         # Re-validate ids against the curated shortlist so hallucinations can't reach the frontend.
@@ -151,7 +151,7 @@ def _decorate(llm_suggestion: dict, available: list[CuratedEntry]) -> dict | Non
     }
 
 
-async def _call_classifier(settings, prompt: str, available: list[CuratedEntry]) -> dict:
+async def _call_classifier(settings, prompt: str, available: list[CuratedEntry], task_id: str | None = None) -> dict:
     """One aux-model call, returns validated JSON {is_vague, suggestions}."""
     aux_model, _base = await resolve_aux_model(settings, preferred_tier="haiku")
     client = get_anthropic_client_for_model(settings, aux_model)
@@ -191,6 +191,8 @@ async def _call_classifier(settings, prompt: str, available: list[CuratedEntry])
         max_tokens=300,
         system=system,
         messages=[{"role": "user", "content": user_turn}],
+        # Rides on its query's free-trial run instead of opening its own; ignored off the free lane.
+        extra_headers={"X-Openswarm-Task-Id": task_id} if task_id else {},
     )
 
     text = ""