multi_agent_chat/main_agent: one specialist per task; advertise write_todos for multi-turn plans

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/prompts/routing.md

@@ -9,27 +9,36 @@ simulate one with the other.
   the workspace KB).
 - `scrape_webpage` — fetch the body of a specific public URL.
 - `update_memory` — curate persistent memory (see `<memory_protocol>`).
+- `write_todos` — maintain a structured plan when the turn series spans
+  multiple specialists or steps. Mark each item
+  `in_progress` **before** the `task` call that handles it, `completed`
+  once the call returns. Skip for single-step requests.
 
 **You have NO filesystem tools.** Any read, write, edit, move, rename, or
 search inside the user's workspace goes through `task(knowledge_base, …)` —
 never via `write_file`, `ls`, or any direct file operation.
 
 ### 2. `task(<specialist>, …)` — specialist subagents
-Use `task` for anything beyond the four direct tools above. See
+Use `task` for anything beyond the direct tools above. See
 `<specialists>` for the live roster.
 
 Rules for `task`:
-- **One `task` call per turn.** Bundle related work for the same specialist
-  into a single invocation — the parent graph can't coordinate human
-  approvals across parallel subagents.
+- **One specialist per `task` call.** A single `task` invocation must
+  describe work that one specialist can do end-to-end. Never bundle work
+  for two specialists into one task prompt — the specialist you route to
+  will silently drop the other half.
+- **One `task` call per turn.** If the user's request spans multiple
+  specialists, handle them one at a time across consecutive turns: invoke
+  the first this turn, return, then invoke the next on your next turn (no
+  user input required between). Use `write_todos` to keep the plan alive
+  across those turns.
+- Within a single specialist, bundle every related step into the same task
+  prompt (read + write + summary go together).
 - Put the **full instructions inside the task prompt** — the specialist
   cannot see this thread.
 - Don't claim to already know what a specialist's source contains; invoke
   the specialist and use what it returns.
 
-Parallelism applies to **direct tool calls** (e.g. two `web_search` calls
-for independent queries can go in parallel). It does **not** apply to `task`.
-
 <example>
 user: "Save these meeting notes to my KB: …"
 → task(knowledge_base, "Save the meeting notes below to a new document
@@ -54,4 +63,22 @@ user: "Find my Q2 roadmap and summarise the milestones."
   and summarise its milestones. Use glob or grep if the path isn't
   obvious from the workspace tree.")
 </example>
+
+<example>
+user: "Create a ClickUp ticket and a Linear ticket for the new feature flag."
+→ This turn:
+    write_todos([
+      {content: "Create ClickUp ticket for feature flag rollout", status: "in_progress"},
+      {content: "Create Linear ticket for feature flag rollout",  status: "pending"},
+    ])
+    task(clickup, "Create a ClickUp ticket titled 'Feature flag rollout'
+      in the default list. Description: <…>. Tell me the ticket URL.")
+→ Next turn:
+    write_todos([
+      {content: "Create ClickUp ticket for feature flag rollout", status: "completed"},
+      {content: "Create Linear ticket for feature flag rollout",  status: "in_progress"},
+    ])
+    task(linear, "Create a Linear ticket titled 'Feature flag rollout'
+      in the default team. Description: <…>. Tell me the ticket URL.")
+</example>
 </routing>

surfsense_backend/app/agents/multi_agent_chat/main_agent/system_prompt/prompts/tools/task/description.md

@@ -11,9 +11,5 @@
       thread, so include all context and constraints, plus what you need
       back. The specialist will respond in its own format — don't dictate
       one.
-  - Rules:
-    - One `task` call per turn. Bundle related work for the same specialist
-      into one invocation; the parent graph cannot coordinate human
-      approvals across parallel subagents.
-    - Don't claim to already know what a specialist's source contains;
-      invoke it and use what it returns.
+  - Routing rules (when to call, how often, how to scope) live in
+    `<routing>`.

surfsense_backend/app/agents/multi_agent_chat/middleware/shared/todos.py

@@ -5,5 +5,8 @@ from __future__ import annotations
 from langchain.agents.middleware import TodoListMiddleware
 
 
-def build_todos_mw() -> TodoListMiddleware:
-    return TodoListMiddleware()
+def build_todos_mw(*, system_prompt: str | None = None) -> TodoListMiddleware:
+    """Pass ``system_prompt=""`` to suppress the upstream prompt append. We use a custom system prompt in the main agent."""
+    if system_prompt is None:
+        return TodoListMiddleware()
+    return TodoListMiddleware(system_prompt=system_prompt)

surfsense_backend/app/agents/multi_agent_chat/middleware/stack.py

@@ -132,7 +132,7 @@ def build_main_agent_deepagent_middleware(
     stack: list[Any] = [
         build_busy_mutex_mw(flags),
         build_otel_mw(flags),
-        build_todos_mw(),
+        build_todos_mw(system_prompt=""),
         memory_mw,
         build_anonymous_doc_mw(
             filesystem_mode=filesystem_mode, anon_session_id=anon_session_id