spawn_agent creates child thread but initial task is treated as context #20543

What version of Codex are you using?

codex-cli rust-v0.120.0-osso, built from a fork whose origin/main is a9c111da54 ([rollout_trace] Trace sessions and multi-agent edges (#18879)). The failing path is in the current MultiAgentV2 spawn_agent/InterAgentCommunication implementation.

What subscription do you have?

Pro / ChatGPT OAuth.

What platform is your computer?

Linux (Arch).

What issue are you seeing?

spawn_agent creates a child thread and reports success, but the child does not execute the assigned task. It starts a turn and immediately returns a generic final answer like:

Context loaded for `/home/osso/Repos/ask`. Ready for task.

The parent then receives that as the subagent completion notification, so wait_agent/close_agent report the agent completed without doing the requested work.

This is not the same as the model refusing to call spawn_agent: the tool is called, a child thread is created, and the child turn runs. The assigned task text appears in the child rollout as an inter-agent/developer-style message, but not as an actual user task the child follows.

What steps can reproduce the bug?

From a Codex session, ask the model to run a small subagent task, or call the tool directly with equivalent arguments:

{
  "agent_type": "explorer",
  "task_name": "subagent_exec_smoke",
  "message": "Run exactly this read-only command in /home/osso/Repos/ask and report the output: pwd. Then run: test -f AGENTS.md; echo AGENTS=$?. Do not modify files. Final answer must include both command outputs.",
  "fork_turns": "none",
  "reasoning_effort": "low"
}

Observed parent rollout:

{"type":"function_call","name":"spawn_agent","arguments":"{...}","call_id":"call_..."}
{"type":"function_call_output","call_id":"call_...","output":"{\"task_name\":\"/root/subagent_exec_smoke\",\"nickname\":\"Chandrasekhar\"}"}

Observed child rollout:

{"type":"event_msg","payload":{"type":"task_started", ...}}
{"type":"response_item","payload":{"type":"message","role":"developer","content":[{"type":"input_text","text":"... {\"recipient\":\"/root/subagent_exec_smoke\",\"content\":\"Run exactly this read-only command ...\",\"trigger_turn\":true}"}]}}
{"type":"event_msg","payload":{"type":"agent_message","message":"Context loaded for `/home/osso/Repos/ask`. Ready for task.","phase":"final_answer"}}
{"type":"event_msg","payload":{"type":"task_complete","last_agent_message":"Context loaded for `/home/osso/Repos/ask`. Ready for task.", ...}}

What is the expected behavior?

The spawned agent should treat the message passed to spawn_agent as the task for its first turn and execute it. In the reproduction above, it should run/report pwd and the AGENTS check, or at least answer based on the assigned task instead of only acknowledging context initialization.

Additional technical notes

Current code path appears to be:

• core/src/tools/handlers/multi_agents_v2/spawn.rs converts the initial message into Op::InterAgentCommunication for thread-spawn subagents.
• core/src/agent/control.rs calls send_input(new_thread.thread_id, initial_operation) after creating the thread.
• core/src/session/handlers.rs handles Op::InterAgentCommunication by enqueueing mailbox communication and calling maybe_start_turn_for_pending_work_with_sub_id when trigger_turn is true.
• InterAgentCommunication::to_response_input_item serializes the communication as an assistant-role message containing JSON.

The child turn does start, but the model-visible prompt shape seems to be interpreted as context rather than a user/task instruction. A direct user turn or a clearer task wrapper for the initial spawn message may be needed.