`collab_tool_call` "spawn_agent" item missing from `exec --experimental-json` stream; "wait" item's `receiver_thread_ids` stays empty despite a real, successful delegation underneath

What version of Codex CLI is running?

0.144.5 (installed via @openai/codex-sdk / @openai/codex npm packages)

Which model were you using?

gpt-5.6-sol

What platform is your computer?

Linux (Ubuntu 20.04 focal), invoked via @openai/codex-sdk's Codex.startThread().runStreamed(), which shells out to codex exec --experimental-json (see node_modules/@openai/codex-sdk/src/exec.ts).

What happened?

Using native subagent support ([features] multi_agent = true, a custom
agent defined at <CODEX_HOME>/agents/<name>.toml), I asked the root agent
to spawn a named custom agent and relay its exact reply. The
--experimental-json event stream only ever shows:

{"type":"item.started","item":{"id":"item_1","type":"collab_tool_call","tool":"wait","sender_thread_id":"...","receiver_thread_ids":[],"prompt":null,"agents_states":{},"status":"in_progress"}}
{"type":"item.completed","item":{"id":"item_1","type":"collab_tool_call","tool":"wait","sender_thread_id":"...","receiver_thread_ids":[],"prompt":null,"agents_states":{},"status":"completed"}}
{"type":"item.completed","item":{"id":"item_2","type":"agent_message","text":"SPECIALIST SAID: PONG"}}

Note: no collab_tool_call with tool: "spawn_agent" (or similar) ever
appears in this stream, and the wait item's receiver_thread_ids is
always [], even though the final agent_message content is correct.

This reproduced identically across 6 separate runs, with and without
[features.multi_agent_v2].

Reading the FULL session rollout (~/.codex/sessions/**/*.jsonl) for the
same turn tells a different, complete story — it contains the actual
function_call/function_call_output pair for the spawn, a
sub_agent_activity event ("kind": "started"), the child thread's own
rollout file (session_meta subagent.thread_spawn with the correct
agent_path/agent_nickname), and the real agent_message reply routed
back from the child. state_5.sqlite's thread_spawn_edges table also
correctly records the parent→child edge for every one of the 6 runs.

So: delegation genuinely works end-to-end. The bug is that the
--experimental-json stream — the one @openai/codex-sdk consumes and
exposes to SDK users — is missing the spawn_agent collab_tool_call event
entirely, and the wait item it does emit never reflects the real
receiver thread. This is different from #28318 (missing TypeScript type for
collab_tool_call — that reporter's runtime JSON did include a populated
spawn_agent item with real receiver_thread_ids); here the data itself
is missing from this event source, not just its type.

Expected behavior

exec --experimental-json (and therefore @openai/codex-sdk's public
event stream) should include the spawn_agent collab_tool_call item, and
wait's receiver_thread_ids should reflect the actual spawned child
thread(s) — matching what the full session rollout already records
correctly.

Why this matters

Any SDK consumer trying to observe or verify subagent delegation
(logging, tracing, debugging, deciding whether to trust a delegated
result) via the public event stream currently gets a false negative —
delegation looks like a silent no-op even when it succeeded. We only
caught this by cross-referencing the full rollout JSONL and
state_5.sqlite directly, which isn't part of the SDK's public surface.