`codex exec --ephemeral` fails to spawn a custom subagent with `no thread with id`; non-ephemeral succeeds after global agent placement

Summary

I am seeing a custom-subagent spawn failure in Codex CLI on Windows. The failure is deterministic in --ephemeral mode and reports:

ERROR codex_core::tools::router: error=collab spawn failed: no thread with id: <root-session-id>

The message resembles the thread-lifecycle symptom described in #35781, but my reproduction uses a user-requested custom agent rather than a memory consolidation agent.

Environment

• Codex CLI: codex-cli 0.150.0-alpha.8
• OS: Windows 10 Home, 10.0.19045, x64
• Model: gpt-5.6-terra
• Command mode: codex exec --strict-config --skip-git-repo-check
• The diagnostic prompt forbids networking, external services, secret access, writes, and deletion.

Reproduction

1. Define chief-secretary.toml and researcher.toml with name, description, developer_instructions, and sandbox_mode = "read-only".
2. Ask the root session to delegate once to chief-secretary, which is instructed to delegate once to researcher.
3. Run the command in each mode below.

Project-scoped agents only (<project>/.codex/agents/)

• Normal mode: failed previously with collab spawn failed: no thread with id: <root-session-id>.
• --ephemeral: failed previously with the same error.

Global agents also present (~/.codex/agents/)

• Normal mode: completed with final output responded; no no thread with id router error appeared. The CLI output did not expose a child-agent identity, so this only confirms the requested diagnostic completed without the router error.
• --ephemeral: failed with the same error, before the delegated agent started.

Example failing invocation:

codex exec --strict-config --skip-git-repo-check --ephemeral -C <project> "<read-only prompt requesting chief-secretary -> researcher delegation>"

Expected behavior

The custom subagent should start in both normal and ephemeral modes, or ephemeral mode should reject the workflow before it exposes a spawn tool that cannot resolve the caller thread.

Actual behavior

Non-ephemeral behavior changes when the agent definition is globally available, but ephemeral mode still fails with a lookup for its own root session ID.

Request

Could this be related to the thread-registration or teardown path described in #35781? If not, please advise which redacted diagnostic fields or rollout details would help isolate the custom-agent/ephemeral path.

Start by reproducing the custom-agent delegation with the failing codex exec --ephemeral invocation on Windows, then compare it with normal mode and the project-scoped versus global agent setups. Read the thread-registration and teardown discussion in #35781. Done means ephemeral custom-agent spawning succeeds, or the workflow is rejected before exposing an unusable spawn path.