Desktop app child role intermittently fails with `stream disconnected before completion` during project-local multi-agent runs

What version of the Codex App are you using (From “About Codex” dialog)?

Codex desktop app using codex-cli 0.120.0

What subscription do you have?

ChatGPT subscription with Codex desktop access

What platform is your computer?

macOS (workspace local, trusted project)

What issue are you seeing?

When using project-local child roles for multi-agent work in a trusted workspace, a child role can fail with:

stream disconnected before completion

In one reproduced case, the child role error returned:

stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID ff2696ef-f755-487c-ba09-90bb7940846b in your message.

This happened while using project-local child roles configured under .codex/config.toml and .codex/agents/*.toml for a bounded orchestration flow.

The important distinction is that the overall role bridge appears to be working: the parent thread can spawn child roles and start delegated work. The failure is that a child role can disconnect mid-run even when the task packet is small and scoped.

What steps can reproduce the bug?

1. Open a trusted local workspace in Codex desktop.
2. Enable multi-agent project-local roles under .codex/config.toml.
3. Configure project-local child roles (for example: planning, implementation, review, or docs roles) under .codex/agents/*.toml.
4. Run a bounded multi-agent task that causes the parent thread to spawn a child role.
5. Observe that the child role may fail with stream disconnected before completion even though the parent thread and project config are otherwise valid.

What is the expected behavior?

A spawned child role should either:

• complete normally, or
• fail with a more specific runtime error that makes clear whether the problem is transport, sandbox inheritance, auth, or path permissions.

It should not silently degrade into a generic stream disconnect during normal multi-agent orchestration.

Additional information

Relevant nearby issues that may be related but are not identical:

• #15305 — review-mode subagent ignores runtime sandbox override and falls back to config defaults
• #17535 — workspace-write cannot edit .agents/skills/** inside trusted workspace, but danger-full-access can

From local testing, the parent thread can successfully start project-local child roles, so this does not look like a simple configuration failure.

It appears more like a child-role runtime/transport stability issue in desktop multi-agent execution.

If useful, I can provide a sanitized example of the project-local role setup and task packet shape used in the reproduction.