Bug: Invalid Value: 'input.call_id'. Function call output requires call_id. #43944

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

Version 26.901.51231

What subscription do you have?

Using Azure OAI where I hit this issue - also using a pro licence where the issue doesn't occur

What platform is your computer?

Macos

What issue are you seeing?

Sending a prompt from one Codex task to another can cause the destination task to fail with:

Invalid Value: 'input.call_id'. Function call output requires call_id.

After this occurs, every subsequent prompt in the destination task fails with the same error. Creating a new task and submitting the same prompt normally succeeds.

What steps can reproduce the bug?

Steps to reproduce

1. Create or open two Codex tasks: Task A and Task B.
2. In Task A, ask Codex to continue work in Task B.
3. Task A invokes the built-in send_message_to_thread tool.
4. Open or inspect Task B.
5. Task B fails before executing the requested work.
6. Send any additional prompt directly to Task B.

Actual result

Task B returns the following error for the delegated prompt and all subsequent prompts:

Invalid Value: 'input.call_id'. Function call output requires call_id.

Retrying inside Task B does not recover it. A newly created task works.

What is the expected behavior?

Expected result

The delegated prompt should appear in Task B as a normal incoming delegation/message and start a valid turn.

A failed delegation should not persist malformed tool state that prevents every subsequent turn from running.

Additional information

Suspected cause

The destination transcript appears to receive the cross-task delegation as a function_call_output without the originating function call or its call_id.

The invalid item is then retained in conversation history and replayed with every subsequent request.

Suggested remediation

• Store cross-task deliveries as delegation/user-input events, rather than orphaned function-call outputs.
• Preserve the original call_id and corresponding call when function-output semantics are required.
• Validate incoming transcript items before persisting them.
• Roll back or quarantine malformed items when request validation fails.
• Add recovery logic for tasks that already contain orphaned function outputs.
• Add a regression test covering Task A sending a message to an idle Task B.

Impact

The destination task becomes unusable and its working context is effectively stranded. Users must create a replacement task and manually transfer relevant context.

Workaround

Create a new task and submit the prompt directly rather than using cross-task messaging.