GPT-5.6 Terra repeatedly stops working

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

Version 26.831.21537

What subscription do you have?

Pro

What platform is your computer?

Darwin 25.3.0 arm64 arm

What issue are you seeing?

Description

GPT-5.6 Terra repeatedly stops working in the middle of a task even when it has been explicitly instructed not to respond until a clearly defined part of the task is completed.

This is happening consistently for me in Codex.

For example, I give Terra an explicit instruction such as:

Complete [specific part of the task] entirely before responding to me. Do not stop after intermediate steps.

Terra starts working on the task, usually works for around one minute, and then ends the turn even though the requested work is clearly not finished.

The strange part is that Terra itself acknowledges that the task is incomplete. The final response may explicitly say that there is still work remaining.

When I then tell it:

You were explicitly told to finish this before responding. Please complete it now.

Terra acknowledges the mistake, says that it should not have stopped and that it will continue until the task is complete.

However, it then works for another short period of time and stops again — sometimes after doing almost nothing — while the task is still incomplete.

Expected behavior

If the user gives a clear completion condition such as:

Do not respond until X is completed.

and Codex does not require additional information, approval, or clarification from the user, Terra should continue executing the task until X is actually completed.

Intermediate milestones should not cause the agent run to terminate.

Actual behavior

The current behavior looks roughly like this:

1. User defines a larger task.
2. User explicitly defines a completion condition.
3. Terra starts working.
4. Terra stops after approximately 1 minute at an intermediate point.
5. Terra acknowledges that the requested task is not finished.
6. User tells Terra to continue and reminds it of the completion condition.
7. Terra acknowledges that it stopped incorrectly and says it will finish the task.
8. Terra stops again shortly afterwards without reaching the completion condition.

This can repeat multiple times.

Why this appears to be a bug

This does not appear to be a misunderstanding of the prompt.

Terra correctly understands:

• what the remaining work is,
• that the task is not finished,
• that it does not require additional user input,
• and that it was explicitly instructed not to stop yet.

It will even explicitly acknowledge that terminating the previous turn was incorrect.

Despite this, the agent terminates the next turn again.

It therefore looks like there may be an issue with the agent's continuation / completion logic, where an intermediate milestone is incorrectly treated as a reason to end the turn even though the active goal has not been completed.

For longer development tasks this makes GPT-5.6 Terra extremely difficult to use, because I have to repeatedly tell the agent to continue even though there is no decision or input required from me.

Model

GPT-5.6 Terra

Product

Codex

Frequency

Very frequent / reproducible in my current longer-running tasks.

What steps can reproduce the bug?

The current behavior looks roughly like this:

1. User defines a larger task.
2. User explicitly defines a completion condition.
3. Terra starts working.
4. Terra stops after approximately 1 minute at an intermediate point.
5. Terra acknowledges that the requested task is not finished.
6. User tells Terra to continue and reminds it of the completion condition.
7. Terra acknowledges that it stopped incorrectly and says it will finish the task.
8. Terra stops again shortly afterwards without reaching the completion condition.

What is the expected behavior?

Expected behavior

If the user gives a clear completion condition such as:

Do not respond until X is completed.

and Codex does not require additional information, approval, or clarification from the user, Terra should continue executing the task until X is actually completed.

Intermediate milestones should not cause the agent run to terminate.

Additional information

No response

No source files, tests, or entry points are named. First reproduce the repeated early termination in Codex with GPT-5.6 Terra using a task with an explicit completion condition, then trace the continuation and completion logic. Done means the agent continues until the condition is met without unnecessary user prompts.