The agent yields a final or idle state while the active request remains incomplete and at least one relevant next step is available within its authority. It does not identify a genuine blocker, requested pause, or completed stopping condition.
Establish the unfinished acceptance criteria and a concrete next action the configured system could perform. Distinguish an actual ended turn from a progress message, active background work, or the expected wait for a tool result.
A content-free continuation prompt that resumes the same task is useful corroboration: it supplies no missing task information or new authorization. It is not proof of cause, because the prompt may also restart a failed scheduler or restore a tool connection.
The agent need not falsely claim completion. An accurate partial-progress report followed by an unnecessary stop can fit. Completion–artifact mismatch instead concerns what the agent claims about the delivered result.
A necessary approval, credential, choice, or unavailable input genuinely requires the user.
A background operation remains active and the system is correctly waiting for its result.
The user requested only the completed stage, a time or step budget has been reached, or an actual runtime limit prevents continuation.
Record the active request, completed and remaining work, available next step, final message, tool/job states, stop reason, and runtime limits. Preserve handoff or compaction events and the exact follow-up that resumed work. Distinguish model termination from a transport, scheduler, or user interruption.
Synthetic example: asked to implement and document a change, the agent finishes implementation and ends with an offer to write the documentation. No input is missing. A user message saying continue causes it to write the already-requested documentation.
Reported examples: Codex issue #33398 and its commenters describe unfinished tasks stopping after context transitions and resuming after short user nudges. The reports leave model and orchestration causes unresolved.
Synthetic comparison: the next step would publish to one of two accounts, and the user has not specified which. Stopping to obtain that choice is a real dependency.