You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
gpt-5.3-codex intermittently leaks malformed tool-call text and stops mid-run #15164
When using openai/gpt-5.3-codex in OpenCode, the assistant intermittently emits malformed internal tool-call text into the user-facing response and appears to stop mid-task. The same workflow is stable with openai/gpt-5.2-codex.
Environment
OpenCode: 1.2.14
OS: macOS 26.2 (Build 25C56)
Node: v25.6.1
Model: openai/gpt-5.3-codex
What I expected
The assistant should finish the task and return only the final requested output.
Internal tool-call payloads should never appear in the answer text.
What actually happens
During multi-step fetch/extract tasks, the response sometimes contains internal-looking tool call text (example below), then appears to stop:
... `config.json`에서 `reasoningEffort` 키 위치를 좁혀서 허용값(enum)만 확인하겠습니다. +#+#+#+#+#+assistant to=functions.grep մեկնաբանություն ...
{"pattern":"reasoningEffort","path":"/Users/.../tool-output",...}
In affected runs, the conversation shows repeated finish=tool-calls then a premature finish=stop with incomplete intermediate text.
This was observed both before and after removing MCP servers from config.
Reproduction (example)
Set model to openai/gpt-5.3-codex.
Run:
opencode run "다음 3개 공식 문서 URL만 사용해서 reasoningEffort 허용 값만 추출해줘. 1) https://opencode.ai/docs/models/ 2) https://opencode.ai/config.json 3) https://platform.openai.com/docs/guides/reasoning 규칙: - 위 3개 URL 외 다른 출처는 사용하지 마. - 중간 과정, 내부 실행 로그, 도구 호출 문자열은 출력하지 마. - 최종 결과만 출력해."
Observe occasional malformed tool-call text leakage + apparent mid-run stop.
Control / workaround
Switching only the model to openai/gpt-5.2-codex removes the issue in my tests.
Forcing a single bash-only execution path also reduces occurrence.
Notes
I can provide sanitized session exports and exact timestamps if helpful.
If your issue is distinct (e.g. the malformed tool-call text leakage is a separate symptom from just early termination), please clarify the difference so maintainers can track them separately.
Activity
github-actions commentedon Feb 26, 2026
This issue might be a duplicate of existing issues. Please check:
If your issue is distinct (e.g. the malformed tool-call text leakage is a separate symptom from just early termination), please clarify the difference so maintainers can track them separately.
tibrnui commentedon Feb 26, 2026
Any update here? I have the same issue
PandelisZ commentedon Mar 7, 2026
Happens on gpt-5.4 too via Azure
github-actions commentedon Jul 4, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.