Skip to content

Kiro turns sometimes stop after a progress message #531

Closed

Description

@winehares

Client or integration

Codex App

Area

Provider adapter

English

Summary

Hi. I use Codex Desktop on Windows through OpenCodex 2.7.41 with kiro/gpt-5.6-sol.

During longer tasks, the agent sometimes posts a progress update and then stops, even though the message clearly says it is about to continue working. For example:

I haven't applied the fix yet. I'll update it now and run the focused tests.

No tool call follows. The turn is recorded as final_answer / task_complete, so I have to type “continue” to resume the unfinished task. This happens often enough that it regularly interrupts real work. I do not see the same behavior in the Claude app.

I have also seen internal continuation messages such as [system: conversation continues] and [system: The preceding assistant output did not explicitly complete the turn ...] appear directly in the conversation. It looks like adapter-owned state is being inserted as ordinary user-visible text, which then allows the model to read and echo it.

Steps to reproduce

  1. Run OpenCodex 2.7.41 on Windows and connect Codex Desktop through it.
  2. Use kiro/gpt-5.6-sol for a task that requires several tool calls.
  3. Continue until the assistant outputs only a progress message describing the next action.
  4. Observe that no tool call follows and the turn ends as final_answer / task_complete.
  5. Type “continue” and observe that the unfinished task resumes.

Expected behavior

  • Ordinary assistant text should remain progress/commentary while tools are available.
  • A real tool call should continue the turn.
  • The turn should end only when the dedicated final-answer path is used.
  • Internal continuation state should never be exposed as user-visible prose.

I do not think matching phrases such as “I'll fix it” is a reliable solution. The same failure occurs in Korean and could occur in any language. Completion should be determined from tool and protocol events, not from the wording of the message.

Environment

  • OpenCodex: 2.7.41
  • OS: Windows 11
  • Provider/model: Kiro / kiro/gpt-5.6-sol

한국어

요약

안녕하세요. Windows에서 OpenCodex 2.7.41을 통해 Codex Desktop을 사용하고 있으며, 모델은 kiro/gpt-5.6-sol입니다.

작업이 길어지면 에이전트가 진행 상황만 말한 뒤 그대로 멈추는 일이 반복됩니다. 예를 들면 이런 식입니다.

아직 수정 전입니다. 지금 바로 반영하고 집중 테스트까지 확인하겠습니다.

문장만 보면 곧바로 다음 작업을 이어갈 것처럼 보이지만, 실제 도구 호출은 발생하지 않습니다. 그런데도 해당 턴이 final_answer / task_complete로 종료되어, 미완료 작업을 다시 진행하려면 매번 제가 “계속”이라고 입력해야 합니다. 실제 작업 흐름을 자주 끊을 정도로 반복되고 있으며, Claude 앱에서는 같은 현상을 겪지 않았습니다.

또한 [system: conversation continues], [system: The preceding assistant output did not explicitly complete the turn ...] 같은 내부 continuation 문구가 대화창에 그대로 노출된 적도 있습니다. 어댑터 내부에서만 처리되어야 할 상태가 일반 메시지처럼 전달되고, 모델이 그 문구를 다시 읽어 출력하는 것으로 보입니다.

재현 방법

  1. Windows에서 OpenCodex 2.7.41을 실행하고 Codex Desktop을 연결합니다.
  2. kiro/gpt-5.6-sol로 여러 차례의 도구 호출이 필요한 작업을 요청합니다.
  3. 작업 도중 에이전트가 다음 행동을 예고하는 진행 상황만 출력할 때까지 진행합니다.
  4. 실제 도구 호출 없이 턴이 final_answer / task_complete로 끝나는지 확인합니다.
  5. “계속”이라고 입력하면 미완료 작업이 다시 시작됩니다.

기대 동작

  • 도구를 사용할 수 있는 상태라면 일반 텍스트는 진행 상황 또는 commentary로 처리되어야 합니다.
  • 실제 도구 호출이 있으면 같은 턴에서 작업이 이어져야 합니다.
  • 전용 final-answer 경로를 사용했을 때만 턴이 종료되어야 합니다.
  • 내부 continuation 상태가 사용자에게 보이는 문장으로 노출되면 안 됩니다.

“수정하겠습니다” 같은 특정 표현을 찾아 차단하는 방식은 언어마다 표현이 다르고 오탐 가능성도 있어 근본적인 해결책이 되기 어렵다고 생각합니다. 문장 내용이 아니라 도구 호출과 프로토콜 이벤트를 기준으로 완료 여부를 판단하는 편이 안전해 보입니다.

환경

  • OpenCodex: 2.7.41
  • 운영체제: Windows 11
  • 제공자/모델: Kiro / kiro/gpt-5.6-sol

Checks

  • I searched existing issues and documentation.
  • I removed secrets, tokens, account details, request credentials, and personal data.
Translated Message

Original language: Korean

Summary

Hello. I am using Codex Desktop on Windows through OpenCodex 2.7.41 with the model kiro/gpt-5.6-sol.

When tasks are lengthy, the agent sometimes only states progress and then simply stops. For example:

I haven't applied the fix yet. I'll update it now and run the focused tests.

The phrasing suggests that the next task will follow immediately, but no actual tool call occurs. Consequently, the turn ends as final_answer / task_complete, and I need to type "continue" every time to resume the unfinished task. This issue interrupts real work frequently enough. I haven't experienced this behavior in the Claude app.

I have also seen internal continuation messages like [system: conversation continues] and [system: The preceding assistant output did not explicitly complete the turn ...] appear directly in the conversation. It seems that the state meant for internal handling is being transmitted as ordinary user-visible text, allowing the model to read and output it.

Steps to reproduce

  1. Run OpenCodex 2.7.41 on Windows and connect Codex Desktop through it.
  2. Use kiro/gpt-5.6-sol for a task requiring several tool calls.
  3. Continue until the assistant outputs only a progress message describing the next action.
  4. Check that no tool call follows and that the turn ends as final_answer / task_complete.
  5. Type "continue" and observe that the unfinished task resumes.

Expected behavior

  • Ordinary assistant text should remain as progress/commentary while tools are available.
  • A genuine tool call should further the turn.
  • The turn should only end when the dedicated final-answer path is used.
  • Internal continuation state should never be exposed as user-visible prose.

I do not believe that matching phrases like "I'll fix it" is a reliable solution. The same failure occurs in Korean and could happen in any language. Completion should be determined based on tool and protocol events, not from the wording of messages.

Activity

  1. github-actions commented on Jul 27, 2026

    @github-actions
    Contributor
  2. changed the title [-]Kiro turns sometimes stop after a progress message / 진행 멘트 뒤 작업이 끝나는 문제[/-] [+]Kiro turns sometimes stop after a progress message[/+] on Jul 27, 2026
  3. Ingwannu commented on Jul 27, 2026

    @Ingwannu
    Owner

    Confirmed on current dev. The private completion tool is advertised, but the parser still promotes ordinary text to final_answer when terminal metadata says END_TURN or STOP_SEQUENCE; the adapter continuation prompts are also literal marker-shaped user messages.

    I am fixing this at the protocol boundary: on tool-enabled turns, ordinary text will remain commentary and END_TURN / STOP_SEQUENCE without codex_kiro_final_answer will enter the one bounded completion validation instead of completing directly. I will also remove the marker-shaped continuation prose and pin the behavior with stream and public-endpoint tests. The related empty retry-state defect is tracked in #534.

  4. Ingwannu commented on Jul 27, 2026

    @Ingwannu
    Owner

    Fixed on dev in e347916.

    Root cause: Kiro terminal metadata only proves that inference stopped; it does not reliably distinguish progress prose from a completed tool-enabled turn. The adapter was promoting ordinary text on END_TURN and STOP_SEQUENCE directly to final_answer.

    Tool-enabled Kiro turns now keep ordinary text as commentary and enter one bounded private-completion validation unless codex_kiro_final_answer was actually called. Marker-shaped continuation strings were replaced with natural adapter instructions, and an exact private-completion answer is always emitted in the final phase.

    Validated with stream and public-endpoint regressions, the focused Kiro suite (157 passing), the full repository suite (4,951 passing), typecheck, docs build, and privacy scan.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions