Describe the bug
The attempt_completion tool fails with a 400 error stating thinking is enabled but reasoning_content is missing in assistant tool call message at index 2. This occurs when an agent completes a task and tries to submit the result via attempt_completion. The error is reproducible across multiple retry attempts and appears to be a platform-level issue rather than a user-code issue. The underlying task work (file changes, commits) succeeds, but the final handoff via attempt_completion is blocked.
To Reproduce
Steps to reproduce the behavior:
- Complete a task that requires file edits and a final
attempt_completion call
- Call
attempt_completion with a valid result string
- Observe the tool returns
Verification Subagent Failed: 400 thinking is enabled but reasoning_content is missing in assistant tool call message at index 2
- Retry the same
attempt_completion call — the error persists
Expected behavior
attempt_completion should accept the result and complete the task handoff successfully, or provide a meaningful error related to the task content rather than an internal platform thinking/reasoning_content mismatch.
Screenshots
N/A — error is returned as a tool result message.
Surface and Version:
- IDE: VSCode with Dirac extension
- Dirac extension version: 0.3.29
- LLM API provider: Moonshot, api.moonthot.ai
- Model: kimi-k2.6
Additional context
- The error appears to originate from a verification subagent that inspects the assistant's tool call message.
- The
thinking is enabled part suggests the subagent expects a reasoning_content field in the assistant message when that mode is active, but the field is absent.
- The error index (
at index 2) may refer to the third message in the conversation context passed to the verification subagent.
- This blocks task completion even though all actual work (file writes, git commits) was completed successfully.
Describe the bug
The
attempt_completiontool fails with a400error statingthinking is enabled but reasoning_content is missing in assistant tool call message at index 2. This occurs when an agent completes a task and tries to submit the result viaattempt_completion. The error is reproducible across multiple retry attempts and appears to be a platform-level issue rather than a user-code issue. The underlying task work (file changes, commits) succeeds, but the final handoff viaattempt_completionis blocked.To Reproduce
Steps to reproduce the behavior:
attempt_completioncallattempt_completionwith a validresultstringVerification Subagent Failed: 400 thinking is enabled but reasoning_content is missing in assistant tool call message at index 2attempt_completioncall — the error persistsExpected behavior
attempt_completionshould accept the result and complete the task handoff successfully, or provide a meaningful error related to the task content rather than an internal platformthinking/reasoning_contentmismatch.Screenshots
N/A — error is returned as a tool result message.
Surface and Version:
Additional context
thinking is enabledpart suggests the subagent expects areasoning_contentfield in the assistant message when that mode is active, but the field is absent.at index 2) may refer to the third message in the conversation context passed to the verification subagent.