Reasoning models fail every turn: tools + reasoning_effort rejected on /v1/chat/completions #58

Activity

1. DevJoaoLopes commented on Oct 6, 2026

   

   DevJoaoLopes

   on Oct 6, 2026

   More actions

   I can confirm the same failure when running OpenDots locally on macOS with npm run dev, using OPENAI_MODEL=gpt-6.1-sol.

   The local checkout is on main at 565bf781d654339ee1ce83b17ee00d76d679608e (after the originally reported c2569bb). Both Vite and the application server start successfully; the failure occurs when a Dot executes a model turn.

   The runtime reports:

   400 Function tools with reasoning_effort are not supported for gpt-6.1-sol in /v1/chat/completions. To use function tools, use /v1/responses or set reasoning_effort to 'none'.
   

   The error originates from openai-compatible.chatStream and propagates through convertTanStackStream in the CopilotKit runtime. The current src/server/dot-agent.ts still explicitly selects api: 'chat-completions'.

   This provides another reproduction outside Docker, on the newer main revision. Switching models, setting reasoning_effort to none, and using the Responses API have not yet been tested in this environment.

   Posted with AI assistance on behalf of the user who reported this local failure.
2. DevJoaoLopes commented on Oct 6, 2026

   

   DevJoaoLopes

   on Oct 6, 2026

   More actions

   Update: I have prepared two linked contributions for this issue:

   • CopilotKit runtime fix: CopilotKit/CopilotKit#7672 preserves TanStack encrypted reasoning events, correctly associates step/message IDs, and replays reasoning/tool metadata across human review and history continuations.
   • OpenDots integration (Draft): #104 adds explicit DOT_MODEL_API, DOT_MAX_OUTPUT_TOKENS, and optional DOT_REASONING_EFFORT settings, with native options for Chat Completions and Responses. Chat Completions/2200 and the provider's unspecified effort remain the defaults.

   Validation

   Using the reviewed, locally built runtime package on macOS/Node 24, I verified gpt-6.1-sol with:

   DOT_MODEL_API=responses
   DOT_MAX_OUTPUT_TOKENS=8192
   DOT_REASONING_EFFORT=medium

   The final UI flow made 9 model requests returning HTTP 200, with no HTTP 400. It covered actual page-tool execution, approval and decline of review cards, and reopening the conversation in a fresh browser context. The provider reported 34 reasoning tokens, and the outgoing request after reopening included an encrypted reasoning item. Owner pause also aborted an active provider request; a new turn succeeded after unpausing.

   Offline checks: 2,831 runtime tests and 344 OpenDots tests passed, along with formatting, lint, typechecks and builds in the validated local configurations.

   Release dependency

   The OpenDots PR is Draft, not ready to merge. Published runtime 1.77.0 still lacks the SDK fix. A clean install from its current public lock reproduces 7 failures / 337 passes: three SDK compatibility gates and four review-resume cases. They are retained as regression guards, not skipped.

   Once the SDK fix is merged and published, the OpenDots lock needs updating to that corrected release and normal clean-install CI needs to pass. Until then, the two fixes can be tested together using a locally built runtime package; a merge alone does not update installed dependencies.

   The budget/effort above are explicit test choices, not requirements or new defaults. Model/provider support varies. No credentials, encrypted values or private page content are included in the PRs or this comment.

   Prepared with AI assistance.