openclaw - 💡(How to fix) Fix [Feature]: Support qwen-chat-template when hosted from vLLM -- accepted by schema but ignored -- blocks audio workflows because of thinking time. [1 participants]

Summary

Expected Behavior When a self-hosted Qwen model has compat.thinkingFormat: "qwen-chat-template", OpenClaw should inject chat_template_kwargs: { enable_thinking: <bool> } into the chat completions request body, matching the official Qwen + vLLM documentation.

Actual Behavior buildOpenAICompletionsParams in provider-stream-COLujAAo.js only handles two thinking formats:

"openrouter" → injects params.reasoning = { effort: ... } Default (OpenAI) → injects params.reasoning_effort = ... There is no branch for "qwen-chat-template", so the value is silently ignored.

Root Cause In provider-stream-COLujAAo.js:

if (compat.thinkingFormat === "openrouter" && model.reasoning && resolvedCompletionsReasoningEffort) params.reasoning = { effort: resolvedCompletionsReasoningEffort }; else if (resolvedCompletionsReasoningEffort && model.reasoning && compat.supportsReasoningEffort) params.reasoning_effort = resolvedCompletionsReasoningEffort; // No handling for "qwen-chat-template"

The Zod schema (zod-schema.core-BO_PdpIg.js) accepts "qwen-chat-template" as a valid thinkingFormat, and pi-ai internally has a code path that handles it, but OpenClaw's own request builder never reaches that code or injects the necessary fields.

Reproduction

1. Serve Qwen/Qwen3.6-27B via vLLM on a self-hosted endpoint
2. Configure OpenClaw provider: "5090voice": { "baseUrl": "http://<vllm-host>:1234/v1", "apiKey": "notneeded", "api": "openai-completions", "models": [{ "id": "qwen/qwen3.6-27b", "reasoning": true, "compat": { "thinkingFormat": "qwen-chat-template" } }] }
3. Use the model — model still generates thinking tokens

Environment, OpenClaw version: 2026.4.9, vLLM version: 0.19.1rc1.dev278+ge64b39ea7, Model: Qwen/Qwen3.6-27B, Hardware: RTX 5090, 32GB VRAM

Problem to solve

Allow qwen models hosted on vllm to respect the /think flag in discord and other channels.

Proposed solution

Add a branch in buildOpenAICompletionsParams to handle qwen-chat-template:

else if (compat.thinkingFormat === "qwen-chat-template" && model.reasoning) { params.chat_template_kwargs = { enable_thinking: !!resolvedCompletionsReasoningEffort, preserve_thinking: true }; }

Alternatively, the documented extra_body / extraBody parameter in model params should merge arbitrary fields into the outbound request body, but that feature does not appear to be implemented either (not referenced in the built code).

Alternatives considered

No response

Impact

Anyone using openclaw with a qwen model hosted on vllm. blocks any sort of voice workflows as the thinking forced on makes realtime voice impossible always can't use openclaw and voice with this setup.

Evidence/examples

See example above on how to reproduce, only openrouter and openai are compatible, not qwen.

Additional information

Thanks :)