Bug Description

In openviking/models/vlm/backends/openai_vlm.py, all four completion methods (get_completion, get_completion_async, get_vision_completion, get_vision_completion_async) accept a thinking: bool = False parameter, but this parameter is never used to set enable_thinking in the API request body.

Models like qwen3.5-plus and qwen3.5-flash (DashScope) default to thinking mode (chain-of-thought reasoning). Since enable_thinking is never disabled, every memory extraction call triggers full CoT reasoning, causing severe timeouts.

Steps to Reproduce

1. Configure openviking with qwen3.5-plus or qwen3.5-flash as the VLM model via DashScope
2. Integrate with OpenClaw with autoCapture: true
3. Send a message through the channel (e.g., Feishu)
4. Observe OpenClaw logs after the agent replies

Expected Behavior

When thinking=False (the default), the backend should pass extra_body={"enable_thinking": False} to the API, disabling unnecessary chain-of-thought reasoning for simple memory extraction tasks.

Actual Behavior

auto-capture failed: AbortError: This operation was aborted appears in OpenClaw logs after 15~60 seconds. The root cause is that qwen3.5-plus/flash defaults to thinking mode, and each /extract API call spends 60+ seconds on CoT reasoning before timing out.

Minimal Reproducible Example

# Fix: add to each kwargs block in openai_vlm.py before the API call
if not thinking:
    kwargs["extra_body"] = {"enable_thinking": False}

Error Logs

2026-03-24T15:13:27 openviking: auto-capture failed: AbortError: This operation was aborted
2026-03-24T15:30:02 openviking: auto-capture failed: AbortError: This operation was aborted
# Timing: capture-check triggers at T+0, AbortError at T+15s (default) or T+60s (after raising timeoutMs)

OpenViking Version

0.2.9

Python Version

3.11

Operating System

Linux

Model Backend

OpenAI

Additional Context

Measured API latency on DashScope:

• qwen3.5-plus (thinking ON, default): ~9s for a simple "hello" → 60s+ for conversation extract
• qwen3.5-flash (thinking ON, default): ~5s for "hello"
• qwen-turbo (no thinking): ~0.8s

The thinking parameter already exists in the method signature — it just needs to be wired to extra_body.