Cannot disable thinking for Qwen3.5-plus via Anthropic API compatibility

[lightgao]

What happened?

When integrating the latest Qwen3.5-plus model from Aliyun into Pi Coding Agent via the Anthropic API-compatible endpoint, the thinking function cannot be disabled.

Even when the user explicitly sets “reasoning: false” in the custom models configuration or uses thinking=off in the UI, the model continues to run in thinking mode. This is because the request sent to the Aliyun endpoint does not include 'thinking' field, causing the model to fall back to its default state (thinking enabled by default).

Steps to reproduce

1. Configure the provider in Pi Coding Agent with the Aliyun endpoint:

{
  "providers": {
    "anthropic_aliyun": {
      "baseUrl": "https://dashscope.aliyuncs.com/apps/anthropic",
      "api": "anthropic-messages",
      "apiKey": "sk-sp-tester-001",
      "models": [
        {
          "id": "qwen3.5-plus",
          "reasoning": false,
          "input": ["text", "image"],
          "contextWindow": 960000,
          "maxTokens": 65000
        }
      ]
    }
  }
}

2. Set thinking to off in the Pi coding agent interface.

3. Say 'hi' with Qwen3.5-plus model.

4. Observe the response: The model still outputs detailed reasoning steps, indicating thinking is active.

Expected behavior

When the user sets “reasoning=true && thinking=off”, the Pi Coding Agent should explicitly send the following field in the Anthropic messages request to override the provider's default:

"thinking": { "type": "disabled" }

Version

No response

[lightgao]

A quick fix has been done here.

If you consider this a minor issue worth fixing, please approve this issue. I can then submit a PR to help resolve it.

[deltoid9818]

I'd like to fix this. The root cause is in buildParams() in anthropic.ts — the thinking conditional checks options?.thinkingEnabled && model.reasoning, so when thinking is disabled, the thinking field is simply omitted. Native Anthropic models default to thinking-off, but Anthropic-compatible endpoints (like Aliyun Qwen) may default to thinking-on.

Fix: restructure the conditional so the outer check is model.reasoning only. Inside, if thinkingEnabled is true, existing adaptive/budget logic is preserved unchanged. If false, explicitly set params.thinking = { type: "disabled" }. Non-reasoning models remain unaffected (no thinking field sent).

8 test cases covering all combinations. npm run check clean.

[steflsd]

Same issue for openrouter thinking - maybe a broader fix is needed

[badlogic]

Fixed in packages/ai.

Pi now sends thinking: { type: "disabled" } for anthropic-messages reasoning models when thinking is explicitly off, instead of omitting the field and letting Anthropic-compatible backends fall back to their default.

Also added coverage for this in:

• packages/ai/test/anthropic-thinking-disable.test.ts
• payload assertions for budget-based and adaptive Anthropic models
• env-gated Anthropic E2E coverage for thinking-off

[badlogic]

same of for openai-completions, which affects openrouter models.