[BUG] Missing thinking output with "Enable Reasoning Effort" when using OpenAI Compatible

[marivk]

Problem (one or two sentences)

Problem (one or two sentences)
When the API provider sets to "OpenAI Compatible" and "Enable Reasoning Effort" option high. I don't see any thinking output in chat box. Checking the API with curl when adding "thinking": {"type": "enabled"}, the response include reasoning_content info.

Context (who is affected and when)

None

Reproduction steps

API Provider: OpenAI Compatible
Model: GLM-5
Base URL: https://modelservice.jdcloud.com/coding/openai/v1
Enable Reasoning Effort: high

Expected result

thinking output

Actual result

no thinking output

Variations tried (optional)

No response

App Version

3.51.1 (7c9722b)

API Provider (optional)

None

Model Used (optional)

No response

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause

The OpenAI Compatible provider (OpenAiHandler in src/api/providers/openai.ts) sends reasoning_effort: "high" in the request when "Enable Reasoning Effort" is turned on. However, some OpenAI-compatible models (like GLM-5 on JD Cloud) use a different protocol -- they require thinking: { type: "enabled" } in the request body to activate reasoning output. The reasoning_effort parameter is silently ignored by these models, so no reasoning_content is returned in the response.

The response-side handling is already correct -- the handler checks for reasoning_content in streaming deltas and yields it as reasoning output. The gap is only on the request side.

Proposed Fix

1. Add a "Supports Thinking" checkbox in the OpenAI Compatible settings UI (webview-ui/src/components/settings/providers/OpenAICompatible.tsx), visible when "Enable Reasoning Effort" is checked. This stores supportsReasoningBinary: true on openAiCustomModelInfo.

2. In OpenAiHandler.createMessage() (src/api/providers/openai.ts), when enableReasoningEffort is true AND modelInfo.supportsReasoningBinary is true, include thinking: { type: "enabled" } in the streaming request params. This mirrors the existing pattern in BaseOpenAiCompatibleProvider (line 101-104) used by Z.ai and other providers.

3. Add tests covering the new behavior: verify that thinking: { type: "enabled" } is included when both flags are set, and that it is absent otherwise.

This approach is backward-compatible -- existing users who rely on reasoning_effort are unaffected. Users with models like GLM-5 can check the new toggle to enable the thinking protocol.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.