feat(ai): support thinking/reasoning models in OpenAI-compatible strategy

[ImIvanGil]

Problem

OpenAI-compatible thinking / reasoning models — Kimi K2.5, K2.6, kimi-*-thinking variants, GPT o1 family — fail silently in the AI panel with the generic message "No response received from the API.", regardless of prompt.

There are two distinct upstream causes, both in OpenAiApiStrategy.qml:

Cause 1: hardcoded temperature: 0.7

getBody() hardcodes temperature: 0.7. Thinking models reject any temperature other than 1 with HTTP 400:

{
    "error": {
        "message": "invalid temperature: only 1 is allowed for this model",
        "type": "invalid_request_error"
    }
}

The error is a single-line JSON, so the SSE-parsing SplitParser ignores it (no data: prefix), and curl exits 0 with the buffer never populated → the panel falls into its "no streaming data received" branch and shows the generic placeholder.

Cause 2: parser ignores reasoning_content

Thinking models stream their chain-of-thought as delta.reasoning_content before emitting the actual answer as delta.content. The existing parser only checks delta.content:

if (delta && delta.content)
    return { content: delta.content, done: false, error: null };

All reasoning chunks are dropped. By the time the actual delta.content chunks arrive (sometimes hundreds of reasoning chunks later), various downstream timing issues kick in — buffer's still empty, parser misses the small content tail, etc. Either way: user sees nothing.

Reproduced empirically against api.moonshot.ai/v1: a "hola" prompt to kimi-k2.6 produced 196 reasoning_content chunks then 2 content chunks ("Hola"). The 196-to-2 ratio is typical.

What this PR does

Two minimal additive changes in OpenAiApiStrategy.qml:

1. Dynamic temperature

+let temp = 0.7;
+if (model.model && /k2\.(5|6)|thinking|^o1(-|$)/.test(model.model)) {
+    temp = 1;
+}
 let body = {
     model: model.model,
     messages: _formatMessages(messages),
-    temperature: 0.7
+    temperature: temp
 };

Regex covers:

• k2.5, k2.6 (Kimi K2.5/K2.6 vision-and-thinking)
• *thinking* (kimi-thinking-preview, kimi-k2-thinking, kimi-k2-thinking-turbo, …)
• o1, o1-preview, o1-mini, etc. (OpenAI reasoning family)

2. reasoning_content accumulation

parseStreamChunk():

 if (delta && delta.content)
     return { content: delta.content, done: false, error: null };

+if (delta && delta.reasoning_content)
+    return { content: delta.reasoning_content, done: false, error: null };

parseResponse() (non-stream):

+let outContent = msg.content || "";
+if (msg.reasoning_content && !outContent) {
+    outContent = msg.reasoning_content;
+}
-return { content: msg.content };
+return { content: outContent };

Surfaces the model's chain-of-thought as it streams in, then the final answer concatenated at the end — same flow as ChatGPT-o1 and Claude thinking UIs. Non-thinking models are unchanged.

Tested with

• Kimi K2.6 (heavy thinking, 196+ reasoning chunks per short prompt) — works.
• Kimi K2 (0905 preview) (no thinking, direct content stream) — unchanged, works.
• Moonshot v1 family (no thinking) — unchanged, works.

Visual note

Reasoning and final answer arrive in the same chat bubble, concatenated. Could be polished in a follow-up that renders them in separate styled sections (greyed-out reasoning block + answer below), but that's a UX call worth its own PR.

Diff stats

modules/services/ai/strategies/OpenAiApiStrategy.qml | 20 +++++++++++++++++---
1 file changed, 17 insertions(+), 3 deletions(-)

Related

• Depends on feat(ai): wire Config.ai.extraModels for user-defined OpenAI-compat providers #176 to register custom OpenAI-compatible providers (Kimi/Moonshot) via Config.ai.extraModels — otherwise users can't even select a thinking model to test against. Without feat(ai): wire Config.ai.extraModels for user-defined OpenAI-compat providers #176, the temperature fix still helps any built-in provider that ever adds thinking variants.