fix(local): disable thinking on Ollama's OpenAI-compat endpoint (hybrid reasoning models)

[alekseikondratenko]

Problem

Local-engine analysis silently degrades ~5-10x when the configured Ollama model is a hybrid reasoning model (any of the Qwen3.5 family, and other recent releases that ship with thinking enabled).

Dayflow talks to Ollama through the OpenAI-compatible /v1/chat/completions endpoint, and on that endpoint Ollama enables the thinking stream by default for these models. There is no model-level or server-level way to turn it off on the Ollama side:

• PARAMETER think false in a Modelfile is not supported (Feature request: PARAMETER think false in Modelfile for Qwen 3.5 / thinking models ollama/ollama#14809, How to completely disable the thinking mode by configuring the modelfile in Qwen3.5 ollama/ollama#14617)
• the Qwen3-era /no_think soft switch is ignored by Qwen3.5
• there is no env var / server setting

Since ChatRequest is a fixed struct, users have no workaround short of switching models.

Measured on an M4 MacBook Air (24 GB), qwen3.5:9b, real Dayflow screenshots:

Request	Completion tokens	Wall time per frame
Dayflow's current request shape	602-603 (mostly reasoning)	57-105 s
Same + reasoning_effort: "none"	33-72	~25 s

At 15 sampled frames per batch, that is the difference between a batch finishing in ~6 minutes and a batch taking longer than the 15-minute window it covers - the analysis queue falls behind permanently and the machine stays hot.

Fix

Send reasoning_effort: "none" when (and only when) the local engine is Ollama. Ollama's OpenAI-compat layer honors this field (verified against Ollama 0.32.14); non-thinking models such as the recommended qwen3-vl:4b preset simply ignore it.

The field is an optional on ChatRequest that stays nil for LM Studio and custom endpoints, and JSONEncoder never encodes it in that case - so those request bodies are byte-identical to today's. LM Studio has its own per-model thinking controls, and some OpenAI-compatible providers reject "none", which is why the field is gated to the Ollama engine rather than sent unconditionally.

Testing

• Verified reasoning_effort: "none" disables the reasoning stream and answer quality is unchanged, via direct /v1/chat/completions calls against Ollama 0.32.14 with qwen3.5:9b on real Dayflow screenshot batches (numbers above).
• Verified plain-JSON encoding: with the field nil, the encoded body is unchanged from the current one.
• Note: I authored this patch on a machine without Xcode, so it has not been compiled locally - the change is confined to one struct field and one guarded assignment in makeChatURLRequest. Happy to adjust if CI or a maintainer build turns anything up.

🤖 Generated with Claude Code

[alekseikondratenko]

Updated: extended the same guard to the LM Studio engine after verifying LM Studio 0.4.21 also honors reasoning_effort: "none" (Qwen3.5-9B MLX: 26s / 70 completion tokens with the field vs 120s / 1800+ thinking tokens without; chat_template_kwargs is not honored via the server API, so the request field is the only client-side lever there as well). Custom endpoints remain untouched.