Reasoning toggle (thinking.type=enabled) not emitted for OpenAI-compatible models with a models.dev reasoning toggle (DeepSeek V4 Flash) #39933

Bug: opencode doesn't enable reasoning for OpenAI-compatible models that declare a reasoning toggle in models.dev (e.g. DeepSeek V4 Flash via Zen).

Model: opencode/deepseek-v4-flash — its models.dev entry declares reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["high","max"] }] with [interleaved] field = "reasoning_content".

Expected: reasoning engages (e.g. via --variant max or options.thinking), like the Anthropic thinking option does.

Actual: reasoning never engages. --variant max/high, --thinking, options.reasoningEffort, and options.thinking = {type:"enabled"} all yield zero reasoning_content and 0 reasoning tokens; per-request cost stays flat. providerOptions passthrough of thinking returns HTTP 400.

Repro: the gateway itself honours the toggle — same key/endpoint via curl returns full reasoning_content:

curl https://opencode.ai/zen/v1/chat/completions -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" -d '{"model":"deepseek-v4-flash","thinking":{"type":"enabled"},"reasoning_effort":"max","stream":true,"messages":[{"role":"user","content":"9.11 vs 9.8? reason step by step."}]}'

Through opencode run -m opencode/deepseek-v4-flash --thinking the stream has no reasoning parts.

For comparison, OpenAI-style models on the same gateway (e.g. opencode/gpt-5.6-luna --variant max) DO engage reasoning correctly via reasoning_effort, so the gap is specific to the toggle capability.

Likely cause: the OpenAI-compatible adapter forwards reasoning_effort but not the models.dev toggle field (thinking.type=enabled).

Env: opencode 1.18.5, Linux.