fix: ClinePass thinking-effort levels are wrong for all models: registry shows none/low/medium/high/xhigh, but DeepSeek V4.1 Flash officially supports off/low/high/max

[github-actions]

Summary

ClinePass is a multi-model OpenAI-compatible relay whose model entries mostly carried a copy-pasted GPT-style effort list (none / low / medium / high / xhigh). That blocked real top levels such as max on DeepSeek and advertised levels the underlying labs do not support. reasoning_options are now aligned with each lab’s first-party controls (and established same-surface peers), so clients can offer the correct picker values.

Note on off: models.dev’s effort enum does not include off; the off state is none. Clients that label it “Off” should map UI off → catalog none.

Changes

Model	Old reasoning_options	New
deepseek-v4.1-flash	none/low/medium/high/xhigh	none/low/high/max
deepseek-v4-flash	same	none/low/high/max
deepseek-v4-pro	same	none/high/max
kimi-k3	same	none/low/high/max
glm-5.2	same	none/high/max
qwen3.8-max	minimal/low/medium/high/xhigh	none/low/medium/xhigh
kimi-k2.6	GPT L/M/H list	toggle only
kimi-k2.7-code	GPT L/M/H list	[] (always-on, no control)
mimo-v2.5, mimo-v2.5-pro	GPT L/M/H list	toggle only
minimax-m3	GPT L/M/H list	toggle only
qwen3.7-plus, qwen3.7-max	GPT L/M/H list	toggle only
glm-5.3, glm-5.3-flash	already low/high/max	unchanged

Evidence

• DeepSeek Flash / V4.1 Flash effort low|high|max + disable via thinking toggle / Responses none: DeepSeek Thinking Mode — OpenAI Chat Completions uses reasoning_effort low|high|max (medium/xhigh map to high); thinking can be disabled. First-party catalog: providers/deepseek/models/deepseek-flash.toml / deepseek-v4-flash.toml. Relay peers with off-as-none: e.g. providers/kilo/models/deepseek/deepseek-v4.1-flash.toml, providers/nano-gpt/models/deepseek/deepseek-v4.1-flash.toml.
• DeepSeek Pro effort high|max: same Thinking Mode guide + first-party providers/deepseek/models/deepseek-v4-pro.toml.
• Kimi K3 low|high|max + disable: lab providers/moonshotai/models/kimi-k3.toml; OpenRouter peer providers/openrouter/models/moonshotai/kimi-k3.toml.
• GLM-5.2 effective high|max, off via none/minimal: lab providers/zhipuai/models/glm-5.2.toml and Zhipu thinking docs (cited in that file).
• Qwen3.8 Max graded low|medium|xhigh: lab providers/alibaba/models/qwen3.8-max.toml and Qwen thinking docs cited there.
• Toggle-only labs (K2.6, MiMo, MiniMax-M3) and always-on K2.7 Code: first-party provider TOMLs under moonshotai/, xiaomi/, minimax/; OpenRouter peers match.
• Qwen3.7 Plus/Max on this host as toggle-only: lab uses toggle + budget_tokens; ClinePass is OpenAI-compatible Chat Completions (Cline API); OpenRouter peers for the same models use toggle only without budget_tokens.
• ClinePass model list / host kind: ClinePass docs; catalog API https://api.cline.bot/api/v1/ai/cline/recommended-models.

Validation

• Cross-checked each changed model against lab first-party TOMLs, DeepSeek/Zhipu/Qwen docs where cited above, and OpenRouter (or kilo/nano-gpt) peers.
• Confirmed schema effort enum is none|minimal|low|medium|high|xhigh|max|default (no off) in packages/core/src/schema.ts.
• Did not run bun validate or live ClinePass API calls (workflow constraints / no Bash).

Review notes

• Issue UI label off is encoded as catalog none so validation passes; consumers should map display labels.
• Broadened beyond the two DeepSeek rows because the same incorrect GPT template was on almost every ClinePass model; left glm-5.3 / glm-5.3-flash alone (already correct).
• Did not claim ClinePass forwards native thinking.type / enable_thinking / thinking_budget wire fields without a live probe; shapes follow lab + OpenAI-compat peer practice for this host kind.
• Did not add missing catalog models from the live recommended list (e.g. muse-spark-1.3-contributor) — out of scope for this reasoning fix.

Closes #7531

Automated by the issue fixer: https://github.com/anomalyco/models.dev/actions/runs/35494668733