[siliconflow] Audit models and reasoning controls

[rekram1-node]

Summary

• audit both SiliconFlow catalogs against first-party API schemas, model pages, and dated global/China release notes
• remove 38 expired global entries and 6 expired China entries; add 10 active global entries
• preserve only positively documented reasoning controls and explicitly leave unresolved reasoning controls omitted
• add exact SiliconFlow control regression coverage and support symlink baselines in compare:migrations

Final catalogs

Provider	Total	Reasoning	Non-reasoning	With controls	Toggle	Budget	Effort	Fixed []	Unresolved/omitted
SiliconFlow global	48	23	25	17	9	16	0	0	6
SiliconFlow China	46	27	19	16	16	0	0	0	11

Combined verified controls: 25 toggles and 16 token budgets. No non-reasoning model has options.

Evidence decisions

• the current global text ChatCompletionRequest documents thinking_budget for reasoning models with accepted request range 128..32768; budgets are retained only for the 16 exact current enum-listed endpoint IDs covered by that schema
• the reasoning guide says only Qwen3 strictly enforces termination at thinking_budget; for other covered models this metadata represents request-field support, not guaranteed strict enforcement
• the separate VLM request schema does not expose thinking_budget; all three Qwen Thinking VLMs omit options
• China has no token budgets or effort controls; shared aliases were replaced where necessary to prevent global options from leaking into the China catalog
• Hy3 retains the documented global request budget but no inferred effort option
• global zai-org/GLM-4.7 was removed on its announced June 11, 2026 deprecation date; China Pro GLM-4.7 remains independently listed with its verified toggle

Unresolved controls

Global, reasoning=true with options omitted:

• Qwen/Qwen3-VL-235B-A22B-Thinking
• Qwen/Qwen3-VL-30B-A3B-Thinking
• Qwen/Qwen3-VL-32B-Thinking
• deepseek-ai/deepseek-v4-flash
• deepseek-ai/deepseek-v4-pro
• stepfun-ai/Step-3.5-Flash

China, reasoning=true with options omitted:

• Pro/deepseek-ai/DeepSeek-R1
• Pro/moonshotai/Kimi-K2.5
• Pro/moonshotai/Kimi-K2.6
• Pro/zai-org/GLM-5.1
• Qwen/Qwen3-235B-A22B-Thinking-2507
• Qwen/Qwen3-VL-235B-A22B-Thinking
• Qwen/Qwen3-VL-30B-A3B-Thinking
• Qwen/Qwen3-VL-32B-Thinking
• deepseek-ai/DeepSeek-R1
• deepseek-ai/DeepSeek-V4-Pro
• stepfun-ai/Step-3.5-Flash

First-party sources

• https://docs.siliconflow.com/en/api-reference/chat-completions/chat-completions
• https://docs.siliconflow.com/cn/api-reference/chat-completions/chat-completions
• https://docs.siliconflow.com/en/userguide/capabilities/reasoning
• https://docs.siliconflow.com/cn/userguide/capabilities/reasoning
• https://docs.siliconflow.com/en/release-notes/overview
• https://docs.siliconflow.com/cn/release-notes/overview
• official model pages under https://www.siliconflow.com/models

Validation

• bun validate: pass
• bun compare:migrations: pass
• bun test packages/core/test/siliconflow.test.ts packages/core/test/sync-runner.test.ts: 10 pass, 0 fail
• git diff --check: pass
• live API tests not run: no safe SiliconFlow credentials are present