feat: add Token River, Volcano Engine Agent Plan, and Volcano Engine Coding Plan providers

[yoyo636]

New Providers

Token River (20 models)

• kimi-k3, kimi-k2.6, kimi-k2.7-code
• deepseek-v4-pro, deepseek-v4-flash
• glm-5.1, glm-5.2
• qwen3.7-flash/max/plus, qwen3.8-max
• minimax-m2.7, minimax-m3
• mimo-v2.5, mimo-v2.5-pro
• hy3, hy3-preview
• LongCat-2.0
• doubao-seed-2-1-pro, doubao-seed-2-1-turbo

Volcano Engine Agent Plan (10 models)

• doubao-seed-1-6, doubao-1-5-pro-256k/32k
• doubao-1-5-thinking-pro, doubao-1-5-vision-pro
• doubao-pro-32k/128k, doubao-lite-32k/128k
• deepseek-r1-250120

Volcano Engine Coding Plan (7 models)

• doubao-pro-code, doubao-seed-1-6
• doubao-1-5-pro-256k/32k
• doubao-pro-32k/128k, doubao-lite-32k

[github-actions]

Action items

• [critical] [violation] providers/token-river/provider.toml:1 - Check: New provider must ship a compliant logo.svg. Why: token-river, volcengine-agent, and volcengine-coding are new providers with no logo.svg (must use currentColor, no fixed size/hardcoded colors, prefer square viewBox). Action: Add providers/<id>/logo.svg for each of the three providers.
• [critical] [violation] providers/token-river/models/doubao-seed-2-1-pro.toml:1 - Check: Non-lab hosts must use base_model; add complete models/ lab metadata when missing. Why: Token River is a multi-lab relay. Doubao Seed 2.1 Pro/Turbo are authored as full inline ByteDance models with no base_model and no models/<lab>/… entries, despite being nameable lab models (also hosted elsewhere). Action: Add complete lab TOMLs (e.g. under models/bytedance/ or the chosen lab id) and make both Token River files override-only with base_model, cost, and correct reasoning_options.
• [critical] [violation] providers/volcengine-agent/models/deepseek-r1-250120.toml:1 - Check: Third-party hosts of lab models must use base_model. Why: DeepSeek R1 is not created by Volcano Engine; models/deepseek/deepseek-r1.toml already exists, but this entry is a full inline definition (attachment = true, image input, zero cost, empty reasoning controls) instead of inheriting the lab model. Action: Set base_model = "deepseek/deepseek-r1" and keep only host-specific fields/real deltas (cost, reasoning_options, limit/status overrides).
• [high] [violation] providers/token-river/models/deepseek-v4-pro.toml:6 - Check: Relay reasoning_options must match lab + same-surface peer controls, not a invented universal enum. Why: Almost every Token River reasoner uses bare effort = ["low","high","max"]. That contradicts first-party baselines: DeepSeek V4 Pro is toggle + high/max (not low); Qwen 3.7/3.8 are toggle + budget and/or different effort; GLM-5.1 is toggle-only; GLM-5.2 is high/max; LongCat is toggle-only; Hy3 is toggle + low/medium/high; MiMo is toggle-only; MiniMax-M2.7 is []; MiniMax-M3 is toggle; Kimi K2.6 is toggle; Kimi K2.7 Code is []. Blanketing L/H/max misrepresents caller controls. Action: For each model, copy the lab/peer option set this host can actually expose (including toggle + leading wire comment when needed); do not invent a shared L/H/max list.
• [high] [violation] providers/volcengine-agent/models/doubao-seed-1-6.toml:8 - Check: reasoning = true requires accurate host reasoning_options; relays must not use [] from uncertainty. Why: Multiple Volcengine Agent/Coding reasoners set reasoning_options = [] (Doubao Seed 1.6, Doubao 1.5 Pro variants, Thinking Pro, Pro Code, DeepSeek R1) with no evidence the ARK API exposes no controls. Empty means affirmatively no caller control. Action: Document the ARK request surface and set real options (or keep [] only with evidence of always-on/no control); add a leading wire comment for any toggle.
• [high] [violation] providers/token-river/models/deepseek-v4-pro.toml:19 - Check: Provider overrides must be real deltas, not incorrect copied capabilities. Why: Token River restates [modalities] input = ["text","image"] and often attachment = false on models whose lab metadata is text-only (DeepSeek V4, GLM, MiniMax-M2.7, LongCat, Hy3, etc.). That invents vision support and fights inheritance. Action: Drop unchanged/incorrect name/description/attachment/[modalities]/[limit] copies; override only true host deltas (e.g. smaller context) and provider fields (cost, reasoning_options).
• [medium] [possible mistake] providers/token-river/provider.toml:5 - Check: OpenAI-compatible api base URL should be the chat-completions root clients expect. Why: api = "https://api.tokenriver.cn" omits a version path; nearly all @ai-sdk/openai-compatible providers use a …/v1 (or equivalent) base. A wrong base breaks SDK routing. Action: Verify Token River’s OpenAI-compatible base URL and set the exact documented path (likely including /v1 if that is what the provider publishes).
• [medium] [possible mistake] providers/volcengine-coding/models/doubao-1-5-pro-256k.toml:6 - Check: attachment must stay consistent with declared input modalities. Why: Several entries set attachment = false while [modalities].input includes "image" (e.g. coding-plan Doubao 1.5 Pro 256K / Pro Code; several Token River files). That is internally inconsistent. Action: Align attachment with whether image/file inputs are actually accepted on that host.
• [low] [possible mistake] .pr-review/pull-request.json:4 - Check: Data-changing PRs should cite first-party pricing/docs/API evidence. Why: 37 models across three providers change catalog facts (limits, modalities, reasoning, zero costs) with no mapped sources in the PR body. Zero cost can be valid for subscription plans, but is unverifiable as written. Action: Add direct provider docs/pricing/API citations stating what each supports (endpoints, plan pricing, context limits, reasoning controls).