feat: add Volcengine Coding Plan provider

[alwinlin23]

Summary

Add the Volcengine Ark Coding Plan (火山方舟 Coding Plan) provider requested in anomalyco/opencode#40203, alongside the existing kimi-for-coding, alibaba-coding-plan-cn, zhipuai-coding-plan, minimax-cn-coding-plan, and tencent-coding-plan entries.

• Provider ID: volcengine-coding-plan
• OpenAI-compatible endpoint: https://ark.cn-beijing.volces.com/api/coding/v3 (subscription usage must go through /api/coding/*; the pay-as-you-go /api/v3 endpoint bills separately)
• Auth: ARK_API_KEY (Ark API key from https://console.volcengine.com/ark)

Models (current plan lineup)

Model	Definition
ark-code-latest	Auto-routing alias for the plan's latest coding model (inline)
doubao-seed-2.0-lite	ByteDance first-party Doubao Seed coding model (inline)
doubao-seed-2.1-turbo	ByteDance first-party Doubao Seed coding model (inline)
deepseek-v4-pro	base_model = "deepseek/deepseek-v4-pro"
deepseek-v4-flash	base_model = "deepseek/deepseek-v4-flash"
glm-5.2	base_model = "zhipuai/glm-5.2"
kimi-k2.6	base_model = "moonshotai/kimi-k2.6"
kimi-k2.7-code	base_model = "moonshotai/kimi-k2.7-code"
minimax-m3	base_model = "minimax/MiniMax-M3"
minimax-m2.7	base_model = "minimax/MiniMax-M2.7"

All models are priced 0 (subscription plan), matching the other coding-plan providers. Deep-thinking models declare a toggle (thinking.type = enabled|disabled) with reasoning_content interleaved output, per the Ark deep-thinking docs. Third-party models inherit limits/modalities from their lab entries; only host-specific fields (cost, reasoning_options, interleaved) are overridden.

Sources

• Coding Plan overview/models: https://www.volcengine.com/docs/82379/1925114
• Codex/Coding Plan integration (Responses support on /api/coding/v3): https://www.volcengine.com/docs/82379/2556056
• Deep thinking format: https://www.volcengine.com/docs/82379/1449737
• Plan landing page: https://www.volcengine.com/activity/codingplan

Test plan

• bun validate passes
• bun test — no new failures (the 5 failing SDK snapshot tests also fail on clean dev; they need a generated artifact)
• Maintainer verification of the two inline Doubao Seed model limits/dates against the console/docs if needed

[github-actions]

Action items

• [high] [possible mistake] models/deepseek/deepseek-v3.2.toml:7 - Check: Lab metadata must be provider-agnostic facts about the model the lab built. Why: This new shared base sets reasoning = false, but established deepseek-v3.2 catalog entries (e.g. OpenRouter, Novita, Kilo, Bedrock, Alibaba/SiliconFlow) treat the model as reasoning-capable with a thinking toggle. Encoding non-reasoning on the lab entry will poison future base_model consumers. Action: Verify against DeepSeek’s V3.2 docs and set lab reasoning to match the model (likely true). If the Coding Plan host truly has no thinking control, override reasoning = false on providers/volcengine-coding-plan/models/deepseek-v3.2.toml (or add the host’s real reasoning_options if thinking.type applies).
• [medium] [violation] models/deepseek/deepseek-v3.2.toml:12 - Check: New lab entries used for base_model must be complete, including usable [limit] defaults. Why: [limit] only sets context and omits output. Provider models require limit.output after merge; sibling DeepSeek lab files ship both. A cost-only host of this base would fail validation. Action: Add a canonical lab limit.output (and keep the Volcengine provider override only if the plan’s max output differs).
• [medium] [violation] providers/volcengine-coding-plan/models/deepseek-v3.2.toml:12 - Check: After base_model, provider files must be override-only—no restated identical base values. Why: context = 128_000 duplicates the lab value; only output is a real delta. Action: Drop the identical context key and keep provider-only fields (cost, and output if it differs from the completed lab default).

[github-actions]

Action items

• [high] [violation] providers/volcengine-coding-plan/models/minimax-m2.7.toml:6 - Check: Reasoning options must match this host’s real controls, using lab + same-surface peers as baseline (AGENTS.md → Reasoning options). Why: First-party MiniMax, both MiniMax coding-plan providers, and other relays author reasoning_options = [] for MiniMax-M2.7 (always-on, no caller control). This PR invents { type = "toggle" }, which misrepresents API capability. Action: Change to reasoning_options = [] unless Ark Coding Plan docs/live API show a real on/off field for this model; if so, cite the exact wire path.
• [high] [violation] providers/volcengine-coding-plan/models/kimi-k2.7-code.toml:6 - Check: Reasoning options baseline for Moonshot Kimi K2.7 Code. Why: Lab/first-party providers/moonshotai/models/kimi-k2.7-code.toml and kimi-for-coding use reasoning_options = [] (thinking always on). Authoring toggle claims a disable control those entries deny. Action: Use reasoning_options = [] unless Volcengine-specific docs prove thinking.type can turn reasoning off for kimi-k2.7-code; cite that evidence if keeping toggle.
• [medium] [possible mistake] providers/volcengine-coding-plan/models/deepseek-v4-pro.toml:6 - Check: DeepSeek V4 baseline is toggle + effort high/max (lab providers/deepseek/). Why: Both DeepSeek V4 entries only declare toggle. Lab and several relays expose graded effort; dropping effort is only correct if Ark Coding Plan does not forward reasoning_effort (or equivalent). Action: Verify Ark Coding docs/API for DeepSeek V4 effort; if supported, add { type = "effort", values = ["high", "max"] } (and document the wire field). Apply the same fix to deepseek-v4-flash.toml.
• [medium] [possible mistake] providers/volcengine-coding-plan/models/glm-5.2.toml:6 - Check: GLM-5.2 lab/peer reasoning controls. Why: Zhipu/Z.AI first-party and coding-plan peers use effort high/max (off via effort none/minimal mapping), not a bare toggle. This file only has toggle, which may under- or mis-describe Ark’s GLM surface. Action: Confirm whether Ark Coding Plan exposes reasoning_effort/output_config.effort for GLM-5.2, only thinking.type, or both; author the matching option set and cite the wire fields.