Add Volcengine Ark Agent Plan

[Cing-self]

Summary

Adds the Volcengine Ark Agent Plan as a distinct OpenAI-compatible provider.

Agent Plan uses dedicated API keys and https://ark.cn-beijing.volces.com/api/plan/v3; generic Ark and Coding Plan keys target different subscription surfaces. Keeping it separate avoids billing/entitlement confusion and mirrors the existing Coding Plan provider structure.

Models

Adds the console-visible IDs:

• ark-code-latest (console-managed routing alias)
• doubao-seed-evolving
• doubao-seed-2.1-turbo
• doubao-seed-2.0-lite
• doubao-seed-2.0-mini
• glm-5.3-flash
• glm-5.3
• deepseek-v4-pro
• deepseek-v4-flash
• kimi-k2.7-code
• minimax-m3

Third-party models use base_model references. Subscription-billed provider models set per-token cost to zero. The routing alias has no caller-controlled target through the model ID, so it declares no reasoning options.

Sources and validation

• Agent Plan quick start/configuration docs: https://www.volcengine.com/docs/82379/2373738
• Agent Plan console subscription/config page
• One-token smoke calls against /api/plan/v3/chat/completions succeeded for the listed direct models and ark-code-latest
• bun install
• bun validate

Notes

• No API keys or private account data are included.
• kimi-k3 was not added because the current Small-plan smoke call returned UnsupportedModel; plan-aware handling can be added later.

[github-actions]

Action items

• [high] [violation] providers/volcengine-agent-plan/models/glm-5.3-flash.toml:7 - Check: Relay reasoning_options must match lab/same-surface peer controls; [] only when there is affirmatively no caller control. Why: This file sets reasoning_options = [] and claims that matches the Z.ai lab entry, but first-party and peer hosts (providers/zai, providers/zhipuai, providers/zai-coding-plan, providers/zhipuai-coding-plan) expose effort low/high/max for zhipuai/glm-5.3-flash. Empty options here advertise no control and rest on an incorrect baseline. Action: Author effort ["low", "high", "max"] (or the measured Agent Plan subset), or keep [] only with host-specific evidence that /api/plan/v3 rejects/ignores those levels; fix the leading comment accordingly.
• [high] [violation] providers/volcengine-agent-plan/models/deepseek-v4-flash.toml:9 - Check: Every toggle requires a leading top-of-file wire-path comment. Why: type = "toggle" is declared without an exact request field (e.g. thinking.type = enabled|disabled on /api/plan/v3/...), which AGENTS.md requires so callers know the control surface. Action: Add a leading # Toggle: … comment with the exact Agent Plan wire path (and endpoint if known).
• [high] [violation] providers/volcengine-agent-plan/models/deepseek-v4-pro.toml:9 - Check: Every toggle requires a leading top-of-file wire-path comment. Why: Same missing wire documentation as Flash while declaring toggle-only control. Action: Add a leading # Toggle: … comment with the exact Agent Plan wire path.
• [high] [violation] providers/volcengine-agent-plan/models/minimax-m3.toml:8 - Check: Every toggle requires a leading top-of-file wire-path comment. Why: type = "toggle" is present without documenting the request field path. Action: Add a leading # Toggle: … comment with the exact Agent Plan wire path (Coding Plan peer uses thinking.type = enabled|disabled).

[github-actions]

No actionable findings.