feat(volcengine-agent-plan): add Volcengine Ark Agent Plan

[liusencomic-cyber]

Summary

• add volcengine-agent-plan as an OpenAI-compatible provider at https://ark.cn-beijing.volces.com/api/plan/v3
• 13 models, all via base_model against existing lab entries (no new lab files needed)
• cost = 0 throughout — Agent Plan is a subscription tier (Small/Medium/Large/Max)

Follow-up to #5534, which added the volcengine-coding-plan subscription provider. Same structure, different plan surface.

Why a separate provider

Agent Plan is a subscription tier with its own base URL and its own key space. A plan key is rejected with 401 on the pay-as-you-go /api/v3 (verified 2026-09-21), and per Volcengine's warning on the coding plan, the two billing surfaces must stay distinct — this cannot be a baseURL override on the existing volcengine entry.

Model naming

Short console labels (doubao-seed-2.1-turbo, glm-5.3), which this endpoint accepts; the pay-as-you-go /api/v3 requires dated IDs and 404s on short labels. Same convention as volcengine-coding-plan.

Limits are live-measured, not copied from docs

The official plan docs disagree with the wire in several places, so every limit was probed against POST /api/plan/v3/chat/completions on 2026-09-21:

Output caps come from the exact max_tokens rejection message ("expected a value <= N"):

model	docs	measured
doubao-seed-2.0-mini / -lite	128K	131072
doubao-seed-2.1-turbo / -evolving	256K	262144
minimax-m3, glm-5.3, glm-5.3-flash, kimi-k3, kimi-k2.8-preview	128K	131072
kimi-k2.7-code	32K	32768
deepseek-v4-pro	384K	393216
deepseek-v4-flash / v4.1-flash	384K	accepted max_tokens > 8M (not separately enforced) — lab value kept

Context caps were bisected with oversized prompts until the request failed with "Total tokens of image and text exceed max message tokens":

model	docs	measured (bisect)
doubao-seed-2.1-turbo	256K	fails above ~229K prompt tokens → 228000
doubao-seed-evolving	1M	OK at 916,875, fails by 928,750 → 916000
minimax-m3	1M	OK at 1,039,437, fails by 1,043,718 → 1039000
kimi-k3	1M	OK at 1,043,718 → 1043000
deepseek-v4-flash	1M	OK at 1,035,156, fails by 1,039,437 → 1035000

(The >1M readings on minimax/kimi are the 1M window plus request overhead, consistent with 1_048_576 total.)

Roster & verification

• All 13 models returned HTTP 200 on a minimal chat call (2026-09-21).
• Roster cross-checked against the Agent Plan overview and the management-API model list (ListArkAgentPlanModel, docs).
• bun validate passes.

Deliberately omitted

• ark-code-latest — a console-controlled routing pointer with no stable lab model (same reasoning as feat(volcengine-coding-plan): add Volcengine Ark Coding Plan #5534). Happy to add it if maintainers prefer.
• Non-text models (seedream/seedance/embedding/TTS/ASR) — image/video/audio endpoints, out of scope for a chat-completions provider entry.

Reasoning options

Not wire-probed on this endpoint (unlike #5534's n=4-6 per-level probing); each file mirrors the lab entry / coding-plan peer per the relay convention, and says so in a comment. kimi-k2.7-code declares reasoning_options = [] matching its coding-plan peer.

[github-actions]

Action items

• [high] [violation] providers/volcengine-agent-plan/provider.toml:2 - Check: Subscription/plan providers must not share the pay-as-you-go auth env with volcengine. Why: env = ["ARK_API_KEY"] is the same key as providers/volcengine. Coding Plan deliberately uses ARK_CODING_PLAN_API_KEY so a single env does not enable both surfaces at once; this PR’s own notes say Agent Plan has a separate key space and the wrong key type fails with 401. Sharing ARK_API_KEY still lights up both providers whenever that var is set, so clients show a catalog entry that cannot work for that key. Action: Use a dedicated env (e.g. ARK_AGENT_PLAN_API_KEY) and document why it must not be ARK_API_KEY, matching volcengine-coding-plan.
• [medium] [violation] providers/volcengine-agent-plan/models/glm-5.3.toml:15 - Check: After base_model, provider files must be override-only (no identical restated fields). Why: [limit] restates context = 1_000_000 and output = 131_072, which already match models/zhipuai/glm-5.3.toml. Same-surface peer volcengine-coding-plan omits [limit] and inherits. Action: Remove the identical [limit] block (keep cost / reasoning_options / interleaved).
• [medium] [violation] providers/volcengine-agent-plan/models/glm-5.3-flash.toml:15 - Check: Override-only base_model files. Why: [limit] is identical to models/zhipuai/glm-5.3-flash.toml (1_000_000 / 131_072). Action: Drop the redundant [limit] section.
• [medium] [violation] providers/volcengine-agent-plan/models/deepseek-v4.1-flash.toml:16 - Check: Override-only base_model files. Why: [limit] restates lab models/deepseek/deepseek-v4.1-flash.toml exactly (context = 1_000_000, output = 384_000) with no measured delta. Action: Remove the identical [limit] block; keep only real overrides.

[github-actions]

Action items

• [medium] [violation] providers/volcengine-agent-plan/models/deepseek-v4-pro.toml:105 - Check: Override-only base_model files (no restated identical lab fields; [limit] deep-merges). Why: Several models still copy lab-identical limit keys after the glm/v4.1-flash cleanup: deepseek-v4-pro context 1_000_000; deepseek-v4-flash output 384_000; doubao-seed-2.0-lite / doubao-seed-2.0-mini context 256_000; kimi-k2.7-code / kimi-k2.8-preview context 262_144 / 1_048_576; kimi-k3 output 131_072. Action: In each file, author only the limit keys that differ from the lab entry (e.g. output-only or context-only overrides).

[github-actions]

No actionable findings.