fix(volcengine-coding-plan): add glm-5.3-flash and kimi-k3

[cifangyiquan]

Summary

The official model configuration list for Coding Plan (https://www.volcengine.com/docs/82379/1928261, "支持配置的 Model Name") currently includes 10 models; this provider modeled only 8. Adding the two missing ones:

• glm-5.3-flash — listed explicitly alongside glm-5.3 (glm-latest)
• kimi-k3 — listed alongside kimi-k2.7-code

Both are zero-cost subscription entitlements, matching the existing entries. glm-5.3-flash mirrors the effort levels already declared in providers/zai-coding-plan/models/glm-5.3-flash.toml; for kimi-k3 I kept reasoning_options = [] like the kimi-k2.7-code relay entry and noted that endpoint-level effort behavior wasn't measured — happy to adjust if maintainers have observations.

Source accessed 2026-09-11.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] providers/volcengine-coding-plan/models/kimi-k3.toml:8 - Check: Relay reasoning_options must match lab/same-surface peers, not [] from uncertainty. Why: Lab providers/moonshotai/models/kimi-k3.toml and peers (e.g. OpenRouter, kimi-for-coding) expose toggle + effort low/high/max. This host is a multi-model relay; AGENTS.md / the reasoning audit skill forbid [] when controls exist upstream and you only skipped measuring this endpoint. Copying kimi-k2.7-code is invalid — that entry is empty only after measured no-effect on that model. Action: Author lab/peer baseline (toggle + effort ["low","high","max"]) with a leading wire comment for this host (thinking.type, effort field), or measure on /api/coding/v3 and keep [] only with affirmative no-control evidence like the sibling files.

[cifangyiquan]

Good catch — fixed in the latest push. reasoning_options now mirrors the lab (providers/moonshotai/models/kimi-k3.toml) and the kimi-for-coding relay baseline:

[[reasoning_options]]
type = "toggle"

[[reasoning_options]]
type = "effort"
values = ["low", "high", "max"]

with the /api/coding/v3 wire fields documented in the header (thinking.type = enabled|disabled|adaptive, output_config.effort = low|high|max). I did not measure endpoint-specific level behavior, so the comment notes the baseline is inherited and should be trimmed if probing shows levels collapse as they do on kimi-k2.7-code.

[github-actions]

Action items

• [medium] [violation] providers/volcengine-coding-plan/models/kimi-k3.toml:7 - Check: Toggle/effort wire comments must document this host’s real request fields. Why: The leading comment claims POST /api/coding/v3/chat/completions accepts output_config.effort and thinking.type=adaptive, but this provider’s OpenAI coding path uses reasoning_effort (and typically thinking.type enabled|disabled); output_config.effort is the Anthropic Messages / Moonshot–kimi-for-coding surface, not the chat-completions path in provider.toml. Action: Rewrite the wire comment for volcengine coding-plan (e.g. thinking.type + reasoning_effort on /api/coding/v3/chat/completions, and Anthropic output_config.effort only if documenting that path separately). Keep the lab/peer option set (toggle + low/high/max) unless endpoint probing shows otherwise.

[cifangyiquan]

Fixed — the wire comment now documents this provider's real fields per provider.toml:

POST /api/coding/v3/chat/completions takes
thinking.type = enabled|disabled plus reasoning_effort = low|high|max

with notes that adaptive is the Moonshot lab surface (not accepted on this relay) and that output_config.effort belongs to the separate Anthropic /api/coding/v1/messages path. The declared option set stays toggle + low/high/max; unmeasured endpoint behavior remains explicitly flagged.

[github-actions]

No actionable findings.