v0.8.65: 接入智谱 GLM-5.2 作为 provider route fixture

[413057leebobo]

需求描述

希望在 CodeWhale 中支持智谱 GLM-5.2 模型作为 provider，用于子 Agent 调用。

使用场景

• 当前 DeepSeek 模型表现不错，但某些中文场景（长文档理解、中文创作）GLM-5.2 有明显优势
• 通过 agent 工具指定 model: "glm-5.2" 直接分发子任务到 GLM
• Constitution 中已有 model_strength 对 GLM 家族的映射规划（GLM-5.2 → GLM-5-Turbo）

API 信息

• API Base: https://open.bigmodel.cn/api/paas/v4/
• 模型 ID: glm-5.2（或 glm-5.2-flash）
• 智谱 API 文档: https://open.bigmodel.cn/dev/api

期望配置方式

在 config.toml 中类似现有 provider 格式：

[providers.zhipu]
api_base = "https://open.bigmodel.cn/api/paas/v4/"
api_key = "$ZHIPU_API_KEY"
models = ["glm-5.2", "glm-5.2-flash"]

感谢！

[github-actions]

Thanks @413057leebobo for the report.

This issue is staying open for maintainer triage. CodeWhale gets better because people bring us real edge cases from real machines, providers, regions, and workflows.

If you can add a reproduction, logs, version output, screenshots, or the provider/model involved, that makes it much easier for us to verify and harvest the fix. Maintainers may comment /lgtmi to mark recurring issue reporters as approved so this intake note is skipped next time.

[Artenx]

有个zai好像就是智谱吧

[Hmbown]

Triage note: moving this into the v0.8.65 provider/model-routing lane. Treat GLM-5.2 as a provider-scoped route fixture: provider descriptor, auth/config shape, model facts/offering rows, readiness checks, and docs should route through the same explicit provider/model machinery rather than a one-off model string.

[Hmbown]

Decision: treat Zhipu GLM as a first-class provider-scoped route fixture (not a redirect). @Artenx is correct — z.ai is 智谱/Zhipu.

We'll add a zhipu (aliases glm, bigmodel) provider descriptor modeled on the Qianfan fixture (#3425 / #3357): OpenAI-compatible endpoint with the bigmodel.cn base, provider-scoped auth/config ([providers.zhipu], ZHIPU_* / GLM_* env), model facts + offering rows for the GLM family (incl. GLM-5.2, GLM-4.6, GLM-4-Flash), readiness checks, and docs — routing through the same explicit provider/model machinery. A PR is incoming. Thanks @413057leebobo and @Artenx.

[Hmbown]

Update after checking Models.dev: the catalog already has first-class GLM/Zhipu coverage we should consume instead of hand-growing another provider/model table. Relevant current catalog shape:

• Providers: zhipuai, zai, zhipuai-coding-plan, zai-coding-plan
• BigModel base: zhipuai -> https://open.bigmodel.cn/api/paas/v4
• Z.AI base: zai -> https://api.z.ai/api/paas/v4
• Coding-plan bases: https://open.bigmodel.cn/api/coding/paas/v4 and https://api.z.ai/api/coding/paas/v4
• Canonical model facts include zhipuai/glm-5.2; provider rows expose wire id glm-5.2 with reasoning, tool-call, structured-output, text-in/text-out, 1M context, and provider pricing.

PR #3497 adds the first network-free codewhale_config::models_dev schema foundation so follow-up route/catalog work can hydrate this fixture from Models.dev-shaped data rather than adding a Zhipu-specific match table. Keeping this issue open until the runtime provider/model picker and route resolver actually consume that catalog.

[Hmbown]

Pre-fix code map (2026-06-23)

Disposition: keep as a route fixture, but implement it through Models.dev-backed provider offerings and route-candidate tests rather than a one-off GLM table.

Relevant Models.dev facts already verified:

• Provider keys include zai, zai-coding-plan, zhipuai, and zhipuai-coding-plan.
• Canonical GLM-5.2 exists as a GLM-5 family/series model; provider offerings expose glm-5.2 with context/output limits, reasoning/tool/structured-output flags, and cost.
• Z.AI and Zhipu/BigModel base URLs are separate provider offerings even when serving the same GLM-5 variant.

Code seams:

• crates/tui/src/route_resolver.rs:24 should resolve Z.AI/Zhipu provider + glm-5.2 into one route candidate.
• crates/tui/src/commands/groups/core/provider.rs:63 currently pre-normalizes explicit provider model args; raw GLM provider ids should reach the resolver.
• crates/tui/src/model_routing.rs:51 contains GLM-specific fast-lane logic; this should eventually read route-tier/model-family overlay data.
• crates/tui/src/config.rs:207 still defines hard-coded Z.AI defaults and GLM constants; use these only as local fallback/override once catalog-backed routing exists.

Acceptance checks:

• /provider zai glm-5.2 and equivalent picker/model-route paths resolve without DeepSeek normalization.
• GLM-5.2 and GLM-5-Turbo are represented as GLM-5 family/series variants with route tiers, not as unrelated strings.
• Z.AI and Zhipu offerings stay provider-scoped, including base URL and pricing provenance.