🤖 feat: add Z.ai as a native provider with GLM 5.3 Flash

[ibetitsmike]

Summary

Adds Z.ai (Zhipu) as a first-class native provider (provider id zai, backed by the official @ai-sdk/zai package), starting with one model: zai:glm-5.3-flash (GLM 5.3 Flash, aliases glm / glm-flash).

Background

GLM models were previously reachable only through the OpenRouter gateway (openrouter:z-ai/glm-4.6). Users with a Z.ai API key can now use GLM 5.3 Flash directly: 1M context, 128K max output, vision input, and forced reasoning with Z.ai's reasoning_effort control.

Implementation

• Provider registry: zai direct provider (createZai, ZAI_API_KEY, default base URL https://api.z.ai/api/paas/v4); the generic direct-provider factory path handles it with no custom code. OpenRouter routing is intentionally not wired (OpenRouter uses the hyphenated z-ai vendor slug).
• Known model GLM_53_FLASH with tokenizer override zai/glm-4.5 (closest published GLM encoding in ai-tokenizer).
• Thinking: the GLM 5.3 family cannot disable reasoning and accepts effort low/high/max only. isGlm53Model drives a ["low", "high", "max"] policy, and resolveEffectiveThinkingLevel clamps unset/"off" through the policy the same way Mythos-class Anthropic models do, so tracked levels match wire behavior.
• Provider options: the zai branch sends thinking: enabled, mapped reasoningEffort (high/xhigh to high, max to max, else low), and toolStream: true (Z.ai's recommended streaming-tool-calls flag).
• Cost/limits metadata uses Z.ai list prices ($0.15/M in, $0.03/M cached read, $0.50/M out; a temporary 50% promo ends 2026-09-09).
• UI: Z.ai provider icon, settings key URL, and "GLM" display-name casing.

Validation

• Remote dogfood UAT (Coder Agents) against this exact head: settings UI and key persistence (reload + restart), model picker casing and alias resolution, icon at 1440px and 375px, no-NaN cost surfaces, and the wire contract proven via API debug logs plus a local capture: endpoint /api/paas/v4/chat/completions, model glm-5.3-flash, thinking: enabled, and reasoning_effort mapping low to low, high to high, max to max. A dummy key surfaced Z.ai's genuine auth error cleanly.
• No live Z.ai key exists in dev/CI environments, so tests and UAT validate wiring rather than live completions.

Notes

UAT observation for a product-level call: the composer hides Low for this model by default because the global default Min Thinking floor is medium (existing behavior for all models, see policy.test.ts "medium floor hides off/low"; Grok 4.6 loses Low the same way). With the per-model Min Thinking setting lowered, the composer offers exactly Low/High/Max, and "off" is never offered. If always-reasoning models should be exempt from the floor, that belongs in a follow-up.

Risks

Low regression risk: the change is almost entirely additive (new provider/model entries plus one new providerOptions branch). The one shared-path edit is resolveEffectiveThinkingLevel, which now clamps for GLM 5.3 in addition to Mythos-class Anthropic models; behavior for all other models is unchanged and covered by existing policy tests.

---

Generated with xum • Model: anthropic:claude-fable-5 • Thinking: xhigh • Cost: $40.18

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
Mux	🟢 Ready	View Preview	Aug 29, 2026, 8:28 AM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[ibetitsmike]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: e9c4983eba

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[ibetitsmike]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 82e96fa0e3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".