[nicho810]

关联 Issue

Closes #115

变更说明

Adds cheapaimodel as a built-in API-key provider preset. Mirrors #101 (Cheaper Inference).
I work on cheapaimodel. It is an OpenAI-compatible LLM gateway billed per token from a prepaid USD balance.

• config/provider/zcode-builtin.json: template cheapaimodel (openai-chat-completions, base URL https://api.cheapaimodel.ai/v1), placed right after cheaperinference, with deepseek-v4-1-flash and glm-5.3 enabled in templateModelRules.
• No new providerSiteRules: both models already match generic modelRules, and the limits match the gateway catalog (https://cheapaimodel.ai/models): deepseek-v4-1-flash 1000000 context / 384000 max output, image input; glm-5.3 1000000 / 128000, text only.
• Logo: light and dark SVG of the official mark from https://cheapaimodel.ai/favicon.svg (background tile removed, square 30x30 viewBox), registered in ProviderLogo.tsx and model-provider-logo-sources.json.
• The config revision is not changed, same as feat(provider): add Cheaper Inference preset #101.
• No OAuth flow and no catalog service. The user pastes an API key (keys at https://console.cheapaimodel.ai/keys).

验证

• zcode-builtin.json and model-provider-logo-sources.json parse as JSON.
• Checked that deepseek-v4-1-flash matches the .*deepseek-v4[.-]1-flash rule and glm-5.3 the .*glm-5\.3(?:-flash)? rule, and that their context and max output equal the catalog values above.
• GET https://api.cheapaimodel.ai/v1/models without a key returns 401 (endpoint is up and requires auth).
• Both SVG logos are square.
• Not run: pnpm verify:pre-push, Vitest, oxfmt and oxlint; I did not install the repo toolchain.
• Live check against https://api.cheapaimodel.ai/v1 with a real key, using the same OpenAI-compatible Chat Completions requests the existing path sends: GET /v1/models lists the models, and all five (deepseek-v4-1-flash, glm-5.3, glm-5.3-flash, kimi-k3, grok-4.7) answer non-streaming and streaming requests and make tool calls. Image input works on all of them except glm-5.3, which is text only. The two built-in models, deepseek-v4-1-flash (text and image) and glm-5.3 (text only), match the context and modality limits of the existing rules they fall under. I did not run ZCodium itself.