Switch generation to RunPod completions endpoint and make model optional

[deepfates]

Motivation

• Use the provided RunPod completions endpoint without requiring a provider-specific model string.
• Remove OpenRouter-specific header/credential assumptions for the new completions path.
• Allow generation requests to omit model and still produce output by falling back to a configured model.

Description

• Updated the OpenAI-compatible client to point at https://xob3rm6bnyl1j1-8000.proxy.runpod.net/v1 and to use config.completionsApiKey instead of the previous OpenRouter key and headers (server/apis/openaiClient.ts).
• Replaced the required OpenRouter config flow with a new LOOMPAD_COMPLETIONS_API_KEY backed completionsApiKey field in server/config.ts (defaulted to a safe string for non-production use).
• Made model optional in GenerateRequestBody and changed validation so model is only validated when present (server/apis/validators.ts).
• Updated generateText to resolve a model config from the provided model or fall back to the first configured model and stopped sending model in the upstream /completions request payload (server/apis/generation.ts).
• Aligned the judge code to read the renamed completionsApiKey config field (server/apis/judge.ts).
• Added and adjusted tests to cover the model-optional behavior in server/__tests__/validators.test.ts.

Testing

• Ran bun test server/__tests__/validators.test.ts and all validator tests passed.
• Ran the full test suite via bun test and all tests passed (no failures).

---

Codex Task

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1649640c8e

ℹ️ About Codex in GitHub

Codex has been enabled to automatically 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 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Return client error for unknown model IDs

validateGenerateRequestBody only checks that model is a non-empty string, so a request like { model: "does-not-exist" } reaches generateText; with the new fallback logic, selectedModelId is set to that invalid string and getModel returns undefined, which now returns HTTP 500 ("No models configured") instead of a 4xx invalid-model error. This turns a user input mistake into a server error and can trigger misleading alerts/retries for bad client requests.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Fail fast when completions key is missing

validateConfig now always supplies a placeholder ("not-required") when LOOMPAD_COMPLETIONS_API_KEY is unset, but /api/judge still calls OpenRouter (judge.ts) with apiKey: config.completionsApiKey; in production this allows boot to succeed and then makes judge requests fail at runtime with auth errors. The previous behavior guarded against this by requiring a real key outside development.

Useful? React with 👍 / 👎.