fix: implement OpenAI-compatible generation for 5 AI stubs (#450)

[nikos1005]

Summary

Replaces stub placeholder responses with real OpenAI-compatible API calls for 5 AI adapter packages:

Changes

• kimi (Moonshot): https://api.moonshot.cn/v1/chat/completions
• cohere: https://api.cohere.ai/v1/chat/completions
• novita: https://api.novita.ai/v1/chat/completions
• venice: https://api.venice.ai/v1/chat/completions
• parasail: https://api.parasail.io/v1/chat/completions

Implementation details

Each adapter now:

• Uses config.baseUrl with sensible defaults per provider
• Forwards system/user messages, maxTokens, temperature, and opts.extra
• Maps choices[0].message.content, response model, and token usage
• Preserves existing dryRun behavior
• Throws descriptive errors on API failures

Fixes #450

[greptile-apps]

Greptile Summary

This PR replaces stub placeholder responses with real OpenAI-compatible API calls across five AI adapter packages: kimi, cohere, novita, venice, and parasail. The implementation follows the established adapter pattern already used by deepseek, cerebras, and others in the codebase.

• kimi, novita, parasail: Correctly implement OpenAI-compatible generation; parasail also fixes pre-existing stub bugs where the API key name was used as the model and a URL was used as the secret key.
• cohere: The DEFAULT_BASE resolves to https://api.cohere.ai/v1/chat/completions, but Cohere's OpenAI-compatible endpoint lives under a /compatibility prefix — requests will fail at runtime without this correction.
• venice: The DEFAULT_BASE is missing the /api path segment; the correct base is https://api.venice.ai/api/v1, so all requests will hit a non-existent path.

Confidence Score: 3/5

Two of the five adapters (cohere and venice) will fail every live API call because their default base URLs point to non-existent paths; merging as-is means those integrations are broken out of the box.

The kimi, novita, and parasail adapters are solid and match the established codebase pattern. However, cohere uses https://api.cohere.ai/v1/chat/completions when the correct OpenAI-compatible path is https://api.cohere.ai/compatibility/v1/chat/completions, and venice uses https://api.venice.ai/v1/chat/completions when the correct base is https://api.venice.ai/api/v1. Both errors will cause every non-dry-run request to fail immediately.

packages/ai/cohere/src/index.ts and packages/ai/venice/src/index.ts need their DEFAULT_BASE constants corrected before merge.

Important Files Changed

Filename	Overview
packages/ai/cohere/src/index.ts	Replaces stub with real API call, but DEFAULT_BASE points to the wrong path — Cohere's OpenAI-compatible endpoint requires a /compatibility prefix that is missing.
packages/ai/venice/src/index.ts	Replaces stub with real API call, but DEFAULT_BASE is missing the /api path segment — the correct Venice base URL is https://api.venice.ai/api/v1.
packages/ai/kimi/src/index.ts	Replaces stub with OpenAI-compatible call to Moonshot; endpoint, auth, and response mapping look correct and consistent with other adapters.
packages/ai/novita/src/index.ts	Replaces stub with OpenAI-compatible call to Novita; endpoint, auth, and response mapping look correct.
packages/ai/parasail/src/index.ts	Fixes multiple stub bugs (API key used as model name, URL used as secret key) and implements correct OpenAI-compatible generation for Parasail.

Sequence Diagram

sequenceDiagram
    participant C as Caller
    participant A as Adapter (generate)
    participant V as Vault (ctx.secret)
    participant API as Provider API

    C->>A: generate(ctx, prompt, opts, config)
    A->>V: ctx.secret(API_KEY_NAME)
    V-->>A: apiKey
    A->>A: build messages array (system + user)
    alt ctx.dryRun
        A-->>C: "{ text: '[dry-run]', model }"
    else live call
        A->>API: POST /v1/chat/completions (Bearer apiKey)
        alt HTTP 2xx
            API-->>A: "{ choices, model, usage }"
            A-->>C: "{ text, model, inputTokens, outputTokens }"
        else HTTP error
            API-->>A: error status
            A-->>C: throws Error("Provider STATUS: …")
        end
    end

Loading

_{Reviews (1): Last reviewed commit: "fix: implement OpenAI-compatible generat..." | Re-trigger Greptile}

[greptile-apps]

Wrong base URL path for Venice API

The Venice API base URL is https://api.venice.ai/api/v1, not https://api.venice.ai/v1. With the current DEFAULT_BASE, every request will be sent to https://api.venice.ai/v1/chat/completions, which does not exist and will return a 404. The official Venice docs confirm the base URL as https://api.venice.ai/api/v1.

[greptile-apps]

Wrong base URL path for Cohere's OpenAI-compatible endpoint

Cohere's OpenAI-compatible chat completions endpoint is at https://api.cohere.ai/compatibility/v1/chat/completions, not https://api.cohere.ai/v1/chat/completions. The current DEFAULT_BASE will produce requests to the wrong path, which will fail at runtime. The official Cohere compatibility docs confirm this path.