OpenAI-compatible chat/completions should formally support response_format: { type: "json_object" }

[thesithunyein]

Summary

When using Venice's /chat/completions endpoint with a strict JSON-only system prompt and temperature: 0.3, occasional responses arrive wrapped in a markdown code fence:

```json
{ "decision": "buy_eth", ... }

…rather than as a raw JSON string, even when the prompt explicitly forbids fences.

### Why this is annoying

It forces every consumer to write a defensive `stripFence()` parser:

```ts
function stripFence(s: string): string {
  return s.replace(/^```(?:json)?\s*\n?/, "").replace(/\n?```$/, "").trim();
}

This is exactly the kind of boilerplate Venice's "OpenAI-compatible" promise implies you wouldn't need — OpenAI itself supports response_format: { type: "json_object" } to guarantee unfenced JSON.

Reproduction

Send 200 requests with the same JSON-only prompt at temperature: 0.3 to /chat/completions. Approximately 3–5% of responses include the fence wrapper. The rate is low enough that ad-hoc testing won't catch it, but high enough that production agents trip on it within hours of running.

What would help

1. Document the official response_format: { type: "json_object" } parameter if it's already supported, and recommend it explicitly in the docs alongside the JSON-only prompt pattern.

2. If not yet supported, add a note to the quickstart: "callers should expect occasional fence wrapping; here is a sample sanitiser ↓" and ship one in the Venice SDK.

3. Best of both: support response_format: { type: "json_object" } and document it. This is what openai/openai-node callers expect when they swap a base URL.

Self-disclosure

Filing as part of the Best Feedback track of the MetaMask Smart Accounts Kit × 1Shot × Venice AI Dev Cook-Off. Project: https://github.com/thesithunyein/delegate. The defensive stripFence() lives in src/app/api/agent/think/route.ts.