feat(runtime): add OpenAI-compatible HTTP agent for BYOK

[walliai666]

Summary

The current BYOK mode only supports the Anthropic Messages API (anthropic-api). I'd like to add an openai-api HTTP agent that speaks the OpenAI Chat Completions protocol (POST /v1/chat/completions), using the same kind: 'http' extension point already established by anthropic-api.

This single addition covers a large surface of providers without any further changes, since most modern LLM APIs implement the OpenAI-compatible interface: OpenAI, Azure OpenAI, Groq, Together AI, Fireworks, Ollama, LM Studio, and any other OpenAI-shim provider.

Motivation

• Many users already have an OpenAI API key and no Anthropic key.
• Self-hosted / local inference stacks (Ollama, LM Studio, vLLM) expose /v1/chat/completions — this would make html-video usable fully offline without a CLI agent installed.
• The runtime already has a clean AgentDef.httpHandler + httpProbe interface (packages/runtime/src/types.ts), so this is a contained addition with no interface changes.

I'm happy to submit a PR for this.

[lefarcen]

Hi @walliai666! This fits the current runtime shape well — AgentDef.kind = 'http' plus httpHandler / httpProbe already carries anthropic-api, so an OpenAI-compatible sibling looks like an extension rather than an interface change.

A likely implementation path is a new runtime def next to packages/runtime/src/defs/anthropic-api.ts, then wiring it into packages/runtime/src/registry.ts and the BYOK picker/copy so the new agent can be discovered and selected. The main thing we'd want to keep an eye on is streaming compatibility across providers, since the existing HTTP path is currently tailored to Anthropic's SSE format.

A PR would be very welcome here — I've tagged this as enhancement + help wanted. If you want to take it, feel free to open one and we're happy to review.

[lefarcen]

If anyone in the community wants to take this on, comment /claim and we can route it to you for a PR. 🙌

[walliai666]

/claim

[lefarcen]

Awesome — thanks for picking this up, @walliai666 🙌

Roughly how long do you think it'll take: a few hours, 1–3 days, about a week, or longer?

[walliai666]

@lefarcen about a week

[lefarcen]

Heads-up @walliai666 — PR #75 is now open for this work. It follows the implementation path we discussed here: a new openai-api runtime def alongside anthropic-api, wired into the registry plus the BYOK picker/copy, and it also includes the stream-sanitization tweaks needed for OpenAI-style fenced output.

We'll track the implementation and review in #75, and keep this issue open until the PR lands.