opencode-harness

Harnesses for OpenCode 2 — drive Claude Code, Grok Build and Codex through their own CLIs on their own subscriptions, behind one OpenAI-compatible bridge. No API keys.

What it does

A harness is another vendor's coding agent, driven through its own CLI on your own subscription. No API key is used and nothing is billed per token — authentication belongs to claude, grok and codex themselves.

Harness	CLI	Models
Claude Code	claude	claude, opus, sonnet
Grok Build	grok	grok-4.6, grok-4.5
Codex	codex	account default, gpt-5.6-sol

All three are served behind one OpenAI-compatible bridge on 127.0.0.1:3012, so a harness is selected the way a model is.

"grok-build": {
  "name": "Grok Build (Harness)",
  "package": "@opencode-ai/ai/providers/openai-compatible",
  "env": [],
  "settings": { "baseURL": "http://127.0.0.1:3012/v1", "apiKey": "local" },
  "models": { "grok-4.6": { "limit": { "context": 500000, "output": 500000 } } }
}

Declaring limit is required: these provider ids have no models.dev entry, and an omitted limit makes the host invent a default — which will compact a 500k model less than halfway through its window.

Things these CLIs do that will bite you

• npm installs them as .CMD, and Windows refuses to spawn a batch shim without a shell. windowsBatchLaunch routes them through cmd.exe.
• They report API failures in-band and still exit 0. The terminal result event carries is_error / api_error_status while subtype stays "success". Check it before the exit code, or an expired login reads as exited with code 1 — or the error prose is returned as the answer.
• Partial-message streams repeat themselves. The deltas and the terminal result carry the same text; concatenating both doubles every answer.
• SSE needs a blank line between events, or clients report stream ended without finish_reason.

Usage reached

Every way a provider says the plan is spent — 402, 403, 429, resource_exhausted, usage balance exhausted, you've hit your usage limit — collapses to one line:

Usage reached — grok-build/grok-4.6. Falling over to openai/gpt-5.6-luna-fast.

No status code or provider dump reaches the user; the excerpt stays on failure.detail for logs.

License

MIT