Agent CLIs

Every model call in Openworks is dispatched to a command-line agent you are already signed in to. There is no provider API key anywhere in this repository, and adding one would defeat the design.

Why not an API

Metered inference is the wrong shape for a product that acts on its own. When each run bills per token, every autonomous reaction becomes a purchase, and software that spends the operator's money unprompted has to ask first, or batch, or ration. All three close the process back up, which is the thing this is trying to keep open.

Dispatching to a CLI moves the cost onto a subscription that is already paid for. What remains is a rate limit and a wall clock, which bound a run without requiring it to be justified one invocation at a time.

It also means the deployment cannot be hosted for you. Hosting would mean running your work on our logins, which is exactly the metered product this exists to avoid.

The dispatch layer

scripts/actor.mts is the only place a provider is spawned. It builds the argv for each CLI, runs it, and falls through to the next one on failure.

The default order is codex, then antigravity, then claude. Per task type:

pr-fix differs because it is the one task that is a concrete diff against a real checkout.

Override the first provider with AI_PROVIDER. The fallback chain after it is unchanged.

Invocation flags

The argv is pinned by tests, because a wrong flag does not crash. The CLI exits zero having achieved nothing, the job never leaves summarizing, and the failure reads as the model being bad rather than the invocation being wrong.

The one worth knowing: codex exec sandboxes to workspace-write, which denies network. Every prompt the worker builds writes its result back over the network, so network access is enabled explicitly. Without it codex cannot finish a single job.

Embeddings

Embeddings are not dispatched. They run locally on CPU through ONNX, in the worker process. This is why there is no embedding key either.