Atomic Chat inference backend

CoWork OS can optionally use an already-running Atomic Chat instance as an inference backend. This is an inference adapter, not an Atomic agent integration.

What CoWork owns

CoWork continues to own the task loop, system prompt, tool visibility, permissions, approvals, network/file restrictions, tool scheduling, evidence, cancellation, retries, and task persistence. Atomic Chat only receives the OpenAI-compatible model request and returns a completion.

CoWork does not start or stop Atomic Chat, install models, manage its model picker, or import Atomic Agent's tool loop.

Setup

1. Start Atomic Chat and load a model.
2. In CoWork Settings, select Atomic Chat (local).
3. Leave the API key empty unless Atomic Chat proxy authentication is enabled.
4. Keep the default endpoint http://127.0.0.1:1337/v1, or enter the endpoint exposed by the running instance.
5. Refresh models and select the exact model ID returned by Atomic Chat.

The adapter probes GET /v1/models with a five-second discovery deadline. The Settings refresh action preserves the saved model selection and reports whether the endpoint is unreachable, rejected authentication, returned a valid empty catalogue, returned an incompatible payload, timed out, or returned exact model IDs. An auto selection is resolved to the first reported loaded model immediately before inference. Inference requests are non-streaming for this initial integration and have a 60-second deadline. These defaults can be overridden by the provider implementation in tests; they are not a model-installation or server-startup timeout.

Refresh Models updates the draft model list only after a valid response. If the endpoint is unavailable or returns an empty/incompatible catalogue, CoWork keeps the existing saved model selection instead of silently replacing it. Test Connection performs a small non-streaming request against the selected model, so it can reach the server successfully while still reporting that the server is temporarily busy or that the model is unavailable.

Compatibility and failure behavior

Atomic Chat exposes an OpenAI-compatible /v1/chat/completions endpoint. CoWork reuses its existing message and tool-call translation, then sends stream: false. Tool arguments remain subject to CoWork's normal validation and permission pipeline; malformed arguments become rejected tool results and are never dispatched.

Discovery reports distinct outcomes for success, valid empty responses, authentication rejection, invalid payloads, offline endpoints, cancellation, and timeout. Inference errors are typed as cancellation, timeout, unreachable, authentication, model unavailable, context limit, unsupported parameter, invalid response/tool call, or temporary busy state. Cancellation is terminal for the request and is not silently replaced with a fallback answer.

Atomic Chat's API surface can change between releases. Treat the selected Atomic Chat version as part of qualification and re-run the local fixture/live probe before enabling it for unattended tasks.

Troubleshooting

• No models or endpoint unavailable: start Atomic Chat, confirm its local API port, then use Refresh Models. CoWork does not start the server or install a model for you.
• Authentication rejected: leave the key empty for an unauthenticated local server, or configure the proxy key in CoWork. The adapter sends the key as both bearer and X-Api-Key authentication.
• Temporarily busy: the server was reachable but could not accept the inference probe at that moment. Retry after the active generation finishes.
• Invalid response or tool call: the response did not match the OpenAI-compatible contract. CoWork rejects malformed tool arguments rather than dispatching them.

The initial integration has been validated for model discovery and typed connection errors. A successful full-generation benchmark, streaming support, and universal tool/model compatibility remain qualification work; do not infer those guarantees from a successful /v1/models response alone.