Anthropic Claude provider

This Enterprise preview plugin provides a separate LLM generation interface. The plugin itself does not execute resource actions or save generated rules. The Enterprise 0.14.0 host integrates it with the shared editor's preview AI rule assistance. The host enforces connection grants, resource transmission policies, budgets and candidate validation. Operators review and manually apply a valid draft, then save it separately.

Configure the API base URL (https://api.anthropic.com/v1 for the direct API), an authorized model ID or alias and the secret API key. The key belongs to connection settings, never to model context. HTTPS verifies the server certificate; configure a trusted CA for a private gateway. HTTP requires the explicit connection option. Redirects and environment proxies are disabled.

Choose text, prompt-constrained json, or native_schema for the configured model. Native mode uses Claude's output_config.format schema transport and requires verified model/gateway support. It does not downgrade to prompt JSON. JSON syntax checks protect the transport boundary; the host must still validate the full schema and rule semantics. Unsupported schema features remain errors.

Connection checking reads the model metadata endpoint without generating text or making a paid sample request. Metadata access does not demonstrate actual generation quality or costs. Billable generation requires a separate authorized call. There are no automatic retries, streaming, model tools, retrieval or embeddings in this protocol. Calls stop on cancellation/deadline or the 60-second upper limit.

Refusals and incomplete responses remain failures. Authentication, permission, model availability, rate-limit, context-limit and network failures have distinct typed outcomes without provider error bodies or credentials. Usage omission is distinct from a reported zero. Local tests use mock servers; actual models, private gateways, schema support and billing still need live verification.