[nosam1998]

The LLM provider contract now has an Anthropic Messages implementation for direct structured actions and strict game tools. Requests record explicit effort and thinking policy, cache the immutable briefing, and optionally use server-side tool-result clearing while preserving local history and signed blocks.

The client uses a fixed HTTPS endpoint with sensitive credential headers, bounded bodies and timeouts, and no automatic retries, redirects, or ambient proxies. Responses retain provider content and separate uncached input, output, cache reads, and both cache-write lifetimes. Refusal/output exhaustion remain completions with usage; transport errors contain no raw provider diagnostics. Hosts must reserve budget before calls and retain uncertain billing reservations after failures.

Validation: 15 tests pass, including loopback HTTP tests for both modes, history preservation, cache accounting, invalid input/response, large and chunked responses, body deadlines, cancellation, no hidden retries, and redirect credential isolation. Clippy with warnings denied, formatting, and workspace dependency checks pass. No paid API calls were made. Match scheduling and durable budget integration follow separately.

SemVer: additive feature (minor).