Anthropic protocol never sends effort variant payloads (built + displayed, dropped at request construction) #46314

For custom providers using @ai-sdk/anthropic, reasoning-effort variants are computed, synced and displayed - but selecting one produces a request body containing neither effort nor thinking. The payload never reaches the wire.

Evidence (2026-08-31, current build)

1. Variants are built for Anthropic-protocol models. packages/opencode/src/provider/transform.ts reasoningEffort() returns anthropicEffort(model, effort) ?? { effort } - for models without adaptive thinking that is { effort: "<value>" }.

2. Config-declared variants load correctly. A custom provider model declaring

   "variants": { "high": { "effort": "high" }, "max": { "effort": "max" } }

   shows them in opencode models <provider> --verbose and in the effort selector of the desktop/TUI.

3. Selecting one sends nothing. With --variant high (CLI) and with high selected in the desktop UI, the outgoing request body contains exactly the usual keys - model, system, messages, tools, tool_choice, stream, max_tokens - and no effort and no thinking. Verified by tracing the HTTP body at a gateway between OpenCode and the upstream, on multiple turns, across both the CLI and desktop paths.

4. Root cause. packages/llm/src/protocols/anthropic-messages.ts fromRequest() builds the body from a fixed key list and only lowers thinking via lowerThinking() - which returns undefined unless thinking.type === "enabled". The effort provider option is never read, and adaptive payloads ({thinking: {type: "adaptive", ...}, effort}) are dropped too, since lowerThinking discards non-enabled types.

Impact

Every effort variant on every @ai-sdk/anthropic model is a silent no-op - the UI says the effort changed; the model always runs at its provider default. OpenAI-protocol models are unaffected (openai-chat/openai-responses lower reasoningEffort), which makes the asymmetry easy to miss.

Suggested fix

In fromRequest(), lower the effort provider option into the request body (pass-through is enough for gateway-style upstreams), and/or map effort → thinking budget for native Anthropic models. Passing the option through would also un-break adaptive payloads, which currently lose both halves.

Drafted and filed by Claude.