OpenCode Go: add MiMo V2.6 Flash and Pro model configurations #1997

Name of model: MiMo V2.6 Flash (mimo-v2.6-flash) and MiMo V2.6 Pro (mimo-v2.6-pro), through the built-in OpenCode Go provider.

Model creator: Xiaomi / MiMo.

Link to model: OpenCode Go model/endpoint table, Go pricing, and Xiaomi MiMo deep-thinking documentation.

Problem and user benefit

Both V2.6 IDs are returned by the live Go catalog, but neither has an entry in the native Go registry. They can be discovered and use the documented Chat Completions route, but receive generic limits and lack model-specific pricing, reasoning controls, and reasoning-preservation configuration.

The goal is accurate, usable V2.6 support through Go, without requiring users to manually guess capabilities or change to the separate Xiaomi provider. This is one task for the two variants of the same model generation, not two independent model issues.

Concrete evidence

Audited upstream main at ecab5198, October 10, 2026. Relevant sources and tests are byte-identical to the locally executed checkout at 50e7aea3.

• Direct execution of the native metadata lookup returned no entry for either ID.
• The live catalog observed during the audit returned IDs and administrative fields only. The unknown-model parser therefore supplies 200,000-token context, 32,768-token output, no images, no prompt-cache capability, and no curated pricing/reasoning settings.
• Go request construction gates reasoning-preserving message conversion on model metadata. Xiaomi's linked documentation explicitly includes both V2.6 variants and requires complete reasoning pass-back for tool-call continuations in thinking mode. Confirm the gateway behavior; no authenticated Go rejection is claimed here.
• The catalog-completeness test lists V2.5 but not either V2.6 ID.

Desired behavior and acceptance criteria

• Add both IDs to the Go-native configuration registry with independently confirmed context/output limits and input capabilities. Do not copy V2.5 limits solely because the generation names are similar.
• Configure reasoning capabilities to match the Go contract; preserve prior reasoning in tool-call continuations where required. A focused provider regression test verifies the outbound history after reasoning, a tool call, and a tool result.
• ID-only catalog responses resolve those configurations rather than generic fallbacks, with tests for both models in the registry/fetcher suites.
• Keep Chat Completions routing; do not migrate these models to Messages or Responses.
• Use current Go pricing per 1M tokens, checked October 10, 2026: Flash input/output/cache-read = $0.14/$0.28/$0.0028; Pro = $0.435/$0.87/$0.003625. Do not invent a cache-write rate or copy direct Xiaomi PAYG prices.
• Existing Go models and the focused provider/fetcher/registry suites continue to pass.

Related work and scope boundaries

Issue #1761 and PR #1748 cover the separate direct MiMo provider, its model discovery, and its default model. This request is specifically for the OpenCode Go configuration and gateway contract. Reuse confirmed capabilities when appropriate, but keep Go pricing separate. Do not assume Xiaomi's V2.5 retirement schedule applies to Go; changing Go's default model and adding ultraspeed are outside this task.

Searched open and closed issues/PRs by the V2.6 IDs/family and OpenCode Go; no direct Go-specific duplicate was found.

Verification performed

Direct offline registry execution confirmed both IDs are uncurated. Existing focused tests passed: 41 registry tests and 93 provider/fetcher tests. This is baseline verification, not new regression coverage. No authenticated inference request, installed-extension reproduction, or V2.6 token-limit measurement was performed.