Add Kimi K3 support across applicable providers

[navedmerchant]

Problem

Kimi K3 is not available as a correctly configured model in the Moonshot and OpenCode Go providers. Users need the model's long context, multimodal input, reasoning, and automatic prompt caching to work without provider-specific failures.

Context

This affects users selecting Kimi K3 directly through Moonshot or through the curated OpenCode Go catalog. Model parameters were researched from Moonshot's platform docs (https://platform.moonshot.ai/docs) and the OpenCode Go model table (https://opencode.ai/docs/go).

Desired behavior

Add Kimi K3 to the Moonshot and OpenCode Go model registries with its confirmed capabilities and pricing, and keep related preserve-reasoning detection in sync. Do not add speculative static entries for providers whose Kimi K3 IDs have not been confirmed.

Request checklist

• Existing issues and discussions were searched for duplicates.
• The request describes a specific model-support gap and its user impact.

Acceptance criteria

• Direct Moonshot model ID kimi-k3 is available with a 1,048,576-token context, default maximum output of 131,072 tokens, image support, automatic prompt caching, and pricing of $3 input / $15 output / $0.30 cache read.
• The curated OpenCode Go catalog includes kimi-k3 with the same confirmed capabilities and pricing, routed via the OpenAI-compatible wire format (not Anthropic /v1/messages).
• Reasoning support is declared: always-on thinking with configurable reasoning_effort (low/high/max, default max) and preserveReasoning in both registries, with kimi-k3 added to LITELLM_PRESERVE_REASONING_MODEL_IDS.
• Focused tests cover the model catalogs (opencode-go.test.ts, lite-llm.test.ts); the full pnpm test suite passes (6851 tests).
• Fireworks, Baseten, Bedrock, and Vertex receive no guessed static entries until provider-specific IDs are confirmed.
• Dynamic or user-supplied catalogs such as OpenRouter require no static model addition.

Resolution

Implemented in #996.

[zoomote]

Progress update: #996 (draft) covers the first two acceptance criteria groups.

Done in #996:

• Direct Moonshot kimi-k3 entry: 1,048,576-token context, 131,072 default max output, image support, automatic prompt caching, $3 input / $15 output / $0.30 cache read pricing
• Curated OpenCode Go kimi-k3 entry with the same capabilities and pricing (OpenAI-compatible wire format, not Anthropic /v1/messages), plus supportsMaxTokens for the output slider
• kimi-k3 added to LITELLM_PRESERVE_REASONING_MODEL_IDS (preserveReasoning: true in both registries)
• No speculative entries added for Fireworks, Baseten, Bedrock, or Vertex; OpenRouter untouched
• Full pnpm test green (6851 passed)

Still open from the acceptance criteria:

• Vercel AI Gateway moonshotai/kimi-k3 normalization (and Zoo Gateway inheritance)
• Request-serialization requirements: omitting temperature, always requesting max reasoning effort, and verifying reasoning_content preservation through tool-call continuations on the applicable request paths (the PR currently sends temperature: 1.0 via defaultTemperature and defaults reasoningEffort to max rather than forcing it)
• Real-SDK HTTP serialization tests

Note: the Go-plan context window in the PR is 1,048,576 to match Moonshot's documented spec; the live Go /v1/models fetch merges the gateway-advertised value (1,000,000) on top at runtime.