fix(kimi-for-coding): update metadata for K2.8 Preview - #7036

Summary

Kimi Code upgraded kimi-for-coding in place to K2.8 Preview on September 11, 2026. The published models.dev entry still describes K2.7 Code with a 256K context window and no selectable reasoning options.

• Add K2.8 Preview lab metadata and point the existing provider entry at it, without changing the API model ID.
• Set the provider display name to kimi-for-coding, matching the stable service alias.
• Update context to 1,048,576 tokens and expose thinking toggle plus low / high / max effort.
• Preserve text/image/video input and subscription pricing conventions.
• Leave K3, K3-256K, and K2.7 Code HighSpeed unchanged.

Sources

• Kimi Code model configuration
• K2.8 Preview release announcement, September 11, 2026
• Official OpenCode integration guide

Kimi documents an unchanged kimi-for-coding request ID, up to 1M context across membership tiers, and low/high/max effort with max as the service default. Thinking-off requests are served by K2.8 Preview without thinking.

This is the first-party Kimi Code host, using the existing Anthropic-compatible provider surface: thinking.type controls thinking and output_config.effort selects effort, matching the existing K3 provider entry. No SDK, endpoint, or request implementation is changed.

Metadata boundaries

• The official documentation does not specify a maximum output-token limit. This PR retains the existing provider limit of 32,768 rather than inferring an output limit from the 1M context window. This retained value is not a newly verified K2.8 maximum.
• The new lab slug is a metadata identity, not a new Moonshot API model ID. No new provider model or endpoint is introduced.
• K2.7-specific knowledge cutoff, structured-output/temperature claims, and open-weight metadata are not inherited by the new model. No public K2.8 weight release was found; the new entry does not claim open weights or add speculative weights links.
• HighSpeed remains on its existing K2.7 metadata.

Validation

• bun validate — passed with Bun 1.3.0.
• Assertions against the generated provider JSON — passed for ID/display name, context/output limits, reasoning options, modalities, and release/update dates.
• Compared canonical generated provider JSON before and after, excluding only the updated kimi-for-coding model: SHA-256 is identical, confirming all other provider entries are unchanged.
• git diff --check — passed.
• bun test packages/core/test/generate.test.ts packages/core/test/schema.test.ts — 18 passed, 1 failed. The failing test is repository open-weight model metadata includes weights links, reporting 42 unrelated existing lab entries. The same 18/1 result and missing-links list were reproduced in a separate, unmodified worktree at base commit 06501753deb2f85a4913efbc8b0b7837108bdacc.

No authenticated inference requests or long-context runtime tests were performed.