lsm / HyperNeo / 29602329024

81%

push

github

feat: add Kimi K3 and K2.7 high-speed support (#2243)

* feat: add Kimi K3 and K2.7 high-speed support

- Add kimi-k3 (1M context) and kimi-k2.7-code-highspeed (256k) to Kimi provider catalogue

- Route selected IDs to upstream Kimi Code IDs with correct context windows

- Support global (api.moonshot.ai) and China (api.moonshot.cn) regions with env overrides

- Update registry, frontend family/display logic, and unit tests

* fix(kimi): address review feedback on region-aware routing

- Use region-specific upstream model ID for K2.7 (global → kimi-k2.7-code, china → kimi-for-coding)

- Honor KIMI_REGION env when probing models in getModels()

- Reject OpenAI-compatible /v1 base URL overrides in buildSdkConfig

- Route moonshot-k3 alias to K3 fixed ID

- Update unit tests for new region-aware behavior

* fix(kimi): K3 upstream ID, highspeed thinking, and China endpoint

- Resolve K3 aliases to the upstream fixed ID 'kimi-k3' instead of 'k3'.
- Force thinking enabled for 'kimi-k2.7-code-highspeed' when the effective
  thinking level is off so the upstream request succeeds.
- Point the China region Anthropic-compatible endpoint to
  https://api.moonshot.cn/anthropic.
- Honor KIMI_BASE_URL when probing credentials in getModels().
- Update unit tests for the new China endpoint, K3 upstream ID, base URL
  override, and highspeed thinking override.

* fix(kimi): K3 prefix, K2.7 thinking, and no-thinking credential probe

- Probe Kimi credentials with 'kimi-k3' in getModels() so the health check
  avoids the K2.7 'invalid thinking' requirement.
- Force thinking enabled for all Kimi K2.7 models (including aliases) when the
  effective thinking level is off; keep K3 and moonshot-k3 prefixes exempt.
- Treat any ID starting with 'moonshot-k3' as K3 before the generic moonshot-2.7
  fallback in canonicalizeModelId and resolveUpstreamModelId.
- Update unit tests for probe model, regular K2.7 thinking override, and
  moonshot-k3 prefix routing.

* fix(kimi): infer ... (continued)

11626 of 16737 branches covered (69.46%)

Branch coverage included in aggregate %.

301 of 340 new or added lines in 13 files covered. (88.53%)

96640 of 117495 relevant lines covered (82.25%)

295.48 hits per line