lsm / neokai / 25616806840

82%

push

github

fix: set autoCompactWindow for non-native providers with large context windows (#1839)

The SDK defaults to a 200k context window for auto-compact when it can't
discover the actual model capacity. This caused DeepSeek v4 (1M context)
to compact at ~17% usage (~170k tokens) instead of the expected ~80%.

Changes:
- QueryOptionsBuilder.build() now fetches the session's ModelInfo and
  passes contextWindow to buildProviderSettings().
- buildProviderSettings() sets autoCompactWindow for non-native providers
  (OpenRouter, Ollama, GLM, etc.) when the model's contextWindow is known.
- Native Anthropic providers (anthropic, anthropic-copilot) are skipped
  since the SDK already knows their correct context windows.
- The Codex bridge's hardcoded 1M autoCompactWindow is preserved.

Tests updated to verify:
- OpenRouter 1M model gets autoCompactWindow=1_000_000
- OpenRouter 128k model gets autoCompactWindow=128_000
- Native anthropic provider leaves settings undefined
- Unknown model context window leaves settings undefined

8213 of 11402 branches covered (72.03%)

Branch coverage included in aggregate %.

24 of 24 new or added lines in 1 file covered. (100.0%)

64121 of 76066 relevant lines covered (84.3%)

289.54 hits per line