lsm / neokai / 27182873492

82%

push

github

fix: prefer provider metadata context window for non-native providers (#2082)

* fix: prefer provider metadata context window for non-native providers

Invert the default in ContextFetcher so non-Anthropic providers use
their own model metadata contextWindow instead of the SDK's assumed
capacity (which reports 1M or other generic values for GLM, Kimi,
Ollama, MiniMax, custom endpoints, etc.).

Only anthropic and anthropic-copilot continue to trust SDK-reported
values by default. Explicit preferContextWindowMetadata still overrides.

Add tests for non-native provider preference, native provider fallback,
and explicit opt-out via the flag.

* fix: use metadata for translated SDK model IDs on non-native providers

Non-native providers translate their model IDs to SDK placeholders like
'default', so getContextUsage() reports the generic SDK model name while
getSessionModelInfo() returns metadata for the real session model. This
meant metadataMatchesResponse was false and the fix from the previous
commit still fell back to SDK capacity.

Broaden metadataMatchesResponse to also match when the response model is
a generic SDK tier name (default, haiku, sonnet, opus) and the provider
is non-native. Native providers continue to trust SDK values in this
case.

Add tests for generic placeholder model matching.

* fix: update copilot isolation test to allow provider IDs in generic infra

10124 of 14808 branches covered (68.37%)

Branch coverage included in aggregate %.

9 of 0 new or added lines in 0 files covered. (NaN%)

86851 of 103132 relevant lines covered (84.21%)

291.41 hits per line