You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fixed stalled output-limit recovery in the shared native-provider agent loop.
Empty, reasoning-only, or replayed capped responses can grow the next chunk
by up to 2× within model, configured, and context limits. Useful output has no
fixed total continuation count; two unproductive attempts stop the loop only
after no larger usable budget remains. Actual request/usage metadata prevents
repeated failed increases.
Corrected replay fingerprints for visible output, split stream blocks, code
whitespace, and equal-sized opaque reasoning. Ollama reasoning-only history
now survives normalization in its native thinking field without becoming
answer text.
Preserved partial output during context-window compaction and aligned
streamed/buffered boundary handling, including Anthropic stop reasons and
fallback request budgets. Added deterministic provider and real-query
regressions covering recovery requests and provider response formats.
Kept Homebrew, standalone, and npm native distribution free of a separate Bun
requirement; Windows native distribution remains deferred.