v2.3.0

2.3.0 (2026-09-15)

Bug Fixes

• llm: report truncated responses instead of parsing them (0035c5d)

  • max_tokens and model_context_window_exceeded both arrive as successful responses

• llm: surface refusals and read the text block by type (a35ea35)

  • a refusal returns HTTP 200, so report its explanation instead of a missing-output error
  • select the first text block, which thinking blocks can now precede

• prompt: align PR review output format with the tool schema (030ade9)

  • cross-file issues use message and files, not description and filesInvolved
  • recommendations use category, suggestion, priority and impact

Features

• llm: default to Claude Sonnet 5 (64a80f1)

  • omit temperature for models that reject sampling parameters

• llm: support strict schema enforcement on return_json (105f8de)

  • enable it for project file selection, whose schema already complies