Skip to content

v3.17.0 — Multi-LLM Review SkillSet

Latest

Choose a tag to compare

@masaomi masaomi released this 22 Apr 13:10
· 313 commits to main since this release

Added

  • Multi-LLM Review SkillSet — Parallel dispatch to heterogeneous LLMs (Claude Opus 4.6/4.7, Codex GPT-5.4, Cursor Composer-2) via CLI subprocesses.
    Consensus verdict with aggregated findings. Convergence rule configurable.
  • LLM adapter extensions — codex_adapter + cursor_adapter with SafeSubprocess wrapper, provider_override routing.
  • Agent SkillSet integration — OODA loop can invoke multi_llm_review as ACT step.

Fixed

  • Adapter model label (codex/cursor no longer mislabeled as claude-opus-4-6)
  • Runtime integration (CLI auth, effort control, auto-complexity)
  • MMP keypair persistence to .kairos/keys/
  • paused_risk skip transition
  • llm_call AuthError fallback

Review

Design 2R × 4LLM + Implementation 2R × 4LLM. Runtime test passed.

See CHANGELOG for full detail.