Skip to content

CritiqueExecutor

Role: Owned execution backend — adversarial and reflexion critique loops GitHub: ProtocolWarden/CritiqueExecutor License: AGPL-3.0-or-later

What It Does

CritiqueExecutor runs iterative critique loops for AI task refinement.

Topologies:

  • Adversarial: proposer (Claude Code or Codex CLI subprocess) generates output; critic reviews in strict isolation; proposer revises until ACCEPT or round limit
  • Reflexion: agent executes; independent critic grades against criteria; agent receives only the rejection reason (never critic identity)

Worker Backend

worker_backend selects the CLI used for proposer/agent subprocesses:

  • claude_code (default) — claude --message … --no-auto-commits --output-format json
  • codex_cli — codex --approval-mode full-auto -q …

The critic always receives only the prompt and verdict JSON — never the backend identity.

Subprocess Safety

All subprocess calls use core_runner.process.safe_run():

  • start_new_session=True — child owns its own process group
  • os.killpg(SIGKILL) on timeout — kills full descendant tree
  • Timeout returns {"status": "reject", "reason": "critic timed out"} rather than raising

Invariants

  • max_rounds hard-capped at 10 (enforced in CritiqueConfig.__post_init__)
  • Critic isolation: agent/proposer identity never reaches the critic prompt
  • Topology routing: only adversarial and reflexion — no fallback topology

New Capability

This is a new executor with no direct predecessor. It provides critique-driven quality enforcement that was previously unavailable in the platform.

Dependencies

  • CoreRunner — safe_run() subprocess primitive
  • RxP — RuntimeInvocation/RuntimeResult contract
  • CxRP — BackendName vocabulary