gsd-plan-review-convergence

To re-plan until HIGH risks converge, you repeatedly review an AI-generated plan with external AI CLIs and trigger re-planning when HIGH-severity concerns are detected. This loop continues until no HIGH concerns remain.

Yes, you can use local LLMs for plan review. The workflow supports local OpenAI-compatible servers like Ollama, LM Studio, and llama.cpp, alongside external backends like Codex, Gemini, and Claude.

The best way to gate plan approval on HIGH severity issues is to loop planning and review phases, checking workflow flags for HIGH concerns. The process stops when no HIGH concerns remain or a maximum cycle limit is reached.

Yes, multi-agent plan review supports cycle limits and stall detection. It converges when no HIGH concerns remain or stops after a configurable maximum number of cycles, while preserving workflow gates like stall detection and escalation.

Your LLM reviewer might keep flagging the same HIGH concerns during replanning if the feedback is not properly integrated into the next planning phase. The loop requires re-running the planning phase using the review feedback to resolve issues.