plan-improve

To improve a plan confidence score automatically, you can run an autonomous iteration loop that applies deterministic fixes like removing loopholes and injecting TDD templates, followed by LLM-driven semantic updates. The system re-scores after each iteration until the target verdict is reached.

A ralph-loop autonomous iteration is an orchestration pattern that repeatedly applies fixes to a plan and re-scores it until a target confidence verdict is achieved. It wraps state handling, pre-flight guards, and post-promise sanity checks to converge NON_SHIPPABLE or INVALID plans toward a passing score.

To fix weak imperatives and loopholes in a project plan, apply deterministic fixes that replace ambiguous language and close gaps in task definitions. This process lifts the plan's M2 confidence score by ensuring tasks meet strict validation criteria before human review.

You should use automated plan improvement when a plan is judged NON_SHIPPABLE or INVALID and you want the system to attempt auto-improvement before human intervention. It coordinates ADR alternatives and strategic guidance to converge toward a target verdict without manual edits.

Yes, the plan improvement loop validates scores after each automated fix by re-scoring the plan during the ralph-loop iteration. It only emits a PLAN_IMPROVED promise and stops the autonomous cycle when the target verdict is successfully reached.

Yes, you can coordinate ADR alternatives automatically through LLM-driven semantic improvements during the autonomous iteration loop. This process adjusts architectural decision records and strategic guidance to converge the plan toward a SHIPPABLE_WITH_CAVEATS or higher verdict.