self-healing

Invoke the self-healing workflow with a target scope. It runs a stress-interview pass with verifier, reviewer, and challenger subagents, sends only evidence-backed findings to a worker for fixes, then repeats the review once more before stopping.

It runs at most two review-and-repair cycles. It stops earlier if no actionable findings remain, a decision requires user input, the worker cannot stay in scope, or verification cannot be completed.

P0/P1 issues with safe mechanical fixes are applied immediately, while P0/P1 issues requiring judgment are escalated to the user. P2/P3 items are fixed only when small and behavior-preserving; weakly evidenced items are reported as remaining risk.

No. Only verified, reproducible, and narrowly actionable findings are sent to the worker. Speculative challenger questions and weakly evidenced items are reported as remaining risks instead of being treated as confirmed defects.

Avoid it for large refactors, product decisions, or security tradeoffs, since these require explicit human approval. The workflow is designed for bounded, evidence-backed fixes on a defined scope, not open-ended redesign.