• Add a /refine command and globally persistent rlm.harness state for continual learning across sessions, with CRUD methods, a refinement log, and rollback.
• Inject a compact harness overview into the system prompt and run refine over the agent connection with a longer timeout.
• Add an edit built-in Python skill for targeted single-occurrence file edits from the kernel or shell.
• Require %%bash to be the first line of a shell cell to match the rlm-harness.