LoopFlow

LoopFlow addresses common limitations in unattended agent workflows by enforcing a distinct reviewer persona that must output an explicit pass verdict, applying hard USD ceilings on every iteration, and maintaining a shared Markdown memory file that is appended after each run and injected into subsequent prompts. Optional git worktree isolation further prevents concurrent loops from interfering with the main working tree while every step records a resumable session identifier for later inspection. The tool runs entirely locally with no additional services or keys required beyond an existing Claude Code installation. Users initialize a project with starter loop templates for tasks such as test fixing, technical debt auditing, and documentation synchronization, then execute them directly from the command line with optional dry-run previews. Each loop step invokes Claude Code in headless mode, passing the current memory, prior outputs, and any gate feedback so the process becomes progressively more informed without manual intervention.