kungfu

Kungfu targets the cost of switching coding agents: copying context, re-explaining decisions and checking what got lost. It keeps a Work outside any agent session, recording objective, progress, evidence, next action and completion state, under a Project. Each agent run is an Attempt recorded under that Work, including failures, so a disconnect or crash does not erase what happened.

You can stay inside your current agent and ask it to run kungfu agent brief, or start one through Kungfu with kungfu run codex, claude, opencode or amp and keep its native console. Kungfu stops a second live agent from writing to a Work another one owns, and it separates finishing from approving: an agent produces a result and evidence, while independent review and Kungfu settlement decide completion. A deterministic mock agent demonstrates the recovery path without provider keys.

The project is labelled Alpha and the README says regular onboarding still needs a real agent for independent review. It stores state in a project-local .kungfu/ directory that should not be deleted or committed wholesale. Optional TUI and GUI views show Work across sessions.