Kungfu
An alpha-stage tool that keeps one unit of work moving across different coding agents, such as Codex, Claude and OpenCode, so switching agents does not mean re-explaining context or losing progress.
Kungfu keeps a single piece of work, which it calls a Work, moving across different coding agents so a developer can use the best agent for a task and a cheaper one when it does not matter, without copying context or re-explaining decisions on every switch. A per-user installer adds the kungfu command on macOS and Linux without using sudo or editing the shell profile.
A built-in Mock Agent, which needs no provider credentials, demonstrates the recovery story by running one Work through a disconnect, a crash and a recovered delivery, and a shorter scenario exercises a question, an approval and a ready-for-review result. For real work, a developer can either stay in their current agent and ask it to use Kungfu as the durable Work layer, or start the agent through Kungfu directly with a command such as running Codex, OpenCode or Amp, with Kungfu keeping the Work behind it in a project-local workspace.