Handover writes the note the next session reads first, when a session or its context is ending and work continues later or with another agent. You start it by name: /agent-harness:handover in Claude Code or $agent-harness:handover in Codex.
It first reconciles the session’s beliefs with reality: git status, the branch, background jobs and project state files. Then it writes objective, where things stand, evidence, work left, open decisions, locations and one concrete next action.
What You Get
- A handover note with objective, current state, evidence, work left, open decisions, locations and a next action
- State files corrected where the disk disagreed with them
- A HANDOVER.md at the repository root when the project has no handover convention
- The exact line to give the next session, plus a list of anything left running
Common Use Cases
- Ending a long session before the context fills up
- Passing work from one agent vendor to another
- Pausing a build overnight with background jobs still running
- Resuming a stage after a break with the evidence of what was run
What Makes It Unique
The note is built from what the disk and the processes show, with this session’s own memory treated as a claim to check. The evidence section records commands, exit codes and numbers, and says so when something is unverified. The next action must be a concrete command or file, never the word continue. The skill runs only when you name it, so it never fires by itself.
Under The Hood
The skill is a single instruction file with no scripts. When it runs, the agent runs shell and git commands such as git status to check the branch and uncommitted files, reads logs of background jobs, and writes or rewrites a handover file or the project’s status file in the repository. It fetches no web pages and calls no APIs, and it tells the agent to name where secrets are kept without copying them into the note.