Shared project memory for agents
Keep current state, decisions and validation history in separate project files
Shared project memory for agents should be easy to recover without losing important decisions. The agent-handoff package creates an entry document and separate files for current state, risks, backlog, decisions and work history, so the next session can read what matters.
The workflow operates on an authorised project. Existing state should not be replaced with blank templates, and project-rule changes need review. Maintenance can archive eligible history and compact current state; it does not establish that tests passed or replace independent backups.
What it does
- Create a clear project entry point
- Separate current state from long history
- Preserve decisions and unresolved risks
- Record validation results and limitations
- Merge marked project guidance blocks
- Archive eligible records to control file size
Install and use
- Load the skill in Codex, Claude Code or DSH with authorised project access. Bundled helpers need Python 3.10 or newer; documented manual templates are an alternative.
- For example: “Create a continuation pack, preserve existing rules and label checks that have not run.”
- Review generated files and guidance blocks, replacing unknowns with evidence from the project.
- Start the fresh session from the entry document and current state. Optional Claude reminders need Node and separate setup; they do not update the handoff files themselves.
FAQ
Source
Made by WeirdSky924. Original skill: agent-handoff, published on GitHub under the MIT licence.
View on GitHub