cekrause/harness

Agent work often spans several sessions, participants, and folders. Harness gives them one canonical project record for identity, knowledge, task claims, contributions, and checkpoints, stored outside the project itself.

Git projects, non-Git folders, and worktrees are supported. Worktrees share project context while keeping their workspace provenance distinct, so an agent can find prior results without assuming that another checkout contains the same files.

I designed five agentic skills for project identity, selective context, cooperative ownership, knowledge writing and maintenance. Agents use ordinary Markdown and their existing tools; each independently installable skill carries the same small transactional helper.

The task model records responsibilities and checkpoints for participating agents. Claims help them detect overlapping work, while acceptance, commits, and publication remain separate decisions backed by their own evidence.

Agents search and read relevant Markdown documents, then decide what to retain or rewrite. Sources, scope, uncertainty and meaningful dates remain readable prose, without a ranking engine, classifier or metadata schema for knowledge.

The helper handles project identity, atomic reservations, current handoff publication and document writes checked against observed content. Its success confirms those operations; the agent still investigates, judges and verifies the outcome. A short host instruction connects the workflow without hooks, a daemon or an external API dependency.