Harness init

Harness Init

Initialize the global, file-native Harness without adding metadata to the project. Treat this skill as setup, binding, and repair. Normal recall is agent-initiated and does not depend on hooks.

Initialize

Run:

python3 scripts/init.py init --project <project-path> --json

The command:

• uses ${HARNESS_HOME:-~/.harness};
• assigns an opaque UUID;
• stores all Harness state under projects/<uuid>/;
• registers the project path as a machine-local binding;
• can also register host identity and retain a legacy Git link when available;
• creates no project files;
• is idempotent when the path already resolves to a valid project.

Use link --project-id <uuid> only when an existing Harness project has been selected. Use resolve to inspect identity without mutation. Resolution priority is explicit UUID, host binding, nearest registered path ancestor, then legacy Git identity.

--repo remains a compatibility alias for --project.

Optional lifecycle adapters

Harness does not require hooks. When a host supports silent lifecycle maintenance, install or refresh its adapters explicitly:

python3 scripts/install_hooks.py --host codex --host claude-code --json

The explicit command is the only supported installation path. It is idempotent and preserves unrelated host configuration.

Adapters must fail open and remain model-silent. They may maintain timestamps or clean temporary workspace files. They must not inject memory, create empty sessions, initialize an unrelated path, persist raw host messages, or promote durable memory.

Boundaries

• Never create .harness, .project-harness, or any other Harness file in the project.
• Never infer two projects are identical when multiple remote matches exist.
• Do not require Git, a remote, or a lifecycle hook.
• Never ask routine lifecycle questions; report an actionable warning and let the host continue.
• Read references/layout.md when inspecting or repairing storage.