Agent workflow

Before editing

1. Read the repository’s AGENT.md, .har/README.md, and .har/stages.json.
2. Check status before choosing a slot.
3. Launch once and record the returned work directory.
4. Read .har/CLAUDE.agent.md in that worktree for resolved URLs and the repository-specific definition of done.

har env status
har env preflight 2
har env launch 2

Use one slot per task. Separate parallel tasks use separate slots.

Occupied and failed slots

A normal launch never silently replaces an active session.

har env launch 2 --replace

Only use --replace after reviewing the current branch and worktree. If that worktree has uncommitted changes, replacement remains blocked until the owner explicitly approves --force.

A launch that failed partway is different. Preserve and resume its existing state:

har env recover 2
# equivalent to: har env launch 2 --resume

During implementation

• Edit only under the slot’s workDir.
• Let the harness own service startup, ports, databases, and process management.
• Use preview URLs returned by status or launch instead of hardcoding ports.
• Run focused project commands through .har/agent-cli.sh <id> exec ... when needed.
• Use quick verification as the feedback loop.

har env verify 2
har env status --json

Before handoff

1. Add or update automated tests.
2. Run full verification.
3. Stage exactly the state that passed.
4. Commit inside the session worktree.
5. Complete the environment.

har env verify 2 --full
git add -A
git commit -m "feat: describe the change"
har env complete 2

Any edit after full verification changes the tree hash and requires another full verify. Completion keeps the branch, so the user can push it or open a pull request.

When to teardown

Use plain teardown for abandoned work, manual cleanup, or a task that should not record a completion validation:

har env teardown 2

Branch deletion is a separate, explicit operation.