Skip to content

Workflows

Follow a whole task from start to finish instead of one feature at a time. Each section below is a goal, the commands in order, and what to expect. Each links to the page that explains the details.

Leave an agent working and come back to approvals

Section titled “Leave an agent working and come back to approvals”

Goal: give an agent a task, walk away, and answer only what it could not decide alone.

  1. Commit your work, so you can undo anything the agent changes.

  2. Choose how much to be asked. The defaults suit this: the AI judge decides routine requests and holds the rest. Make commands stricter if the project can run anything dangerous on your machine:

    Terminal window
    odette gates commands manual
  3. Start the agent with its task:

    Terminal window
    cd ~/code/my-project
    odette --prompt "make the failing tests in ./api pass"

    Claude Code starts in plan mode. Read its plan and accept it in the terminal; after that it works on its own.

  4. Open the dashboard (odette dashboard) and click the bell in the rail to allow desktop notifications. You are alerted when an approval is waiting and when the session finishes its turn (Dashboard).

  5. When you come back, the Overview lists what is waiting. Allow or deny each request. Allow a host In this project if the agent will need it again (Approvals).

  6. Agents can also ask you a question without stopping. Those land in the Inbox; your reply reaches the agent when it is idle (Approvals).

What to expect: a request nobody answers is denied after approval_timeout seconds (300 by default), so a long absence turns into denials, not into permission. Raise the value if you want agents to wait longer (Configuration).

Goal: work on separate changes at once without the agents overwriting each other.

  1. Start each agent in its own git worktree, in separate terminals:

    Terminal window
    odette --worktree=fix-login --prompt "fix the login redirect"
    odette --worktree=add-export --prompt "add CSV export to the report page"

    Each gets a worktree under .claude/worktrees/ in the project, on its own branch.

  2. See them all:

    Terminal window
    odette sessions

    The dashboard’s Overview shows one tile per session.

  3. Answer approvals as they arrive. Approvals from every session land in the same Inbox.

  4. When an agent is done, review its branch on your machine and merge it as you normally would. A session cannot push: it holds no GitHub credential. Let the agent run git push on your machine through the command gate, or push yourself (Security).

  5. Stop a session you no longer need with odette sessions stop <name>. Its conversations are kept.

To have one agent direct the others, use a pod (Pods).

Goal: every teammate gets the same sandbox, egress rules and tools from the repository, and nobody shares a credential.

  1. In the project, create the shared files and apply the setup your team needs:

    Terminal window
    odette init --kit go # or any kit you use; see Kits

    Review what it wrote (Kits).

  2. Commit .odette/, including kits.lock, but not the *.local.json files (odette init adds them to .gitignore). Hosts the team always needs go in .odette/domains.allow (Configuration).

  3. Each teammate installs odette, runs odette setup once and then odette in a clone. Each signs in with their own Claude account and sets their own dashboard password. Nothing secret is in the repository.

  4. If your organization restricts Claude Code with managed settings, odette applies that policy to every session and says so at each launch. Check what is in force with:

    Terminal window
    odette policy

    Where the policy forbids what odette needs, odette refuses to start and says what to ask your administrators for (Security).

  5. To change the shared setup later, edit or re-apply the kit, review the diff with git diff, and commit. Teammates get it on their next launch.

What each machine keeps to itself: its Claude homes, its stored secrets and its *.local.json files (Configuration).

Theme