Code mode

Chats are built around files. Code mode is built around a repository. You register a local git repo, open a workspace for a piece of work, and run a coding agent inside it — with the same permission modes, approval cards, and review step the rest of Tidebreak uses.

Switch surfaces with the Work / Code control at the top of the sidebar.

Repos, workspaces, and sessions

Three nouns, and they nest:

• A repo is a local git repository you registered: its root path, the base branch new work starts from, and a branch prefix.
• A workspace is one unit of work on that repo. It owns exactly one git worktree and one branch for its whole life, and it carries the pull-request state for that branch.
• A session is one conversation with one harness inside a workspace.

The worktree is the important part. Each workspace is a separate checkout, so several agents can work on the same repository at once without writing over each other, and none of them touches the copy you have open in your editor.

By default worktrees are created under ~/Tidebreak/workspaces. You can change where the next one goes in settings; existing workspaces stay where they are, because git records absolute paths and moving a checkout is a repair operation rather than a rename.

Harnesses

A harness is the coding-agent CLI Tidebreak drives. Four are supported:

Tidebreak drives the harness you already have installed and signed in. It observes that authentication and never brokers it — your Claude or OpenAI subscription stays between you and the tool it belongs to. Settings → Coding harnesses runs a doctor that reports which harnesses it found and what is missing.

Each harness runs on the model you configured for it, so the model choices from Models and providers apply here too.

A turn, and the review that follows

A turn is one exchange with the agent. While it runs you see tool activity as it happens, and approval cards whenever the agent asks for something its permission mode does not cover. Denying opens a field so you can say why, and that reason goes back to the agent.

When the turn ends you get a review card: what changed, how long it took, and the diffstat. Diffs open per file and are anchored per turn, so you can read one turn's work without untangling it from the rest.

Nothing reaches your branch because a turn finished. The worktree holds the change until you commit it.

Permission modes

Code mode uses the same four modes as chats — Plan, Ask, Auto, and Allow all — with the same meaning: how much the agent may do before it stops to ask. The selector sits in the composer, per session.

Terminals

Every workspace has a terminal, opened as a drawer at the bottom of the surface. It runs in that workspace's worktree, so you are always in the checkout the agent is working in. Use it for the things that are faster to type than to ask for — a test run, a git command, a quick look at a file.

Pull requests

The review sidebar carries git state, the pull request, and its comments. From there you can push the branch, open a pull request, watch its checks, read review comments, and merge — without leaving the workspace that produced the change.

Pull-request operations shell out to your own gh CLI. Tidebreak observes that authentication the same way it observes a harness login. If gh is missing or signed out, you get copyable instructions rather than a button that fails.

What it does not do yet

Code mode is a full surface, not a preview, but some things are deliberately out of scope for now:

• No in-app code editor. Tidebreak reviews changes; it does not replace the editor you write in.
• Checkpoints are recorded, but there is no restore-from-checkpoint surface yet.
• No remote execution — harnesses run on your machine, not in a managed sandbox.
• No mobile client for watching a run.

The longer-term direction is to drop the mode switch entirely: one conversation concept, where binding it to a workspace is what makes it behave like code mode.