Plan Mode
Spec the work together, then hand it to a team of agents that code, review, and stage the PR for your final approval.
You stay focused on the interesting parts — what to build, how it should work, whether it’s good enough to ship. Cate runs the process in between.
Plan Mode → To Do
Section titled “Plan Mode → To Do”Click New Task in the top right (Cmd/Ctrl+N) and you land on Plan Mode. (Press Shift+Tab to swap to Pair for live coding instead.)
Describe what you want — a new feature, a bug fix, or a pointer to an existing issue. Leave it blank and the agent will ask.
The agent reads your codebase and your team’s prior research before proposing anything. It writes a spec with acceptance criteria, dependency diagrams, and scope boundaries. You review and refine together. No code is written until the spec is agreed.
For larger work, the agent creates an epic — decomposing into subtasks with a dependency graph, identifying the critical path, and flagging what can run in parallel. For bugs, it triages quickly: captures symptoms, affected code, and a test plan, then gets the issue on the board so a work agent can dig in with full context.
Connect a tracker like Jira, Linear, or GitHub Issues, and all of this is visible to your team (and all of their agents). Context compounds; your team gets sharper over time.
In Progress
Section titled “In Progress”Work agents claim issues from To Do and move them to In Progress as they work towards a PR.
Each gets its own git worktree — synced fresh from origin/main at the start of the session — its own issue, and the full spec. Agents work in complete isolation from each other and from your working copy. They read the acceptance criteria, implement the change, and must pass your tests, linting, and security gates before they can open a PR.
They document what they tried, what worked, and the rationale behind key decisions in the issue comments. That context is there for whoever comes next: the reviewer, the next agent, or you.
AI Review
Section titled “AI Review”When a work agent opens a PR, the issue moves to AI Review. A separate review agent runs a two-phase review before it reaches you:
- Spec compliance — did it build what was asked, or did it drift?
- Code quality — testing, security, and adherence to your team’s standards.
If either phase requests changes, the issue goes back to the board for another work pass. Nothing reaches you without passing both.
Ready for Review
Section titled “Ready for Review”Issues that pass AI review land here — your queue.
Click any task to load its PR. The issue’s branch is checked out in the worktree, so you can use a run config to launch your app and try it. Ask the agent to explain a design decision, run tests, or make fixes — it commits them directly to the PR branch.
Cate shows whether the PR is mergeable — CI status, required approvals, everything GitHub reports. Merge it yourself, assign reviewers, or send it back for more work.
For work that needs more than one PR, Plan Mode creates an epic.
That means a parent issue with sub-issues, an epic branch, and a draft PR. Sub-issue PRs merge to the epic branch instead of directly to main, with AI reviewers checking each one along the way. When all sub-issues are complete, the epic’s PR is marked Ready for Review and you review the full body of work as one diff before merging.
The agent handles the decomposition, the dependency graph, the branch structure, and the epic’s PR. Cate delegates sub-issues to work agents in dependency order, clearing blockers first. You do one thorough review per epic instead of reviewing each piece separately.