Build
A real tool-using loop over your repository, on your machine. It reads before it writes, verifies before it claims, and leaves a trail you can undo.
How a run goes
- Plan. It writes a task plan you can watch, and ticks steps off. If it notices it is planning without acting, it re-plans instead of grinding to the step limit.
- Act. It reads files, searches, edits, runs commands, drives git — streaming each action into the transcript as it happens.
- Verify. It runs your project's build and test ladder, reads the failures, and fixes them.
- Reflect. Before finishing, it re-reads its own diff against what you actually asked for and fixes what it missed. It will not declare victory while plan steps remain outstanding.
- Remember. Durable facts about the project — the real test command, a convention, a gotcha — persist, so a second run starts already knowing them.
What it can do
| Area | Tools |
|---|---|
| Files | Read, write, targeted snippet edits, directory listing, project-wide search |
| Understanding code | Go to definition, find references, file outline, project-wide rename — tree-sitter backed, not grep |
| Running things | Shell commands, background jobs, dev servers with captured logs, HTTP health checks, git |
| Proving it worked | Runs your verification ladder; drives a real headless browser to assert a feature is actually on the page |
| Research | Web search and fetch, so it reads the current API docs instead of guessing from training |
Things worth knowing
Snippet edits, not rewrites
The agent is told to prefer replacing an exact snippet over rewriting a file. It is dramatically cheaper in tokens, and it means the agent edits your file rather than replacing it with its idea of your file.
Checkpoints and the Diff tab
A git checkpoint is taken before every run, so a whole run is undoable as a unit. The Diff tab shows red/green line counts per file with Keep and Revert, and flags likely leaked credentials in changed files before they land.
Many repositories at once
Link sibling repositories from the Explorer and the agent can reason across them — trace a call from your app into a shared library. Writes stay in the primary project.
Workflows
Named, reusable recipes that run the agent with checkpoints, verification, policy and budget applied: release readiness, security remediation, a migration from one thing to another, an incident assistant that root-causes a pasted log. They are .toml files you can commit and share.