Work modes

For questions and codebase exploration: explanations, pasted snippets and logs, code search with grep and glob.

Explore cannot edit files. It CAN run non-mutating commands and start a dev server, so you can investigate a running app — but a write to your source is refused at the OS layer.

Turns a goal into a structured plan the agent can follow. The plan is stored as a session artifact and shown as a collapsible Plan card in chat: expand it to see the steps, revise in chat, then switch to Build.

Plan can read, search, research the web, and delegate read-only recon to sub-agents. It cannot edit files or run commands.

The full implementation surface. File edits and codegen show a diff you approve inline; terminal commands run inside a per-OS sandbox, and only genuinely dangerous ones pause to ask.

Build verifies with your project's own checks and tests before treating a task as done.

For errors, stack traces, failing tests, and unexpected behaviour. Fix forms a hypothesis — root cause plus the evidence for it — before changing anything, then makes the minimal change that tests it and re-runs your failing checks.

No drive-by refactors and no unrelated feature work. (Older docs and prompts call this Debug mode; it is the same thing.)

For quality and security passes: correctness, edge cases, error handling, consistency with your existing patterns, and test coverage.

Review is read-only by design. Instead of editing, it returns a list of recommended changes — file:line, what is wrong, and the specific fix — so you can apply them in Build or Fix.

Pick the mode in the chat area, then choose a model — an Atlarix Core model or your own provider.

The mode is fixed for the turn. If the agent needs something the current mode cannot do, it will tell you which mode to switch to rather than failing quietly.