Start with a small workspace
Sign in to Aditya Code, select a recoverable project and review a first task before allowing changes.
On this page · 9 sections
Prepare the CLI and runtime
Follow your branch’s source-build setup instructions to install Aditya Code and prepare its agent runtime. This guide does not promise a public one-command installer or a model running locally. The CLI account service and the agent runtime are separate dependencies; a successful sign-in does not prove that project tools or inference are ready.
Check aditya-code --help in your installed version. Use a small test project without production data or secrets for the first run. Keep the model’s data destination and any connector permissions visible while choosing that workspace.
aditya-code --helpRecord the starting state
Inspect the worktree before allowing writes. Existing changes belong to the people who made them; they should not be overwritten or silently included in the agent’s result. A version-control checkpoint or backup makes review and recovery easier, but is not authorization to discard someone else’s work.
Choose one project directory and one small question. In a shared worktree, agree which files are in scope and preserve unrelated changes. A task that only explains code is a useful first check before requesting a patch.
git status --short
git diff --statApprove your own device sign-in
Run aditya-code login and open the link printed in the terminal. Sign in on the platform and check the code, device and request time before approving. Only approve a code you initiated yourself. If someone sent you a code, deny it; a browser sign-in should not silently authorize another person’s terminal.
Return to your terminal after approval and use whoami to confirm the account. If access fails, address that error before starting project work. Do not paste credentials into a task prompt as a workaround.
aditya-code login
aditya-code whoamiSelect the project explicitly
Use --cwd with the directory you intend to inspect, quoting paths that contain spaces. Without it, the CLI uses the current launch directory. Confirm you are not accidentally selecting the runtime repository or a broader directory containing unrelated projects.
The selected workspace controls project context and file tools, but does not certify isolation of arbitrary shell commands. Read each proposed operation and its target before approval. Do not enable automatic execution simply to get past an unexpected permission prompt.
aditya-code --cwd "/path/to/my project"Ask for an explanation first
Begin with a bounded investigation: identify a helper, trace one failing test, or explain how a route is assembled. State that no code should change. An explanation should cite what was actually inspected and distinguish an observed failure from a possible cause.
For one task, pass the prompt as a quoted argument. Review the tool activity and response instead of relying on a completion label alone. If a dependency is unavailable, preserve that limitation and use the runtime or troubleshooting guide to narrow the failure.
aditya-code --cwd /path/to/project "Trace the failing test; explain what you find without editing code."Read the outcome and checks
The task receipt separates status, elapsed time, steps, failed steps, changed files, tokens and reported verification. A file edit is not a passing test. An answered question may have no verification checks at all, and changed files can remain explicitly unverified.
Compare the receipt with your acceptance criteria. If a relevant check did not run, say so before accepting the work. For scripts, --json supplies a structured result and the CLI guide documents exit codes; do not treat exit zero as proof that the requested behavior was tested.
| Fact | Review question |
|---|---|
| Scope | Was the intended project inspected? |
| Operations | What actually executed? |
| Files | Are unrelated edits preserved? |
| Checks | Did the relevant acceptance check run? |
Authorize one bounded change
After reviewing the investigation, request one specific implementation change and its acceptance check. Keep deployment, production data and unrelated cleanup out of scope unless you explicitly intend those actions. Approval should attach to the operation shown, not to an open-ended promise that every future operation is safe.
Inspect changes afterwards and use undo --dry-run before considering recovery. Undo covers journaled file writes, not every command or external action. Later conflicting edits must be reviewed rather than removed to make a rollback look clean.
aditya-code changes
aditya-code undo --dry-runResume or watch deliberately
Use history to inspect earlier tasks and resume [id] to continue a chosen session. Recheck current files and dependencies before relying on stored observations. A restored conversation does not establish that another person has not changed the workspace.
If you want to watch activity on another device, start a session with --share and open CLI sessions on the platform. Sharing is optional and publishes messages and summaries; review their sensitivity first. The watch page is read-only and does not replace approvals in the terminal.
aditya-code history
aditya-code resume
aditya-code --shareKeep the boundaries visible
Read the CLI guide for account commands, safe scripting, plugin trust and exit codes. Read the runtime guide if the agent service cannot be reached. Internal names such as AROXA_PORT and .aroxa remain valid configuration or storage names; they are not the public terminal command.
When you finish, use aditya-code logout to clear this computer’s sign-in. If reporting a problem, include the command, a concise reproduction and redacted output. Preserve missing evidence honestly instead of describing an incomplete run as a verified fix.