GOAT0.1 (Kid) · Source preview

Work on code ​

Use a capable model with native Herder tools to inspect files, make changes and run approved commands. Begin with a small disposable project while learning the permission controls.

Prepare ​

Create a Pen bound to the intended folder. In Settings → GOATed → Extensions → Built-in, enable Herder and the operations you need. Enable tools in the chat. The engine must return structured tool calls; a “Coder” label does not establish that capability.

Make a small, verifiable change ​

  1. Ask the model to list the project files and read the relevant file before proposing an edit. For example: “Read the onboarding page and shorten its introduction. Preserve the layout.”
  2. Inspect the proposed path and arguments in the approval dialog. Use Allow Once for an individual change, or deliberately choose a chat/Pen grant.
  3. Expand the tool activity group to inspect the result. Ask the model to read back the changed file. Creation cannot overwrite an existing file; edits use an exact unique fragment from the current contents.
  4. Ask for the project’s documented build or test command. Review its executable, workspace, arguments and requested networking. Commands have a separate approval system from file edits.
  5. Let the model check pen_command_status with the exact returned job ID until the process finishes. Review the exit code and output before accepting the result.

A successful final message is not proof of a successful build. Check the command result and, where relevant, the changed files or preview.

Guide or stop work ​

Use Lead to add instructions while work is active. Guidance applies after the current action; approvals remain open. Stop interrupts work but does not roll back completed changes.

Commands are non-interactive, bounded and confined to the Pen. Detached development servers are unsupported. Package downloads need separately approved networking, and host credentials or runtime configuration may be unavailable inside confinement. See the permission reference.

If a tool times out or the app loses a result, inspect the affected files before retrying. The outcome may be unknown. Text printed as JSON or a shell command is not executed as a tool; repeated text-only calls point to model-template or parser compatibility. Use Troubleshooting.