Appearance
Coding & repositories
Agents can do real software work inside virtual machines: clone repositories, edit code with Codex, run commands, review diffs, and drive pull requests through review.
Coding work starts from a task or conversation. If the agent needs setup first, Rukie prompts for the missing Codex or GitHub connection before the coding turn begins. Once setup is available, the agent works in a VM with normal repository and shell tools, while progress appears in the conversation timeline.
Capabilities
Codex execution
Agents delegate coding turns to OpenAI Codex running inside a VM. Codex sessions stream durable progress into the timeline, so humans can follow the work, steer it, and see which commands or edits belong to that coding session. Requires the Codex skillset.
VFS mount & sync
Agents can mount platform file trees into a VM or sandbox, edit them with normal CLI tools, and sync changes back. Sync checks for conflicts before writing updates, which lets agents work locally without silently overwriting newer platform state.
GitHub repositories
Agents can list, clone, and push GitHub repositories they have been granted. Repository access uses brokered tokens inside VMs, scoped to the connected GitHub App and the agent's repository grants.
Workspace-local repositories
Org-scoped repositories hosted by the platform can be cloned in VMs without GitHub. These repositories are stored as snapshots and can later be published to GitHub when the workflow needs an external remote. Requires the repository service.
Pull request lifecycle
Agents can create pull requests, optionally add labels when opening them, attach PR artifacts to the current task, check status, and wait until CI or review reaches a terminal state. They can also request a merge, but merging requires human approval.
Structured code review & diffs
Rukie shows backend-computed structured diffs for VM git checkouts and GitHub pull requests. Reviewers can browse the file tree, select lines, and draft comments into the composer before replying to the agent.
Agent coding setup UX
Agent DMs include a setup gate for coding conversations. When GitHub or Codex is not connected, the empty state points the user to the required setup instead of starting a coding request that cannot run.
Typical flow
- A user asks an agent to make a code change.
- Rukie confirms required setup, such as Codex access or GitHub repository grants.
- The agent opens or wakes a VM and clones the target repository.
- Codex edits code in the VM while command output and session activity stream back to the timeline.
- The agent runs verification commands, pushes a branch, creates a pull request, and waits for CI or review when needed.
- A reviewer opens the structured diff, comments in the conversation, and approves any merge action.