Pick an issue. An agent builds it on its own branch.

App guide

Task Implementer at a glance

The package is packages/task-implementer. The window has three tabs: Tasks, Dependencies and Plan.

Read this first

The agent runs as you, with permission prompts turned off. Task Implementer starts

claude -p <prompt> --output-format stream-json --verbose --dangerously-skip-permissions

in the folder the app was started from. The agent can read and write files, run commands and push without asking. The app does not create a sandbox, a separate account or a virtual display. Use it only on repositories and machines where that is acceptable.

Before you start

• The Claude Code CLI, claude, on your PATH and signed in. It is the only agent this app runs; the active provider in RobOS Agents does not change it.
• A task server in RobOS Task Servers. For GitHub, the GitHub CLI gh, signed in. For Jira, the robos-task-client package and your API token in pass at the server's token_pass_path. Without one, the window shows Open Task Servers.
• A local clone of the repository, with origin pointing at it, and push access.
• Issues to work on. Task Planner creates them, or use ones you already have.

Start it in your clone

The agent works in the folder Task Implementer was started from. The window has no folder picker, so start it from your clone:

cd ~/src/my-app
~/robos/scripts/dev-launch.sh task-implementer

Use your own clone and RobOS paths. If you open the app from the launcher instead, put the clone's full path in the extra-context box so the agent knows where to work.

Run a task

1. Load the tasks. In Tasks, choose Open, Closed or All. GitHub lists up to 50 issues with gh issue list; Jira runs a search for up to 50. Each row shows the persona the app picked, and a lock if the task is blocked by unfinished work.
2. Select one. Start with a task nothing else depends on. The header shows its key, the persona and Open, which opens the issue in your browser. A blocked task shows Blocked by with links to the blockers.
3. Check the persona. Change it in the drop-down if it is wrong. Directives shows the guidance the agent gets for that persona. Edit it for this run, or Reset to Role Default.
4. Add hints in the extra-context box: files to look at, the command that runs the tests, things not to touch, the branch name you want.

5. Click Start Agent and watch Agent Output. The agent's messages stream in as it works. Stop ends the process; Clear empties the pane.
6. Read the summary at the end: the branch, the base branch, what it ran to check the change, and where any evidence is.

You can start agents on several tasks at once, one per task. They all work in the same folder, so run one at a time per clone, or start another copy of the app from a second clone.

What the agent is told

The prompt is built from the task, the persona and your context, followed by the RobOS writing and commit rules and this delivery workflow:

The agent also gets ROBOS_AGENT_ROLE and ROBOS_AGENT_MODE in its environment, so scripts it runs can tell which persona started them.

Review the branch

When the agent has finished, click Review branch… and choose the clone the agent worked in. The Agent Code Review window opens on that branch, diffed against your default branch. Uncommitted changes are not part of the diff. It refuses while an agent is still running on the task, so wait for Agent Output to finish.

From there you read the diff, ask for changes in plain language, check the evidence, take a walkthrough, and click Create PR only when you are ready. The steps, with screenshots, are in Review the branch before it becomes a pull request.

Personas and directives

A persona is the role the agent takes. The app picks one for each task, first match wins:

1. a persona or role recorded on the task itself;
2. a role: label, such as role:backend-dev or role:frontend-web-dev;
3. words in the title, body and labels: Godot or sprite for Game Developer, migration or Kafka for Data & Storage Engineer, React or component for Frontend Web Developer, Helm or Dockerfile for DevOps & Cloud Engineer, ADR or architecture for Software Architect, API or endpoint for Backend Systems Developer.

The personas are the same eight that Task Planner uses. Agent Personas edits their system prompt, planning directive, implementation directive and development guidance; + Add Custom Persona adds one, and Restore Defaults puts the originals back. Changes are shared by both apps.

Two personas work with a display. Non-Headless Developer runs the agent with DISPLAY set to :99 (or ROBOS_XVFB_DISPLAY), and Human + Agent Desktop Co-Pilot uses your current display. Their prompts tell the agent to use an isolated virtual display, temporary storage and IDE bridge tools. Task Implementer does not start a virtual display or mount storage itself, so these only work where your environment already provides them.

Dependencies tab

A tree of the loaded tasks: each parent with its children, a lock and Blocked by on tasks waiting for unfinished work, and Blocks on tasks others wait for. Search by key or title, use Expand All and Collapse All, and double-click a task to open it in Tasks.

Parents and blockers come from Jira's parent field and issue links. GitHub issues are listed without them, so on GitHub the tree is a flat list.

Plan tab

The AI Plan Generator suggests an order for the tasks you loaded in Tasks.

1. Write anything the order should respect in Context & Suggestions: a deadline, a task to do first, work to leave for later.
2. Click Generate Plan. The app runs claude -p with each task's key, title, status, priority, type, labels, parent, blockers and the start of its description. The agent's progress shows under Generation Log.
3. Implementation Plan lists every task once, first to last, with one sentence on why it sits there.

The prompt is written for Jira, where priority, issue types, parents and blockers are available. On GitHub it has only titles, labels and descriptions to go on. The plan is saved per task server and shown again the next time you open the tab. This run does not use --dangerously-skip-permissions.

Where things are saved

Limits and fixes

• The agent works in the wrong folder. Start the app from your clone, as in Start it in your clone, or give the full path in the extra-context box.
• Nothing happens on Start Agent. Run claude --version in a terminal and sign in. Another agent may already be running on the same task.
• "No task server configured." Click Open Task Servers, add one and save, then reopen Task Implementer.
• The wrong persona is chosen. Pick the right one in the drop-down, and add a role: label to the issue so it is chosen next time.
• The push fails. The agent pushes with your Git credentials. Check git push works from the clone.
• Only 50 tasks show. The list stops at 50. Close finished issues, or filter by state.

Next

Reference: Task Implementer in the classic docs · Source code