Composer

The graph composer is the main place where you send work to agents from the graph.

Prompt Box

Write the request in the prompt box. Press Enter to send, or Shift+Enter to add a new line.

Use @ to reference files from the repository. Agentree opens a file picker and inserts the selected path into the prompt.

Use $ to reference skills discovered by the selected agent provider. Agentree opens a read-only picker for Codex and Claude skills, then inserts the selected skill name into the prompt as plain text. Agentree does not install, import, delete, or sync skills.

Target

The top left target chip shows where the message will land.

• Reply to sends the message as a reply to the selected graph node.

Orchestration

The orchestration controls choose whether the prompt is sent as a normal agent request or as a natural-language orchestration request.

• Normal send starts one agent request from the selected target.
• Orchestration starts a real orchestrator run that can create and control normal child runs.
• Prompt templates insert saved @node or orchestration prompts into the composer.

Use orchestration when the task benefits from planning, multiple independent attempts, review, or follow-up child runs. Complex orchestration previews its planned child runs before they start.

Model

Click the selected model chip to choose a different model. Use the + button beside the chips to add another model.

Use the Codex and Claude tabs in the model selector to browse one provider at a time.

For normal sends, Agentree starts one run by default. When multiple runs are selected, Agentree starts them in parallel from the same prompt and context. The model chips show how many runs each model will start, with controls to increase or decrease the count.

Use multiple runs when you want alternative implementations or independent debugging attempts. Use a single model when the task is small or you already know which agent should handle it.

Code Base

The code base button chooses the base commit for the run. Agentree infers a base from the selected node, parent output, tracked refs, or workspace base, but you can override it.

Use this when you want the agent to work from a specific branch, commit, or ancestor instead of the inferred base. The picker supports searching branches, commits, and SHAs.

Context

The context button allows you to filter conversation history before sending to agent.

Use inferred context for normal continuation. Use fresh context when earlier conversation would distract the agent. Use start and end context when a long or merged branch needs manual trimming.

Send, Steer, and Cancel

The primary button changes based on the current state:

• Send starts a normal agent run.
• Confirm orchestration preview starts the planned child runs after preview.
• Steer sends guidance to an assistant that is already running.
• Edit and run reruns from an edited user message.

When the selected assistant is running, the cancel button appears. Use it to stop that agent run before sending new steering or choosing a different branch.