Conduct your coding agents.

New in 0.1.1. An orchestrated run adds a Main task that manages every other task. Each task runs only the steps it needs, each step can use its own specialist agent, executor, model and effort, and the whole plan shows up as a nested dependency flowchart in the terminal and in VS Code.

Guided setup (maestrun init → Orchestrator) uses agent coordination. Either strategy runs tasks in sequence or in parallel, and coordinator calls use your executor account like any other step.

### Task report: Research and write a report

- Depends on: none
- Steps: custom:Research, custom:Write report
- Agents: {"custom:Research":".claude/agents/research.md"}

Gather evidence, then write the report.

### Task check: Fact-check the report

- Depends on: report
- Steps: custom:Fact check

Check every claim against its source.

---
model: your-exact-model-id
executor: claude
effort: high
---

Research using primary sources and cite the evidence.

• Depends on schedules, Parent groups. A task starts only after all its prerequisites finish.Parent only places it in the navigation tree; without one, a task appears under its longest prerequisite chain.
• Coordinated specialist repairs. With agent coordination, a task that still needs changes after local repairs goes back to the coordinator, which adds specialist repair tasks with independent verification. The original task retries after them, within execution.maxRepairs.
• Saved, resumable state. The source plan, generated tasks, coordination rounds and step results live in .maestrun/run.json. maestrun run or a background reconnect resumes with completed tasks kept.
• Sequential or parallel worktrees. Orchestration and scheduling are independent: parallel tasks keep their own Git worktrees and concurrency limit.

F

Dependency flowchart: Main task first, every task once after all its prerequisites

D

Task boxes only, or every task with its complete step flowchart inside

V

Horizontal (left to right) or vertical (top to bottom), steps included

Shift+←/→

Pan across the tree; PgUp/PgDn scroll

Enter

Results and logs of the selected task

Try it: maestrun sample orchestrator-plan (6 predefined tasks, 11 worker stages) · maestrun sample orchestrator-agent (the agent creates the tasks; it targets 2 coordination calls plus 11 worker calls, but its decisions set the real count). Both run in sequence with no automatic repairs, on your executor account. Navigating the diagram makes no model calls.

Let Planning think hard while Tests stay quick. Each stage gets its own executor, exact model and effort level, or you use one for the whole workflow. Edit the stages below; the config.json Maestrun reads updates as you go. Add as many steps as you need: QA, Review, or your own written in Markdown.

{
  "version": 1,
  "stages": [
    {
      "id": "planning",
      "executor": "claude",
      "model": "your-claude-model",
      "effort": "high"
    },
    {
      "id": "implementation",
      "executor": "codex",
      "model": "your-codex-model",
      "effort": "medium"
    },
    {
      "id": "testing",
      "executor": "copilot",
      "model": "your-copilot-model"
    }
  ],
  "execution": {
    "mode": "sequential"
  }
}

Same thing from the terminal: maestrun config stages planning effort high

The plan is a Markdown task list you can read and edit before anything executes. Each task then walks through its stages; when a model reports a failure, Maestrun restarts at Implementation, up to five times by default. If a task still fails, independent tasks keep going, dependants wait, and the failed task stays on screen with its error. When a task needs you (a decision, an approval, a login), it pauses with its question while the rest keep running; answer it and it resumes the same stage. Completed work is checkpointed, and replanning keeps finished tasks, so a stopped run picks up where it left off.

Maestrun doesn't replace your coding agent; it conducts it. Install and sign in to at least one executor. Planning, execution, and samples run on your executor account.