run

---
name: run
description: Use when executing a full lifecycle: initialize, capture baseline, spawn agents, evaluate, and merge results in one command.
category: Engineering
version: 1.0.0
tools: []
---

# One-Shot Lifecycle

Run the full AgentHub lifecycle in one invocation: initialize, capture baseline, spawn agents, evaluate results, and merge the winner.

## Parameters

| Parameter | Required | Description |
|-----------|----------|-------------|
| `--task` | Yes | Task description for agents |
| `--agents` | No | Number of parallel agents (default: 3) |
| `--eval` | No | Eval command to measure results (skip for LLM judge mode) |
| `--metric` | No | Metric name to extract from eval output (required if `--eval` given) |
| `--direction` | No | `lower` or `higher` — which direction improves metric (required if `--metric` given) |
| `--template` | No | Agent template: `optimizer`, `refactorer`, `test-writer`, `bug-fixer` |

## Execution Steps

### Step 1: Initialize

Set up session with provided task, agent count, evaluation criteria, and template if given. Display session ID to user.

### Step 2: Capture Baseline

If `--eval` provided:
- Run the evaluation command in current directory
- Extract the metric value from output
- Store baseline for delta comparison later
- Display: `Baseline captured: {metric} = {value}`

If no `--eval`, skip baseline.

### Step 3: Spawn Agents

Launch all agents in parallel. If `--template` provided, use template-specific dispatch prompt with eval command, metric, and baseline passed as variables. Otherwise use default dispatch.

### Step 4: Wait and Monitor

After all agents complete, display summary of each agent's work before proceeding to evaluation.

### Step 5: Evaluate

Rank results using:
- Metric-based ranking if `--eval` was provided (with baseline delta if captured)
- LLM judge mode if no `--eval` (coordinator reads diffs and ranks)

Display ranked results table.

### Step 6: Confirm and Merge

Present results to user and ask for confirmation before merging. User can:
- Accept and merge top result
- Decline and pick different agent manually
- Re-evaluate with LLM judge
- Inspect branches before deciding

## Critical Rules

- Sequential execution — each step depends on previous completion
- Stop on failure — report error and halt if any step fails
- User confirms merge — never auto-merge without explicit approval
- Template is optional — without it, agents use default dispatch prompt