One-Shot Lifecycle

The One-Shot Lifecycle skill (invoked via /hub:run) chains the full AgentHub competition flow into a single command: it initializes a session, captures a baseline metric, spawns parallel agents, evaluates their results, and merges the winning branch after user confirmation. It produces a ranked results table and merges the chosen agent's work into the main branch.

✨ What it does

  • Chains init → baseline → spawn → eval → merge in one command.
  • Captures a baseline metric from an eval command and stores it in the session config.
  • Spawns multiple agents in parallel using the specified template or default dispatch prompt.
  • Evaluates results using metric-based ranking (result_ranker.py) or LLM judge mode.
  • Displays a ranked results table and asks for user confirmation before merging.
  • Supports optional agent templates: optimizer, refactorer, test-writer, bug-fixer.

🎯 When to use it

  • When you want to run a complete AgentHub competition end-to-end without manually invoking each step.
  • When you have a well-defined task (e.g., performance optimization, refactoring, test writing) and want multiple agents to propose solutions
  • When you need to compare agent outputs against a measurable baseline metric (e.g., latency, coverage) and automatically rank them.
  • When you want to use an LLM judge to rank agent results when no quantitative eval command is available.

🚀 How to use

Trigger: /hub:run

Trigger the skill by running the slash command /hub:run with the required --task parameter. Optionally provide --agents, --eval, --metric, --direction, and --template. The command executes the full lifecycle sequentially and prompts for merge confirmation at the end.

Example prompts:

/hub:run --task "Reduce p50 latency" --agents 3 --eval "pytest bench.py --json" --metric p50_ms --direction lower --template optimizer
/hub:run --task "Write 3 email subject lines for spring sale campaign" --agents 3 --judge

📄 Output: A ranked results table and, after user confirmation, a merged branch containing the winning agent's changes.

📦 Add this skill to Claude Code

# 1. Get the skills repo
git clone --depth 1 https://github.com/alirezarezvani/claude-skills /tmp/claude-skills

# 2. Copy this skill into your project (or ~/.claude/skills for all projects)
mkdir -p .claude/skills
cp -r /tmp/claude-skills/engineering/agenthub/skills/run .claude/skills/run

Skill source: engineering/agenthub/skills/run/SKILL.md

⚠️ Good to know

Requires the AgentHub environment and dependencies (e.g., Python scripts, git branches). The eval command must output the metric in a parseable format; if no eval is provided, LLM judge mode is used. The skill never auto-merges without explicit user confirmation.

❓ FAQ

What happens if I don't provide an --eval command?

If no --eval is provided, the skill skips baseline capture and uses LLM judge mode during evaluation, where the coordinator reads the diffs and ranks the agents.

Can I change the winning agent after the merge prompt?

Yes, if you decline the merge, you can run /hub:merge --agent agent-{N} to pick a different winner, or use /hub:eval --judge to re-evaluate with the LLM judge.

What templates are available for agents?

The available templates are optimizer, refactorer, test-writer, and bug-fixer. If no template is specified, agents use the default dispatch prompt from /hub:spawn.

🤖 Overview, features, install steps and FAQ were generated from the project's SKILL.md on Sep 4, 2026. Always check the original source before running commands.