agenthub
Multi-agent collaboration plugin that spawns N parallel subagents competing on the same task via git worktree isolation. Agents work independently, results are evaluated by metric or LLM judge, and the best branch is merged. Use when: user wants multiple approaches tried in parallel — code optimization, content variation, research exploration, or any task that benefits from parallel competition. Requires: a git repo.
- License 许可
- MIT
- 最近更新
- 2026-10-28
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- agenthub
- license
- MIT
- metadata
- { "author": "Alireza Rezvani", "updated": 1773705600000, "version": "2.1.2", "category": "engineering" }
- description
- Multi-agent collaboration plugin that spawns N parallel subagents competing on the same task via git worktree isolation. Agents work independently, results are evaluated by metric or LLM judge, and the best branch is merged. Use when: user wants multiple approaches tried in parallel — code optimization, content variation, research exploration, or any task that benefits from parallel competition. Requires: a git repo.
02 资源文件 9
- 参考文档references/agent-templates.md7.6 KB
- 参考文档references/coordination-strategies.md5.8 KB
- 参考文档references/dag-patterns.md4.0 KB
- 脚本scripts/board_manager.py8.4 KB
- 脚本scripts/dag_analyzer.py8.7 KB
- 脚本scripts/dry_run.py11.1 KB
- 脚本scripts/hub_init.py8.3 KB
- 脚本scripts/result_ranker.py10.3 KB
- 脚本scripts/session_manager.py9.8 KB
03 SKILL.md 正文
AgentHub — Multi-Agent Collaboration
Spawn N parallel AI agents that compete on the same task. Each agent works in an isolated git worktree. The coordinator evaluates results and merges the winner.
Slash Commands
| Command | Description |
|---|---|
/hub:hub-init |
Create a new collaboration session — task, agent count, eval criteria |
/hub:spawn |
Launch N parallel subagents in isolated worktrees |
/hub:hub-status |
Show DAG state, agent progress, branch status |
/hub:eval |
Rank agent results by metric or LLM judge |
/hub:merge |
Merge winning branch, archive losers |
/hub:board |
Read/write the agent message board |
/hub:run |
One-shot lifecycle: init → baseline → spawn → eval → merge |
Agent Templates
When spawning with --template, agents follow a predefined iteration pattern:
| Template | Pattern | Use Case |
|---|---|---|
optimizer |
Edit → eval → keep/discard → repeat x10 | Performance, latency, size |
refactorer |
Restructure → test → iterate until green | Code quality, tech debt |
test-writer |
Write tests → measure coverage → repeat | Test coverage gaps |
bug-fixer |
Reproduce → diagnose → fix → verify | Bug fix approaches |
Templates are defined in references/agent-templates.md.
When This Skill Activates
Trigger phrases:
- "try multiple approaches"
- "have agents compete"
- "parallel optimization"
- "spawn N agents"
- "compare different solutions"
- "fan-out" or "tournament"
- "generate content variations"
- "compare different drafts"
- "A/B test copy"
- "explore multiple strategies"
Coordinator Protocol
The main Claude Code session is the coordinator. It follows this lifecycle:
INIT → DISPATCH → MONITOR → EVALUATE → MERGE
1. Init
Run /hub:hub-init to create a session. This generates:
.agenthub/sessions/{session-id}/config.yaml— task config.agenthub/sessions/{session-id}/state.json— state machine.agenthub/board/— message board channels
2. Dispatch
Run /hub:spawn to launch agents. For each agent 1..N:
- Post task assignment to
.agenthub/board/dispatch/ - Spawn via Agent tool with
isolation: "worktree" - All agents launched in a single message (parallel)
3. Monitor
Run /hub:hub-status to check progress:
dag_analyzer.py --status --session {id}shows branch state- Board
progress/channel has agent updates
4. Evaluate
Run /hub:eval to rank results:
- Metric mode: run eval command in each worktree, parse numeric result
- Judge mode: read diffs, coordinator ranks by quality
- Hybrid: metric first, LLM-judge for ties
5. Merge
Run /hub:merge to finalize:
git merge --no-ffwinner into base branch- Tag losers:
git tag hub/archive/{session}/agent-{i} - Clean up worktrees
- Post merge summary to board
Agent Protocol
Each subagent receives this prompt pattern:
You are agent-{i} in hub session {session-id}.
Your task: {task description}
Instructions:
1. Read your assignment at .agenthub/board/dispatch/{seq}-agent-{i}.md
2. Work in your worktree — make changes, run tests, iterate
3. Commit all changes with descriptive messages
4. Write your result summary to .agenthub/board/results/agent-{i}-result.md
5. Exit when done
Agents do NOT see each other's work. They do NOT communicate with each other. They only write to the board for the coordinator to read.
DAG Model
Branch Naming
hub/{session-id}/agent-{N}/attempt-{M}
- Session ID: timestamp-based (
YYYYMMDD-HHMMSS) - Agent N: sequential (1 to agent-count)
- Attempt M: increments on retry (usually 1)
Frontier Detection
Frontier = branch tips with no child branches. Equivalent to AgentHub's "leaves" query.
python scripts/dag_analyzer.py --frontier --session {id}
Immutability
The DAG is append-only:
- Never rebase or force-push agent branches
- Never delete commits (only branch refs after archival)
- Every approach preserved via git tags
Message Board
Location: .agenthub/board/
Channels
| Channel | Writer | Reader | Purpose |
|---|---|---|---|
dispatch/ |
Coordinator | Agents | Task assignments |
progress/ |
Agents | Coordinator | Status updates |
results/ |
Agents + Coordinator | All | Final results + merge summary |
Post Format
---
author: agent-1
timestamp: 2026-03-17T14:30:22Z
channel: results
parent: null
---
## Result Summary
- **Approach**: Replaced O(n²) sort with hash map
- **Files changed**: 3
- **Metric**: 142ms (baseline: 180ms, delta: -38ms)
- **Confidence**: High — all tests pass
Board Rules
- Append-only: never edit or delete posts
- Unique filenames:
{seq:03d}-{author}-{timestamp}.md - YAML frontmatter required on all posts
Evaluation Modes
Metric-Based
Best for: benchmarks, test pass rates, file sizes, response times.
python scripts/result_ranker.py --session {id} \
--eval-cmd "pytest bench.py --json" \
--metric p50_ms --direction lower
The ranker runs the eval command in each agent's worktree directory and parses the metric from stdout.
LLM Judge
Best for: code quality, readability, architecture decisions.
The coordinator reads each agent's diff (git diff base...agent-branch) and ranks by:
- Correctness (does it solve the task?)
- Simplicity (fewer lines changed preferred)
- Quality (clean execution, good structure)
Hybrid
Run metric first. If top agents are within 10% of each other, use LLM judge to break ties.
Session Lifecycle
init → running → evaluating → merged
→ archived (if no winner)
State transitions managed by session_manager.py:
| From | To | Trigger |
|---|---|---|
init |
running |
/hub:spawn completes |
running |
evaluating |
All agents return |
evaluating |
merged |
/hub:merge completes |
evaluating |
archived |
No winner / all failed |
Proactive Triggers
The coordinator should act when:
| Signal | Action |
|---|---|
| All agents crashed | Post failure summary, suggest retry with different constraints |
| No improvement over baseline | Archive session, suggest different approaches |
| Orphan worktrees detected | Run session_manager.py --cleanup {id} |
Session stuck in running |
Check board for progress, consider timeout |
Installation
# Copy to your Claude Code skills directory
cp -r engineering/agenthub ~/.claude/skills/agenthub
# Or install via ClawHub
clawhub install agenthub
Scripts
| Script | Purpose |
|---|---|
hub_init.py |
Initialize .agenthub/ structure and session |
dag_analyzer.py |
Frontier detection, DAG graph, branch status |
board_manager.py |
Message board CRUD (channels, posts, threads) |
result_ranker.py |
Rank agents by metric or diff quality |
session_manager.py |
Session state machine and cleanup |
Related Skills
- autoresearch-agent — Single-agent optimization loop (use AgentHub when you want N agents competing)
- self-improving-agent — Self-modifying agent (use AgentHub when you want external competition)
- git-worktree-manager — Git worktree utilities (AgentHub uses worktrees internally)
04 同仓库更多 Skills
syllabus
Generates a curated supplementary reading list from any course syllabus using Consensus academic search. Grill-me intake (syllabus input format + course audience + year range) plus a grouping forcing-options checkpoint before any search runs — so the reading list matches the course's level and recency need. Parses the syllabus to extract topics and learning outcomes, searches Consensus for recent peer-reviewed papers per topic, and produces a professionally formatted .docx with clickable Consensus links, plain-language summaries calibrated to audience level, and Bloom-higher-order discussion questions tied to course learning goals. Use when the user uploads a syllabus, course outline, or curriculum document and wants supplementary readings (e.g., 'create a reading list from this syllabus', 'find recent papers for my course') — even casual mentions with a syllabus attached should trigger this skill.
pulse
Multi-source recency research skill that takes the pulse of any topic across Reddit, Hacker News, the open web, and optionally X/Twitter within a configurable recent window (default 30 days). Forcing intake clarifies topic specificity, angle (trend/sentiment/problems/opportunities/comparison), time window, and platform scope before searching. Returns a synthesized briefing with citations, engagement metrics, and cross-platform pattern analysis. Use when the user requests multi-source recency intelligence on a topic (e.g., 'pulse on [topic]', 'what's happening with [topic]', 'what are people saying about [topic]', 'current conversation about [topic]', 'take the pulse of [topic]', 'trending: [topic]', 'find me info on [topic]'), and for competitor research, trend discovery, tool comparisons, and audience sentiment analysis.
patent
Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search / freedom-to-operate / competitive landscape / acquisition diligence / litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Use when the user asks for patent searching or analysis (e.g., 'prior art search for [invention]', 'freedom to operate analysis for [product]'). Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope.
notebooklm
Browser automation skill for controlling Google's NotebookLM. Use when the user wants anything done in NotebookLM (e.g., 'open NotebookLM', 'check my [name] notebook', 'ask my notebook about X', 'add [source] to NotebookLM', 'generate a Video Overview from my notebook', 'use NotebookLM Studio'). Handles reading and querying notebooks, adding sources (URLs, text, files, YouTube links, synthesized content), generating Studio outputs (Audio/Video Overviews, Mind Maps, Reports incl. Briefing Doc/Study Guide/FAQ, Flashcards, Quiz, slide decks, infographics — discover the exact set from the live Studio panel; the UI evolves fast), and creating new notebooks. Requires browser automation environment — fails gracefully when unavailable.
litreview
Academic literature orientation skill that searches papers via free keyless APIs (PubMed E-utilities + OpenAlex) by default — with the Consensus MCP as an optional enhancement lane when connected — builds a strategic search plan using PICO (default) or SPIDER / Decomposition / hybrid as fallbacks, and synthesizes findings into a formatted Word (.docx) research guide. Grill-me intake (research question specificity + framework hint + tentative depth) before the recon search; a second forcing checkpoint after Phase 2 confirms framework + sub-areas + depth before searches consume budget. Configurable depth (5/10/20 queries) controls coverage vs. speed. Output is a 'launching pad' — an orientation guide that lets a researcher dive in confidently, not a finished review. Use when the user starts literature-oriented research (e.g., 'litreview on [topic]', 'literature review on [topic]', 'I'm starting a literature review on X', 'I'm writing a paper on X', 'help me research X', 'I'm doing research on X', 'can you help me research X'). Do NOT use for single one-off paper searches wanting a quick list — that's a plain PubMed/OpenAlex (or Consensus) query.
grants
NIH grant research skill for clinical researchers. Grill-me intake (research idea + career stage + preliminary data + environment + submission posture + known institute targets) locks down the funding strategy before any search runs. Runs a 5-facet Consensus positioning analysis (with draft Significance/Innovation language), maps the research to the right NIH institutes and study sections via RePORTER, finds NOSIs and funded overlap, and produces an editable Word document (.docx) with budget/scope-aware mechanism recommendations, submission timelines, and a mandatory program officer recommendation. Use when the user asks about research funding or makes any grant-related request (e.g., 'grants for [topic]', 'find grants for my research idea', 'what grants match my research', 'help me find NIH funding', 'grant opportunities for my research'). NIH-only scope — non-NIH funders (PCORI, DOD CDMRP, VA, foundations) are out of scope and flagged at intake.