ai-kit

Multi-agent collaboration plugin suite for AI coding tools — task-driven workflow via tmux, code review, and learning aids.

Overview

ai-kit provides a collection of plugins that coordinate multiple AI agents (Claude Code, Codex, OpenCode, etc.) working in separate tmux panes through a structured task-driven workflow. Instead of ad-hoc communication, agents exchange structured messages with task labels and report tags, creating a traceable collaboration loop. Additional plugins provide code review, learning, and self-reflection capabilities.

┌─────────────┐    task from     ┌─────────────┐
│  Agent A     │ ───────────────→ │  Agent B     │
│  (Sender)    │                  │  (Receiver)  │
│              │ ←─────────────── │              │
└─────────────┘   report from    └─────────────┘

Installation

Claude Code

Register this repository as a plugin marketplace, then install:

/plugin marketplace add fingergohappy/ai-kit

Install the plugins you need:

/plugin install agentflow@ai-kit
/plugin install tmux@ai-kit
/plugin install git@ai-kit
/plugin install code-kit@ai-kit
/plugin install learning@ai-kit
/plugin install self-learn@ai-kit

After installation, restart Claude Code. Skills will be available with the plugin prefix:

/agentflow:task login-system
/agentflow:dispatch %7 docs/tasks/login_feature.md
/code-kit:evaluate "use postgres vs mysql"
/learning:learn rust lifetimes
/git:commit
/tmux:tmux-send %7 "hello"

<details> <summary>Alternative: local development</summary>

claude --plugin-dir /path/to/ai-kit

</details>

Codex (OpenAI)

This repository includes Codex plugin manifests for agentflow, tmux, and git.

Register this repository as a Codex plugin marketplace:

codex plugin marketplace add https://github.com/fingergohappy/ai-kit

Quick AI-driven install/update:

Paste this URL into Codex and ask it to follow the runbook:

follow this install https://github.com/fingergohappy/ai-kit/blob/main/docs/codex_plugin_install_update.md

Detailed guide:

• docs/codex_plugin_install_update.md

Plugins

agentflow

Agent collaboration loop: task → dispatch → evaluate → report → review → redo.

| Skill | Purpose | |-------|---------| | agentflow:task | Generate structured task documents (feature / change / task) | | agentflow:dispatch | Send task or fix instructions to a tmux pane for execution | | agentflow:gate-evaluate | Receiver-side input guard — evaluate incoming tasks before execution | | agentflow:report | Report execution results back to the sender | | agentflow:gate-review | Sender-side output guard — review work results, decide pass or redo |

code-kit

Code review and evaluation utilities for single-agent workflows.

| Skill | Purpose | |-------|---------| | code-kit:evaluate | Rigorous evidence-based evaluation (tech selection, architecture, claim verification) | | code-kit:review-init | Analyze project tech stack and generate customized review skills | | code-kit:review-report | Generate structured review report from audit results | | code-kit:fix-review | Fix a specific issue from a review report and update its status | | code-kit:fix-review-all | Batch-fix all pending issues from a review report in parallel | | code-kit:nvim-lsp-init | Generate Neovim LSP environment setup scripts for the project |

learning

Personal learning and note-taking aids.

| Skill | Purpose | |-------|---------| | learning:learn | Explain concepts with minimal examples, code analogies, and simplification | | learning:take-note | Generate structured learning notes with runnable code examples |

self-learn

AI self-reflection and lesson capture.

| Skill | Purpose | |-------|---------| | self-learn:learn-from-mistake | After AI is corrected, propose solidifying the lesson as a guardrail rule |

tmux

Tmux infrastructure utilities for inter-pane communication and long-running service management.

| Skill | Purpose | |-------|---------| | tmux:tmux-send | Send text content to a tmux pane | | tmux:agent-tmux | Start/restart/stop long-running commands in shared tmux session (auto-isolates by project/branch) |

git

Git worktree and branching utilities.

| Skill | Purpose | |-------|---------| | git:rebase-to-root | Rebase worktree feature branch back to root's current branch | | git:commit | Create atomic git commits with validation and conventional commit messages |

Workflow

Agent Collaboration (agentflow)

1. Design Phase

/agentflow:task <task-name>

Enter design discussion mode — discuss without writing code, generate document when ready. Outputs to docs/tasks/.

2. Dispatch Phase

/agentflow:dispatch [loop] <pane_id> <doc-path>

Send the task document to another agent's tmux pane. The receiving agent gets a [task from ...] labeled message. Add loop to enable automatic review-fix cycling.

3. Execution & Report

The receiver evaluates the task via gate-evaluate, executes, then calls report to send results back:

[report from Claude Code, pane_id: %5, loop: true: completed 3 tasks]

4. Review & Fix

The sender receives the report and gate-review triggers:

• Reviews work against original design
• If loop: true + issues found → auto-dispatches fix instructions (up to 3 rounds)
• If loop: false + issues found → outputs conclusions, user decides next step
• If all passed → done

Code Review (code-kit)

1. Initialize Review

/code-kit:review-init

Analyze the project tech stack and generate customized review skills into .claude/skills/.

2. Run Review

Use the generated review skills, then generate a report:

/code-kit:review-report

3. Fix Issues

Fix individual issues or batch-fix all:

/code-kit:fix-review docs/review/2026-04-26_full_review.md:69
/code-kit:fix-review-all docs/review/2026-04-26_full_review.md

Evaluation (code-kit)

/code-kit:evaluate "use postgres vs mysql for this project"

Collects dual-source evidence (project facts + external best practices) and produces a rigorous evaluation with cited sources.

Message Protocol

Task Dispatch

[task from {agent-name}, pane_id: {pane_id}, loop: {true|false}: {task-summary}]

Fix Dispatch

[fix from {agent-name}, pane_id: {pane_id}, loop: {true|false}: {fix-summary}]

Execution Report

[report from {agent-name}, pane_id: {pane_id}, loop: {true|false}: {result-summary}]

Agents use these tags to identify message types and route responses correctly.

Requirements

• tmux session with multiple panes
• AI coding tool running in each pane (Claude Code, Codex, OpenCode, etc.)
• tmux:tmux-send skill available for inter-pane communication

License

MIT