jons-plan-marketplace

jons-plan Plugin

A Claude Code plugin inspired by Anthropic's Effective Harnesses for Long-Running Agents blog post for managing complex, multi-session coding tasks.

Prerequisites

• Claude Code v2.1.3+ (the Anthropic CLI tool)
• uv - Python package runner (curl -LsSf https://astral.sh/uv/install.sh | sh)
• graphviz (optional) - For workflow viewer (brew install graphviz)
• Codex CLI (optional) - For planning panel (npm install -g @openai/codex)
• Gemini CLI (optional) - For planning panel (npm install -g @anthropic/gemini-cli)

Installation

Install via Claude Code's plugin marketplace:

/plugin marketplace add jonmmease/jons-plan
/plugin install jons-plan@jonmmease-jons-plan

Then restart Claude Code and test with /jons-plan:status.

---

Overview

When AI agents work on complex projects that span multiple context windows, each new session starts with no memory of what came before. This plugin solves that problem by providing:

• Structured task management with dependencies and status tracking
• Progress logging to orient Claude at the start of each session
• Task-level progress for fine-grained resumption context
• Automatic session resumption when tasks were interrupted
• PreCompact hook to preserve state during context compaction
• Parallel task execution via subagents for independent work
• Workflow-based execution with phase transitions and state machine
• Planning panel — 3-agent parallel planning (Opus + Codex CLI + Gemini CLI) with senior-architect synthesis
• Artifact system with phase-level and plan-level artifact tracking
• Dead-end tracking to prevent repeating failed approaches

For detailed architecture documentation, see docs/architecture.md.

How It Works

The plugin uses a two-agent architecture:

Planning Agent (/jons-plan:new)

Creates the infrastructure for a new plan:

• Analyzes your request and suggests an appropriate workflow
• Conducts pre-research and records a research brief
• Sets up the plan structure and initial phase
• Creates request.md with the refined/approved request

Coding Agent (/jons-plan:proceed)

Works incrementally on tasks across sessions:

• Reads progress log to understand current state
• Resumes any in-progress tasks first
• Handles phase transitions automatically
• Updates task status and logs progress for the next session

Planning Panel

Workflows with planning_panel = true use three independent agents to generate plans in parallel, then a senior synthesis agent reviews all three and produces the definitive plan:

1. Opus — via Claude Code Task tool
2. Codex — via Codex CLI (background)
3. Gemini — via Gemini CLI (background)
4. Synthesis — Opus reviews all three, investigates disagreements, dismisses weak ideas, and produces the final plan

The synthesis agent has full authority to investigate the codebase, dismiss incorrect concerns, add missing elements, and restructure the plan. It is not a mechanical merger — it is the decision-maker.

Slash Commands

Command	Purpose
/jons-plan:new [topic]	Create new plan (auto-suggests workflow type)
/jons-plan:plan [feedback]	Refine the active plan
/jons-plan:proceed	Start/continue implementing tasks
/jons-plan:switch [name]	Switch to a different plan
/jons-plan:status	Show all plans and task progress
/jons-plan:viewer	Open the workflow viewer

Use --workflow <name> with /new to specify workflow type explicitly.

Workflow Types

Workflow	Purpose
implementation	Build features, fix bugs with research and validation
direct-implementation	Simple plan-implement-verify for familiar code (no research)
design	Research, explore, produce design.md
design-and-implementation	Design first, optionally implement after approval
deep-implementation	Thorough research + external review before implementation
dynamic	Research-first, phases generated based on codebase exploration
iteration	Iterative execute-evaluate-generate loops for long-horizon goals
review-tour	Generate guided PR review tour from a GitHub PR URL
code-review	Review code changes + generate PR description
pr-review	Review existing PR description for quality
deslop-pr	Detect and remove AI-generated patterns from PR descriptions
tech-docs	Technical documentation with multi-source research
tech-docs-review	Review RFCs, design docs, proposals with structured criteria

Auto-selection: When no --workflow is specified, the plugin analyzes your request and suggests the most appropriate workflow.

Plan Directory Structure