pstack-claude

Product Introduction

1. Definition: pstack-claude is a multi-agent workflow implementation and port of the open-source "pstack" methodology, originally designed for the Cursor IDE, to major AI coding assistants including Claude Code, Codex, Pi, Gemini, and Prime Agent. It functions as a plugin or extension that injects standardized, rigorous agentic workflows into different AI harnesses.
2. Core Value Proposition: It exists to solve AI vendor lock-in and inconsistent agent behavior by providing a unified, high-quality framework for AI-assisted software development. Its core value is enabling reliable, reproducible AI coding patterns and rigorous agent workflows that work identically across Claude, Codex, Pi, and other platforms, ensuring developer teams can maintain consistent quality and processes regardless of their chosen AI tool.

Main Features

1. Multi-Harness Plugin Architecture: The product is delivered as native plugins/extensions for specific AI environments. For Claude Code, it uses the official Claude plugin system. For Codex, it installs via a CLI. For Pi, it loads as a Pi extension. This architecture ensures deep integration with each harness's unique slash-command system and runtime.
2. Automated Skill Routing (poteto-mode): A core intelligence layer that analyzes a user's stated goal (e.g., "fix the search filter resetting") and automatically invokes the correct pre-defined workflow or "skill." This removes the need for developers to manually orchestrate agents, routing tasks to specialized sub-agents for bug reproduction, investigation, delegation, and verification.
3. Standardized pstack Workflow Library: The port includes a comprehensive translation of Poteto's pstack methodology into executable skills. These cover specific playbooks for bug fixing, feature planning, refactoring, performance investigations, PR maintenance, and long-term projects. Each workflow enforces steps like reproducing failures, using "how and why" investigation, delegating fixes, and providing passing/failing evidence.

Problems Solved

1. Pain Point: Inconsistent AI Agent Outputs. Different AI models and harnesses (Claude vs. Copilot vs. Gemini) produce varying code quality and follow different processes, making it hard to standardize team workflows.
2. Target Audience: Development Teams and Lead Engineers who adopt multiple AI coding assistants and need consistent outcomes; Open-Source Maintainers seeking reliable, auditable AI contributions; Individual Developers who want to apply rigorous, systematic problem-solving to their AI-aided work.
3. Use Cases: Cross-Platform Bug Resolution – diagnosing and fixing a complex bug using Claude Code one day and Codex the next with identical methodology. Onboarding Standardization – ensuring all team members use the same AI-assisted code review or refactoring process. Project Portability – moving a development project from a Claude-based environment to a Gemini-based one without losing the structured agentic workflow.

Unique Advantages

1. Differentiation: Unlike single-vendor AI coding tools (GitHub Copilot) or generic prompt chains, pstack-claude is a vendor-agnostic workflow engine. It does not replace the AI model but orchestrates it, providing a consistent meta-layer on top of disparate systems. It is more structured and opinionated than simple prompt templates.
2. Key Innovation: The translation of Cursor-specific "skills" and agent primitives (like subagent delegation, question loops, and wake-up tools) into a generic framework that can be injected into other AI harnesses via their plugin APIs. Its runtime configuration (setup-pstack) allows per-role reasoning effort control (e.g., "opus @xhigh"), letting users optimize cost/performance for different workflow steps across models.

Frequently Asked Questions (FAQ)

1. How does pstack-claude handle data privacy and telemetry? pstack-claude has no server, backend, or telemetry. All skills execute locally on your machine. Scripts run in your local environment, and any data (like session transcripts) read by the agents is sent only to your configured AI model provider (e.g., Anthropic, OpenAI).
2. Can I use pstack-claude with GitHub Copilot? The current direct plugin support is for Claude Code, Codex, and Pi. For GitHub Copilot or other harnesses, you can use the "skills-only" installation method detailed in the shared installation notes, which may require more manual integration.
3. What is the difference between pstack-claude and the original Cursor pstack? pstack-claude is a port that maintains the core pstack methodology but adapts its implementation to work outside the Cursor IDE. It tracks updates from the original (upstream) pstack and also maintains declared policy forks, ensuring it works seamlessly with the API and interface constraints of Claude, Codex, Pi, and others.
4. Does pstack-claude require an internet connection to work? While the AI models themselves require an internet connection, the pstack-claude plugin/extension and its workflow definitions are installed locally. Its operation depends on the connectivity of the underlying AI harness it is extending.
5. Is pstack-claude suitable for beginner programmers? Its primary design is for developers familiar with AI-assisted coding who seek rigor and reproducibility. The automated routing (poteto-mode) simplifies invocation, but effectively using and configuring the various workflows benefits from an understanding of software development processes and the specific AI harness being used.