Med Marrouchi developed Slack2PR, an AI-powered coding assistant integrated directly into Slack to streamline software development. The tool allows users to transform ideas or bug reports into active GitHub pull requests without leaving the messaging platform. Slack2PR classifies user inputs into three specific categories: features, which it plans and implements; bugs, which require investigative steps and human approval; and questions, which function in a read-only mode for code explanation.

The architecture relies on Hexabot to manage the conversational workflow, using YAML-based definitions to handle intent classification, memory, and approval cycles. Google Gemini serves as the core intelligence for intent analysis and coding tasks. Developers can swap the underlying model through TanStack AI adapters, ensuring compatibility with Claude Code, Codex, and Grok Build. Each interaction triggers a TanStack AI Sandbox, an isolated Docker environment where the agent performs file inspection, component implementation, unit testing, and git operations. By utilizing the GitHub CLI within these sandboxes, the system creates branches and opens pull requests while maintaining credential security.

The workflow emphasizes modularity by breaking down tasks into distinct, controllable stages. Rather than using a single monolithic prompt, the process proceeds sequentially from intent classification to requirements gathering, planning, implementation, testing, and final pull request submission. This stage-gated design provides developers with transparency and granular control over the agent's actions during the automation lifecycle.

Med Marrouchi developed Slack2PR, an AI-powered coding assistant integrated directly into Slack to streamline software development. The tool allows users to transform ideas or bug reports into active GitHub pull requests without leaving the messaging platform. Slack2PR classifies user inputs into three specific categories: features, which it plans and implements; bugs, which require investigative steps and human approval; and questions, which function in a read-only mode for code explanation.

The architecture relies on Hexabot to manage the conversational workflow, using YAML-based definitions to handle intent classification, memory, and approval cycles. Google Gemini serves as the core intelligence for intent analysis and coding tasks. Developers can swap the underlying model through TanStack AI adapters, ensuring compatibility with Claude Code, Codex, and Grok Build. Each interaction triggers a TanStack AI Sandbox, an isolated Docker environment where the agent performs file inspection, component implementation, unit testing, and git operations. By utilizing the GitHub CLI within these sandboxes, the system creates branches and opens pull requests while maintaining credential security.

The workflow emphasizes modularity by breaking down tasks into distinct, controllable stages. Rather than using a single monolithic prompt, the process proceeds sequentially from intent classification to requirements gathering, planning, implementation, testing, and final pull request submission. This stage-gated design provides developers with transparency and granular control over the agent's actions during the automation lifecycle.