SethGammon/Citadel

What it solves

Citadel provides an operating layer for AI coding agents like Claude Code and OpenAI Codex. It solves the problem of lost context and fragmented workflows when agent-led coding tasks extend beyond a single prompt, ensuring that repository state, decisions, and handoffs are preserved across different sessions.

How it works

Citadel integrates as a plugin for coding agents, introducing a /do entry point for natural-language requests. It manages the process through a structured loop of five states: Request, Run, Evidence, Needs You (for approvals or conflicts), and Resume (to identify the next action for new sessions). It maintains project-local state in dedicated directories (like .planning/ and .citadel/) to track campaigns, evidence, and telemetry without interfering with the application code.

Who it’s for

Developers using Claude Code or OpenAI Codex who are performing complex, multi-step coding tasks that require durable state, explicit verification boundaries, and coordination across multiple agents or branches.

Highlights

• Durable Session State: Preserves repository-local decisions and discoveries so work can be resumed across sessions.
• Unified Entry Point: Uses a single /do command to route natural-language requests to specific workflows.
• Verification Evidence: Records and reports outcomes as passed, failed, blocked, or unknown to make agent evaluations inspectable.
• Safety Boundaries: Implements approval gates and explicit verification to manage risky or multi-step changes.
• Cross-Platform Support: Compatible with Node.js 22+ on Linux, macOS, and Windows.