Autonomous work mode for Claude Code, andylizf/nonstop addresses the common issue of AI assistants halting prematurely on complex tasks due to blockers or the need for human approval. It provides an "autonomous work mode" that performs pre-flight risk assessments, intelligently handles blockers, and ensures tasks are completed or properly marked as blocked upon the user's return. This benefits users who need AI to work through multi-step tasks overnight or while they are away, ensuring progress and minimizing manual intervention.

How It Works

Before execution, the /nonstop command triggers a four-phase sequence: 1. Mental Simulation: Claude plans task steps and flags potential blockers. 2. Dangerous Ops Manifest: Explicit user approval is required for actions like killing processes, deleting files, force pushing, or using paid APIs. 3. Permissions Check: Warns if permission prompts might halt progress while the user is absent. 4. Confirm and Go: Final confirmation before entering nonstop mode. While the user is away, a stop hook implements a decision framework: if blocked, Claude attempts to solve it, then work around it (noting the method), or mark it blocked and move to the next task if truly stuck. Long-running operations are delegated to background agents. The system uses two files (SKILL.md for protocol, nonstop.sh for the stop hook) and a session-scoped flag file, integrating intelligence directly into Claude's workflow.

Quick Start & Requirements

• Primary install / run command:
  • Tell Claude Code: Paste into Claude Code and follow instructions at https://raw.githubusercontent.com/andylizf/nonstop/main/INSTALL.md.
  • Plugin Marketplace: /plugin marketplace add andylizf/nonstop followed by /plugin install nonstop@nonstop.
  • Manual: curl -fsSL https://raw.githubusercontent.com/andylizf/nonstop/main/install.sh | bash.
• Non-default prerequisites and dependencies: Requires a Claude Code environment.
• Links: https://raw.githubusercontent.com/andylizf/nonstop/main/INSTALL.md

Highlighted Details

• Comprehensive pre-flight risk assessment, including a "mental simulation" and a "dangerous ops manifest" requiring explicit user sign-off.
• Intelligent blocker handling: Claude attempts to solve, work around (documenting the approach), or mark tasks as blocked before moving on.
• Session-scoped stop hook with configurable nudge limits (NONSTOP_MAX) to prevent infinite loops.
• Delegation of long-running operations (builds, tests) to background agents, preventing Claude from idling.

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or roadmaps were provided in the README excerpt.

Licensing & Compatibility

• License type: MIT.
• Compatibility notes: The MIT license is permissive and generally compatible with commercial use and linking within closed-source projects.

Limitations & Caveats

The effectiveness relies on Claude's ability to accurately perform the "mental simulation" and make sound decisions regarding workarounds. The "Permissions check" phase aims to mitigate issues with interactive prompts, but complex permission scenarios might still pose challenges. The interpretation and execution of "dangerous ops" and "workarounds" are dependent on Claude's current capabilities.