AnandChowdhary/continuous-claude

What it solves

Continuous Claude solves the problem of AI coding tools that halt after a single task or lack the ability to self-criticize and iterate. It enables the completion of large, multi-step projects (such as increasing test coverage from 0% to 80% across a massive codebase) by running an AI coding agent in a continuous loop, maintaining context across iterations without exhausting the context window.

How it works

The tool acts as a conductor that repeatedly invokes an AI coding agent (like Claude Code or Codex CLI) and manages the GitHub pull request lifecycle. For each iteration, it:

1. Creates a new git branch and runs the AI agent to generate a commit.
2. Pushes the changes and creates a GitHub pull request.
3. Monitors CI checks and reviews via the GitHub CLI (gh).
4. Merges the PR on success or discards it on failure.
5. Pulls the updated main branch and repeats the process.

To maintain continuity, it uses a shared markdown file (SHARED_TASK_NOTES.md) as external memory. The AI agent is instructed to treat the process as a "relay race," recording progress and handoff notes for the next iteration to prevent context drift and enable self-improvement.

Who it’s for

Developers who need to automate mundane, large-scale refactoring tasks, dependency updates with breaking changes, or comprehensive test suite expansion that would otherwise be too tedious for humans to perform manually.

Highlights

• Automated PR Lifecycle: Fully automates the branch-commit-PR-merge cycle using GitHub CLI.
• Context Continuity: Uses a persistent markdown file to pass knowledge between stateless AI queries.
• Flexible Constraints: Supports limits based on maximum runs, USD cost budgets, or time-boxed durations.
• Cros-Platform Support: Includes both a Bash runner and a native PowerShell runner for Windows users.
• Reviewer Pass: Optional integration of a second AI agent to validate and fix changes before merging.
• Parallel Execution: Supports git worktrees for running multiple agents simultaneously on different parts of a project.