Skip to content
Search

Continuous Claude

lastmod 2025-12-09created 2025-12-091 edits

A workflow tool that runs Korean: 클로드 코드 repeatedly to automate everything through PR creation, checks, and merging

About

  • A CLI tool for handling repetitive multi-step development work, such as implementing a large project, with an AI agent loop
  • Calls Claude Code continuously from a Bash script, making small code changes in each iteration
    • Automatically creates a new branch, modifies code, then commits and pushes
    • Creates PRs via GitHub CLI and monitors CI status and review results with gh pr checks
    • Repeats a cycle: merges if all specified checks and reviews pass, and if they fail, closes the PR and discards the branch and changes
  • Uses a shared markdown file such as SHARED_TASK_NOTES.md as external memory to keep context between iterations
    • Summarizes what was done in each iteration and what to do next, recording it in a “relay race” style
    • E.g. leaving a note like “need to handle null input in function Y” makes the next iteration handle it first, forming a self-improvement loop
    • Prompts are designed to leave a handoff package that later developers and agents can understand immediately, instead of needlessly long logs
  • Provides a fully automated pipeline covering the entire PR lifecycle
    • Create branch → run Claude Code → commit → create PR → wait for CI and review → merge if passed → update main branch → clean up, then next iteration
    • Uses the existing repo's code owner rules, required checks, and preview environments as they are, naturally including human review in the workflow
  • Various flags for execution control can limit cost, time, and number of attempts
    • --max-runs sets the maximum number of iterations; 0 runs an infinite loop
    • --max-cost sets a cost cap in dollars, and --max-duration sets a time limit in the form 2h, 30m
    • Two or more can be specified together to build compound constraints like “10 runs or fewer, $5 or less, 1 hour or less”
  • Also supports finely controlling branch strategy and storage structure through GitHub integration options
    • --merge-strategy selects among squash / merge / rebase
    • --git-branch-prefix sets the branch naming rule; can change continuous-claude/ to feature/ etc.
    • --owner, --repo flags specify the repo explicitly even in environments where the remote is not GitHub or automatic extraction is difficult
  • Context storage method and termination conditions can be customized
    • --notes-file allows a different file name instead of SHARED_TASK_NOTES.md
    • With --completion-signal and --completion-threshold, exits early when agents output a “project complete” phrase a certain number of times or more
  • Includes safe mode and dry-run features for testing, debugging, and experimentation
    • --disable-commits disables actual commits, PR creation, and merging so only local changes can be tried
    • --dry-run simulates the whole flow, letting you check in the logs which commands would run
  • Supports running multiple tasks in parallel using the git worktree feature
    • --worktree <name> and --worktree-base-dir create independent worktrees to run tasks such as tests and documentation work simultaneously in different directories
    • Provides management features: --cleanup-worktree cleans up the worktree after the task finishes, and --list-worktrees shows currently active worktrees
  • Requires Claude Code CLI, GitHub CLI, and jq as dependencies, and the environment can be set up quickly with a simple install script
    • A one-line install script installs continuous-claude to ~/.local/bin or /usr/local/bin for use
  • Real-world usage scenarios : suited to highly repetitive tasks such as expanding test coverage, large-scale refactoring, and automatically fixing code broken after dependency updates
    • Existing Dependabot only handles version updates, but this tool behaves like a “souped-up Dependabot” that automatically creates follow-up fix PRs based on release notes and failed tests
    • Can also be used for long-running work that creates and merges 20 or more PRs in a row, such as splitting a monolithic codebase into multiple modules or converting callbacks to async/await
  • A concept similar to GitHub Next's Continuous AI and agentics research, designed with running multiple specialized agents simultaneously in mind
    • Supports a pattern of running agents with different roles, such as testing, refactoring, and adding features, in parallel to advance multiple areas of a monorepo at once
    • Lets you experiment with an iterative strategy that tolerates waste on the premise of falling costs, relying overall on a “probability distribution heading in the right direction” even when individual runs fail
  • Overall, a tool that provides an agent layer where AI takes on repetitive tasks and leftover work while keeping human developers' PR-based workflow intact, and it can be attached to real-world work repos right away

See also

Favorite site

author Zer0 sitemap