Continuous Claude – run Claude Code in a loop (github.com)

Continuous Claude is a new CLI that runs Claude Code in a persistent loop to autonomously drive multi-step software tasks: it generates code changes, creates feature branches, opens GitHub pull requests, waits for CI and reviews, merges when green, and repeats until the job is done or a budget/iteration limit is reached. Implemented as a Bash “conductor” script, it maintains continuity across runs by writing compact handoff notes to a shared markdown file (SHARED_TASK_NOTES.md), so each Claude invocation sees prior context instead of starting statelessly. The tool integrates with Claude Code CLI, GitHub CLI (gh), and jq; accepts a task prompt plus max-runs or max-cost, repo owner/name, merge strategy and branch-prefix, and can run infinitely, on a schedule, or within CI workflows. For the AI/ML community this demonstrates a practical pattern for long-running agentic work: iterative, incremental edits mirror human development and avoid one-shot brittleness while leveraging existing GitHub protections (CI, code-owner reviews, preview environments). Key trade-offs: wasted iterations are possible if a PR is closed and discarded, and the system relies on prompt engineering and Claude’s discipline to keep the external memory concise. Use cases include large refactors, test coverage drives, dependency upgrades, and other technical-debt tasks that benefit from many small, verifiable commits rather than monolithic changes.