rloop

DAG-based task executor for AI coding agents. Define tasks as files, run them in isolated git worktrees, verify with tests, merge on success.

Built as an experiment in agentic engineering — I designed the architecture, drove the implementation through agent loops, and use it daily. Held to a strict test + clippy-pedantic bar. The structure may show its origins; the behavior is as specified in VISION.md, PRD.md, and SPEC.md.

What it does

For each task in a DAG:

1. Spawn an isolated git worktree
2. Run a Claude agent against the task description
3. Run verification (tests, clippy, build — whatever you configure)
4. On success: commit and merge back; on failure: capture learnings and retry
5. Move to the next ready task

Install

git clone https://github.com/jhartquist/rloop
cd rloop
cargo install --path .

Requires a working cargo toolchain and the claude CLI on PATH.

Quickstart

# In your project:
rloop init                              # create .rloop/ directory and config

# Define a task (LLM-enhanced from a rough description):
rloop task new "Add JSON output to the CLI"

# Or add a task with explicit fields:
rloop task add --title "Add JSON output" --verify "cargo test"

# Run a specific task:
rloop run 31

# Run all ready tasks in DAG order:
rloop run --all

Tasks live in .rloop/tasks/<id>.md as markdown with frontmatter. Edit them directly — they're just files.

Other commands

rloop task list                # list tasks and status
rloop task show <id>           # show full task details
rloop session show <id>        # replay a session log
rloop run --all --auto-merge   # run everything, auto-merge to parent branch

Layout

.rloop/
├── config.toml      # parent_branch, verify command, model, max_parallel
├── tasks/           # task definitions (markdown + frontmatter)
├── sessions/        # per-run JSONL logs
└── worktrees/       # isolated worktrees per task (transient)

CLAUDE.md snippet for projects using rloop

Drop this into the CLAUDE.md of a project that uses rloop, so Claude knows the workflow when you're working in that repo:

## rloop

This project uses rloop for task-driven development.

- Tasks live in `.rloop/tasks/<id>.md`. Each is markdown body + frontmatter
  (`title`, `verify`, `deps`, optional `model`).
- `rloop run` executes the next ready task in an isolated worktree, runs
  the verification command, and merges on success.
- `rloop task new "<description>"` creates a new task via LLM enhancement;
  `rloop task add --title ...` skips the LLM and is deterministic.
- Do not edit `.rloop/sessions/` — those are run logs.
- Do not commit `.rloop/worktrees/` — they are transient.

Docs

• VISION.md — MDP framing, design philosophy
• PRD.md — Product requirements
• SPEC.md — Technical design, module layout
• ROADMAP.md — Future direction
• research/ — Long-form notes on parallel DAGs, daemon, LLM verification, human queues, context/cost tracking

License

MIT.