🚧 Warning

This project is experimental and not production-ready. Expect breaking changes and incomplete features. AND all your data and code to be compromised, if you don't know what you are doing, pls avoid for now.

laun

laun is a Rust CLI that orchestrates a two-agent implementation loop against a PRD checklist.

• loop_agent (fast, cheaper): chooses the next PRD item and writes scoped instructions.
• worker_agent (slower, stronger): makes code changes for that item.
• Orchestrator: runs tests, retries fixes when tests fail, commits on success, and marks PRD items complete.

Why this exists

This pattern separates planning cadence from implementation quality:

• Fast model keeps momentum and task ordering.
• Strong model handles code-heavy tasks.
• Each agent can have different visible file/test scope in prompts.

Current capabilities

• Config-driven agent commands (works with OpenCode, Codex, or custom wrappers).
• PRD markdown checkbox parsing (- [ ] / - [x]).
• Test gate after each worker run.
• Retry loop with failure feedback to the worker.
• Optional auto-commit (git add -A && git commit).
• Optional auto-mark PRD item as done.

Installation

Run from source

cargo run -- --help

Build binary

cargo build --release
./target/release/laun --help

Quick start

1. Initialize config and PRD files:

laun init

2. Edit laun.toml:

• Set your agent command + args for your local CLI setup.
• Set models for loop/worker.
• Set test commands in workflow.execution_tests.

3. Edit PRD.md with checklist items:

- [ ] Implement feature A
- [ ] Add tests for feature A
- [ ] Ship docs

4. Validate config:

laun validate

5. Run a dry simulation first:

laun run --dry-run --max-iterations 1

6. Run for real:

laun run

How the loop works

For each iteration:

1. Load unchecked PRD items.
2. Ask loop_agent for next action (JSON contract).
3. Ask worker_agent to implement selected item.
4. Run workflow.execution_tests.
5. If tests fail, retry worker with failing output (max_fix_attempts).
6. If tests pass:

• optional commit (workflow.auto_commit)
• optional PRD item check-off (prd.auto_mark_completed)

Config overview

Generated defaults currently target OpenCode CLI. You can switch either agent to Codex or any custom command.

[prd]
file = "PRD.md"
auto_mark_completed = true

[workflow]
max_iterations = 12
max_fix_attempts = 2
auto_commit = true
execution_tests = ["cargo test"]

[loop_agent]
provider = "opencode"
command = "opencode"
args = ["run", "--model", "{model}", "--thinking", "{prompt}"]
model = "google/gemini-3-flash-preview"

[worker_agent]
provider = "opencode"
command = "opencode"
args = ["run", "--model", "{model}", "--thinking", "{prompt}"]
model = "google/gemini-3-pro-preview"

Template placeholders supported in args:

• {model}
• {prompt}
• {prompt_file} (absolute temp file path containing the prompt)

Safety notes

• workflow.auto_commit = true stages and commits all current workspace changes.
• Run with --dry-run first to inspect flow without calling agents/tests/commits.
• Start with a dedicated git branch.

Usage docs

Detailed command reference and advanced examples:

• docs/USAGE.md