loopty

Start an iterative development loop. Each iteration spins up a fresh, isolated Claude agent that works toward a goal, writes a timestamped journal entry, and commits — then hands off to the next agent.

Usage

The user may provide arguments to customize the run. Parse $ARGUMENTS for any of:

• A time interval (e.g., "10m", "30m", "600") — defaults to 15m
• A max iteration count (e.g., "3 iterations", "5x", "n=3") — defaults to unlimited
• A prompt file path (ending in .md) — defaults to .loopty/prompt.md
• An inline goal in quotes (e.g., "improve test coverage") — creates/updates .loopty/prompt.md
• "status" — show journal history summary instead of starting a loop
• "resume" or "--resume" — continue iteration numbering from last journal
• "no-spin-check" or "--no-spin-check" — disable spin/stall detection
• "dry-run" or "--dry-run" — show config without starting
• "quiet" or "--quiet" — suppress banner output
• "verbose" or "--verbose" or "-V" — show detailed journal content in status
• "json" or "--format json" — machine-readable JSON output for status

Steps

Bootstrap (first time only):

If loopty.sh is not present in the project root, the skill needs to be bootstrapped.

To find the plugin source files, resolve this path relative to this SKILL.md:

• SKILL_DIR is the directory containing this file (skills/loopty/)
• PLUGIN_ROOT is two levels up from SKILL_DIR (the repo root)

Run the install script from the plugin root:

bash "$PLUGIN_ROOT/install.sh" "$(pwd)"

If install.sh is not found, bootstrap manually:

1. Copy $PLUGIN_ROOT/loopty.sh to the project root
2. Copy $PLUGIN_ROOT/.claude/commands/loopty.md to the project's .claude/commands/
3. Create .loopty/ directory and copy $PLUGIN_ROOT/.loopty/prompt.md.example
4. Tell the user that loopty has been bootstrapped into their project

If "status" mode:

1. Run bash loopty.sh status and present the output
2. Read .loopty/last-run-summary.md if it exists for additional context
3. Do NOT start a loop

If normal run mode:

1. Check that .loopty/prompt.md exists. If not:

   • If the user provided an inline goal in quotes, create the prompt file with that goal
   • Otherwise, ask the user what their iterative goal is and create it for them
   • The prompt should describe: the goal, success criteria, and constraints

2. Parse arguments to build CLI flags:

   • Time like "10m" → -i 600
   • Time like "30m" → -i 1800
   • Time like "600" (bare number > 59) → -i 600
   • Iterations like "3 iterations" or "3x" or "n=3" → -n 3
   • "resume" or "--resume" → --resume
   • "--dry-run" or "dry-run" → --dry-run
   • "quiet" or "--quiet" → --quiet
   • Work turns like "w=50" or "--work-turns 50" → -w 50
   • Inline goal in quotes → -g "the goal text"
   • Model override → -m MODEL
   • Wrapup timeout like "wrapup-timeout=60" → --wrapup-timeout 60
   • Cooldown like "cooldown=5" → --cooldown 5
   • "no-cooldown" or "--no-cooldown" → --no-cooldown
   • "no-commit" or "--no-commit" → --no-commit
   • "no-spin-check" or "--no-spin-check" → --no-spin-check
   • "verbose" or "--verbose" or "-V" → --verbose
   • "json" or "--format json" → --format json

3. Confirm the settings with the user before starting:

   • Show the prompt content (first ~10 lines if long)
   • Show the interval and max iterations
   • Show how many previous journal entries exist in .loopty/journal/
   • If --dry-run, run the script with --dry-run and show output, then stop
   • Otherwise, ask for confirmation to begin

4. Run the loop script:

   bash ./loopty.sh [flags]
   

   This will run in the foreground. Each iteration:

   • Launches an isolated claude -p agent with the goal + previous journals
   • After the time budget, a wrap-up agent writes a journal entry
   • All changes are committed with git

5. When the loop finishes (max iterations reached or interrupted), summarize:

   • How many iterations ran
   • Key findings from the journal entries
   • What the next agent would work on
   • Reference .loopty/last-run-summary.md for full details