loop-launch-a

Action skill — Pre-flight, setup, and launch for agent loops.

Input

$ARGUMENTS

---

Launch an agent loop in the current repo. This skill handles everything from validation to tmux launch.

Pre-Flight Checks

Run these checks before launching. If any fail, report the issue and stop — do not launch a broken loop.

1. Git State

# Check for uncommitted changes
git status --porcelain

• If there are uncommitted changes, warn the user. Uncommitted work can be lost or cause conflicts during the loop.
• Suggest committing or stashing before launch.

2. Branch State

# What branch are we on?
git rev-parse --abbrev-ref HEAD

• Note the current branch — this becomes the base branch for the loop.

3. Existing Loops

# Check for active loops in this repo
ls agents/loops/ 2>/dev/null

• If other loops are running, warn about potential conflicts (especially if on the same branch).
• Check ~/.claude-loops/registry.jsonl for start events without matching end events in this repo.

4. Tooling Setup

# Verify agents/ directory exists with required files
ls agents/prompts/session.md agents/prompts/worker.md agents/prompts/reviewer.md agents/run.sh 2>/dev/null

• If agents/ is not set up, run the setup script:

  ~/claude-tooling/agents-src/setup.sh
  

• If setup script is not available, tell the user to set up the agents directory first.

Loop ID

Derive the loop-id from the user's input:

• If the user provided a loop-id, use it directly
• If the user described a task, derive a short kebab-case slug (e.g., "fix auth flow" → fix-auth-flow)
• Append timestamp: {slug}_{YYYYMMDD_HHMM} (e.g., fix-auth-flow_20260315_1430)

Session Prompts

Check if session prompts exist for this loop:

ls agents/loops/{loop-id}/session/SESSION_WORKER.md agents/loops/{loop-id}/session/SESSION_REVIEWER.md 2>/dev/null

If they don't exist:

1. Ask the user if they want to write them now or have you generate them
2. If generating, invoke the loop-author-a skill with the task description
3. If writing manually, create the directory structure and tell them where to write:

   agents/loops/{loop-id}/session/SESSION_WORKER.md
   agents/loops/{loop-id}/session/SESSION_REVIEWER.md
   

Launch

Once pre-flight passes and session prompts exist:

# Launch the loop
./agents/run.sh --loop-id {loop-id} [additional flags]

run.sh automatically wraps itself in a tmux session named loop-{loop-id}.

Additional Flags

Ask the user or infer from context:

• --max-iter N — default is 5
• --worker-spec FILE — specialization (e.g., planner.md, python.md, rust.md)
• --pr-per-iteration — enable PR-per-iteration branching
• --claude-effort LEVEL — low, medium (default), or high

Post-Launch Output

After launching, tell the user:

1. The tmux session name: loop-{loop-id}
2. How to attach: tmux attach -t loop-{loop-id}
3. How to detach without stopping: Ctrl+B, then D
4. How to check status: point them to the loop-observe-a skill
5. Where state files will live: agents/loops/{loop-id}/state/
6. Where git-committed history will be: .agent-loops/{loop-id}/

Common Patterns

User says	What to do
"Start a loop for X"	Full pre-flight → generate prompts → launch
"Launch loop {id}"	Pre-flight → check prompts exist → launch
"Resume loop {id}"	Check if loop dir exists, session prompts present → launch with existing dir