/execute

Execute an approved plan with Ralph Loop hard constraint. Agent stops don't matter — bash loop keeps going until all checklist items are done.

Step 1: Load Plan

Resolve which plan to execute: 1. Read `docs/plans/.active` — if it exists, use that path 2. If not found, find the most recently modified `docs/plans/*.md` and write it to `docs/plans/.active` 3. If multiple plans modified within the last hour, list them and ask the user to pick

Verify the plan has reviewer APPROVE verdict. If not approved, tell the user to run @plan first.

Step 1b: Detect Work Dir

Check if the plan header contains `**Work Dir:**`:

WORK_DIR=$(grep -oE '^\*\*Work Dir:\*\*\s*.+' "$PLAN_FILE" | sed 's/^\*\*Work Dir:\*\*\s*//' | tr -d '[:space:]')

If Work Dir is set: 1. Resolve to absolute path relative to project root 2. If path doesn't exist, create worktree (infer submodule and branch from plan slug) 3. Launch ralph loop with isolation env vars:

PLAN_POINTER_OVERRIDE=<plan_file_path> RALPH_WORK_DIR=<work_dir_abs> python3 scripts/ralph_loop.py

If Work Dir is absent, proceed normally (backward compatible).

Step 2: Verify Checklist

The plan MUST contain a `## Checklist` section with at least one `- [ ]` item. If missing, STOP and tell the user the plan needs a checklist.

Step 3: Launch Ralph Loop

Run **foreground** (NEVER use `nohup &` — you need to see the exit summary):

python3 scripts/ralph_loop.py

This bash script will:

• Loop until all `- [ ]` items become `- [x]`
• Each iteration spawns a fresh Kiro CLI instance with clean context
• Circuit breaker: exits if 3 consecutive rounds make no progress
• Agent stopping is fine — the loop restarts a new instance
• On exit (success or failure), prints a full summary to stdout

Step 4: Report Results

The script prints a summary block on exit. Use that output to report:

• How many checklist items completed vs total
• Any `- [SKIP]` items with reasons
• Read skills/omk-finishing/SKILL.md for merge/PR/cleanup options