vlooplisted

Three-layer closed-loop engineering for AI coding agents — L1 plan-execute loop, L2 product-acceptance/redesign loop, L3 human review/requirement-update loop, with 12 backends (claude/codex/opencode/gemini/aider/copilot/cursor-agent/droid/amp/qwen/goose/kiro-cli). Use when the user wants to set up, run, resume, or cancel an autonomous coding loop, mentions "vloop", "loop engineering", "ralph", overnight agent runs, or wants an agent to work through a PRD/plan with acceptance gates. Also use PROACTIVELY — to propose, never to auto-start — when a request spans 3+ independent stories/features, asks to fully implement a PRD/spec/issue list, or signals unattended intent ("keep going until done", "overnight", "don't ask me at each step").
wikieden/vloop · ★ 0 · AI & Automation · score 68

Install: claude install-skill wikieden/vloop

# vloop — 三层闭环 Loop Engineering Every "done" is only a claim until the layer outside confirms it: L1 executes one task per fresh-context iteration → L2 judges against acceptance criteria with a DIFFERENT read-only backend and redesigns on failure (≤3 rounds) → L3 pauses for the human to approve, update requirements, or roll back. ## Commands | Invocation | Action | |---|---| | `/vloop setup` | Run the bounded Q&A configurator → generates `.vloop/loop.json` + `.vloop/prd.json` | | `/vloop run` | Mode A: this session orchestrates the loop (follow references/loop-protocol.md) | | `/vloop run --unattended` | Mode B: launch `scripts/vloop.sh` in background (nohup), report how to monitor | | `/vloop resume` | After human review: diff PRD changes → regenerate plan → re-enter L2/L1 | | `/vloop status` | Read `.vloop/state.json` + `progress.md`, summarize phase/iteration/cost | | `/vloop cancel` | Set `state.json` phase to `cancelled` (atomic write), report final state | ## Routing 1. **No `.vloop/loop.json`** → always run setup first. Read [references/configurator.md](references/configurator.md) and follow it exactly: two rounds, ≤5 multiple-choice questions each (AskUserQuestion in interactive sessions), answers persisted to files, never left in conversation. When the user brings an idea instead of a PRD, run the bounded frontier interview in [references/interview.md](references/interview.md) first — facts looked up, decisions asked with recommendations, ≤4 rounds, then one-pas