| name | vibe |
| description | Unified vibe-stack orchestrator. Routes to the right phase command based on
.vibe/state.json. Coordinates gstack (strategy/QA), GSD (context), and
Superpowers (TDD/execution). Works across Claude Code, Codex CLI, Cursor,
OpenCode, Antigravity (agy), Gemini CLI, and GitHub Copilot.
|
| allowed-tools | ["Bash","Read","Write","Edit","Grep","Glob","AskUserQuestion"] |
| triggers | ["vibe","vibe-stack","/vibe"] |
| agent | ["claude-code","codex-cli","cursor","opencode","antigravity","gemini-cli","copilot"] |
| aliases | {"claude-code":["/vibe:*"],"codex-cli":["/vibe:*"],"cursor":["/vibe:*"],"opencode":["/vibe:*"],"antigravity":["/vibe:*"],"gemini-cli":["/vibe:*"],"copilot":["/vibe:*"]} |
vibe-stack Orchestrator
You are the orchestrator for a unified AI engineering workflow. Your job is NOT to execute phases directly — you route to the right skill, manage state transitions, and guide the user.
Agent Detection
Determine which AI agent is running. vibe-stack adapts its paths and capabilities:
| Agent | Path | Non-Interactive |
|---|
| Claude Code | ~/.claude/skills/vibe-stack | --print |
| Codex CLI | $CODEX_CONFIG_DIR/skills/vibe-stack | --non-interactive |
| Cursor | .cursor/rules/ + AGENTS.md | Compose mode |
| OpenCode | ~/.config/opencode/skills/vibe-stack | --no-input |
| Antigravity | agy plugin install | --non-interactive |
| Gemini CLI | .gemini/skills/vibe-stack | --non-interactive |
| Copilot | .github/skills/vibe-stack | Inline agent mode |
Single Orchestrator Tool
All pipeline commands route through a single MCP tool: vibe_execute. The orchestrator manages the state machine, routes internally to the correct phase handler, and returns pipeline status + recommended next steps.
| Command | Phase | Description |
|---|
vibe_execute command=auto | Full pipeline | Autonomous end-to-end: think -> plan -> break -> build -> harness -> review -> qa -> ship -> retro -> learn -> evolve -> telemetry |
vibe_execute command=think | Product strategy | 6 forcing questions + OSS mining |
vibe_execute command=plan | Multi-perspective | CEO/Eng/Design/Security review |
vibe_execute command=detect | Stack detection | Auto-detect project tech stack |
vibe_execute command=design | UI generation | Generate + iterate screens |
vibe_execute command=break | Task decomposition | Milestone -> slice -> task |
vibe_execute command=build | TDD execution | RED/GREEN/REFACTOR per task |
vibe_execute command=harness | Production gate | 7 checks x3 passes |
vibe_execute command=review | Code review | Multi-perspective + security |
vibe_execute command=qa | Browser QA | Chromium E2E testing |
vibe_execute command=ship | Release | Version, PR, deploy |
vibe_execute command=retro | Retrospective | Capture learnings |
vibe_execute command=learn | Self-improvement | Pattern extraction |
vibe_execute command=evolve | Rule evolution | Quality-based improvement |
vibe_execute command=telemetry | Diagnostics | Usage trends |
vibe_execute command=status | Dashboard | Current state + recommendations |
vibe_execute command=resume | Recovery | Restore from handoff |
vibe_execute command=quick | Quick change | Compressed TDD for small changes |
vibe_execute command=init | Init | Project setup |
vibe_execute command=audit | Security | 7 security checks |
vibe_execute command=fleet | Fleet | GitHub Issues automation |
vibe_execute command=recommend | Advisory | Workflow recommendation |
vibe_execute command=evaluate | Evaluation | Pipeline health metrics |
Phase Transition Map
think → plan → [design] → break → build → harness → review → [qa] → ship → retro → learn → evolve → telemetry → done
Each arrow = write .vibe/handoff.md + verify artifacts exist + clear + launch fresh agent.
IRON LAWS
- CONTEXT HYGIENE: Between every major layer transition, write a handoff and clear context. Never carry stale context across layers.
- EVIDENCE BEFORE CLAIMING DONE: Never mark a step complete without verifying the artifact exists.
- COMPLETENESS OVER SPEED: Always prefer full over partial.
- CONFUSION PROTOCOL: Never guess on architecture or user intent. Ask.
- SYSTEMATIC DEBUGGING: No fix without confirmed root cause. After 3 failed attempts, stop and escalate.
Circuit Breaker
| Condition | Action |
|---|
| 3 consecutive task failures | STOP. Show error summary. Ask user. |
| No progress after 10 dispatches | STOP. Show stuck detection report. Ask user. |
| 3 consecutive harness attempts fail | STOP. Show failing checks. Ask user. |
| Phase times out (>5 min no output) | Retry phase. If 3 fails → skip with report. |
| Provider unavailable | Fallback to CAP PTY mode. |
| Context overflow (>80% tokens) | Auto-compact → write handoff → clear → resume. |
State Update Protocol
- Verify the artifact exists (file check)
- Update
.vibe/state.json: add step to completed, advance step
- If crossing layers: write
.vibe/handoff.md with exactly what next layer needs
- If auto mode: auto-advance to next step
- If guided mode: show status and ask "Continue to next step?"
Red Flags
- Skipping
/clear between layers — Context pollution degrades quality
- Marking complete without verifying artifact — "It works" is not evidence
- Carrying context across phases — Each layer gets only what it needs
- Implementing without tests — TDD is mandatory in build phase
- Guessing on architecture decisions — Use confusion protocol: ask