# Antigravity > --- - Skill: `vibhasdutta/antigravity` (Agent Skill, multi-file: 43 files) - Install (CLI): `npx skillmds@latest add vibhasdutta/antigravity` - Raw SKILL.md: https://api.skillmd.com/api/skills/vibhasdutta/antigravity/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: vibhasdutta (https://skillmd.com/u/vibhasdutta) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/vibhasdutta/antigravity --- --- name: loop-engineer description: > Loop engineering wizard for Antigravity. Asks 3 questions, then orchestrates a fully autonomous parallel agent team (resource-scout, researcher, planner, agent-factory, executor, verifier, auditor, memory-keeper) until the goal is met. Uses invoke_subagent for true parallel dispatch. Researcher agents use the built-in research TypeName. Modes: build (from scratch), research (investigate only), patch (fix/extend existing code), audit (review only, no changes). Persistent memory, git integration. --- # Loop Engineer (Antigravity) You are running a loop engineering wizard. Follow these phases in order. **Prerequisite — reduce permission friction:** confirmed via antigravity.google/docs/permissions — file reads/writes inside your active project are auto-allowed by default, so that part shouldn't interrupt you. `command(...)` actions (every terminal/bash call — including the init script, and every `git commit` the executor makes) default to **Ask** unless you've added an allow rule. Before running an autonomous loop, add an Allow rule for the commands this loop needs (at minimum `command(bash)`, or `command(*)` for zero friction) via Antigravity's permission settings — either the in-editor approval-card "expand scope" option the first time a command prompts, or your project/user settings directly. Without this, the loop will pause on every single Bash call, which defeats "fully autonomous." --- > **To update loop-engineer:** re-run `install.sh --update` / `install.ps1 -Update -Antigravity`. Updates are never applied automatically mid-loop. --- ## Phase 1 — Core Wizard **Q1 — Mode:** if invoked with an argument matching `build`/`research`/`patch`/`audit`, use it as MODE and skip this question. Otherwise ask: "Mode? build (new from scratch) / research (investigate and report, no code changes) / patch (fix or add a feature using the existing codebase) / audit (review existing code/output only, no changes)". Default to `build` if unclear. **Q2:** "What do you want the loop to accomplish? (1-2 sentences)" Generate LOOP_ID: lowercase slug, first 4 meaningful words, max 24 chars. Auto-set: `STOP_CONDITION` = "all tasks in loop-stack//PLAN.md checked", `MAX_TURNS` = 20. **Q3:** "Should the loop auto-commit after each verified task? (yes / no)" --- ## Phase 2+3 — Initialize Loop Run the init script — tries all three Antigravity surface paths automatically: **Bash (macOS/Linux):** ```bash INIT_SCRIPT="$(ls \ "$HOME/.gemini/antigravity-cli/skills/loop-engineer/scripts/init-loop.sh" \ "$HOME/.gemini/antigravity/skills/loop-engineer/scripts/init-loop.sh" \ "$HOME/.gemini/config/skills/loop-engineer/scripts/init-loop.sh" 2>/dev/null | head -1)" bash "$INIT_SCRIPT" \ --loop-id \ --goal "" \ --stop "all tasks in loop-stack//PLAN.md checked" \ --git \ --mode \ --platform antigravity ``` **PowerShell (Windows):** ```powershell $init = @( "$env:USERPROFILE\.gemini\antigravity-cli\skills\loop-engineer\scripts\init-loop.ps1", "$env:USERPROFILE\.gemini\antigravity\skills\loop-engineer\scripts\init-loop.ps1", "$env:USERPROFILE\.gemini\config\skills\loop-engineer\scripts\init-loop.ps1" ) | Where-Object { Test-Path $_ } | Select-Object -First 1 & $init -LoopId "" -Goal "" ` -Stop "all tasks in loop-stack//PLAN.md checked" ` -Git -Mode -Platform antigravity ``` If the script is missing, install the skill first: `git clone https://github.com/vibhasdutta/loop-engineer && bash install.sh --antigravity` The script creates `loop-stack//`, `.agents/` with all agent .md files + knowledge-sources/, and `verifier.md` with the actual stop condition substituted. --- ## Phase 4 — Startup Sequence **How parallel dispatch actually works in Antigravity:** confirmed via official docs (antigravity.google/docs/subagents) — there is no "Subagents array" parameter. The real mechanism: calling `invoke_subagent` spawns a subagent that "immediately begins executing its task" and "runs asynchronously in the background, allowing the parent agent to delegate a task and immediately resume its own work." So to get true concurrency, call `invoke_subagent` multiple times back-to-back in the same turn — each call returns right away without blocking, so N calls in a row dispatch N subagents running concurrently. Then wait for all of them to send their completion messages before proceeding to the next step. **Nesting limit:** a maximum depth of 10 levels of subagents-of-subagents is enforced. Not a concern for loop-engineer's flat dispatch pattern, but worth knowing if an agent-factory-created specialist tries to spawn its own subagents. **IMPORTANT:** Subagents start with a clean slate — no parent context. Every `Prompt` must be fully self-contained with the loop directory path, what to read, what to write, and a reference to the agent's instruction file in `.agents/`. ### Step 1 — Parallel RESEARCHERS (dynamic count) Determine researcher count by goal complexity: - Simple/single-domain → 2 researchers - Medium/multi-domain → 3 researchers - Large/multi-system → 4 researchers Call `invoke_subagent` once per researcher, back-to-back in the same turn (each call returns immediately, so this dispatches all of them concurrently — see note above). Use `TypeName: "research"` — the built-in type optimized for codebase exploration. Domains to distribute: - **Context & Prior Work**: source structure, patterns, package files, existing tests - **External Knowledge & Resources**: README, docs/, external APIs, .env.example, configs - **Requirements & Constraints**: DB schema, data models, state management (for 3+ researchers) - **Environment & Integration**: CI/CD, infrastructure, build, Docker (for 4 researchers) Each researcher Prompt: ``` Loop directory: loop-stack// Focus: {ASSIGNED_DOMAIN} — {specific files and concerns} GLOBAL DATA FIRST: read loop-stack/.global/MEMORY.md and loop-stack/.global/TOOLS.md. Write findings to loop-stack//RESEARCH.md under "## {Domain Name}". Update loop-stack//STATUS.md "Last Researcher Result". Read .agents/researcher.md for full instructions. ``` **Stuck-agent check (no watcher subagent needed):** Antigravity already exposes `manage_subagents(Action: "list")` — use it directly if one researcher is taking noticeably longer than its peers, to see whether it's still running or has stalled. If stalled, proceed without it and note `{agent}: STUCK — no output` in STATUS.md `## Active Heartbeats`. Wait for all researcher subagents to complete before Step 2. ### Step 2 — RESOURCE SCOUT Call `invoke_subagent` with one entry, `TypeName: "self"`: ``` Loop directory: loop-stack// Check loop-stack/.global/TOOLS.md — if < 7 days old, reuse it. Otherwise discover all tools. Write to loop-stack//TOOLS.md AND loop-stack/.global/TOOLS.md. Read .agents/resource-scout.md for full instructions. ``` Wait for resource-scout to send its completion message. ### Step 3 — PLANNER Call `invoke_subagent` with one entry, `TypeName: "self"`: ``` Loop directory: loop-stack// Read RESEARCH.md (all sections) and TOOLS.md. Task type depends on MODE: build/patch → implementation tasks; research → research/writing tasks only, no code changes; audit → review tasks only, no code changes. Create 3–7 tasks with parallel group tags [G1], [G2], etc. Same group = parallel (independent files/modules). Different group = sequential dependency. Replace "## Tasks" in PLAN.md. Update STATUS.md. Read .agents/planner.md for full instructions. ``` Wait for planner to send its completion message. Write `loop-stack//AGENTS.md` with `# Specialized Agents\n## Status\nNONE CREATED YET`. (Agent-factory is on-demand, not a fixed step — see Rules.) Proceed to Phase 5. --- ## Phase 5 — Outer Loop **FULLY AUTONOMOUS. Never pause for user input.** Initialize: `turns_used = 0`, `skipped_tasks = []`. **Mode gating:** build (default) — full flow. patch — same steps, but every researcher/executor prompt adds "existing codebase is ground truth — fix/extend, don't rewrite from scratch." research — skip step 5 (executors) and steps 6–7 (audit) entirely; the researcher (step 3) writes each task's final deliverable directly; the verifier checks that instead of built code. audit — skip step 5; the auditor step IS the task (read-only review, findings to RESEARCH.md); a BLOCK verdict is just recorded, never auto-fixed, always proceeds to the verifier. For parallel steps: call `invoke_subagent` once per agent, back-to-back in the same turn (each call returns immediately without blocking — see Phase 4 note). Wait for all to send completion messages before proceeding. ### Iteration steps: 1. **Budget check** — over MAX_TURNS → Phase 6. 2. **Read state** — identify current parallel group (all unchecked [GN] tasks). 3. **Parallel RESEARCHERS** — one per task (2 if batch=1). Call `invoke_subagent` once per task, back-to-back in the same turn, `TypeName: "research"`: ``` Loop directory: loop-stack// GLOBAL DATA FIRST — read loop-stack/.global/MEMORY.md AND loop-stack/.global/TOOLS.md. Focus: {ASSIGNED_DOMAIN} for task: {this_task} Append findings to RESEARCH.md under "## Task-Specific Research — {this_task}". Update STATUS.md "Last Researcher Result". Read .agents/researcher.md for full instructions. ``` Wait for all. Increment turns_used. 4. **Before executors, check AGENTS.md.** For every task in the batch that clearly needs domain expertise beyond a generic executor and has no specialist yet, call `invoke_subagent` once per such task, back-to-back in the same turn (`TypeName: "self"`, each reads `.agents/agent-factory.md`) to create their agent files and update AGENTS.md — same parallel-first rule as researchers/executors. Wait for all to finish. Skip this for most tasks. 5. **Parallel EXECUTORS** — one per task. Call `invoke_subagent` once per task, back-to-back in the same turn, `TypeName: "self"`: ``` Loop directory: loop-stack// GLOBAL DATA FIRST — read loop-stack/.global/MEMORY.md AND loop-stack/.global/TOOLS.md. Read MEMORY.md, TOOLS.md, PLAN.md, STATUS.md. READ: RESEARCH.md "## Task-Specific Research — {this_task}". Current task: {this_task}. Scope: only files for this task. Implement. Append discoveries to MEMORY.md directly. Update STATUS.md. Goal output (code, documents, files) goes to the project directory, NOT inside loop-stack/. loop-stack/ is state-only. Read .agents/executor.md for full instructions. ``` Wait for all. Increment turns_used. 6. **Parallel AUDITORS** — one per task just built. `TypeName: "self"`. Wait for all. 7. **Process audit results**: - CLEAN/WARN → proceed to verifier - BLOCK → auto-fix (one executor with BLOCK context, `TypeName: "self"`, re-run auditor once). Still BLOCK → auto-skip. 8. **Parallel VERIFIERS** (tasks that passed audit only) — one per task. Verifier is the final gate: checks the task against RESEARCH.md's Verification Criteria (right place, satisfies criteria, no placeholders), then runs the stop condition. `TypeName: "self"`. Wait for all. 9. **Process verifier results**: - PASS → memory-keeper - FAIL < 3 → retry from step 3 - FAIL ≥ 3 → auto-skip 10. **MEMORY-KEEPER consolidation** — single `invoke_subagent`, `TypeName: "self"`, local + global write. This is the only memory-keeper call per batch — executors already appended their raw learnings inline during step 5. Wait. 11. **Advance** — mark [x], git commit if enabled. Find next group. None → ALL DONE → rename `loop-stack//` → `loop-stack/_DONE/` → Phase 6. --- ## Phase 6 — Completion Report Write `REPORT.md` inside the renamed loop directory and print summary. --- ## Rules - **File copy**: try CLI path (`~/.gemini/antigravity-cli/skills/`), then IDE path (`~/.gemini/antigravity/skills/`), then 2.0 path (`~/.gemini/config/skills/`). Never write manually. - **Global data first**: every agent reads `.global/MEMORY.md` + `.global/TOOLS.md` before acting. - **invoke_subagent**: parallel = multiple calls back-to-back in the same turn (each returns immediately, no "Subagents array" parameter exists). Sequential = deliberately waiting for one to finish before calling the next. Wait for completion messages between steps. - **TypeNames**: use `"research"` for researcher agents (built-in, codebase-optimized). Use `"self"` for all others. - **Self-contained prompts**: subagents have a clean context slate — every Prompt must include loop dir, files to read/write, and `.agents/{role}.md` reference. - **Researcher before executor**: always. Dynamic count based on goal complexity. - **Agent-factory is on-demand, not a fixed phase step.** Invoke it only right before executing a task that clearly needs a specialist. Most loops never call it. - **knowledge-sources.md is a reference file researchers consult on demand**, not a phase step. - **No watcher agent.** Use `manage_subagents(Action: "list")` directly if you need to check whether an agent is still running; never spawn a dedicated watcher subagent. - **No separate evaluator.** Verifier is the final gate: checks the task against RESEARCH.md's Verification Criteria, then runs the stop condition. One agent, one call, same rigor — not two sequential subagents for what's really one quality gate. - **Audit before verify.** Auditor reviews the build first (step 6); verifier runs last as the final pass/fail gate (step 8) and is what triggers retry. - **Memory-keeper runs once per task batch** (after verify, local + global) — not a separate mid-batch checkpoint. Executors already append their own learnings to MEMORY.md directly as they work, so there's nothing left for a checkpoint pass to do that isn't already written. Its only job is capturing learnings/context — never executes the goal or writes goal output. - **Executors append to MEMORY.md directly** during work. - **Planner**: once at startup. Tasks MUST include [G1]/[G2] parallel group tags. - **Fully autonomous**: no pauses. Audit BLOCK → auto-fix once → skip. 3 verifier fails → auto-skip. - **Modes**: `build` (default), `research`, `patch`, `audit` — set once in Phase 1, gates Phase 5 (see above). - **HARD RULE — no plan-approval gate**: after Phase 1's questions, proceed through Phase 2 onward without presenting a plan for approval or waiting for a "click proceed" confirmation. Do not use any native plan-then-approve UX for this loop. - No resume support: every invocation starts a fresh loop. On completion: rename to `_DONE/` (bookkeeping only). - **Monitor**: use `manage_subagents(Action: "list")` to check running subagents if needed. - **AGENTS.md**: ensure `AGENTS.md` is in the project root for workspace context. - **MCP config**: remote servers require `serverUrl` field (not `url` or `httpUrl`).