agy-native

What this skill does

# agy-native

Drive the AgentOps loop on the **Antigravity image** (AGY): the `agy` CLI backed by the brain/knowledge store at `~/.gemini/antigravity-cli/`. This is the third harness alongside the Claude image and the Codex image — same loop laws, AGY-native primitives. **Invoke `agy`, never rebuild it.**

## Overview / When to Use

AGY is Google's Antigravity agent harness exposed as a local CLI (`~/.local/bin/agy`). It speaks the Claude-style packaging shape (plugins, skills, subagents, hooks, MCP) and a **portable `SKILL.md`** — AGY reads skills from `~/.gemini/skills/` directly, so the AgentOps corpus already loads on it. Use this skill when you need the claim->work->validate->close->persist loop running on AGY (driving Gemini, or Claude/GPT models *through* AGY — `agy models` exposes Gemini 3.5/3.1, Claude Sonnet/Opus 4.6, GPT-OSS), as a parallel or fallback to the Claude/Codex images.

**AGY ≠ gemini-cli.** The retired `gemini` CLI lane (`gemini -p`, `gemini skills`, `gemini extensions`, `--approval-mode`, `--worktree`) is gone. Everything here is an AGY affordance; §"Distribution" + the reference file name each AGY equivalent for an old gemini habit.

Verified primitives on this host (`agy --help`, `agy plugin help`, `agy models`):
- **Headless run:** `agy -p "<prompt>"` / `agy --print` (one-shot, prints, exits; `--print-timeout` default 5m). `-c`/`--continue` resumes the most recent conversation; `--conversation <id>` resumes by ID.
- **Plugins:** `agy plugin {list,import,install,uninstall,enable,disable,validate,link}`. `import [gemini|claude]` pulls existing plugin trees in. `install <target>` reads a `plugin.json` (supports `plugin@marketplace`).
- **Permissions:** `--dangerously-skip-permissions` auto-approves tool calls (loop/headless lane); `--sandbox` restricts the terminal.
- **Workspace scope:** `--add-dir <dir>` (repeatable) scopes which repos a run can touch (AGY's write-isolation primitive — it scopes by directory, not by spawning worktrees).
- **Brain/knowledge:** durable agent memory + user-facing artifacts under `~/.gemini/antigravity-cli/{brain,knowledge}/` (per-conversation dirs; `*.md` + `*.md.metadata.json` with `{summary, updatedAt, userFacing}`).

## ⚠️ Critical Constraints

- **Rule 1 — Never `claude -p` for workers (LAW 0).** AGY runs on Gemini OAuth (and proxied Claude/GPT). Drive AGY workers with `agy --print` or `agy -i`, Codex with `codex exec`, Claude only via NTM panes / subagents. **Why:** `claude -p` bills the API per-token, not the Max sub; the overnight factory burned API this exact way (banned).
- **Rule 2 — author != judge, always two contexts.** The subagent that closes a bead must NOT be the one that validates it. Spawn the judge as a separate async subagent with a clean context (or a separate `agy --print` invocation; **never** `-c`/`--continue` across roles — it shares context). **Why:** a self-grading worker is a flatterer; independent verdict is the membrane (control-plane LEARNINGS: a tie-break quorum caught a false-FAIL).
- **Rule 3 — evidence-gated close.** A bead closes only against a persisted verdict artifact (a `brain/*.md` with `userFacing:true` or a committed repo file), never against chat text alone. **Why:** agents are ephemeral; the system carries state. Consume an agent's *published compression*, never its live session.
- **Rule 4 — worktree / `--add-dir` isolation.** Concurrent author and judge get isolated worktrees or non-overlapping `--add-dir` scopes. No two roles edit the same file. **Why:** prevents swarm races and clobbered work.
- **Rule 5 — `dcg` guard stays on.** `~/.gemini/settings.json` wires a `BeforeTool` hook on `run_shell_command` to `dcg`. Do not remove it even under `--dangerously-skip-permissions`. **Why:** it blocks destructive commands the auto-approve flag would otherwise let through.
- **Rule 6 — match permission to role.** Author = `--dangerously-skip-permissions` with a **tight** `--add-dir`; judge = **default** (no auto-approve) with a read-mostly scope; full-auto only inside `--sandbox`. **Why:** auto-approve is a blast-radius choice — a validator that can edit is a false-close path.
- **Rule 7 — operator-side; invoke-never-rebuild.** This drives the flywheel harness. Do NOT write under `~/dev/agentops`, do NOT git push agentops, do NOT treat AGY as something to re-author. **Why:** AGY is Emanuel's substrate (ACFS doctrine) — own a thin adapter, not the tool.

## Distribution — exposing skills/plugins to AGY

AGY discovers capability through three layers (weight ascending): a **portable `SKILL.md`** under `~/.gemini/skills/<name>/` (read directly — no packaging), a **plugin** tree with `plugin.json` (`skills`/`subagents`/`hooks`/`mcpServers`), and **MCP servers** (the `agy-mcp-plugins` lane). The retired gemini split of `skills` vs `extensions` collapses into the single AGY **plugin** unit — there is no `agy extensions` surface; treat a former gemini "extension" as an AGY plugin.

Dev discipline (folded from the retired extension lane): **`agy plugin link <path>`** for local development (live edits, never a stale copy); **`agy plugin install <dir|name@marketplace>`** for released/remote artifacts; don't run both for one plugin. **Source of truth stays in AgentOps** — do not hand-edit the managed runtime copies under `~/.gemini/skills/`. Every mutation: **validate → apply → list → record rollback** (see references). Full verb list, install-vs-link table, and the permission×output×scope matrix: **[references/distribution-and-run-control.md](references/distribution-and-run-control.md)**.

## Workflow / Methodology

### Phase 1: Verify the image is live
```bash
which agy && agy models | head        # CLI present, models reachable
ls ~/.gemini/antigravity-cli/{brain,knowledge}   # brain store exists
agy plugin list                       # what's already imported/enabled
```
**Checkpoint:** confirm `agy` resolves, a model lists, and the brain dir exists before dispatching any tick.

### Phase 2: Package + expose the laws
A bare `SKILL.md` under `~/.gemini/skills/` is portable — no `plugin.json` needed just to expose a skill. To bundle rules + workflows + subagents + hooks + MCP, lay out a plugin tree (the `agy-control-plane` unit):
```
agy-control-plane/
  plugin.json   # { name, version, skills, subagents, hooks, mcpServers }
  rules/        # invariant law (author!=judge, evidence-gated close, scoped commit)
  workflows/    # slash-command loop trajectories (claim->work->validate->close->persist)
  subagents/    # worker.md, validator.md, tie-break.md, scout.md
  hooks.json    # pre/post-tool guardrails (close gate, format/lint, dcg)
  skills/       # or rely on portable ~/.gemini/skills/
```
```bash
agy plugin import claude               # pull an existing plugin tree in (optional)
agy plugin validate ./agy-control-plane
agy plugin link ./agy-control-plane    # dev (live edits) — or `install` for released
agy plugin enable agy-control-plane
agy plugin list                        # confirm enabled (rollback: disable/uninstall)
```
**Checkpoint:** `agy plugin validate` passes and `agy plugin list` shows the plugin enabled.

### Phase 3: One headless tick (author)
Spawn the author in a tight scope with scoped-auto-edit permission; let it claim and work one ready bead:
```bash
agy --print --add-dir "$REPO" --dangerously-skip-permissions \
  "Claim one ready bead via br. Implement only it in this worktree. \
   Commit scoped. Write evidence to brain as userFacing. Do NOT close it — a judge will."
```
**Checkpoint:** a scoped commit exists and an evidence artifact landed in `brain/`; the bead is implemented but still OPEN.

### Phase 4: Independent verdict (judge — separate context)
Spawn the judge as an async subagent / second `agy --print` with a **fresh context and a read-mostly scope** (default permissions, no auto-approve):
```bash
agy --print --add-dir "$REPO" \
  "Validate bead <id> against its evidence artifact ONLY. You did not author it. \
   Emit PASS/WARN/