agy-rules-workflows

# agy-rules-workflows

Land the flywheel's invariant **laws** and its **operating loop** as Google Antigravity (`agy`) native Rules and Workflows — so an AGY agent crank obeys the same membrane as the Claude and Codex turnouts, expressed in AGY's own primitives rather than ported wrappers.

## Overview / When to Use

The flywheel runs the same five laws on every harness; only the *skin* changes per vendor. On Claude that skin is `operating-loop-workflow` + `bead-crank` + hooks; on AGY it is **markdown Rules** (invariant law, always-active) + a **markdown slash Workflow** (the loop trajectory) + **`BeforeTool`/`AfterTool` hooks** (mechanical guardrails) + **dynamic subagents** (author!=judge). Use this skill when you are bootstrapping or repairing the AGY node and need those four surfaces to carry the laws below.

The `agy` CLI is real and headless-capable: `agy -p "<prompt>"` (or `--print`) runs one prompt non-interactively; `agy plugin {list,import,install,validate}` packages skills/rules/hooks/subagents; config lives under `~/.gemini/` (Antigravity's on-disk root — `settings.json` for hooks/permissions, `skills/` for skills, `antigravity-cli/{brain,knowledge}/` for durable memory). Project rules live in `.agents/rules/`; loop workflows in `.agents/workflows/` invoked as slash commands.

AGY also exposes two **native loop controls** the contract targets: the AGY goal command pins a durable objective the agent steers toward across turns (the loop's north-star — set it to the bead/epic intent), and the AGY schedule command registers a recurring tick so the loop self-cranks headless without an external cron. Express the trajectory as the `agy-loop` workflow, set the objective with the goal command, and drive the cadence with the schedule command — these three AGY-native primitives replace the gemini-cli wrappers the legacy lane used.

## ⚠️ Critical Constraints

These five are the **laws** — they MUST become always-active AGY Rules, not advisory prose:

- **Rule 1 — Author != Judge.** The agent that wrote a change MUST NOT be the one that closes its bead. Validation runs in a **clean-context AGY subagent** (or worktree-isolated agent). **Why:** a self-grading author reliably rationalizes its own work to PASS — the single biggest false-close vector.
- **Rule 2 — Evidence-gated close.** A bead closes only when each Given/When/Then maps to a captured proof surface (test output, diff, artifact). **Why:** "looks good" closes leave the loop carrying broken state; evidence is the only thing the next ephemeral agent can trust.
- **Rule 3 — No self-grade.** An agent never asserts its own verdict as authority. It reports; a separate judge decides; ties go to a tie-break subagent. **Why:** self-report is biased toward "done" — external verification beats self-report (the core flywheel axiom).
- **Rule 4 — Scoped commits.** Each commit touches only the bead's declared write scope; no drive-by edits outside the slice. **Why:** scope collisions are what make parallel agent waves unsafe; tight scopes are what let waves run at all.
- **Rule 5 — Persist before done.** A loop turn is not complete until evidence + learning are written to shared state (the bead, `.agents/ratchet/`, an AGY artifact). **Why:** an ephemeral agent's knowledge dies with its context unless it is published to the system.

Mechanical enforcement: encode Rules 1/3 as subagent-dispatch policy, Rules 2/5 as an `AfterTool` close-guard hook, and Rule 4 as a pre-commit scope check. Rules are the law; hooks are the immune system.

## Workflow / Methodology

### Phase 1: Verify the AGY surface
Confirm the harness before writing law into it.
```bash
command -v agy && agy plugin list
ls ~/.gemini/settings.json ~/.gemini/skills 2>/dev/null
```
**Checkpoint:** `agy` resolves and `~/.gemini/` exists. If not, stop — the AGY image is not installed; this skill has nothing to enforce against.

### Phase 2: Write the invariant Rules
Materialize each law as an always-active markdown rule at project scope. Create `.agents/rules/flywheel-laws.md`:
```markdown
# Flywheel Laws (always active)
1. Author != Judge — the writer of a change never closes its own bead; close runs in a clean-context subagent.
2. Evidence-gated close — close only when every Given/When/Then maps to a captured proof surface.
3. No self-grade — report, never self-verdict; a separate judge decides; ties -> tie-break subagent.
4. Scoped commits — commit only the bead's declared write scope; no drive-by edits.
5. Persist before done — write evidence + learning to the bead, .agents/ratchet/, and an AGY artifact before closing the turn.
```
**Checkpoint:** the rules file exists and reads as imperative law (not advice). Confirm AGY loads project rules from `.agents/rules/`; if your install uses a different rules path, write there instead and note it.

### Phase 3: Write the loop Workflow
Create `.agents/workflows/agy-loop.md` — a slash workflow expressing claim->work->validate->close->persist (the seven-move operating loop in AGY skin). See `references/agy-loop-workflow.md` for the full workflow body to paste. The trajectory:
1. **Claim** — `br ready` / `bd ready`, pick one bead, claim it (`--claim`). One bead per turn.
2. **Work** — implement the smallest vertical slice; TDD per slice; record evidence onto the bead. Stay inside the declared write scope (Rule 4).
3. **Validate** — dispatch a **clean-context subagent** as judge (Rule 1/3). It re-derives PASS/FAIL from the proof surfaces only, never from the author's claim. Tie -> tie-break subagent.
4. **Close** — only on judge PASS with evidence mapped to each acceptance example (Rule 2). Otherwise reopen with the gap.
5. **Persist** — write evidence + learning to the bead + `.agents/ratchet/` + an AGY artifact, then commit the scoped change (Rule 5).

Bind the trajectory to AGY's native controls: set the durable objective with the goal command so every turn steers toward the same north-star, and register the recurring tick with the schedu