aphrody-perfect-grind

name	aphrody-perfect-grind
version	2.0.0
description	Forced-loop coding mode. Wraps /loop 30s /aphrody-yolo-grind with an explicit perfection oracle — does NOT exit until the codebase meets all objective metrics. Every shell + every agent runs in background to maximize parallelism.
when_to_use	User types "/aphrody-perfect-grind", says "code en boucle", "force le grind", "go full autonomous until perfect", or wants the project driven to ship-ready state without any handhold. Use whenever the goal is "don't stop until the project is production-perfect".

Universal perfect grind — Loop until objectively perfect

Mode /goal permanent: persistent objective, zero confirmation, do not exit until the perfection oracle returns PASS.

Forced-loop mode. Every tick is a /aphrody-yolo-grind invocation on a 30-second cron; every shell + agent spawned from the orchestrator runs with run_in_background: true.

1. The Two Hard Rules

1. run_in_background: true on EVERY Bash call AND every Agent dispatch. Foreground tool calls block the orchestrator and break parallelism. The only exception is short read-only tools (Read, Glob, Grep, TaskUpdate) — those are inherently fast and don't need backgrounding.

2. Do NOT exit the loop until the perfection oracle returns PASS. Default Claude wants to hand back after a clean tick. Not here.

2. Dynamic Perfection Oracle

Runs after every tick. All gates relevant to the detected project type must pass:

2.1 Workspace & Compilation Gates

• Rust: cargo check --workspace --all-targets --locked and cargo clippy --workspace --all-targets --locked -- -D warnings must exit 0.
• Bun / Node.js: bun run lint / npm run lint or oxlint must exit 0.
• Python: ruff check / ruff format --check must exit 0.
• Go: go vet ./... must exit 0.

2.2 Testing Gates

• Rust: cargo test / cargo nextest run passes all tests.
• Bun / Node.js: bun test or npm test passes all tests.
• Python: pytest passes all tests.
• Go: go test ./... passes all tests.

2.3 Dependency & Supply Chain Gates

• Checks security/bans/licenses where available (e.g. cargo deny check / npm audit / uv pip-audit).
• Detects unused dependencies where applicable (e.g. cargo machete or equivalent).

2.4 Plan Gate (0 Unblocked Items)

• The detected plan/todo file must contain 0 open/unblocked items (e.g. searching for ⏳ or open checkbox - [ ] returns no unblocked tasks).

3. The Loop

loop forever:
  1. Tick fires (/aphrody-yolo-grind via cron or manual invocation).
  2. Wait for background agent notifications (orchestrator yields, does not poll).
  3. When done:
       - Run project validation suite (background).
       - Git add + commit staged changes (foreground OK, fast).
       - Run the perfection oracle (every gate in background, in parallel).
       - If any gate FAILS: do not break, the next tick will pick up.
       - If all gates PASS: delete the cron/loop timer and exit.
  4. ScheduleWakeup({ delaySeconds: 30, prompt: "/aphrody-perfect-grind continue" })

4. When to Break

ONLY break when:

• Perfection oracle PASSES, OR
• 3 consecutive ticks ship 0 FAIT (plan truly exhausted of autonomous-actionable tasks), OR
• A destructive remote operation is required (force-pushing to main, publishing package) — end the loop and document the gate; do not perform it autonomously.

NEVER break when:

• A single agent reports INCOMPLET — next tick covers it.
• A single oracle gate fails — next tick fixes it.
• "I think the user is happy" — wrong; perfect is objective, not vibes.

5. A2A Peer Coordination (If Enabled)

If A2A directory .coord/ is detected:

• Bumps heartbeat file .coord/heartbeat-<agent>.txt on every tick.
• Appends to .coord/inbox-from-<agent>.jsonl with tick states.
• When the oracle passes: posts a final fact envelope to peer containing the full oracle results, commit SHA, and timestamp.s/aphrody-yolo-grind/SKILL.md`** — the actual tick logic.