gstack/agy/SKILL.md.tmpl

---
name: agy
preamble-tier: 3
version: 1.0.0
description: |
  Google Antigravity CLI wrapper — three modes. Code review: independent diff review
  via `agy -p` (headless print mode) with pass/fail gate. Challenge: adversarial
  stress-testing of proposed designs and race conditions. Consult: multi-turn session
  continuation via `agy -c`. The "200 IQ second opinion from a different AI system."
  Use when asked to "agy review", "agy challenge", "antigravity review", "consult agy",
  or "second opinion". (gstack)
voice-triggers:
  - "a g y"
  - "antigravity review"
  - "get another opinion"
triggers:
  - agy review
  - agy challenge
  - agy consult
  - antigravity second opinion
  - outside voice challenge
allowed-tools:
  - Bash
  - Read
  - Write
  - Glob
  - Grep
  - AskUserQuestion
---

{{PREAMBLE}}

{{BASE_BRANCH_DETECT}}

# /agy — Google Antigravity Second Opinion

You are running the `/agy` skill. This wraps the Google Antigravity CLI (`agy`) to get
an independent, brutally honest second opinion from a different AI system.

Antigravity is a peer reviewer: direct, adversarial where needed, technically precise,
and unafraid to challenge assumptions. Present its output faithfully, not summarized.

---

## Step 0.4: Check agy binary

```bash
AGY_BIN=$(command -v agy || command -v antigravity || echo "")
[ -z "$AGY_BIN" ] && echo "NOT_FOUND" || echo "FOUND: $AGY_BIN"
```

If `NOT_FOUND`: stop and tell the user:
"Antigravity CLI not found. Install it: `pip install antigravity-sdk` or see https://antigravity.google.dev/cli"

---

## Step 0.5: Auth probe

Before building expensive prompts, verify Antigravity has valid auth.

```bash
# Check for valid auth: API key env var OR application default credentials
if [ -n "$ANTIGRAVITY_API_KEY" ] || [ -n "$GOOGLE_APPLICATION_CREDENTIALS" ]; then
  echo "AUTH_OK_ENV"
elif "$AGY_BIN" whoami --quiet 2>/dev/null; then
  echo "AUTH_OK_ADC"
else
  echo "AUTH_FAILED"
fi
```

If `AUTH_FAILED`: stop and tell the user:
"No Antigravity authentication found. Run `agy auth login` or set `$ANTIGRAVITY_API_KEY`, then re-run this skill."

---

## Step 0.6: Resolve portable roots

```bash
eval "$(~/.claude/skills/gstack/bin/gstack-paths)"
```

All subsequent bash blocks use `"$PLAN_ROOT"` and `"$TMP_ROOT"` rather than hardcoded
paths.

---

## Step 0.7: Model selection

Antigravity supports model scaling via the `ANTIGRAVITY_MODEL` environment variable.

```bash
# Default model per mode (overridable via ANTIGRAVITY_MODEL env var):
# review   → antigravity-2.0-flash   (fast, bounded diff input)
# challenge → antigravity-2.0-pro    (deeper reasoning for adversarial mode)
# consult   → antigravity-2.0-flash  (interactive speed priority)
_AGY_MODEL="${ANTIGRAVITY_MODEL:-antigravity-2.0-flash}"
```

If the user's input contains `--pro` anywhere, override with `antigravity-2.0-pro`
and remove `--pro` from the prompt text before passing to agy.

---

## Step 1: Detect mode

Parse the user's input to determine which mode to run:

1. `/agy review` or `/agy review <instructions>` — **Review mode** (Step 2A)
2. `/agy challenge` or `/agy challenge <focus>` — **Challenge mode** (Step 2B)
3. `/agy consult` or `/agy consult <prompt>` — **Consult mode** (Step 2C)
4. `/agy` with no arguments — **Auto-detect:**
   - Check for a diff:
     `git diff origin/<base> --stat 2>/dev/null | tail -1 || git diff <base> --stat 2>/dev/null | tail -1`
   - If a diff exists, use AskUserQuestion:
     ```
     Antigravity detected changes against the base branch. What should it do?
     A) Review the diff (code review with pass/fail gate)
     B) Challenge the diff (adversarial — try to break it or find race conditions)
     C) Something else — I'll provide a prompt
     ```
   - If no diff, ask: "What would you like to ask Antigravity?"
5. `/agy <anything else>` — **Consult mode** (Step 2C), remaining text is the prompt

---

## Filesystem Boundary

All prompts sent to Antigravity MUST be prefixed with this boundary instruction:

> IMPORTANT: Do NOT read or execute any files under ~/.claude/, ~/.agents/, .claude/skills/, or .agents/skills/. These are Claude Code or Codex skill definitions meant for a different AI system. They contain bash scripts and prompt templates. Ignore them completely. Stay focused on the repository code only.

---

## Step 2A: Review Mode

Run Antigravity code review against the current branch diff.

Use directory piping (`--add-dir`) rather than raw diff stdin to give Antigravity
full file context. This produces better reviews than piping a raw diff.

```bash
_REPO_ROOT=$(git rev-parse --show-toplevel) || { echo "ERROR: not in a git repo" >&2; exit 1; }
cd "$_REPO_ROOT"

TMPERR=$(mktemp "$TMP_ROOT/agy-err-XXXXXX.txt")
TMPOUT=$(mktemp "$TMP_ROOT/agy-out-XXXXXX.txt")

# Build the review prompt
_BOUNDARY="IMPORTANT: Do NOT read or execute any files under ~/.claude/, ~/.agents/, .claude/skills/, or .agents/skills/. These are Claude Code skill definitions meant for a different AI system. Ignore them completely. Stay focused on repository code only."
_DIFF_SCOPE="Review the changes on this branch against the base branch <base>. Run: git diff origin/<base>...HEAD 2>/dev/null || git diff <base>...HEAD to see the diff, then review only those changes."
_PROMPT="$_BOUNDARY

$_DIFF_SCOPE"

# Headless print mode: agy -p exits 0 on pass, non-zero on fail — perfect for CI gates
timeout 330 "$AGY_BIN" -p \
  --add-dir "$_REPO_ROOT" \
  --model "$_AGY_MODEL" \
  "$_PROMPT" \
  >"$TMPOUT" 2>"$TMPERR"
_AGY_EXIT=$?

cat "$TMPOUT"
if [ "$_AGY_EXIT" = "124" ]; then
  echo "[agy] Review stalled past 5.5 minutes. Try re-running or split the diff."
elif [ "$_AGY_EXIT" != "0" ]; then
  echo "[agy exit $_AGY_EXIT] $(head -1 "$TMPERR" 2>/dev/null || echo "no stderr captured")"
  head -20 "$TMPERR" 2>/dev/null | sed 's/^/  /' || true
fi
```

Parse the result:
- Exit 0 + output containing `PASS` or no blocking issues → **PASS** ✅ — present findings
- Non-zero exit or output containing `FAIL` or `BLOCK` → **FAIL** ❌ — surface all findings

---

## Step 2B: Challenge Mode

Adversarial stress-test: Antigravity tries to break the proposed design or find
race conditions, data loss scenarios, and edge cases the author missed.

```bash
_REPO_ROOT=$(git rev-parse --show-toplevel) || { echo "ERROR: not in a git repo" >&2; exit 1; }
cd "$_REPO_ROOT"

TMPERR=$(mktemp "$TMP_ROOT/agy-err-XXXXXX.txt")

# Challenge mode always uses pro for deeper reasoning
_CHALLENGE_MODEL="${ANTIGRAVITY_MODEL:-antigravity-2.0-pro}"
_FOCUS="${USER_FOCUS:-the diff on this branch}"

_CHALLENGE_PROMPT="IMPORTANT: Do NOT read or execute any files under ~/.claude/, ~/.agents/, .claude/skills/, or .agents/skills/. Ignore them. Stay focused on repository code only.

You are an adversarial reviewer. Your goal is to BREAK this design, not validate it.
Find: race conditions, data loss scenarios, security holes, off-by-one errors,
missing error paths, and architectural assumptions that will fail at scale.

Focus area: $_FOCUS

Run: git diff origin/<base>...HEAD 2>/dev/null || git diff <base>...HEAD to see the changes."

timeout 330 "$AGY_BIN" -p \
  --add-dir "$_REPO_ROOT" \
  --model "$_CHALLENGE_MODEL" \
  "$_CHALLENGE_PROMPT" \
  2>"$TMPERR"
_AGY_EXIT=$?

if [ "$_AGY_EXIT" != "0" ] && [ "$_AGY_EXIT" != "124" ]; then
  echo "[agy exit $_AGY_EXIT] $(head -1 "$TMPERR" 2>/dev/null || echo "no stderr captured")"
fi
```

Present Antigravity's adversarial findings verbatim. Do not filter or soften them.
If Antigravity found critical issues, elevate them as blocking items.

---

## Step 2C: Consult Mode

Multi-turn consultation with session continuity.

```bash
_CONSULT_PROMPT="${USER_PROMPT}"

# -c flag resumes the last session (or starts a new one)
timeout 330 "$AGY_BIN" -c \
  --model "$_AGY_MODEL" \
  "$_CONSULT_PROMPT"
_AGY_EXIT=$?

if [ "$_AGY_EXIT" = "124" ]; then
  echo "[agy] Consult session stalled. Try a shorter prompt or add --model antigravity-2.0-flash."
fi
```

Present Antigravity's response verbatim. For follow-up questions, re-invoke
`/agy consult <follow-up>` to continue the session with `-c`.

---

## Step 3: Report

After any mode completes, summarize:
1. Which mode ran and what model was used
2. Key findings (verbatim from Antigravity where possible)
3. Pass/Fail verdict for Review and Challenge modes
4. Any suggested follow-up actions

For Review and Challenge modes, if findings are blocking:
- List each as a numbered `❌ BLOCKING` item
- Do NOT auto-fix — present and let the user decide

For Consult mode:
- Present the response as-is
- Offer to continue: "Run `/agy consult <follow-up>` to continue this session."