The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
feat(commands): add Antigravity CLI (agy) as santa-loop Reviewer B option
Leone Martins3fed6f3cfe29
diff --git a/commands/santa-loop.md b/commands/santa-loop.mdindex 111087966..015d9ea24 100644--- a/commands/santa-loop.md+++ b/commands/santa-loop.md@@ -76,6 +76,7 @@ First, detect which CLIs are available:```bashcommand -v codex >/dev/null 2>&1 && echo "codex" || truecommand -v gemini >/dev/null 2>&1 && echo "gemini" || true+command -v agy >/dev/null 2>&1 && echo "agy" || true```Build the reviewer prompt (identical rubric + instructions as Reviewer A) and write it to a unique temp file:@@ -86,7 +87,7 @@ cat > "$PROMPT_FILE" << 'EOF'EOF```-Use the first available CLI:+Use the first available CLI, in this order:**Codex CLI** (if installed)```bash@@ -100,7 +101,14 @@ gemini -p "$(cat "$PROMPT_FILE")" -m gemini-2.5-prorm -f "$PROMPT_FILE"```-**Claude Agent fallback** (only if neither `codex` nor `gemini` is installed)+**Antigravity CLI** (if installed and neither codex nor gemini is)+```bash+agy -p "$(cat "$PROMPT_FILE")" --model gemini-3.6-flash-high --sandbox+rm -f "$PROMPT_FILE"+```+Despite the "flash" name, this outranks `gemini-3.1-pro-high` on every published coding/agentic benchmark (SWE-Bench Pro, Terminal-Bench, MLE-Bench) — Pro only leads on PhD-level reasoning benchmarks (GPQA, HLE), which aren't relevant to code review. Don't "correct" this back to a `-pro-` model by name alone; check current benchmarks first, since generation-over-tier ordering shifts release to release. Run `agy models` to see the current catalog before assuming this is stale.++**Claude Agent fallback** (only if none of `codex`, `gemini`, or `agy` is installed)Launch a second Claude Agent (subagent_type: `code-reviewer`, model: `opus`). Log a warning that both reviewers share the same model family — true model diversity was not achieved but context isolation is still enforced.In all cases, the reviewer must return the same structured JSON verdict as Reviewer A.@@ -166,9 +174,10 @@ Result: [PUSHED / ESCALATED TO USER]## Notes- Reviewer A (Claude Opus) always runs — guarantees at least one strong reviewer regardless of tooling.-- Model diversity is the goal for Reviewer B. GPT-5.4 or Gemini 2.5 Pro gives true independence — different training data, different biases, different blind spots. The Claude-only fallback still provides value via context isolation but loses model diversity.-- Strongest available models are used: Opus for Reviewer A, GPT-5.4 or Gemini 2.5 Pro for Reviewer B.-- External reviewers run with `--sandbox read-only` (Codex) to prevent repo mutation during review.+- Model diversity is the goal for Reviewer B. GPT-5.4, Gemini 2.5 Pro, or Antigravity's Gemini 3.6 Flash (via `agy`) all give true independence — different training data, different biases, different blind spots. The Claude-only fallback still provides value via context isolation but loses model diversity.+- Strongest available models are used: Opus for Reviewer A, GPT-5.4, Gemini 2.5 Pro, or Gemini 3.6 Flash High (`agy`) for Reviewer B, in that priority order.+- Never point `agy` at a Claude model (`claude-sonnet-4-6`, `claude-opus-4-6-thinking`) — Reviewer A is already Claude Opus, so that would eliminate model diversity entirely.+- External reviewers run with `--sandbox read-only` (Codex) or `--sandbox` (`agy`) to prevent repo mutation during review.- Fresh reviewers each round prevents anchoring bias from prior findings.- The rubric is the most important input. Tighten it if reviewers rubber-stamp or flag subjective style issues.- Commits happen on NAUGHTY rounds so fixes are preserved even if the loop is interrupted.