The Deterministic Verification & Context Harness for AI Coding Agents
AI agents recommend. Deterministic gates decide. Scaffold, govern, and verify multi-agent development for Claude Code, Cursor, Codex, Antigravity, and OpenCode with zero hallucinated completion.
One-command setup for Claude Code, Cursor, Codex, Antigravity & OpenCode
Agent Harness Generator
One-Click Production Stacks
npx @sg/harness init my-saas-app \
--framework=nextjs \
--mode=PRODUCTION \
--agents=claude-code,cursor,antigravity \
--domains=frontend,backend,database,api-contracts,auth \
--context=graphify,hydradb,context7,framework-docsThe 16-Step GAUNTLET Verification Loop
GAUNTLET mode runs an adversarial zero-trust review loop across 8 critic roles, ensuring that no feature can be certified until all measurable acceptance criteria and fault injections are defended.
Active Agent Skills Catalog
29 specialized skills routed via .sg/routing.json for maximum relevant context.
User requests real-world consensus, recent 2026 opinions, or discussions from Reddit/HN/X/GitHub.
Working with an unfamiliar API, third-party SDK, or recently updated library syntax.
User asks for interactive requirement discovery, product scoping, or trade-off evaluation before execution.
Initial phase of open-ended features requiring multiple structural variations or trade-offs.
Understanding codebase structure, dependency paths, AST nodes, or blast radius before refactoring.
Major refactoring, structural decoupling, code bloat elimination, or harness evolution.
Building or refactoring React, Astro, or TypeScript UI components, islands, and layouts.
Writing CSS tokens, @theme definitions, dynamic utilities, or responsive layout styles.
Polishing UI typography, micro-details, optical alignment, layout rhythm, and interactive states.
Creating visually stunning pages, landing screens, color schemes, and avoiding generic AI templates.
Tuning spring curves, damping, layout transitions, exit/enter animations, or hover micro-interactions.
Complex scroll-driven parallax, scrubbing, timeline sequencing, or canvas animations.
Writing Go services, goroutines, cancellation contexts, or backend APIs.
Writing Rust services, CLI tools, async tokio runtimes, or WASM modules.
Querying or recording architectural decisions, ADR conclusions, rejected options, or debugging discoveries.
Working with LLM prompt engineering, structured extraction, JSON rescue, or resume parsing.
Verifying code against security rules, OWASP Top 10, secrets leaks, and input sanitization.
Visual UI rendering verification, end-to-end user flows, layout checks, or DOM inspection.
Debugging blob creation, file downloads, PDF generation, or clipboard copy failures.
Preparing production release summaries, git tag changelogs, or customer-facing updates.
Substantial UI changes, responsive layout refactors, design-system tokens, theme parity, or GAUNTLET mode visual critiques.
Creating or modifying UI components, forms, navigation dialogs, color themes, or running PRODUCTION/GAUNTLET gates.
Implementing auth flows, database models, RLS policies, team memberships, or API endpoints handling tenant resources.
Performing security audits, dependency vulnerability checks, injection defense, or secret leak protection.
Creating/altering tables, writing SQL migrations, configuring RLS, adding indexes, query tuning with EXPLAIN ANALYZE, or pgvector similarity search.
Writing or reviewing schema migrations, table locks, column additions, batched backfills, or rollback plans.
Optimizing Core Web Vitals (LCP, CLS, INP), evaluating asset bundles, auditing dependencies, or tuning hydration.
Configuring Sentry, structured logging, request correlation IDs, error boundaries, or health endpoints.
Major features, architecture shifts, security changes, pre-release certification, or explicit gauntlet invocation.
SG Command Cheatsheet
Universal multi-agent command interface for humans, Claude Code, Cursor, Codex, Antigravity, OpenCode, and CI.
Dynamically resolves applicable gates and executes deterministic checks.
Multi-Agent Compatibility Matrix
Single canonical source of truth in AGENTS.md. Thin adapters translate into each harness natively without rule drift.
| Coding Agent | Native Config | AGENTS.md Support | Skill Routing | MCP Integration | Completion Gate |
|---|---|---|---|---|---|
| Claude CodeAnthropic | CLAUDE.md | Full Root Pointer | Via .sg/routing.json | .claude/mcp.json | ./sg verify |
| OpenAI CodexOpenAI | CODEX.md | Full Root Pointer | Progressive disclosure | Tool-calling protocol | ./sg verify |
| Cursor IDECursor / Anysphere | .cursor/rules/*.mdc | Full MDC Pointer | Rule Triggers | .cursor/mcp.json | ./sg verify |
| Google AntigravityGoogle DeepMind | .agents/AGENTS.md | Native Kernel | Native .agents/skills/ | Native eager/lazy | ./sg verify |
| OpenCode / OpenHandsOpen Source | opencode.json | instructions field | Via schema | Standard MCP | ./sg verify |