Skip to content

AGENTS.md

Latest commit

 

History

History
144 lines (115 loc) · 8.25 KB

AGENTS.md

File metadata and controls

144 lines (115 loc) · 8.25 KB

SINT Protocol — AI Agent Guide

This file is optimized for AI agents (LLMs) working in this codebase. For human contributor guidance, see CONTRIBUTING.md. For implementation details and pitfalls, see CLAUDE.md.

What This Repo Does

SINT Protocol is a security enforcement layer that sits between AI agents and physical systems (robots, PLCs, drones, tool calls), ensuring every action is authorized via capability tokens, constrained to safe physical limits, and recorded in a tamper-evident audit log. It is a pnpm monorepo with 30 workspace members built on TypeScript 5.7 + Node.js 22.

Key Invariants (Never Violate These)

  1. Every authorization decision flows through PolicyGateway.intercept() — no bridge or route handler makes authorization decisions directly
  2. Tokens are attenuation-only: scope(child) ⊆ scope(parent) — permissions only narrow, never expand (invariant I-T1)
  3. EvidenceLedger is append-only and SHA-256 hash-chained — never modify or delete emitted events (invariant I-G3)
  4. UUID v7 is required for requestId — crypto.randomUUID() gives v4 and WILL fail schema validation. Use generateUUIDv7() from @sint/gate-capability-tokens
  5. Physical constraints live in the token, not in config files — velocity, force, geofence are enforced cryptographically
  6. E-stop is unconditional — estop event transitions any non-terminal DFA state to ROLLEDBACK, bypassing all token checks (invariant I-G2)

Quick Orientation

Package Dependency Graph

@sint/core
  ↓
@sint/gate-capability-tokens   @sint/persistence
  ↓                               ↓
@sint/gate-evidence-ledger
  ↓
@sint/gate-policy-gateway
  ↓
@sint/bridge-*   @sint/engine-*   @sint/avatar
  ↓
@sint/gateway-server   @sint/mcp
  ↓
@sint/conformance-tests   @sint/dashboard

Workspace Layout

packages/core/              → Shared types, Zod schemas, tier constants, DFA states
packages/capability-tokens/ → Ed25519 token issuance, delegation, revocation
packages/policy-gateway/    → THE choke point: tiers, constraints, rate limiting, M-of-N
packages/evidence-ledger/   → SHA-256 hash-chained append-only audit log
packages/bridge-*/          → Protocol adapters (MCP, ROS2, A2A, IoT, OPC-UA, MAVLink, ...)
packages/engine-*/          → AI execution layer (System1, System2, HAL, capsule sandbox)
packages/avatar/            → Behavioral identity profiles, CSML-driven tier escalation
packages/persistence/       → Storage interfaces + in-memory implementations
packages/persistence-postgres/ → PostgreSQL adapters
packages/client/            → TypeScript HTTP client for the gateway API
packages/conformance-tests/ → Security regression suite (must pass on every PR)
apps/gateway-server/        → Hono HTTP API (port 3100)
apps/sint-mcp/              → Security-first multi-MCP proxy
apps/dashboard/             → Real-time approval dashboard
sdks/typescript/            → Zero-dependency public SDK
capsules/navigation/        → Reference capsule: waypoint navigation
capsules/inspection/        → Reference capsule: visual anomaly detection
capsules/pick-and-place/    → Reference capsule: gripper control

How to Run Tests

pnpm run test                                       # all 815 tests, whole workspace
pnpm --filter @sint/gate-policy-gateway test        # single package
pnpm --filter @sint/bridge-mcp test                 # single bridge
pnpm run build && pnpm run test                     # full clean run

Before Starting Work

git pull --rebase
pnpm run build
pnpm run test   # must be 0 failures before you begin

Multi-Agent Collaboration Rules

Several agents may work on this repo concurrently. Follow these rules:

  1. Before starting: git pull --rebase && pnpm run build && pnpm run test — must be 0 failures
  2. Package ownership: see CLAUDE.md for the package ownership table
  3. Never amend published commits — always create new commits
  4. Test gate: your PR/commit must maintain 0 failures on all 815+ tests
  5. Branch naming: feat/<topic> or fix/<topic> — describe what changes, not who made it
  6. No force-push to main/master — create a PR instead

Common Mistakes

  • UUID v4 vs v7: crypto.randomUUID() produces v4. Use generateUUIDv7() from @sint/gate-capability-tokens — see CLAUDE.md "Common Name Collision Risks"
  • CircuitBreaker: trip() sets manualTrip=true permanently preventing auto-HALF_OPEN. Use recordDenial() to test auto-recovery — see CLAUDE.md
  • SintDeploymentProfile exists in both policy.ts (site profiles) and was renamed in engine.ts to SintHardwareDeploymentProfile. Do not re-add generic names in engine packages
  • Throwing errors: All fallible operations must return Result<T, E> using ok() / err() helpers from @sint/core. Never throw or use try/catch for control flow
  • Modifying the ledger: Evidence ledger events are immutable once written. If you need to correct a record, append a new correction event

Entry Points

If you're asked to:

  • Add a new bridge → create packages/bridge-<name>/, follow packages/bridge-iot/ as template. Register in apps/gateway-server/src/routes/.
  • Add a security plugin → see packages/policy-gateway/src/goal-hijack.ts as template. Plugins implement PolicyPlugin and are registered via PolicyGateway constructor.
  • Add conformance tests → see packages/conformance-tests/src/ for patterns. Every new security invariant should have a conformance test.
  • Add an API route → see apps/gateway-server/src/routes/ for patterns. Routes use Hono; all requests go through PolicyGateway.intercept().
  • Add a reference capsule → create capsules/<name>/, follow capsules/navigation/ as template.
  • Add a new engine component → see packages/engine-system1/ (perception) or packages/engine-system2/ (reasoning) for patterns.

Architecture in One Paragraph

SINT Protocol implements a T0–T3 tiered authorization model where every agent request is classified into one of four tiers by consequence severity: T0 (observe/read), T1 (low-impact write), T2 (physical state change, e.g. robot movement), or T3 (irreversible/high-value commitment). All requests flow through PolicyGateway.intercept(), the single choke point that validates Ed25519 capability tokens, enforces physical constraints (velocity, force, geofence) embedded in the token, runs per-token rate limiting, detects forbidden action sequences, and either auto-approves T0/T1 or escalates T2/T3 to a human approval queue. Every decision is written to the EvidenceLedger, an append-only SHA-256 hash-chained log that provides cryptographic tamper evidence. Bridge adapters (for MCP, ROS2, MAVLink, OPC-UA, A2A, etc.) translate protocol-specific calls into SintRequest objects and forward them to the gateway — they never make authorization decisions themselves.

Approval Tiers Reference

Tier Auto-approved? Example actions
T0 — OBSERVE Yes (logged) Read sensor, query DB, status check
T1 — PREPARE Yes (audited) Write file, save waypoint, stage plan
T2 — ACT No — requires review Move robot, operate gripper, /cmd_vel
T3 — COMMIT No — requires human + M-of-N quorum Execute trade, novel env entry, irreversible

Key Types (Quick Reference)

// The request entering the gate
interface SintRequest {
  requestId: UUIDv7;          // MUST be v7 — use generateUUIDv7()
  agentId: Ed25519PublicKey;
  tokenId: UUIDv7;
  resource: string;           // "ros2:///cmd_vel" | "mcp://filesystem/*"
  action: string;             // "publish" | "call" | "subscribe"
  params: Record<string, unknown>;
  physicalContext?: { humanDetected?: boolean; currentVelocityMps?: number; currentForceNewtons?: number };
  recentActions?: string[];   // For forbidden combo detection
}

// The gateway's decision
type PolicyDecision =
  | { action: "allow"; assignedTier: ApprovalTier }
  | { action: "deny"; assignedTier: ApprovalTier; denial: { reason: string; policyViolated: string } }
  | { action: "escalate"; assignedTier: ApprovalTier; escalation: { requiredTier: string; timeoutMs: number } }
  | { action: "transform"; assignedTier: ApprovalTier; transformations: { constraintOverrides?: unknown } };