Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
feat(sandbox): resolve a sandboxed script's tool tier so financial tools stay out of scripts
nirholas8a27395d314f
diff --git a/api/_lib/sandbox/tiers.js b/api/_lib/sandbox/tiers.jsnew file mode 100644index 000000000..dbcc06e51--- /dev/null+++ b/api/_lib/sandbox/tiers.js@@ -0,0 +1,59 @@+// Which tier a tool name belongs to when a sandboxed script asks for it.+//+// The bridge (api/_lib/sandbox/bridge.js) holds a script to the same tool+// policy as a model (docs/prompts/03, packages/mcp-policy): read and write+// tools the run may use are callable, and a financial tool is never callable+// from inside a script. A financial tool needs its preview shown to the user+// and the user's yes in the conversation, and a script has no conversation.+//+// The tier comes from, in order:+// 1. the server agent registry (api/_lib/agent-tools.js): read, always;+// 2. the policy table, across every server: the most restrictive tier any+// server gives the name (a name that is financial anywhere is financial);+// 3. for a name nobody classified, its shape: names that execute, transfer,+// send, buy, sell, launch, withdraw, sign or pay are treated as financial,+// so an unclassified fund-moving tool fails closed instead of open.++import { POLICY } from '@three-ws/mcp-policy';+import { AGENT_TOOLS } from '../agent-tools.js';++const RANK = { read: 0, write: 1, financial: 2 };++const FINANCIAL_SHAPE = /(^|_)(execute|transfer|send|buy|sell|swap|launch|withdraw|sign|pay|mint|bridge|borrow|repay|supply|redeem|stake|unstake|close_position|open_position)(_|$)/;++let _table = null;++function policyIndex() {+ if (_table) return _table;+ const out = new Map();+ for (const rows of Object.values(POLICY)) {+ for (const [name, row] of Object.entries(rows)) {+ const tier = row?.tier || 'read';+ const prev = out.get(name);+ if (!prev || RANK[tier] > RANK[prev.tier]) {+ out.set(name, { tier, previewTool: row.previewTool || null, confirmFlag: row.confirmFlag || null });+ }+ }+ }+ _table = out;+ return out;+}++/**+ * The tier of a tool name, and for a financial tool the preview it requires.+ * @param {string} name+ * @returns {{ tier: 'read'|'write'|'financial', previewTool: string|null, source: 'registry'|'policy'|'shape'|'default' }}+ */+export function toolTier(name) {+ const n = String(name || '');+ if (Object.hasOwn(AGENT_TOOLS, n)) return { tier: 'read', previewTool: null, source: 'registry' };+ const row = policyIndex().get(n);+ if (row) return { tier: row.tier, previewTool: row.previewTool, source: 'policy' };+ if (FINANCIAL_SHAPE.test(n)) return { tier: 'financial', previewTool: null, source: 'shape' };+ return { tier: 'read', previewTool: null, source: 'default' };+}++/** Whether a script may never call this tool, whatever the run allows. */+export function isFinancialTool(name) {+ return toolTier(name).tier === 'financial';+}