/** * The single authoritative tool-execution path. * * Both transports — MCP (`mcp.ts`) and OpenAPI/HTTP (`http.ts`) — route every * tool call through {@link executeTool}. Centralizing here is what makes the * security guarantees of design.md §2.5 hold uniformly across interfaces: * * 1. tool lookup (unknown → `UnknownToolError` → 404) * 2. server-side scope enforcement (`requireScope`; defense-in-depth — handlers * also call it). Tool-hiding in the UI is NOT the boundary (design.md §2.5). * 3. input-schema validation BEFORE the handler runs (ajv) — never pass * unvalidated input to a handler (design.md §2.5 prompt-injection containment). * 4. per-session cap + per-tool rate limit (design.md §7.3). * 5. handler execution. * 6. finance-tier redaction on egress — bank/routing/card/SSN masked before the * value leaves the dispatcher (design.md §2.5, §7.3). Belt-and-braces over * each finance tool's own internal redaction. * 7. structured audit record for every finance (and future physical) call — * args HASHED, never logged raw; no secrets (design.md §2.5, §7.3). * * Tool output is treated strictly as DATA, never as instructions: the dispatcher * inspects/redacts it but never re-enters itself based on its content, so a * prompt-injection payload in a tool response cannot trigger another tool call * (design.md §2.5). */ import AjvModule from 'ajv'; import addFormatsModule from 'ajv-formats'; import type { Ajv as AjvInstance, Options as AjvOptions, ValidateFunction } from 'ajv'; // ajv / ajv-formats ship as CJS; under NodeNext ESM the callable lives on // `.default`. Normalize so both module shapes work, then re-type the runtime // values as the constructable class / callable plugin. type AjvCtor = new (opts?: AjvOptions) => AjvInstance; const Ajv = ((AjvModule as { default?: unknown }).default ?? AjvModule) as unknown as AjvCtor; const addFormats = ((addFormatsModule as { default?: unknown }).default ?? addFormatsModule) as ( ajv: AjvInstance, ) => AjvInstance; import { requireScope, ScopeError } from './auth.js'; import { hashArgs, type AuditLogger, type AuditRecord } from './audit.js'; import { redact } from './redact.js'; import { RateLimitError, type RateLimiter } from './rate-limit.js'; import type { ToolRegistry } from './registry.js'; import type { AuthContext, ToolDef } from './types.js'; // --------------------------------------------------------------------------- // Errors (adapters translate these to status codes) // --------------------------------------------------------------------------- /** Unknown tool name → HTTP 404 / MCP "method not found"-equivalent. */ export class UnknownToolError extends Error { readonly tool: string; constructor(tool: string) { super(`Unknown tool "${tool}".`); this.name = 'UnknownToolError'; this.tool = tool; Object.setPrototypeOf(this, new.target.prototype); } } /** Input failed JSON-Schema validation → HTTP 400. */ export class InputValidationError extends Error { readonly tool: string; /** Human-readable validation messages; never echoes secrets. */ readonly issues: string[]; constructor(tool: string, issues: string[]) { super(`Input validation failed for "${tool}": ${issues.join('; ')}`); this.name = 'InputValidationError'; this.tool = tool; this.issues = issues; Object.setPrototypeOf(this, new.target.prototype); } } // --------------------------------------------------------------------------- // Dependencies injected into the dispatcher // --------------------------------------------------------------------------- export interface DispatchDeps { /** Audit sink — finance/physical calls emit a record here. */ auditLogger: AuditLogger; /** Per-session + per-tool limiter consulted before each handler runs. */ rateLimiter: RateLimiter; } // --------------------------------------------------------------------------- // AJV — compiled validators cached per tool // --------------------------------------------------------------------------- // One Ajv instance for the process. Schemas are JSON Schema (draft-07 / the // OpenAPI 3.1 subset our tools use). `strict: false` because tool authors use // vocabulary (e.g. `description`) liberally; we only need structural validation. // // `allErrors: false` (the default) is deliberate and security-relevant: the // input is UNTRUSTED, and `allErrors: true` makes ajv enumerate every schema // violation, which an attacker can weaponize into CPU/memory exhaustion by // sending a large/deeply-nested payload that fails many constraints at once // (CodeQL js/resource-exhaustion). Short-circuiting on the first error caps the // work per request; the 400 still names the first failing path, which is enough // for a caller to fix their input. const ajv = new Ajv({ allErrors: false, strict: false, coerceTypes: false }); addFormats(ajv); const validatorCache = new WeakMap(); function getValidator(tool: ToolDef): ValidateFunction { const schema = tool.inputSchema as object; const cached = validatorCache.get(schema); if (cached) return cached; const validate = ajv.compile(schema); validatorCache.set(schema, validate); return validate; } // --------------------------------------------------------------------------- // Finance egress redaction // --------------------------------------------------------------------------- /** Field names whose values are masked wholesale on finance egress. */ const SENSITIVE_FIELD_RE = /(account|routing|card|ssn|tax[_-]?id|iban|swift)/i; /** * Deep-redact a finance tool's output before it leaves the dispatcher. * * Two complementary passes (design.md §2.5): * - String values run through `redact()` (pattern-based: routing/account/card/SSN). * - Any field whose KEY looks sensitive is fully replaced with `[REDACTED]`, * catching isolated values (e.g. a bare `bankAccountNumber: "123456789"`) * that the inline pattern matcher would miss without keyword context. * * Returns a NEW structure; the handler's value is not mutated. */ export function redactDeep(value: unknown): unknown { if (typeof value === 'string') return redact(value); if (Array.isArray(value)) return value.map(redactDeep); if (value !== null && typeof value === 'object') { const out: Record = {}; for (const [key, v] of Object.entries(value as Record)) { if (SENSITIVE_FIELD_RE.test(key) && v !== null && v !== undefined) { // Key looks sensitive → mask the ENTIRE value wholesale, whether it is a // scalar, an array, or a nested object. Recursing into a non-scalar here // would lose the keyword context redact() needs, letting a bare nested // value (e.g. { account: { number: "021000021" } }) escape unmasked. // Over-masking on the finance tier is the correct trade: a false negative // is a PII leak (design.md §2.5). out[key] = '[REDACTED]'; } else { out[key] = redactDeep(v); } } return out; } return value; } // --------------------------------------------------------------------------- // executeTool — the one path // --------------------------------------------------------------------------- /** * Look up, authorize, validate, rate-limit, run, redact, and audit a single * tool call. Used identically by both transports. * * @throws {UnknownToolError} unknown tool name (→ 404) * @throws {ScopeError} caller lacks the required scope (→ 403) * @throws {InputValidationError} input failed schema validation (→ 400) * @throws {RateLimitError} limiter rejected the call (→ 429) * @throws {Error} handler threw (→ 500; message not leaked verbatim) */ export async function executeTool( registry: ToolRegistry, ctx: AuthContext, toolName: string, rawInput: unknown, deps: DispatchDeps, ): Promise { const tool = registry.get(toolName) as ToolDef | undefined; if (!tool) { throw new UnknownToolError(toolName); } const audited = tool.tier === 'finance'; let decision: AuditRecord['decision'] = 'deny'; let result: AuditRecord['result'] = 'error'; try { // 2. Server-side scope enforcement (authoritative; not UI tool-hiding). requireScope(ctx, tool.requiredScope); decision = 'allow'; // 3. Input-schema validation BEFORE the handler sees the input. const validate = getValidator(tool); if (!validate(rawInput)) { const issues = (validate.errors ?? []).map( (e) => `${e.instancePath || '(root)'} ${e.message ?? 'is invalid'}`, ); throw new InputValidationError(toolName, issues.length ? issues : ['invalid input']); } // 4. Rate limit / session cap. deps.rateLimiter.check(ctx.sub, toolName); // 5. Run the handler. Its output is data only. const output = await tool.handler(rawInput, ctx); // 6. Finance egress redaction (belt-and-braces over internal redaction). const safeOutput = tool.tier === 'finance' ? redactDeep(output) : output; result = 'ok'; return safeOutput; } finally { // 7. Audit every finance/physical call regardless of outcome. if (audited) { deps.auditLogger.log({ sub: ctx.sub, tool: toolName, argsHash: hashArgs(rawInput), decision, result, ts: new Date().toISOString(), }); } } } // Re-export the error types adapters need to translate outcomes. export { ScopeError, RateLimitError };