| @@ -0,0 +1,286 @@ |
| 1 | +// Cross-target: when the authorization server definitively rejects a |
| 2 | +// refresh-token grant, the sandbox must see an auth failure it can act on — |
| 3 | +// never the scrubbed "Internal tool error [hex]" defect. |
| 4 | +// |
| 5 | +// Prod regression (2026-07-16): an AS answered refresh grants with a 400 |
| 6 | +// whose RFC 6749 error code was NOT `invalid_grant`. Executor classified |
| 7 | +// that as a storage failure, and the opaque-defect boundary scrubbed it to |
| 8 | +// "Internal tool error [id]" — so agents (and the humans driving them) had |
| 9 | +// no signal that the connection simply needed to be re-authenticated. |
| 10 | +// |
| 11 | +// The journey: an OpenAPI integration completes a real authorization-code |
| 12 | +// flow against a live test AS that mints instantly-expiring access tokens |
| 13 | +// and rejects every refresh grant with `invalid_request`. The first tool |
| 14 | +// call must therefore refresh, the AS says a definitive no, and the failure |
| 15 | +// the sandbox sees carries code `oauth_refresh_failed`, the AS's own error |
| 16 | +// code and description, and connection-recovery guidance. |
| 17 | +import { randomBytes } from "node:crypto"; |
| 18 | +import { createServer } from "node:http"; |
| 19 | + |
| 20 | +import { expect } from "@effect/vitest"; |
| 21 | +import { Effect } from "effect"; |
| 22 | +import { composePluginApi } from "@executor-js/api/server"; |
| 23 | +import { openApiHttpPlugin } from "@executor-js/plugin-openapi/api"; |
| 24 | +import { |
| 25 | + AuthTemplateSlug, |
| 26 | + ConnectionName, |
| 27 | + IntegrationSlug, |
| 28 | + OAuthClientSlug, |
| 29 | +} from "@executor-js/sdk/shared"; |
| 30 | +import { serveOAuthTestServer } from "@executor-js/sdk/testing"; |
| 31 | + |
| 32 | +import { scenario } from "../src/scenario"; |
| 33 | +import { Api, Mcp, Target } from "../src/services"; |
| 34 | + |
| 35 | +const api = composePluginApi([openApiHttpPlugin()] as const); |
| 36 | + |
| 37 | +const unique = (prefix: string) => `${prefix}_${randomBytes(4).toString("hex")}`; |
| 38 | + |
| 39 | +/** Upstream on 127.0.0.1: `GET /issues` is 200 for any bearer. The refresh is |
| 40 | + * rejected before any upstream call, so this only proves the failure came |
| 41 | + * from the token endpoint, not from here. */ |
| 42 | +const serveUpstream = () => |
| 43 | + Effect.acquireRelease( |
| 44 | + Effect.callback<{ readonly url: string; readonly close: () => void }>((resume) => { |
| 45 | + const server = createServer((request, response) => { |
| 46 | + if (request.method === "GET" && (request.url ?? "").startsWith("/issues")) { |
| 47 | + response.writeHead(200, { "content-type": "application/json" }); |
| 48 | + response.end(JSON.stringify({ issues: [] })); |
| 49 | + return; |
| 50 | + } |
| 51 | + response.writeHead(404, { "content-type": "application/json" }); |
| 52 | + response.end(JSON.stringify({ error: "not_found" })); |
| 53 | + }); |
| 54 | + server.listen(0, "127.0.0.1", () => { |
| 55 | + const address = server.address(); |
| 56 | + const port = typeof address === "object" && address ? address.port : 0; |
| 57 | + resume( |
| 58 | + Effect.succeed({ |
| 59 | + url: `http://127.0.0.1:${port}`, |
| 60 | + close: () => { |
| 61 | + server.close(); |
| 62 | + server.closeAllConnections(); |
| 63 | + }, |
| 64 | + }), |
| 65 | + ); |
| 66 | + }); |
| 67 | + }), |
| 68 | + (server) => Effect.sync(server.close), |
| 69 | + ); |
| 70 | + |
| 71 | +const spec = ( |
| 72 | + baseUrl: string, |
| 73 | + oauth: { readonly authorizationEndpoint: string; readonly tokenEndpoint: string }, |
| 74 | +): string => |
| 75 | + JSON.stringify({ |
| 76 | + openapi: "3.0.3", |
| 77 | + info: { title: "Issues API", version: "1.0.0" }, |
| 78 | + servers: [{ url: baseUrl }], |
| 79 | + paths: { |
| 80 | + "/issues": { |
| 81 | + get: { |
| 82 | + operationId: "listIssues", |
| 83 | + summary: "List issues", |
| 84 | + security: [{ oauth: ["issues.read"] }], |
| 85 | + responses: { "200": { description: "issues" } }, |
| 86 | + }, |
| 87 | + }, |
| 88 | + }, |
| 89 | + components: { |
| 90 | + securitySchemes: { |
| 91 | + oauth: { |
| 92 | + type: "oauth2", |
| 93 | + flows: { |
| 94 | + authorizationCode: { |
| 95 | + authorizationUrl: oauth.authorizationEndpoint, |
| 96 | + tokenUrl: oauth.tokenEndpoint, |
| 97 | + scopes: { "issues.read": "Read issues" }, |
| 98 | + }, |
| 99 | + }, |
| 100 | + }, |
| 101 | + }, |
| 102 | + }, |
| 103 | + }); |
| 104 | + |
| 105 | +const invokeByAddressCode = (address: string, args: unknown) => ` |
| 106 | +const segments = ${JSON.stringify(address)}.split(".").slice(1); |
| 107 | +let node = tools; |
| 108 | +for (const segment of segments) node = node[segment]; |
| 109 | +const result = await node(${JSON.stringify(args)}); |
| 110 | +return JSON.stringify(result); |
| 111 | +`; |
| 112 | + |
| 113 | +type ToolEnvelope = { |
| 114 | + readonly ok: boolean; |
| 115 | + readonly error?: { |
| 116 | + readonly code?: string; |
| 117 | + readonly message?: string; |
| 118 | + readonly details?: { |
| 119 | + readonly category?: string; |
| 120 | + readonly recovery?: Record<string, string>; |
| 121 | + }; |
| 122 | + }; |
| 123 | +}; |
| 124 | + |
| 125 | +scenario( |
| 126 | + "Auth failures · a refresh rejected by the authorization server surfaces as an actionable auth failure, not an internal error", |
| 127 | + {}, |
| 128 | + Effect.scoped( |
| 129 | + Effect.gen(function* () { |
| 130 | + const target = yield* Target; |
| 131 | + const { client: makeClient } = yield* Api; |
| 132 | + const mcp = yield* Mcp; |
| 133 | + const identity = yield* target.newIdentity(); |
| 134 | + const client = yield* makeClient(api, identity); |
| 135 | + const upstream = yield* serveUpstream(); |
| 136 | + // Instantly-expiring access tokens force the first tool call to |
| 137 | + // refresh; the AS rejects every refresh grant with a non-invalid_grant |
| 138 | + // 400, mirroring the prod AS that triggered the regression. |
| 139 | + const oauth = yield* serveOAuthTestServer({ |
| 140 | + scopes: ["issues.read"], |
| 141 | + tokenExpiresInSeconds: 0, |
| 142 | + supportRefresh: false, |
| 143 | + invalidRefreshTokenErrorCode: "invalid_request", |
| 144 | + invalidRefreshTokenDescription: "Refresh token expired", |
| 145 | + }); |
| 146 | + const slug = unique("refreshrej"); |
| 147 | + const clientSlug = OAuthClientSlug.make(unique("refreshrejc")); |
| 148 | + |
| 149 | + yield* Effect.ensuring( |
| 150 | + Effect.gen(function* () { |
| 151 | + yield* client.openapi.addSpec({ |
| 152 | + payload: { |
| 153 | + spec: { kind: "blob", value: spec(upstream.url, oauth) }, |
| 154 | + slug, |
| 155 | + baseUrl: upstream.url, |
| 156 | + authenticationTemplate: [ |
| 157 | + { |
| 158 | + slug: "oauth", |
| 159 | + kind: "oauth2", |
| 160 | + authorizationUrl: oauth.authorizationEndpoint, |
| 161 | + tokenUrl: oauth.tokenEndpoint, |
| 162 | + scopes: ["issues.read"], |
| 163 | + }, |
| 164 | + ], |
| 165 | + }, |
| 166 | + }); |
| 167 | + yield* client.oauth.createClient({ |
| 168 | + payload: { |
| 169 | + owner: "org", |
| 170 | + slug: clientSlug, |
| 171 | + grant: "authorization_code", |
| 172 | + authorizationUrl: oauth.authorizationEndpoint, |
| 173 | + tokenUrl: oauth.tokenEndpoint, |
| 174 | + clientId: "test-client", |
| 175 | + clientSecret: "test-secret", |
| 176 | + originIntegration: IntegrationSlug.make(slug), |
| 177 | + }, |
| 178 | + }); |
| 179 | + |
| 180 | + const started = yield* client.oauth.start({ |
| 181 | + payload: { |
| 182 | + client: clientSlug, |
| 183 | + clientOwner: "org", |
| 184 | + owner: "org", |
| 185 | + name: ConnectionName.make("main"), |
| 186 | + integration: IntegrationSlug.make(slug), |
| 187 | + template: AuthTemplateSlug.make("oauth"), |
| 188 | + }, |
| 189 | + }); |
| 190 | + expect(started.status, "oauth.start redirects to the authorization server").toBe( |
| 191 | + "redirect", |
| 192 | + ); |
| 193 | + if (started.status !== "redirect") return yield* Effect.die("no redirect"); |
| 194 | + |
| 195 | + // Drive the test IdP's consent by hand (authorize → login → code). |
| 196 | + const code = yield* Effect.promise(async () => { |
| 197 | + const authorize = await fetch(started.authorizationUrl, { redirect: "manual" }); |
| 198 | + const loginUrl = authorize.headers.get("location"); |
| 199 | + if (!loginUrl) throw new Error(`authorize did not redirect: ${authorize.status}`); |
| 200 | + const login = await fetch(loginUrl, { |
| 201 | + method: "POST", |
| 202 | + headers: { |
| 203 | + authorization: `Basic ${Buffer.from("alice:password").toString("base64")}`, |
| 204 | + }, |
| 205 | + redirect: "manual", |
| 206 | + }); |
| 207 | + const callbackUrl = login.headers.get("location"); |
| 208 | + if (!callbackUrl) throw new Error(`login did not redirect: ${login.status}`); |
| 209 | + const minted = new URL(callbackUrl).searchParams.get("code"); |
| 210 | + if (!minted) throw new Error("callback carried no authorization code"); |
| 211 | + return minted; |
| 212 | + }); |
| 213 | + yield* client.oauth.complete({ payload: { state: started.state, code } }); |
| 214 | + |
| 215 | + const tools = yield* client.tools.list({ query: {} }); |
| 216 | + const address = tools |
| 217 | + .filter((tool) => String(tool.integration) === slug) |
| 218 | + .map((tool) => String(tool.address)) |
| 219 | + .find((addr) => addr.endsWith("listIssues")); |
| 220 | + expect(address, "the listIssues tool is in the catalog").toBeDefined(); |
| 221 | + |
| 222 | + // Call through the real MCP surface — the exact channel the prod |
| 223 | + // regression hid the failure on — so the assertion covers the whole |
| 224 | + // path an MCP client sees: dispatch, classification, rendering. |
| 225 | + const session = mcp.session(identity); |
| 226 | + let called = yield* session.call("execute", { |
| 227 | + code: invokeByAddressCode(address!, {}), |
| 228 | + }); |
| 229 | + // Approval-gated tools pause the execution once per gated call. |
| 230 | + let guard = 0; |
| 231 | + while (called.text.includes("executionId:") && guard < 10) { |
| 232 | + called = yield* session.approvePaused(called.text); |
| 233 | + guard += 1; |
| 234 | + } |
| 235 | + expect( |
| 236 | + called.ok, |
| 237 | + `the MCP execute call itself completed (got: ${called.text.slice(0, 400)})`, |
| 238 | + ).toBe(true); |
| 239 | + const failure = JSON.parse(called.text) as ToolEnvelope; |
| 240 | + |
| 241 | + // THE guarantee: the AS's definitive rejection reaches the agent |
| 242 | + // as an auth failure with the AS's own verdict in the message. |
| 243 | + expect(failure.ok, "the tool call failed").toBe(false); |
| 244 | + expect( |
| 245 | + failure.error?.code, |
| 246 | + "the failure is classified as a rejected refresh, not an internal defect", |
| 247 | + ).toBe("oauth_refresh_failed"); |
| 248 | + expect( |
| 249 | + failure.error?.message ?? "", |
| 250 | + "the message carries the AS's RFC 6749 error code", |
| 251 | + ).toContain("invalid_request"); |
| 252 | + expect( |
| 253 | + failure.error?.message ?? "", |
| 254 | + "the message carries the AS's error description", |
| 255 | + ).toContain("Refresh token expired"); |
| 256 | + expect( |
| 257 | + failure.error?.message ?? "", |
| 258 | + "the opaque defect message never reaches the sandbox", |
| 259 | + ).not.toContain("Internal tool error"); |
| 260 | + expect(failure.error?.details?.category, "the failure is auth-flavored").toBe( |
| 261 | + "authentication", |
| 262 | + ); |
| 263 | + expect( |
| 264 | + failure.error?.details?.recovery?.oauthInstructions ?? "", |
| 265 | + "the recovery tells the agent how to re-connect via OAuth", |
| 266 | + ).toContain("startOAuthTool"); |
| 267 | + }), |
| 268 | + Effect.gen(function* () { |
| 269 | + yield* client.connections |
| 270 | + .remove({ |
| 271 | + params: { |
| 272 | + owner: "org", |
| 273 | + integration: IntegrationSlug.make(slug), |
| 274 | + name: ConnectionName.make("main"), |
| 275 | + }, |
| 276 | + }) |
| 277 | + .pipe(Effect.ignore); |
| 278 | + yield* client.oauth |
| 279 | + .removeClient({ params: { slug: clientSlug }, payload: { owner: "org" } }) |
| 280 | + .pipe(Effect.ignore); |
| 281 | + yield* client.openapi.removeSpec({ params: { slug } }).pipe(Effect.ignore); |
| 282 | + }), |
| 283 | + ); |
| 284 | + }), |
| 285 | + ), |
| 286 | +); |