fix(providers): reject malformed UTF-8 before MCP decoding

This commit is contained in:
morluto committed 2026-10-26 15:41:03 +08:00
1 parent c849a97d2a
commit f8c5a6cc10
5 files changed
+321 -7

No files matched your search

+25 -4
View File
@@ -33,6 +33,10 @@ import {
AnalysisResourceConstraintError,
} from "../domain/analysisErrorCore.js";
import { ProviderAdapterError } from "../domain/providerAdapterError.js";
import {
McpHttpUtf8Error,
validateMcpHttpResponse,
} from "../providers/Utf8McpResponse.js";
/** Injectable MCP boundary used by the provider and transport conformance tests. */
export interface IdaMcpConnection {
@@ -294,9 +298,15 @@ const readBoundedResponseText = async (
};
const diagnosticFetch =
(redactor: IdaDiagnosticRedactor): typeof fetch =>
(
redactor: IdaDiagnosticRedactor,
rejectTransport: (failure: McpHttpUtf8Error) => void,
): typeof fetch =>
async (input, init) => {
const response = await fetch(input, init);
const response = validateMcpHttpResponse(
await fetch(input, init),
rejectTransport,
);
const readText = response.text.bind(response);
let safeBody: string | undefined;
if (!response.ok) {
@@ -356,6 +366,7 @@ export const createIdaMcpConnection = (
)(config);
let stderr = Buffer.alloc(0);
let stderrBytes = 0;
let httpUtf8Failure: McpHttpUtf8Error | undefined;
const capturedTransportOutput = () => {
let start = 0;
if (stderrBytes > stderr.length)
@@ -407,6 +418,13 @@ export const createIdaMcpConnection = (
? /ReadBuffer exceeded maximum size of (\d+) bytes/u.exec(cause.message)
: null;
if (cause instanceof IdaDiagnosticResourceError) throw cause;
if (cause instanceof McpHttpUtf8Error) throw cause;
if (
httpUtf8Failure !== undefined &&
cause instanceof SdkError &&
cause.code === SdkErrorCode.ConnectionClosed
)
throw httpUtf8Failure;
let failure: Error;
if (cause instanceof IdaHttpJsonError)
failure = new Error(cause.message, { cause });
@@ -459,7 +477,7 @@ export const createIdaMcpConnection = (
capturedOutput,
});
};
const transport =
const transport: StdioClientTransport | StreamableHTTPClientTransport =
"command" in config
? new StdioClientTransport({
command: config.command,
@@ -483,7 +501,10 @@ export const createIdaMcpConnection = (
})
: new StreamableHTTPClientTransport(new URL(config.url), {
requestInit: { headers: config.headers, redirect: "error" },
fetch: diagnosticFetch(redactor),
fetch: diagnosticFetch(redactor, (failure) => {
httpUtf8Failure ??= failure;
void transport.close().catch(() => undefined);
}),
});
// Keep a bounded diagnostic tail for genuine failures; never persist ambient variables.
if (transport instanceof StdioClientTransport)
+6 -3
View File
@@ -180,7 +180,7 @@ export const createStreamableHttpJebMcpConnection: JebMcpConnectionFactory = (
} catch (cause: unknown) {
const failure = physicalFailure(
"catalog",
deadline.failure("jeb_mcp_catalog", cause),
owner.responseFailure(deadline.failure("jeb_mcp_catalog", cause)),
);
owner.acquisitionFailure = { cause: failure };
owner.signal.abort(failure);
@@ -237,7 +237,7 @@ export const createStreamableHttpJebMcpConnection: JebMcpConnectionFactory = (
} catch (cause: unknown) {
throw physicalFailure(
"catalog",
deadline.failure("jeb_mcp_catalog", cause),
owner.responseFailure(deadline.failure("jeb_mcp_catalog", cause)),
);
} finally {
deadline.dispose();
@@ -260,7 +260,10 @@ export const createStreamableHttpJebMcpConnection: JebMcpConnectionFactory = (
{ timeout: deadline.remainingMs(), signal: deadline.signal },
);
} catch (cause: unknown) {
throw physicalFailure("tool", deadline.failure(name, cause));
throw physicalFailure(
"tool",
owner.responseFailure(deadline.failure(name, cause)),
);
} finally {
deadline.dispose();
}
+22
View File
@@ -1,11 +1,17 @@
import {
Client,
SdkError,
SdkErrorCode,
StreamableHTTPClientTransport,
} from "@modelcontextprotocol/client";
import { AnalysisCleanupLease } from "../artifacts/AnalysisCleanupLease.js";
import { ProviderCleanupError } from "../domain/providerCleanupError.js";
import { SharedStartupAttempt } from "../process/SharedStartupAttempt.js";
import type { Result } from "../domain/result.js";
import {
McpHttpUtf8Error,
validateMcpHttpResponse,
} from "../providers/Utf8McpResponse.js";
/** Physical SDK failure captured before any subsequent resource cleanup. */
export class JebMcpPhysicalFailure extends Error {
@@ -61,10 +67,17 @@ export class JebMcpConnectionOwner {
#clientClosed = false;
#transportClosed = false;
#disposal: Promise<void> | undefined;
#httpUtf8Failure: McpHttpUtf8Error | undefined;
constructor(endpoint: URL) {
this.transport = new StreamableHTTPClientTransport(endpoint, {
requestInit: { redirect: "error" },
fetch: async (input, init) =>
validateMcpHttpResponse(await fetch(input, init), (failure) => {
this.#httpUtf8Failure ??= failure;
this.tools = undefined;
void this.transport.close().catch(() => undefined);
}),
});
this.#clientResource = `${endpoint} (JEB MCP client)`;
this.#transportResource = `${endpoint} (JEB MCP transport)`;
@@ -78,6 +91,15 @@ export class JebMcpConnectionOwner {
return this.#disposal !== undefined;
}
/** Attribute only the actual transport close caused by malformed response bytes. */
responseFailure(cause: unknown): unknown {
return this.#httpUtf8Failure !== undefined &&
cause instanceof SdkError &&
cause.code === SdkErrorCode.ConnectionClosed
? this.#httpUtf8Failure
: cause;
}
get cleanupResources(): readonly string[] {
return [
...(this.#clientClosed ? [] : [this.#clientResource]),
+51
View File
@@ -0,0 +1,51 @@
import { AnalysisProtocolError } from "../domain/analysisErrorCore.js";
/** Authored byte-validation failure, distinct from arbitrary upstream protocol diagnostics. */
export class McpHttpUtf8Error extends AnalysisProtocolError {
constructor(cause: unknown) {
super("MCP HTTP response body omitted: not valid UTF-8.", { cause });
}
}
/** Validate accepted MCP message bytes before Fetch or the SDK can replace malformed UTF-8. */
export const validateMcpHttpResponse = (
response: Response,
rejectTransport: (failure: McpHttpUtf8Error) => void,
): Response => {
// Failed HTTP responses belong to each adapter's existing diagnostic policy.
if (!response.ok || response.body === null) return response;
const decoder = new TextDecoder("utf-8", { fatal: true, ignoreBOM: true });
const validate = (bytes?: Uint8Array): void => {
try {
decoder.decode(bytes, { stream: bytes !== undefined });
} catch (cause: unknown) {
const failure = new McpHttpUtf8Error(cause);
// SDK 2.3.1 reports SSE read errors without settling pending tools/call.
// The lifetime owner closes the actual transport to reject that work.
rejectTransport(failure);
throw failure;
}
};
const body = response.body.pipeThrough(
new TransformStream<Uint8Array, Uint8Array>({
transform(bytes, controller) {
validate(bytes);
controller.enqueue(bytes);
},
flush() {
validate();
},
}),
);
const validated = new Response(body, {
status: response.status,
statusText: response.statusText,
headers: response.headers,
});
Object.defineProperties(validated, {
url: { value: response.url },
redirected: { value: response.redirected },
type: { value: response.type },
});
return validated;
};
@@ -0,0 +1,217 @@
import type { ServerResponse } from "node:http";
import { setTimeout as delay } from "node:timers/promises";
import { expect, it } from "vitest";
import { z } from "zod";
import { createVerifierHttpServer } from "../../../scripts/lib/verifier-http-server.mjs";
import { createIdaMcpConnection } from "../../../src/ida/IdaMcpConnection.js";
import { createStreamableHttpJebMcpConnection } from "../../../src/jeb/JebMcpConnection.js";
import { AnalysisCancelledError } from "../../../src/domain/analysisErrorCore.js";
const requestSchema = z.object({
id: z.union([z.string(), z.number()]).optional(),
method: z.string(),
});
const MARKER = "SOURCE_UTF8_BYTES";
const payload = {
success: true,
text: `caf${MARKER}`,
source_url: "file:///selected/a%20b.js?query#fragment",
};
/** Actual HTTP framing and SDK parsing; this server is not an installed IDA/JEB engine. */
const startFixture = async (format: "json" | "sse") => {
let fault: "valid" | "invalid" | "truncated" | "stalled" = "valid";
let announceStalled!: () => void;
let stalled = Promise.resolve();
let release!: () => void;
const gate = new Promise<void>((resolve) => {
release = resolve;
});
const stalledResponses: ServerResponse[] = [];
const { server, close } = createVerifierHttpServer(
async (request, response) => {
if (request.method !== "POST") {
request.resume();
response.writeHead(request.method === "DELETE" ? 200 : 405).end();
return;
}
const chunks: Buffer[] = [];
for await (const chunk of request) chunks.push(Buffer.from(chunk));
const message = requestSchema.parse(
JSON.parse(Buffer.concat(chunks).toString("utf8")),
);
if (message.id === undefined) {
response.writeHead(202).end();
return;
}
const result =
message.method === "initialize"
? {
protocolVersion: "2025-11-25",
capabilities: { tools: {} },
serverInfo: { name: "utf8-http-boundary", version: "1" },
}
: message.method === "tools/list"
? {
tools: ["get_client_information"].map((name) => ({
name,
inputSchema: { type: "object" },
})),
}
: {
content: [{ type: "text", text: JSON.stringify(payload) }],
structuredContent: payload,
};
const json = JSON.stringify({ jsonrpc: "2.0", id: message.id, result });
if (message.method !== "tools/call") {
response.writeHead(200, { "Content-Type": "application/json" });
response.end(json);
return;
}
const selectedFault = fault;
response.writeHead(200, {
"Content-Type":
format === "json" ? "application/json" : "text/event-stream",
"X-Observed-Header": "preserved",
});
response.flushHeaders();
const prefix = format === "sse" ? "event: message\ndata: " : "";
const suffix = format === "sse" ? "\n\n" : "";
const source = `${prefix}${json}${suffix}`;
const start = source.indexOf(MARKER);
if (selectedFault === "stalled") {
stalledResponses.push(response);
response.write(source.slice(0, start));
announceStalled();
await gate;
response.end();
return;
}
const replacement =
selectedFault === "invalid" ? Buffer.from([0xe9]) : Buffer.from("é");
const parts = source.split(MARKER);
const bytes = Buffer.concat(
parts.flatMap((part, index) =>
index === 0 ? [Buffer.from(part)] : [replacement, Buffer.from(part)],
),
);
// Leave the first multibyte character incomplete in one actual HTTP chunk.
response.write(bytes.subarray(0, start + 1));
await delay(10);
if (selectedFault === "truncated") response.end();
else response.end(bytes.subarray(start + 1));
},
);
await new Promise<void>((resolve) => server.listen(0, "127.0.0.1", resolve));
const address = server.address();
if (address === null || typeof address === "string")
throw new Error("Missing HTTP listener");
return {
endpoint: `http://127.0.0.1:${address.port}/mcp`,
setFault: (value: typeof fault) => {
fault = value;
if (value === "stalled")
stalled = new Promise<void>((resolve) => {
announceStalled = resolve;
});
},
get stalled() {
return stalled;
},
async close() {
release();
for (const response of stalledResponses) response.destroy();
await close();
},
};
};
it.each(
(["ida", "jeb"] as const).flatMap((provider) =>
(["json", "sse"] as const).map((format) => ({ provider, format })),
),
)(
"rejects malformed UTF-8 over real $provider $format HTTP without altering valid source or cancellation",
async ({ provider, format }) => {
const fixture = await startFixture(format);
const createConnection = () => {
if (provider === "ida") {
const connection = createIdaMcpConnection(
{
url: fixture.endpoint,
headers: {},
mode: "attached",
protocol: "native",
timeoutMs: 1_000,
},
{},
);
return {
connect: () => connection.connect(),
close: () => connection.close(),
call: (signal?: AbortSignal) =>
connection.call("get_client_information", {}, { signal }),
};
}
const connection = createStreamableHttpJebMcpConnection(
new URL(fixture.endpoint),
{
discoveryTimeoutMs: 1_000,
requestTimeoutMs: 1_000,
},
);
return {
connect: () => connection.connect(),
close: () => connection.close(),
call: (signal?: AbortSignal) =>
connection.call("get_client_information", {}, signal),
};
};
let connection = createConnection();
const call = (signal?: AbortSignal) => connection.call(signal);
try {
await connection.connect();
const expected = { ...payload, text: "café" };
expect(await call()).toEqual(expected);
for (const fault of ["invalid", "truncated"] as const) {
fixture.setFault(fault);
await expect(call()).rejects.toThrow("not valid UTF-8");
if (provider === "ida") {
// IDA's session owner retires failed connections rather than restarting SDK transports.
await connection.close();
connection = createConnection();
await connection.connect();
}
}
fixture.setFault("valid");
expect(await call()).toEqual(expected);
fixture.setFault("stalled");
const sibling = call().catch((failure: unknown) => failure);
await fixture.stalled;
fixture.setFault("valid");
expect(await call()).toEqual(expected);
fixture.setFault("invalid");
await expect(call()).rejects.toThrow("not valid UTF-8");
expect(await sibling).toMatchObject({ _tag: "AnalysisProtocolError" });
if (provider === "ida") {
await connection.close();
connection = createConnection();
await connection.connect();
}
fixture.setFault("stalled");
const controller = new AbortController();
const reason = new Error("caller stopped the incomplete source reply");
const pending = call(controller.signal).catch(
(failure: unknown) => failure,
);
await fixture.stalled;
controller.abort(reason);
const failure = await pending;
expect(failure).toBeInstanceOf(AnalysisCancelledError);
expect(failure).toMatchObject({ cause: reason });
} finally {
await connection.close();
await fixture.close();
}
},
);