oh-my-pi/packages/coding-agent/src/mcp/transports/stdio.ts

/**
 * MCP stdio transport.
 *
 * Implements JSON-RPC 2.0 over subprocess stdin/stdout.
 * Messages are newline-delimited JSON.
 */

import { getProjectDir, readJsonl, Snowflake } from "@oh-my-pi/pi-utils";
import { type Subprocess, spawn } from "bun";
import type {
	JsonRpcError,
	JsonRpcMessage,
	JsonRpcRequest,
	JsonRpcResponse,
	MCPRequestOptions,
	MCPStdioServerConfig,
	MCPTransport,
} from "../../mcp/types";
import { toJsonRpcError } from "../../mcp/types";
import { isMCPTimeoutEnabled, resolveMCPTimeoutMs } from "../timeout";

/** Minimal write surface of `Subprocess.stdin` we need for framed sends. */
interface FrameSink {
	write(chunk: string): unknown;
	flush(): unknown;
}

/** Narrow a value to a thenable so a rejection handler can be attached. */
function isThenable(value: unknown): value is PromiseLike<unknown> {
	return (
		value != null &&
		(typeof value === "object" || typeof value === "function") &&
		typeof (value as { then?: unknown }).then === "function"
	);
}

/**
 * Write a newline-delimited JSON-RPC frame to the subprocess's stdin sink,
 * swallowing both synchronous throws and asynchronous rejections so the caller
 * can decide how to react.
 *
 * Bun's `FileSink.write()`/`flush()` can fail two ways once the read end of the
 * pipe has been closed by a subprocess that exited between read-loop ticks:
 *   - a synchronous throw (most reliably observed on Windows), and
 *   - a *rejected Promise* returned from `write()`/`flush()`, i.e. the EPIPE is
 *     surfaced asynchronously (note the `processTicksAndRejections` frame in the
 *     stack traces on #1710 and the follow-up report).
 *
 * A sibling `async` method's `try/catch` only catches the synchronous case; an
 * un-awaited rejected Promise escapes as a fatal unhandled rejection. So we both
 * catch the throw and neutralize any returned promise's rejection.
 *
 * Returns `true` when the frame was accepted synchronously, `false` when the
 * sink threw — callers signal transport closure on `false`. An asynchronous
 * failure cannot be reflected in the return value; it is neutralized here and
 * the dead transport is detected by the read loop / request timeout instead.
 */
export function writeFrame(stdin: FrameSink, frame: string): boolean {
	try {
		const wrote = stdin.write(frame);
		const flushed = stdin.flush();
		if (isThenable(wrote)) wrote.then(undefined, () => {});
		if (isThenable(flushed)) flushed.then(undefined, () => {});
		return true;
	} catch {
		return false;
	}
}

/**
 * Stdio transport for MCP servers.
 * Spawns a subprocess and communicates via stdin/stdout.
 */
export class StdioTransport implements MCPTransport {
	#process: Subprocess<"pipe", "pipe", "pipe"> | null = null;
	#pendingRequests = new Map<
		string | number,
		{
			resolve: (value: unknown) => void;
			reject: (error: Error) => void;
		}
	>();
	#connected = false;
	#readLoop: Promise<void> | null = null;

	onClose?: () => void;
	onError?: (error: Error) => void;
	onNotification?: (method: string, params: unknown) => void;
	onRequest?: (method: string, params: unknown) => Promise<unknown>;

	constructor(private config: MCPStdioServerConfig) {}

	get connected(): boolean {
		return this.#connected;
	}

	/**
	 * Start the subprocess and begin reading.
	 */
	async connect(): Promise<void> {
		if (this.#connected) return;

		const args = this.config.args ?? [];
		const env = {
			...Bun.env,
			...this.config.env,
		};

		this.#process = spawn({
			cmd: [this.config.command, ...args],
			cwd: this.config.cwd ?? getProjectDir(),
			env,
			stdin: "pipe",
			stdout: "pipe",
			stderr: "pipe",
		});

		this.#connected = true;

		// Start reading stdout
		this.#readLoop = this.#startReadLoop();

		// Log stderr for debugging
		this.#startStderrLoop();
	}

	async #startReadLoop(): Promise<void> {
		if (!this.#process?.stdout) return;
		try {
			for await (const line of readJsonl(this.#process.stdout)) {
				if (!this.#connected) break;
				try {
					this.#handleMessage(line as JsonRpcMessage);
				} catch {
					// Skip malformed lines
				}
			}
		} catch (error) {
			if (this.#connected) {
				this.onError?.(error instanceof Error ? error : new Error(String(error)));
			}
		} finally {
			this.#handleClose();
		}
	}

	async #startStderrLoop(): Promise<void> {
		if (!this.#process?.stderr) return;

		const reader = this.#process.stderr.getReader();
		const decoder = new TextDecoder();

		try {
			while (this.#connected) {
				const { done, value } = await reader.read();
				if (done) break;
				// Log stderr but don't treat as error - servers use it for logging
				const text = decoder.decode(value, { stream: true });
				if (text.trim()) {
					// Could expose via onStderr callback if needed
					// For now, silent - MCP spec says clients MAY capture/ignore
				}
			}
		} catch {
			// Ignore stderr read errors
		} finally {
			reader.releaseLock();
		}
	}

	#handleMessage(message: JsonRpcMessage | JsonRpcMessage[]): void {
		if (Array.isArray(message)) {
			for (const m of message) this.#handleMessage(m);
			return;
		}
		// Server-to-client request: has both method and id
		if ("method" in message && "id" in message && message.id != null) {
			void this.#handleServerRequest(message as JsonRpcRequest);
			return;
		}

		// Response to our request: has id
		if ("id" in message && message.id != null) {
			const response = message as JsonRpcResponse;
			const pending = this.#pendingRequests.get(response.id);
			if (pending) {
				this.#pendingRequests.delete(response.id);
				if (response.error) {
					pending.reject(new Error(`MCP error ${response.error.code}: ${response.error.message}`));
				} else {
					pending.resolve(response.result);
				}
			}
			return;
		}

		// Notification: has method but no id
		if ("method" in message) {
			const notification = message as { method: string; params?: unknown };
			this.onNotification?.(notification.method, notification.params);
		}
	}

	async #handleServerRequest(request: JsonRpcRequest): Promise<void> {
		try {
			if (!this.onRequest) {
				this.#sendResponse(request.id, undefined, { code: -32601, message: "Method not found" });
				return;
			}
			const result = await this.onRequest(request.method, request.params);
			this.#sendResponse(request.id, result);
		} catch (error) {
			this.#sendResponse(request.id, undefined, toJsonRpcError(error));
		}
	}

	#sendResponse(id: string | number, result?: unknown, error?: JsonRpcError): void {
		if (!this.#connected || !this.#process?.stdin) return;
		const response = error
			? { jsonrpc: "2.0" as const, id, error }
			: { jsonrpc: "2.0" as const, id, result: result ?? {} };
		// Silent on failure — a dead subprocess has no use for the response,
		// and the read loop will close the transport on EOF.
		writeFrame(this.#process.stdin, `${JSON.stringify(response)}\n`);
	}

	#handleClose(): void {
		if (!this.#connected) return;
		this.#connected = false;

		// Reject all pending requests
		for (const [, pending] of this.#pendingRequests) {
			pending.reject(new Error("Transport closed"));
		}
		this.#pendingRequests.clear();

		this.onClose?.();
	}

	async request<T = unknown>(
		method: string,
		params?: Record<string, unknown>,
		options?: MCPRequestOptions,
	): Promise<T> {
		if (!this.#connected || !this.#process?.stdin) {
			throw new Error("Transport not connected");
		}

		const id = Snowflake.next();
		const request = {
			jsonrpc: "2.0" as const,
			id,
			method,
			params: params ?? {},
		};

		const timeout = resolveMCPTimeoutMs(this.config.timeout);
		const signal = options?.signal;

		if (signal?.aborted) {
			const reason = signal.reason instanceof Error ? signal.reason : new Error("Aborted");
			return Promise.reject(reason);
		}

		const { promise, resolve, reject } = Promise.withResolvers<T>();
		let timer: NodeJS.Timeout | undefined;
		let settled = false;

		const cleanup = () => {
			if (settled) return;
			settled = true;
			if (timer) {
				clearTimeout(timer);
				timer = undefined;
			}
			if (signal) {
				signal.removeEventListener("abort", onAbort);
			}
			this.#pendingRequests.delete(id);
		};

		const onAbort = () => {
			cleanup();
			const reason = signal?.reason instanceof Error ? signal.reason : new Error("Aborted");
			reject(reason);
		};

		if (signal) {
			signal.addEventListener("abort", onAbort, { once: true });
		}

		this.#pendingRequests.set(id, {
			resolve: (value: unknown) => {
				cleanup();
				resolve(value as T);
			},
			reject: (error: Error) => {
				cleanup();
				reject(error);
			},
		});

		if (isMCPTimeoutEnabled(timeout)) {
			timer = setTimeout(() => {
				cleanup();
				reject(new Error(`Request timeout after ${timeout}ms`));
			}, timeout);
		}

		const stdin = this.#process.stdin;
		const message = `${JSON.stringify(request)}\n`;
		try {
			// Await both: Bun's FileSink can surface a broken pipe either as a
			// synchronous throw or as a rejected Promise (the EPIPE arrives on a
			// processTicksAndRejections tick). Awaiting funnels both into this catch
			// so the request rejects cleanly instead of leaving a floating rejected
			// promise that crashes the process via the unhandledRejection handler.
			await stdin.write(message);
			await stdin.flush();
		} catch (error: unknown) {
			cleanup();
			reject(error instanceof Error ? error : new Error(String(error)));
		}

		return promise;
	}

	async notify(method: string, params?: Record<string, unknown>): Promise<void> {
		if (!this.#connected || !this.#process?.stdin) {
			throw new Error("Transport not connected");
		}

		const notification = {
			jsonrpc: "2.0" as const,
			method,
			params: params ?? {},
		};

		// Bun's FileSink can throw EPIPE synchronously on Windows when the
		// subprocess has exited between the last read-loop tick and this
		// write (e.g. an MCP server that dies after returning `initialize`
		// but before `notifications/initialized` is delivered). Tear the
		// transport down so any wired `onClose` (and reconnect machinery)
		// engages, then surface the failure to the caller so a write that
		// dropped on the floor is never silently treated as delivered —
		// `initializeConnection()` runs before the manager installs its
		// `onClose` handler, so a swallowed failure there would yield a
		// "connected" handle wrapping a dead transport. See #1710.
		if (!writeFrame(this.#process.stdin, `${JSON.stringify(notification)}\n`)) {
			this.#handleClose();
			throw new Error(`Transport closed while sending notification "${method}"`);
		}
	}

	async close(): Promise<void> {
		// `close()` is the authoritative resource teardown. `#handleClose()`
		// may have already run (read-loop EOF, or a notify() write failure
		// that surfaces the dead transport to the caller) and flipped
		// `#connected` to false — but the subprocess and read loop are still
		// alive in that path, so we MUST keep cleaning up regardless. Each
		// step is individually guarded so this remains idempotent across
		// repeat calls.
		if (this.#connected) {
			this.#handleClose();
		}

		if (this.#process) {
			this.#process.kill();
			this.#process = null;
		}

		if (this.#readLoop) {
			await this.#readLoop.catch(() => {});
			this.#readLoop = null;
		}
	}
}

/**
 * Create and connect a stdio transport.
 */
export async function createStdioTransport(config: MCPStdioServerConfig): Promise<StdioTransport> {
	const transport = new StdioTransport(config);
	await transport.connect();
	return transport;
}