@sagentlab/navarch-runtime 0.1.112 → 0.1.117

package/node_modules/@muse-code/sdk/dist/src/connection/spawn.d.ts

@@ -0,0 +1,188 @@
+/** Node stdio binding and type-state handshake for the MSP connection. */
+import { spawn } from "node:child_process";
+import type { ChildProcessWithoutNullStreams, SpawnOptionsWithoutStdio } from "node:child_process";
+import type { InitializeParams, InitializeResult } from "../msp.js";
+import { Connection } from "./connection.js";
+import type { ConnectionOptions, DuplexTransport, NotificationHandler, ProtocolErrorHandler, ServerRequestHandler } from "./connection.js";
+import type { FingerprintWarning } from "../fingerprint.js";
+/**
+ * What `spawnMspConnection` needs to launch a host: the command to run, plus
+ * optional arguments, working directory, environment, connection tuning, a
+ * stderr tap, and the shutdown drain budget.
+ */
+export interface SpawnMspConnectionOptions {
+    readonly command: string;
+    readonly args?: readonly string[];
+    readonly cwd?: string;
+    readonly env?: NodeJS.ProcessEnv;
+    readonly connection?: ConnectionOptions;
+    /** Raw stderr chunks, drained from process birth and never parsed. */
+    readonly onStderr?: (chunk: string) => void;
+    /** See {@link MuseServeChildOptions.shutdownTimeoutMs}. */
+    readonly shutdownTimeoutMs?: number;
+}
+/**
+ * How the host process ended: its exit code, or the signal that ended it.
+ * Exactly one of the two is non-null.
+ */
+export interface ProcessExit {
+    readonly code: number | null;
+    readonly signal: NodeJS.Signals | null;
+}
+/**
+ * The SDK's reading of a host exit. Branch on `kind`: a clean shutdown, a
+ * crash, an unhandled error, or one of the documented error exits. Every arm
+ * but `cleanShutdown` carries its exit evidence (code or signal plus a
+ * bounded stderr tail), and, where the meaning is known, whether retrying
+ * can help.
+ */
+export type ExitClassification = {
+    readonly kind: "cleanShutdown";
+} | {
+    readonly kind: "unhandledError";
+    readonly exitCode: number;
+    readonly stderrTail: readonly string[];
+} | {
+    readonly kind: "usageError" | "configError" | "leaseUnavailable" | "sdkSurfaceUnavailable";
+    readonly exitCode: number;
+    readonly stderrTail: readonly string[];
+    readonly retry: "never" | "fix-config" | "after-lease-release";
+} | {
+    readonly kind: "crash";
+    readonly exitCode: number | null;
+    readonly exitSignal: string | null;
+    readonly stderrTail: readonly string[];
+};
+/**
+ * Options for launching an SDK-owned host process: the binary to run, plus
+ * optional arguments, working directory, environment, a stderr tap, and the
+ * shutdown drain budget.
+ */
+export interface MuseServeChildOptions {
+    readonly museBin: string;
+    readonly args?: readonly string[];
+    readonly cwd?: string;
+    readonly env?: NodeJS.ProcessEnv;
+    /** Raw evidence callback for consumers that also keep a full diagnostic log. */
+    readonly onStderr?: (chunk: string) => void;
+    /**
+     * How long `close()` lets the host drain after stdin EOF before the SDK
+     * ends the process it owns. Defaults to SS2.1.2's 30 s host drain timeout;
+     * set it to match a host configured with a different one. `0` skips the
+     * drain window (SIGTERM at once) — it is NOT an instant kill: the fixed
+     * SIGTERM→SIGKILL grace still runs. Must be an integer in
+     * `0..2147483647`; anything else throws `RangeError` at spawn.
+     */
+    readonly shutdownTimeoutMs?: number;
+    /**
+     * Deterministic-test seam; production always uses `child_process.spawn`.
+     * A stub here observes the exact options the call site passes, so a
+     * `windowsHide` override at the call site fails the pin (FR-45141-1).
+     */
+    readonly spawnFn?: typeof spawn;
+}
+declare const ownedTransport: unique symbol;
+declare const adoptFlushSource: unique symbol;
+/**
+ * Module friend seam: did `initialize` get a usable answer before the
+ * handshake failed? Exported from this MODULE only (never the package
+ * barrel) so the facade can word a host lost AFTER answering truthfully
+ * (#49912) without a new public member on `MspHandshake`.
+ */
+export declare const initializeAnswered: unique symbol;
+/**
+ * Shutdown deadline contract: `expired` resolves and never rejects; after
+ * `clear()` it may remain pending forever.
+ */
+interface ShutdownDeadline {
+    readonly expired: Promise<void>;
+    clear(): void;
+}
+type ShutdownDeadlineFactory = (budgetMs: number) => ShutdownDeadline;
+/**
+ * Is this `stdin.end()` error the benign close race — the child exited
+ * between the liveness check and the end() call, so its stdin is already
+ * gone? That IS the closed end-state close() wants (SS2.11: the caller
+ * still awaits `exited` and gates on the exit code); anything else is a
+ * real failure. `ECANCELED` is the same owned-teardown race one stage later:
+ * the SIGTERM->SIGKILL backstop destroys stdin while a wedged write is still
+ * queued, and libuv cancels that write instead of erroring it. Exported only
+ * for its deterministic unit test.
+ */
+export declare function isBenignCloseRace(error: unknown): boolean;
+/**
+ * Internal transport over an already-spawned child's stdio. Exported from
+ * this MODULE only (never the package barrel) so the deterministic
+ * close()-race tests can construct it around a stub child — the live arm
+ * awaits `exited` first and cannot reach the race branch (PR #15641 review).
+ */
+export declare class ChildStdioTransport implements DuplexTransport {
+    #private;
+    readonly child: ChildProcessWithoutNullStreams;
+    readonly incoming: AsyncIterable<string>;
+    readonly exited: Promise<ProcessExit>;
+    constructor(child: ChildProcessWithoutNullStreams, onStderr?: (chunk: string) => void, shutdownTimeoutMs?: number, ownsProcessGroup?: boolean, 
+    /** Deterministic-test seam; production always uses `makeShutdownDeadline`. */
+    deadlineFactory?: ShutdownDeadlineFactory);
+    write(chunk: string): Promise<void>;
+    /**
+     * The single owner of shutdown (#15943). EOF first, then bounded: SS2.1.2
+     * gives the host a drain window and says the remainder past it has crash
+     * semantics, so past the window the SDK ends the process it owns rather
+     * than waiting on an EOF that may never come. Memoized, so concurrent and
+     * repeated calls share one attempt and one set of timers.
+     */
+    /** Module-private friend seam used by the handshake that owns both sides. */
+    [adoptFlushSource](source: () => Promise<void>): void;
+    close(flushed?: Promise<void>): Promise<void>;
+}
+/**
+ * The exact options `MuseServeChild.spawn` hands to `child_process.spawn`,
+ * factored out so the spawn-options pin can assert on them without spawning.
+ * Exported from this MODULE only (never the package barrel) — the same
+ * module-only-test-seam precedent as `isBenignCloseRace`.
+ */
+export declare function buildSpawnOptions(options: {
+    readonly cwd?: string;
+    readonly env?: NodeJS.ProcessEnv;
+    readonly detached: boolean;
+}): SpawnOptionsWithoutStdio;
+/** One owned MSP host process with SS2.11 diagnostics and total exit mapping. */
+export declare class MuseServeChild {
+    #private;
+    readonly exit: Promise<ExitClassification>;
+    private constructor();
+    static spawn(options: MuseServeChildOptions): MuseServeChild;
+    get stderrTail(): readonly string[];
+    /** Module-private friend seam used by the handshake machine for this child. */
+    [ownedTransport](): ChildStdioTransport;
+    close(): Promise<ExitClassification>;
+}
+/** The only public state before `initialized`: callers cannot send traffic. */
+export declare class MspHandshake {
+    #private;
+    readonly child: MuseServeChild;
+    constructor(options: SpawnMspConnectionOptions);
+    get exited(): Promise<ProcessExit>;
+    onServerRequest(handler: ServerRequestHandler): void;
+    onNotification(handler: NotificationHandler): void;
+    onProtocolError(handler: ProtocolErrorHandler): void;
+    close(): Promise<ProcessExit>;
+    /** See {@link initializeAnswered}. */
+    [initializeAnswered](): boolean;
+    initialize(params: InitializeParams): Promise<SpawnedMspConnection>;
+}
+/** A fully initialized connection plus its owned process boundary. */
+export declare class SpawnedMspConnection {
+    #private;
+    readonly connection: Connection;
+    readonly child: MuseServeChild;
+    readonly initializeResult: InitializeResult;
+    readonly fingerprintWarning: FingerprintWarning | undefined;
+    readonly exited: Promise<ProcessExit>;
+    constructor(connection: Connection, transport: ChildStdioTransport, child: MuseServeChild, initializeResult: InitializeResult, fingerprintWarning: FingerprintWarning | undefined);
+    close(): Promise<ProcessExit>;
+}
+/** Spawn one owned host and begin its SS1.4 handshake state machine. */
+export declare function spawnMspConnection(options: SpawnMspConnectionOptions): MspHandshake;
+export {};

package/node_modules/@muse-code/sdk/dist/src/connection/spawn.js

@@ -0,0 +1,496 @@
+/** Node stdio binding and type-state handshake for the MSP connection. */
+import { spawn } from "node:child_process";
+import { Connection, ProtocolError, requestWithFrame, submissionTail, } from "./connection.js";
+import { checkServedFingerprint } from "../fingerprint.js";
+const DEFAULT_STDERR_MAX_BYTES = 8 * 1024;
+const DEFAULT_STDERR_MAX_LINES = 100;
+/** SS2.1.2's default host drain timeout: the whole EOF sequence's budget. */
+const DEFAULT_SHUTDOWN_TIMEOUT_MS = 30_000;
+/**
+ * SIGTERM → SIGKILL grace. Not a knob: by the time it runs the host has
+ * already had its whole drain window, so this covers only a signal handler's
+ * final flush. Codex's stdio launcher uses the same two seconds.
+ */
+const SIGTERM_GRACE_MS = 2_000;
+/** Node's own `setTimeout` ceiling; past it a delay silently clamps to 1 ms. */
+const MAX_SHUTDOWN_TIMEOUT_MS = 2_147_483_647;
+const ownedTransport = Symbol("MuseServeChild.transport");
+const adoptFlushSource = Symbol("ChildStdioTransport.adoptFlushSource");
+/**
+ * Module friend seam: did `initialize` get a usable answer before the
+ * handshake failed? Exported from this MODULE only (never the package
+ * barrel) so the facade can word a host lost AFTER answering truthfully
+ * (#49912) without a new public member on `MspHandshake`.
+ */
+export const initializeAnswered = Symbol("MspHandshake.initializeAnswered");
+function makeShutdownDeadline(budgetMs) {
+    let timer;
+    const expired = new Promise((resolve) => {
+        timer = setTimeout(resolve, budgetMs);
+    });
+    return {
+        expired,
+        clear: () => {
+            if (timer !== undefined)
+                clearTimeout(timer);
+            timer = undefined;
+        },
+    };
+}
+/** A client-local evidence budget. Its contents never influence behavior. */
+class StderrTail {
+    #text = "";
+    push(chunk) {
+        this.#text += chunk;
+        this.#trimLines();
+        this.#trimBytes();
+    }
+    lines() {
+        if (this.#text.length === 0)
+            return [];
+        const lines = this.#text.split("\n");
+        if (lines.at(-1) === "")
+            lines.pop();
+        return lines;
+    }
+    #trimLines() {
+        let lineCount = this.#text.split("\n").length;
+        if (this.#text.endsWith("\n"))
+            lineCount -= 1;
+        while (lineCount > DEFAULT_STDERR_MAX_LINES) {
+            const firstBreak = this.#text.indexOf("\n");
+            if (firstBreak < 0)
+                break;
+            this.#text = this.#text.slice(firstBreak + 1);
+            lineCount -= 1;
+        }
+    }
+    #trimBytes() {
+        const bytes = Buffer.from(this.#text, "utf8");
+        if (bytes.byteLength <= DEFAULT_STDERR_MAX_BYTES)
+            return;
+        let start = bytes.byteLength - DEFAULT_STDERR_MAX_BYTES;
+        // If the byte budget cuts through a UTF-8 scalar, advance to its first
+        // complete successor rather than manufacturing replacement text.
+        while (start < bytes.byteLength && (bytes[start] & 0xc0) === 0x80)
+            start += 1;
+        this.#text = bytes.subarray(start).toString("utf8");
+    }
+}
+function classifyExit(processExit, stderrTail) {
+    const evidence = [...stderrTail];
+    switch (processExit.code) {
+        case 0:
+            return { kind: "cleanShutdown" };
+        case 1:
+            return { kind: "unhandledError", exitCode: 1, stderrTail: evidence };
+        case 2:
+            return { kind: "usageError", exitCode: 2, stderrTail: evidence, retry: "never" };
+        case 3:
+            return { kind: "configError", exitCode: 3, stderrTail: evidence, retry: "fix-config" };
+        case 4:
+            // Exit 4 is `sessionInUse` arriving early: another live client holds the
+            // lease, and the lease frees itself once that client exits (SS2.11).
+            return {
+                kind: "leaseUnavailable",
+                exitCode: 4,
+                stderrTail: evidence,
+                retry: "after-lease-release",
+            };
+        case 5:
+            return {
+                kind: "sdkSurfaceUnavailable",
+                exitCode: 5,
+                stderrTail: evidence,
+                retry: "never",
+            };
+        default:
+            return {
+                kind: "crash",
+                exitCode: processExit.code,
+                exitSignal: processExit.signal,
+                stderrTail: evidence,
+            };
+    }
+}
+/**
+ * Is this `stdin.end()` error the benign close race — the child exited
+ * between the liveness check and the end() call, so its stdin is already
+ * gone? That IS the closed end-state close() wants (SS2.11: the caller
+ * still awaits `exited` and gates on the exit code); anything else is a
+ * real failure. `ECANCELED` is the same owned-teardown race one stage later:
+ * the SIGTERM->SIGKILL backstop destroys stdin while a wedged write is still
+ * queued, and libuv cancels that write instead of erroring it. Exported only
+ * for its deterministic unit test.
+ */
+export function isBenignCloseRace(error) {
+    const code = error?.code;
+    return code === "ERR_STREAM_DESTROYED" || code === "EPIPE" || code === "ECANCELED";
+}
+/**
+ * Node clamps a `setTimeout` delay that is non-finite or > 2^31-1 down to
+ * 1 ms, so an out-of-range budget would silently become "kill almost
+ * immediately" and fabricate a crash row for a host that was draining
+ * normally — a row FM-002's ephemeral discard then acts on (PR #22819
+ * review). Refuse it at the owning boundary instead of clamping.
+ */
+function assertShutdownTimeout(shutdownTimeoutMs) {
+    if (!Number.isInteger(shutdownTimeoutMs) ||
+        shutdownTimeoutMs < 0 ||
+        shutdownTimeoutMs > MAX_SHUTDOWN_TIMEOUT_MS) {
+        throw new RangeError(`shutdownTimeoutMs must be an integer in 0..${MAX_SHUTDOWN_TIMEOUT_MS}, got ${String(shutdownTimeoutMs)}`);
+    }
+}
+/**
+ * Internal transport over an already-spawned child's stdio. Exported from
+ * this MODULE only (never the package barrel) so the deterministic
+ * close()-race tests can construct it around a stub child — the live arm
+ * awaits `exited` first and cannot reach the race branch (PR #15641 review).
+ */
+export class ChildStdioTransport {
+    child;
+    incoming;
+    exited;
+    #shutdownTimeoutMs;
+    #makeShutdownDeadline;
+    // True only when the boundary that SPAWNED this child made it a process-
+    // group leader (FR-017b). Never inferred from the child itself: a test
+    // handing the transport an ordinary child keeps the direct-kill path.
+    #ownsProcessGroup;
+    #closing;
+    // The default `flushed` for closes routed AROUND the connection — notably
+    // the public `child.close()`. Without it, step 0 existed only for
+    // Connection-routed closes while the contract claimed all three surfaces
+    // behaved identically, and `#closing ??=` let whichever call won the race
+    // discard the other's submission promise (PR #22819 round 4).
+    #defaultFlushed;
+    constructor(child, onStderr, shutdownTimeoutMs = DEFAULT_SHUTDOWN_TIMEOUT_MS, ownsProcessGroup = false, 
+    /** Deterministic-test seam; production always uses `makeShutdownDeadline`. */
+    deadlineFactory = makeShutdownDeadline) {
+        this.child = child;
+        assertShutdownTimeout(shutdownTimeoutMs);
+        this.#shutdownTimeoutMs = shutdownTimeoutMs;
+        this.#ownsProcessGroup = ownsProcessGroup;
+        this.#makeShutdownDeadline = deadlineFactory;
+        this.child.stdout.setEncoding("utf8");
+        this.child.stderr.setEncoding("utf8");
+        this.child.stderr.on("data", (chunk) => onStderr?.(chunk));
+        // Node delivers a failed stdin write to BOTH the write callback and a
+        // later 'error' event; with no listener that event kills the embedding
+        // process (FM-001). The permanent no-op absorbs it — the write/end
+        // callbacks below carry the same error to the real rejection path.
+        this.child.stdin.on("error", () => { });
+        this.incoming = this.#stdoutChunks();
+        this.exited = new Promise((resolve, reject) => {
+            this.child.once("error", reject);
+            this.child.once("close", (code, signal) => resolve({ code, signal }));
+        });
+        // Mark handled at birth: a spawn failure (e.g. ENOENT) rejects `exited`
+        // before anyone can await it, and the default unhandled-rejection policy
+        // would kill the consumer. Later awaiters still receive the rejection.
+        this.exited.catch(() => { });
+    }
+    async write(chunk) {
+        if (this.child.stdin.destroyed)
+            throw new ProtocolError("spawned MSP host stdin is closed");
+        await new Promise((resolve, reject) => {
+            this.child.stdin.write(chunk, "utf8", (error) => (error ? reject(error) : resolve()));
+        });
+    }
+    /**
+     * The single owner of shutdown (#15943). EOF first, then bounded: SS2.1.2
+     * gives the host a drain window and says the remainder past it has crash
+     * semantics, so past the window the SDK ends the process it owns rather
+     * than waiting on an EOF that may never come. Memoized, so concurrent and
+     * repeated calls share one attempt and one set of timers.
+     */
+    /** Module-private friend seam used by the handshake that owns both sides. */
+    [adoptFlushSource](source) {
+        this.#defaultFlushed = source;
+    }
+    async close(flushed) {
+        this.#closing ??= this.#shutdown(flushed ?? this.#defaultFlushed?.());
+        await this.#closing;
+    }
+    async #shutdown(flushed) {
+        // ONE budget covers the whole shutdown, flush wait included, so handing
+        // the connection's submission promise in here cannot double the bound.
+        const deadline = this.#makeShutdownDeadline(this.#shutdownTimeoutMs);
+        try {
+            // Frames the connection accepted before close() must reach the stream
+            // ahead of EOF (PR #22819 round 3); a peer that never drains simply
+            // spends the drain window here instead of after EOF.
+            if (flushed !== undefined)
+                await this.#beforeDeadline(flushed, deadline.expired);
+            return await this.#shutdownAfterFlush(deadline.expired);
+        }
+        finally {
+            deadline.clear();
+        }
+    }
+    async #shutdownAfterFlush(deadline) {
+        // The EOF attempt must NEVER gate the ladder (PR #22819 review, P0). Two
+        // reachable paths leave it unsettled forever, and both are the #15943
+        // wedge itself: a host that stopped READING stdin never flushes a filled
+        // pipe, so `end()`'s callback never fires; and a non-benign end() error
+        // would reject here, before a single signal is sent — memoized, so every
+        // later close() replays that rejection with the child still alive. Start
+        // the clock regardless and await the EOF outcome AFTER the child is gone,
+        // where a pending write settles as the benign EPIPE race and a real error
+        // still surfaces to the caller.
+        const eof = this.#endStdin();
+        eof.catch(() => { });
+        await this.#awaitExitOrTerminate(deadline);
+        await eof;
+    }
+    async #endStdin() {
+        if (this.child.stdin.destroyed || this.child.stdin.writableEnded)
+            return;
+        await new Promise((resolve, reject) => {
+            // The destroyed/writableEnded guard above is racy: the child can exit
+            // between it and end(), and Node then reports ERR_STREAM_DESTROYED
+            // asynchronously (seen on the slower CI runners driving the real
+            // serve-fixture child, PR #15641). That race is the end-state close()
+            // wants, so it resolves; a real error still rejects.
+            this.child.stdin.end((error) => error === null || error === undefined || isBenignCloseRace(error)
+                ? resolve()
+                : reject(error));
+        });
+    }
+    /**
+     * Escalate only as far as the host forces. Every stage ends at the child's
+     * REAL exit, so `exited` — and every classification derived from it — stays
+     * a function of the observed `(code, signal)` pair (INV-011). A signal is a
+     * way to reach an exit, never a substitute for observing one.
+     */
+    async #awaitExitOrTerminate(deadline) {
+        if (await this.#beforeDeadline(this.#settled(), deadline))
+            return;
+        this.#signal("SIGTERM");
+        if (await this.#settledWithin(SIGTERM_GRACE_MS))
+            return;
+        // Past SIGTERM, drop our end of the pipe ourselves (PR #22819 round 3).
+        // Nothing else does, so a write the host never drained keeps `stdin`
+        // open — and the child's `close` event, which `exited` and the trailing
+        // EOF await both wait on, needs the stdio closed. Waiting for that write
+        // to error out on its own is the last unbounded stage; this removes it.
+        this.child.stdin.destroy();
+        this.#signal("SIGKILL");
+        await this.#settled();
+    }
+    /**
+     * One signal per escalation stage, delivered to the whole owned process
+     * group where this spawn created one (FR-017b, #22777). The child PID
+     * alone is not enough: a grandchild that inherited the host's stdout
+     * keeps the `close` event — which `exited` and every stage here wait
+     * on — from firing even after the child itself is gone, and it never
+     * receives a child-targeted signal at all. A group signal that fails for
+     * any reason (expected: `ESRCH`, the whole group already exited between
+     * stages) falls through to the direct path, which Node makes a no-op on
+     * an exited child — the ladder must never rethrow out of a stage.
+     */
+    #signal(signal) {
+        const pid = this.child.pid;
+        if (this.#ownsProcessGroup && pid !== undefined) {
+            try {
+                process.kill(-pid, signal);
+                return;
+            }
+            catch {
+                // Fall through to the direct-child path, safe on every failure here.
+            }
+        }
+        this.child.kill(signal);
+    }
+    /** True once `exited` settles either way — a rejection means no live child. */
+    async #settledWithin(budgetMs) {
+        return await this.#within(this.#settled(), budgetMs);
+    }
+    /** True if `work` settles before the one shared shutdown deadline fires. */
+    async #beforeDeadline(work, deadline) {
+        return await Promise.race([
+            work.then(() => true, () => true),
+            deadline.then(() => false),
+        ]);
+    }
+    /** True if `work` settles inside `budgetMs`; false when the budget expires. */
+    async #within(work, budgetMs) {
+        // Deliberately ambient (NOT this.#makeShutdownDeadline): the SIGTERM grace
+        // window is outside the one shared shutdown budget the injected seam models.
+        const deadline = makeShutdownDeadline(budgetMs);
+        try {
+            return await this.#beforeDeadline(work, deadline.expired);
+        }
+        finally {
+            deadline.clear();
+        }
+    }
+    /**
+     * `exited` REJECTS when the spawn itself failed (ENOENT). There is no
+     * process to end and no exit to observe, so shutdown is complete; the
+     * rejection still reaches the caller through its own `await exited`
+     * (`MuseServeChild.close()`, `MspHandshake.close()`), unchanged.
+     */
+    async #settled() {
+        await this.exited.catch(() => { });
+    }
+    async *#stdoutChunks() {
+        for await (const chunk of this.child.stdout)
+            yield String(chunk);
+    }
+}
+/**
+ * The exact options `MuseServeChild.spawn` hands to `child_process.spawn`,
+ * factored out so the spawn-options pin can assert on them without spawning.
+ * Exported from this MODULE only (never the package barrel) — the same
+ * module-only-test-seam precedent as `isBenignCloseRace`.
+ */
+export function buildSpawnOptions(options) {
+    return {
+        ...(options.cwd === undefined ? {} : { cwd: options.cwd }),
+        ...(options.env === undefined ? {} : { env: options.env }),
+        detached: options.detached,
+        // FR-45141-1: `muse serve` is never meant to be seen. Without this, a
+        // Windows host with no console of its own pops a console window for the
+        // life of the connection; the flag is a no-op on other platforms.
+        windowsHide: true,
+    };
+}
+/** One owned MSP host process with SS2.11 diagnostics and total exit mapping. */
+export class MuseServeChild {
+    #tail = new StderrTail();
+    #transport;
+    exit;
+    constructor(child, onStderr, shutdownTimeoutMs, ownsProcessGroup = false) {
+        this.#transport = new ChildStdioTransport(child, (chunk) => {
+            this.#tail.push(chunk);
+            onStderr?.(chunk);
+        }, shutdownTimeoutMs, ownsProcessGroup);
+        this.exit = this.#transport.exited.then((status) => classifyExit(status, this.#tail.lines()));
+        // A spawn error can reject before a consumer attaches its await. Mark the
+        // derived promise handled at birth without changing what later awaiters see.
+        this.exit.catch(() => { });
+    }
+    static spawn(options) {
+        // Validate BEFORE spawning: a throw from the transport constructor would
+        // otherwise leave the child it was meant to own already running.
+        if (options.shutdownTimeoutMs !== undefined)
+            assertShutdownTimeout(options.shutdownTimeoutMs);
+        // POSIX: own a whole process group, not just the child (FR-017b,
+        // #22777). `detached` makes the host the leader of a new group (and
+        // session), so escalation can end a grandchild holding the inherited
+        // stdout; stdio stays piped, so nothing else about ownership changes.
+        // Windows keeps the child-only path — group signalling via a negative
+        // PID is a POSIX primitive, and pretending otherwise would fake
+        // subtree containment there.
+        const ownsProcessGroup = process.platform !== "win32";
+        const spawnOptions = buildSpawnOptions({
+            cwd: options.cwd,
+            env: options.env,
+            detached: ownsProcessGroup,
+        });
+        const spawnChild = options.spawnFn ?? spawn;
+        return new MuseServeChild(spawnChild(options.museBin, [...(options.args ?? [])], spawnOptions), options.onStderr, options.shutdownTimeoutMs, ownsProcessGroup);
+    }
+    get stderrTail() {
+        return this.#tail.lines();
+    }
+    /** Module-private friend seam used by the handshake machine for this child. */
+    [ownedTransport]() {
+        return this.#transport;
+    }
+    async close() {
+        await this.#transport.close();
+        return await this.exit;
+    }
+}
+/** The only public state before `initialized`: callers cannot send traffic. */
+export class MspHandshake {
+    #connection;
+    child;
+    #transport;
+    #started = false;
+    #answered = false;
+    constructor(options) {
+        this.child = MuseServeChild.spawn({
+            museBin: options.command,
+            ...(options.args === undefined ? {} : { args: options.args }),
+            ...(options.cwd === undefined ? {} : { cwd: options.cwd }),
+            ...(options.env === undefined ? {} : { env: options.env }),
+            ...(options.onStderr === undefined ? {} : { onStderr: options.onStderr }),
+            ...(options.shutdownTimeoutMs === undefined
+                ? {}
+                : { shutdownTimeoutMs: options.shutdownTimeoutMs }),
+        });
+        this.#transport = this.child[ownedTransport]();
+        this.#connection = new Connection(this.#transport, options.connection);
+        // Make step 0 true on ALL THREE close surfaces, not only this class's:
+        // `child.close()` and `spawned.child.close()` are public, and without
+        // this they ended the host's input while a just-accepted frame was still
+        // queued on a microtask (PR #22819 round 4).
+        this.#transport[adoptFlushSource](() => this.#connection[submissionTail]());
+    }
+    get exited() {
+        return this.#transport.exited;
+    }
+    onServerRequest(handler) {
+        this.#connection.onServerRequest(handler);
+    }
+    onNotification(handler) {
+        this.#connection.onNotification(handler);
+    }
+    onProtocolError(handler) {
+        this.#connection.onProtocolError(handler);
+    }
+    async close() {
+        await this.#connection.close();
+        return await this.#transport.exited;
+    }
+    /** See {@link initializeAnswered}. */
+    [initializeAnswered]() {
+        return this.#answered;
+    }
+    async initialize(params) {
+        if (this.#started)
+            throw new ProtocolError("initialize may be sent only once per connection");
+        this.#started = true;
+        const { result: raw, line } = await this.#connection[requestWithFrame]("initialize", params);
+        this.#answered = true;
+        const result = raw;
+        if (typeof result.schema?.fingerprint !== "string") {
+            // The whole refused frame rides on the error, as on every other
+            // refusal: a host that answered with an unusable result was not lost,
+            // and the facade tells the two apart by it (#49912, #50588 review).
+            throw new ProtocolError("initialize result has no schema fingerprint", line);
+        }
+        const fingerprintWarning = checkServedFingerprint(result.schema.fingerprint);
+        this.#connection.notify("initialized");
+        await this.#connection.flush();
+        return new SpawnedMspConnection(this.#connection, this.#transport, this.child, result, fingerprintWarning);
+    }
+}
+/** A fully initialized connection plus its owned process boundary. */
+export class SpawnedMspConnection {
+    connection;
+    child;
+    initializeResult;
+    fingerprintWarning;
+    exited;
+    #transport;
+    constructor(connection, transport, child, initializeResult, fingerprintWarning) {
+        this.connection = connection;
+        this.#transport = transport;
+        this.child = child;
+        this.initializeResult = initializeResult;
+        this.fingerprintWarning = fingerprintWarning;
+        this.exited = transport.exited;
+    }
+    async close() {
+        await this.connection.close();
+        return await this.#transport.exited;
+    }
+}
+/** Spawn one owned host and begin its SS1.4 handshake state machine. */
+export function spawnMspConnection(options) {
+    return new MspHandshake(options);
+}