@moretea-labs/forge 1.7.2 → 1.8.1

package/adapters/mcp/server.ts

@@ -0,0 +1,1092 @@
+import { Server } from "@modelcontextprotocol/server";
+import type { Tool } from "@modelcontextprotocol/server";
+import { Client, SdkError, SdkErrorCode, SdkHttpError, StreamableHTTPClientTransport } from "@modelcontextprotocol/client";
+import { mcpServerInstructions } from './instructions';
+import { buildMcpToolDefinitions, callMcpTool, type CallToolResult, type McpToolContext } from './tool-mapping/tools';
+import { mcpToolDefinitionFromSdk, mcpToolDefinitionToSdk, type McpToolDefinition } from '../../packages/protocols/mcp/tool-contract';
+import { createLegacyMcpToolContext } from './legacy-context';
+import {
+  buildMultiRepositoryToolDefinitions,
+  callMultiRepositoryTool,
+  createMcpToolContext as createMultiRepositoryToolContext,
+  type McpServerOptions,
+  type MultiRepositoryMcpToolContext,
+} from './multi-repository';
+import { callAccessTool } from './tool-mapping/access-tools';
+import { callRepositoryToolWithPostFinalizeAttribution } from './tool-mapping/post-finalize-work-attribution';
+import { callRuntimeTool } from './runtime-gateway/runtime-tools';
+import { callExecutionTool } from './runtime-gateway/execution-tools';
+import { callProcessTool } from './runtime-gateway/process-tools';
+import { injectDurableCommandFields, isGatewayIsolatedTool, routeDurableMcpCall } from './runtime-gateway/router';
+import { controllerExposureSnapshot, isControllerToolExposed } from './toolset';
+import { settleCognitionAfterTool } from './runtime-gateway/cognition-settlement';
+import { randomUUID } from 'crypto';
+import { readFileSync } from 'fs';
+import { join } from 'path';
+import { observeRuntimeStatus } from '../../src/runtime/root/status';
+import { readRuntimeReleaseAuthority } from '../../src/runtime/root/release-store';
+import { getRuntimeWriteClaim } from '../../src/runtime/root/write-fence';
+import { recordMcpIncident, recordMcpTiming, type McpTimingTrace } from '../../src/runtime/diagnostics/mcp-timing';
+import { maybeRegisterMcpIncidentRepair } from '../../src/runtime/diagnostics/incident-repair';
+import { FORGE_VERSION, forgeToolSurfaceFingerprint } from '../../src/cli/controller/runtime-config';
+import { RUNTIME_MCP_CUTOVER_OUTCOME_UNKNOWN } from '../../src/runtime/root/mcp-transport';
+
+export type { McpServerOptions } from './multi-repository';
+export { buildMultiRepositoryToolDefinitions, callMultiRepositoryTool } from './multi-repository';
+
+type ServerToolContext = McpToolContext | MultiRepositoryMcpToolContext;
+
+/** A per-session schema read directly from the Canonical Runtime's tools/list. */
+export interface CanonicalRuntimeToolSchema {
+  definitions: McpToolDefinition[];
+  toolNames: string[];
+  fingerprint: string;
+}
+
+function isMultiRepositoryContext(ctx: ServerToolContext): ctx is MultiRepositoryMcpToolContext {
+  return 'controllerHome' in ctx;
+}
+
+function recordRequestId(args: Record<string, unknown>, _rpcId: string | number | undefined): string {
+  const explicit = typeof args.request_id === 'string' ? args.request_id.trim() : '';
+  // Transport RPC ids are connection-local correlation only. They are not a
+  // semantic idempotency authority and may be recycled across unrelated calls.
+  return explicit || `mcp-call:${randomUUID()}`;
+}
+
+function firstTimingString(...values: unknown[]): string | undefined {
+  for (const value of values) {
+    if (typeof value === 'string' && value.trim()) return value.trim();
+  }
+  return undefined;
+}
+
+function timingIdentity(value: CallToolResult | undefined): {
+  repoId?: string;
+  workId?: string;
+  processId?: string;
+  route?: string;
+} {
+  if (!value?.structuredContent || typeof value.structuredContent !== 'object' || Array.isArray(value.structuredContent)) return {};
+  const structured = value.structuredContent as Record<string, unknown>;
+  const work = structured.work && typeof structured.work === 'object' && !Array.isArray(structured.work)
+    ? structured.work as Record<string, unknown>
+    : undefined;
+  const process = structured.process && typeof structured.process === 'object' && !Array.isArray(structured.process)
+    ? structured.process as Record<string, unknown>
+    : undefined;
+  const resultRef = structured.resultRef && typeof structured.resultRef === 'object' && !Array.isArray(structured.resultRef)
+    ? structured.resultRef as Record<string, unknown>
+    : undefined;
+  return {
+    repoId: firstTimingString(structured.repoId, structured.repo_id, work?.repositoryId, work?.repoId),
+    workId: firstTimingString(structured.workId, structured.work_id, work?.workId),
+    processId: firstTimingString(structured.processId, structured.process_id, process?.processId, resultRef?.processId),
+    route: firstTimingString(structured.route, structured.path, structured.mode, process?.route),
+  };
+}
+
+function tracedResult(
+  value: CallToolResult | undefined,
+  traceId: string,
+  requestId: string,
+  serverTiming?: { startedAt: string; durationMs: number },
+): CallToolResult | undefined {
+  if (!value || !value.structuredContent || typeof value.structuredContent !== 'object' || Array.isArray(value.structuredContent)) return value;
+  const originalStructuredContent = value.structuredContent as Record<string, unknown>;
+  const structuredContent = {
+    ...originalStructuredContent,
+    responseMeta: {
+      ...(originalStructuredContent.responseMeta && typeof originalStructuredContent.responseMeta === 'object'
+        ? originalStructuredContent.responseMeta as Record<string, unknown>
+        : {}),
+      traceId,
+      requestId,
+      ...(serverTiming ? { serverStartedAt: serverTiming.startedAt, serverDurationMs: serverTiming.durationMs } : {}),
+    },
+  };
+  return {
+    ...value,
+    structuredContent,
+    content: value.content.map((block, index) => index === 0 && block.type === 'text'
+      ? { ...block, text: JSON.stringify(structuredContent) }
+      : block),
+  };
+}
+
+async function traceControllerMcpRequest(
+  ctx: MultiRepositoryMcpToolContext,
+  name: string,
+  args: Record<string, unknown>,
+  rpcId: string | number | undefined,
+  handler: (requestId: string, traceId: string, phaseTimings: Partial<McpTimingTrace>) => Promise<CallToolResult>,
+): Promise<CallToolResult> {
+  const startedAtWall = new Date().toISOString();
+  const startedAt = performance.now();
+  const traceId = randomUUID();
+  const requestId = recordRequestId(args, rpcId);
+  let outcome: 'ok' | 'error' | 'exception' = 'ok';
+  let errorCode: string | undefined;
+  let executionIdentity: ReturnType<typeof timingIdentity> = {};
+  const phaseTimings: Partial<McpTimingTrace> = {};
+  try {
+    const value = await handler(requestId, traceId, phaseTimings);
+    executionIdentity = timingIdentity(value);
+    if (value?.isError) {
+      outcome = 'error';
+      const error = value.structuredContent && typeof value.structuredContent === 'object'
+        ? (value.structuredContent as Record<string, unknown>).error
+        : undefined;
+      if (error && typeof error === 'object' && !Array.isArray(error)) {
+        errorCode = typeof (error as Record<string, unknown>).code === 'string'
+          ? (error as Record<string, unknown>).code as string
+          : 'MCP_TOOL_ERROR';
+        const structuredRepoId = typeof (value.structuredContent as Record<string, unknown>).repoId === 'string'
+          ? (value.structuredContent as Record<string, unknown>).repoId as string
+          : undefined;
+        const incident = {
+          traceId,
+          requestId,
+          ...(rpcId === undefined ? {} : { rpcId }),
+          tool: name,
+          kind: 'tool_error' as const,
+          code: errorCode,
+          message: typeof (error as Record<string, unknown>).message === 'string'
+            ? (error as Record<string, unknown>).message as string
+            : 'MCP tool returned an error.',
+          ...(structuredRepoId
+            ? { repoId: structuredRepoId }
+            : typeof args.repo_id === 'string' && args.repo_id.trim()
+              ? { repoId: args.repo_id.trim() }
+              : ctx.explicitRepository?.repoId
+                ? { repoId: ctx.explicitRepository.repoId }
+                : {}),
+        };
+        recordMcpIncident(ctx.controllerHome, incident);
+        try {
+          maybeRegisterMcpIncidentRepair({
+            controllerHome: ctx.controllerHome,
+            runtimeSourceRoot: ctx.runtimeSourceRoot,
+            incident,
+          });
+        } catch {
+          // Incident promotion is best-effort and must never alter the MCP result.
+        }
+      }
+    }
+    const serverDurationMs = Math.round((performance.now() - startedAt) * 100) / 100;
+    return tracedResult(value, traceId, requestId, { startedAt: startedAtWall, durationMs: serverDurationMs }) as CallToolResult;
+  } catch (error) {
+    outcome = 'exception';
+    errorCode = 'MCP_REQUEST_EXCEPTION';
+    const incident = {
+      traceId,
+      requestId,
+      ...(rpcId === undefined ? {} : { rpcId }),
+      tool: name,
+      kind: 'exception' as const,
+      code: errorCode,
+      message: error instanceof Error ? error.message : String(error),
+      ...(typeof args.repo_id === 'string' && args.repo_id.trim()
+        ? { repoId: args.repo_id.trim() }
+        : ctx.explicitRepository?.repoId
+          ? { repoId: ctx.explicitRepository.repoId }
+          : {}),
+    };
+    recordMcpIncident(ctx.controllerHome, incident);
+    try {
+      maybeRegisterMcpIncidentRepair({
+        controllerHome: ctx.controllerHome,
+        runtimeSourceRoot: ctx.runtimeSourceRoot,
+        incident,
+      });
+    } catch {
+      // Incident promotion is best-effort and must never alter the MCP exception.
+    }
+    throw error;
+  } finally {
+    recordMcpTiming(ctx.controllerHome, {
+      tool: name,
+      traceId,
+      requestId,
+      ...(rpcId === undefined ? {} : { rpcId }),
+      layer: ctx.runtimeSourceRoot ? 'canonical_runtime' : 'public_gateway',
+      startedAt: startedAtWall,
+      ...(ctx.hostConversationSessionId?.trim() ? { hostConversationSessionId: ctx.hostConversationSessionId.trim() } : {}),
+      outcome,
+      ...(errorCode ? { errorCode } : {}),
+      ...executionIdentity,
+      ...phaseTimings,
+      totalToolDurationMs: Math.round((performance.now() - startedAt) * 100) / 100,
+    });
+  }
+}
+
+export function createMcpToolContext(
+  opts: McpServerOptions & { profile?: "controller" },
+): MultiRepositoryMcpToolContext;
+export function createMcpToolContext(opts: McpServerOptions): ServerToolContext;
+export function createMcpToolContext(opts: McpServerOptions): ServerToolContext {
+  const profile = opts.profile ?? 'controller';
+  if (profile !== 'controller') return createLegacyMcpToolContext(opts);
+  const repo = opts.repo?.trim() === '.' ? undefined : opts.repo;
+  return createMultiRepositoryToolContext({ ...opts, repo });
+}
+
+function canonicalRuntimeEndpoint(endpoint: string | undefined): URL {
+  if (!endpoint) throw new Error('CANONICAL_RUNTIME_ENDPOINT_UNAVAILABLE');
+  const url = new URL(endpoint);
+  const host = url.hostname.replace(/^\[|\]$/g, '');
+  if (!['127.0.0.1', 'localhost', '::1'].includes(host)) {
+    throw new Error(`CANONICAL_RUNTIME_ENDPOINT_NOT_LOOPBACK: ${url.origin}`);
+  }
+  return url;
+}
+
+function canonicalRuntimeToken(ctx: MultiRepositoryMcpToolContext): string {
+  const path = join(ctx.controllerHome, 'mcp', 'runtime-token');
+  const token = readFileSync(path, 'utf8').trim();
+  if (!token) throw new Error('CANONICAL_RUNTIME_TOKEN_UNAVAILABLE');
+  return token;
+}
+
+export const CANONICAL_RUNTIME_CONNECT_TIMEOUT_MS = 5_000;
+/**
+ * The loopback connection should fail fast, but a valid Canonical Runtime tool
+ * call may legitimately outlive five seconds (for example Work finalization).
+ * Do not let the thin Public Gateway report a false timeout while authoritative
+ * work continues successfully behind it.
+ */
+export const CANONICAL_RUNTIME_TOOL_CALL_TIMEOUT_MS = 120_000;
+export const CANONICAL_RUNTIME_HANDOFF_WAIT_MS = 20_000;
+export const CANONICAL_RUNTIME_HANDOFF_RETRY_INTERVAL_MS = 150;
+const CANONICAL_RUNTIME_HANDOFF_SIGNAL_TTL_MS = 90_000;
+
+export interface CanonicalRuntimeReleaseHandoffObservation {
+  authorityReleaseId?: string;
+  authorityCommittedAt?: string;
+  runtimeReleaseId?: string;
+  runtimeRunning: boolean;
+  runtimeReady: boolean;
+  runtimeStartedAt?: string;
+  runtimeUpdatedAt?: string;
+  recoveryActivationInProgress?: boolean;
+  nowMs: number;
+}
+
+function timestampIsRecent(value: string | undefined, nowMs: number, maxAgeMs: number): boolean {
+  if (!value) return false;
+  const observedAt = Date.parse(value);
+  if (!Number.isFinite(observedAt)) return false;
+  const ageMs = nowMs - observedAt;
+  return ageMs >= -5_000 && ageMs <= maxAgeMs;
+}
+
+/**
+ * Recognize only a bounded whole-Runtime release handoff. An old dead Runtime
+ * must not keep the public Gateway waiting forever, so every transition signal
+ * is time-bounded. This is observability, never release authority.
+ */
+export function canonicalRuntimeReleaseHandoffInProgress(
+  input: CanonicalRuntimeReleaseHandoffObservation,
+): boolean {
+  if (!input.authorityReleaseId) return false;
+  const sameRelease = input.runtimeReleaseId === input.authorityReleaseId;
+  const authorityRecent = timestampIsRecent(
+    input.authorityCommittedAt,
+    input.nowMs,
+    CANONICAL_RUNTIME_HANDOFF_SIGNAL_TTL_MS,
+  );
+
+  // The release authority switches only after the old Runtime has completely
+  // stopped. Until the new Runtime publishes matching status, this mismatch is
+  // the strongest handoff signal. Missing status is equivalent to mismatch only
+  // while the authority commit is recent.
+  if (!sameRelease && authorityRecent) return true;
+
+  // The new Runtime owns the target release but has not completed scheduler +
+  // MCP end-to-end readiness yet.
+  if (sameRelease && input.runtimeRunning && !input.runtimeReady) {
+    return timestampIsRecent(
+      input.runtimeStartedAt ?? input.runtimeUpdatedAt,
+      input.nowMs,
+      CANONICAL_RUNTIME_HANDOFF_SIGNAL_TTL_MS,
+    );
+  }
+
+  // Bridge the intentional bootout -> authority-publish gap only while the
+  // standalone Recovery transaction explicitly owns the activation lock. A
+  // freshly crashed Runtime without that lock is an outage and stays fail-fast.
+  if (sameRelease && !input.runtimeRunning) {
+    return input.recoveryActivationInProgress === true;
+  }
+
+  return false;
+}
+
+export interface CanonicalRuntimeHandoffWaitResult {
+  waited: boolean;
+  settled: boolean;
+  observations: number;
+}
+
+export async function waitForCanonicalRuntimeReleaseHandoff(
+  handoffInProgress: () => boolean,
+  options: {
+    maxWaitMs?: number;
+    intervalMs?: number;
+    now?: () => number;
+    sleep?: (ms: number) => Promise<void>;
+  } = {},
+): Promise<CanonicalRuntimeHandoffWaitResult> {
+  const maxWaitMs = Math.max(0, options.maxWaitMs ?? CANONICAL_RUNTIME_HANDOFF_WAIT_MS);
+  const intervalMs = Math.max(1, options.intervalMs ?? CANONICAL_RUNTIME_HANDOFF_RETRY_INTERVAL_MS);
+  const now = options.now ?? Date.now;
+  const sleep = options.sleep ?? ((ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms)));
+  let observations = 1;
+  if (!handoffInProgress()) return { waited: false, settled: true, observations };
+  const deadline = now() + maxWaitMs;
+  let waited = false;
+  while (now() < deadline) {
+    waited = true;
+    await sleep(Math.min(intervalMs, Math.max(1, deadline - now())));
+    observations += 1;
+    if (!handoffInProgress()) return { waited, settled: true, observations };
+  }
+  observations += 1;
+  return { waited, settled: !handoffInProgress(), observations };
+}
+
+export async function retryCanonicalRuntimeConnectDuringHandoff<T>(
+  connect: () => Promise<T>,
+  handoffInProgress: () => boolean,
+  options: Parameters<typeof waitForCanonicalRuntimeReleaseHandoff>[1] = {},
+): Promise<T> {
+  const initialWait = await waitForCanonicalRuntimeReleaseHandoff(handoffInProgress, options);
+  if (initialWait.waited && !initialWait.settled) {
+    // Preserve the ordinary connect failure shape after the bounded handoff
+    // window rather than manufacturing a new proxy-only error.
+    return await connect();
+  }
+  try {
+    return await connect();
+  } catch (error) {
+    // Retrying is safe only because connect() has not dispatched a tool call.
+    // Once client.callTool() begins, callers must never use this helper.
+    if (!handoffInProgress()) throw error;
+    const waited = await waitForCanonicalRuntimeReleaseHandoff(handoffInProgress, options);
+    if (!waited.settled) throw error;
+    return await connect();
+  }
+}
+
+function recoveryRuntimeReleaseActivationInProgress(controllerHome: string): boolean {
+  try {
+    const lock = JSON.parse(readFileSync(join(controllerHome, 'recovery', 'locks', 'operation.lock'), 'utf8')) as {
+      action?: string;
+      acquiredAt?: string;
+    };
+    if (lock.action !== 'activate_runtime_release' && lock.action !== 'stage_and_activate_runtime_release') return false;
+    return timestampIsRecent(lock.acquiredAt, Date.now(), CANONICAL_RUNTIME_HANDOFF_SIGNAL_TTL_MS);
+  } catch {
+    return false;
+  }
+}
+
+function canonicalRuntimeReleaseHandoffActive(controllerHome: string): boolean {
+  const runtime = observeRuntimeStatus(controllerHome);
+  if (runtime.running && runtime.ready) return false;
+  const authority = readRuntimeReleaseAuthority(controllerHome);
+  return canonicalRuntimeReleaseHandoffInProgress({
+    authorityReleaseId: authority?.active.releaseId,
+    authorityCommittedAt: authority?.committedAt,
+    runtimeReleaseId: runtime.snapshot?.releaseId,
+    runtimeRunning: runtime.running,
+    runtimeReady: runtime.ready,
+    runtimeStartedAt: runtime.snapshot?.startedAt,
+    runtimeUpdatedAt: runtime.snapshot?.updatedAt,
+    recoveryActivationInProgress: recoveryRuntimeReleaseActivationInProgress(controllerHome),
+    nowMs: Date.now(),
+  });
+}
+
+
+const CANONICAL_RUNTIME_FORWARD_META_KEY = 'forgeRuntimeForwarding';
+
+type ForwardedControllerType = NonNullable<McpToolContext['controllerType']>;
+
+interface CanonicalRuntimeForwardingIdentity {
+  principalId?: string;
+  sessionId?: string;
+  controllerType?: ForwardedControllerType;
+  hostConversationSessionId?: string;
+}
+
+function boundedForwardedIdentity(value: unknown): string | undefined {
+  if (typeof value !== 'string') return undefined;
+  const normalized = value.trim();
+  return normalized ? normalized.slice(0, 512) : undefined;
+}
+
+/** ChatGPT supplies this as an anonymized same-conversation correlation fact. It is not an auth token or URL. */
+export function chatgptHostSessionIdFromMcpMeta(meta: unknown): string | undefined {
+  if (!meta || typeof meta !== 'object' || Array.isArray(meta)) return undefined;
+  return boundedForwardedIdentity((meta as Record<string, unknown>)['openai/session']);
+}
+
+export function canonicalRuntimeForwardingIdentity(meta: unknown): CanonicalRuntimeForwardingIdentity {
+  if (!meta || typeof meta !== 'object' || Array.isArray(meta)) return {};
+  const envelope = (meta as Record<string, unknown>)[CANONICAL_RUNTIME_FORWARD_META_KEY];
+  if (!envelope || typeof envelope !== 'object' || Array.isArray(envelope)) return {};
+  const record = envelope as Record<string, unknown>;
+  const controllerType = boundedForwardedIdentity(record.controllerType);
+  const hostConversationSessionId = boundedForwardedIdentity(record.hostConversationSessionId);
+  return {
+    principalId: boundedForwardedIdentity(record.principalId),
+    sessionId: boundedForwardedIdentity(record.sessionId),
+    ...(controllerType && ['chatgpt', 'codex', 'claude', 'grok', 'human'].includes(controllerType)
+      ? { controllerType: controllerType as ForwardedControllerType }
+      : {}),
+    ...(hostConversationSessionId ? { hostConversationSessionId } : {}),
+  };
+}
+
+function canonicalRuntimeForwardingMeta(ctx: MultiRepositoryMcpToolContext): Record<string, unknown> {
+  return {
+    [CANONICAL_RUNTIME_FORWARD_META_KEY]: {
+      ...(ctx.principalId?.trim() ? { principalId: ctx.principalId.trim() } : {}),
+      ...(ctx.sessionId?.trim() ? { sessionId: ctx.sessionId.trim() } : {}),
+      ...(ctx.controllerType ? { controllerType: ctx.controllerType } : {}),
+      ...(ctx.hostConversationSessionId?.trim() ? { hostConversationSessionId: ctx.hostConversationSessionId.trim() } : {}),
+    },
+  };
+}
+
+function canonicalRuntimeRequestContext(baseContext: ServerToolContext, meta: unknown): ServerToolContext {
+  if (!isMultiRepositoryContext(baseContext)) return baseContext;
+  // Host metadata is request-scoped correlation only. The public Gateway carries
+  // it through its private forwarding envelope so Canonical Runtime does not lose
+  // the ChatGPT conversation scope during the proxy hop.
+  const hostConversationSessionId = chatgptHostSessionIdFromMcpMeta(meta);
+  if (!baseContext.runtimeSourceRoot) {
+    return hostConversationSessionId ? { ...baseContext, hostConversationSessionId } : baseContext;
+  }
+  const forwarded = canonicalRuntimeForwardingIdentity(meta);
+  return {
+    ...baseContext,
+    ...(forwarded.principalId ? { principalId: forwarded.principalId } : {}),
+    ...(forwarded.sessionId ? { sessionId: forwarded.sessionId } : {}),
+    ...(forwarded.controllerType ? { controllerType: forwarded.controllerType } : {}),
+    ...(forwarded.hostConversationSessionId ? { hostConversationSessionId: forwarded.hostConversationSessionId } : {}),
+  };
+}
+
+export interface CanonicalRuntimeProxyIdentity {
+  endpoint: URL;
+  token: string;
+  runtimeInstanceId: string;
+}
+
+function canonicalRuntimeProxyIdentity(ctx: MultiRepositoryMcpToolContext): CanonicalRuntimeProxyIdentity {
+  const snapshot = observeRuntimeStatus(ctx.controllerHome).snapshot;
+  if (!snapshot?.runtimeInstanceId) throw new Error('CANONICAL_RUNTIME_IDENTITY_UNAVAILABLE');
+  return {
+    endpoint: canonicalRuntimeEndpoint(snapshot.endpoint),
+    token: canonicalRuntimeToken(ctx),
+    runtimeInstanceId: snapshot.runtimeInstanceId,
+  };
+}
+
+export function sameCanonicalRuntimeProxyIdentity(
+  left: CanonicalRuntimeProxyIdentity,
+  right: CanonicalRuntimeProxyIdentity,
+): boolean {
+  return left.endpoint.href === right.endpoint.href
+    && left.token === right.token
+    && left.runtimeInstanceId === right.runtimeInstanceId;
+}
+
+export function canonicalRuntimeToolCallIsReplaySafe(name: string, args: Record<string, unknown>): boolean {
+  if (name !== 'repository_command_execute') return false;
+  const workId = typeof args.work_id === 'string' ? args.work_id.trim() : '';
+  const requestId = typeof args.request_id === 'string' ? args.request_id.trim() : '';
+  return Boolean(workId && requestId);
+}
+
+export function canonicalRuntimeToolCallFailureIsCutoverOutcomeUnknown(error: unknown): boolean {
+  if (!error || typeof error !== 'object') return false;
+  const record = error as { code?: unknown; message?: unknown; data?: unknown };
+  const message = typeof record.message === 'string' ? record.message : '';
+  if (SdkHttpError.isInstance(error)) {
+    return error.status === 409 && (
+      message.includes(RUNTIME_MCP_CUTOVER_OUTCOME_UNKNOWN)
+      || message.toLowerCase().includes('runtime cutover outcome unknown')
+      || JSON.stringify(record.data ?? '').includes(RUNTIME_MCP_CUTOVER_OUTCOME_UNKNOWN)
+    );
+  }
+  return record.code === -32000 && message === RUNTIME_MCP_CUTOVER_OUTCOME_UNKNOWN;
+}
+
+export function canonicalRuntimeToolCallFailureIsTransient(error: unknown): boolean {
+  if (SdkHttpError.isInstance(error)) return error.status === 404;
+  if (SdkError.isInstance(error)) {
+    return [
+      SdkErrorCode.NotConnected,
+      SdkErrorCode.ConnectionClosed,
+      SdkErrorCode.SendFailed,
+      SdkErrorCode.RequestTimeout,
+    ].includes(error.code);
+  }
+  const record = error && typeof error === 'object' ? error as { code?: unknown; cause?: unknown } : undefined;
+  const directCode = typeof record?.code === 'string' ? record.code : undefined;
+  const cause = record?.cause && typeof record.cause === 'object' ? record.cause as { code?: unknown } : undefined;
+  const causeCode = typeof cause?.code === 'string' ? cause.code : undefined;
+  return ['ECONNRESET', 'ECONNREFUSED', 'ETIMEDOUT', 'EPIPE', 'UND_ERR_SOCKET'].includes(directCode ?? causeCode ?? '');
+}
+
+export async function callCanonicalRuntimeToolWithReplay<T>(input: {
+  name: string;
+  args: Record<string, unknown>;
+  call: () => Promise<T>;
+  reconnect: () => Promise<void>;
+}): Promise<T> {
+  try {
+    return await input.call();
+  } catch (error) {
+    if (canonicalRuntimeToolCallFailureIsCutoverOutcomeUnknown(error)) {
+      throw new Error(`${RUNTIME_MCP_CUTOVER_OUTCOME_UNKNOWN}: Canonical Runtime entered cutover after request admission; reconcile the original request/effect identity before retrying.`);
+    }
+    if (!canonicalRuntimeToolCallIsReplaySafe(input.name, input.args)
+      || !canonicalRuntimeToolCallFailureIsTransient(error)) {
+      throw error;
+    }
+    await input.reconnect();
+    return await input.call();
+  }
+}
+
+export interface CanonicalRuntimeProxy {
+  listTools(): Promise<{ tools: Tool[] }>;
+  callTool(ctx: MultiRepositoryMcpToolContext, name: string, args: Record<string, unknown>, timing?: MutableGatewayProxyTiming): Promise<CallToolResult>;
+  close(): Promise<void>;
+}
+
+/**
+ * The public Gateway has no schema cache authority. It obtains each Connector
+ * session's schema from the Runtime that will execute its calls. When the HTTP
+ * Gateway owns a shared proxy, discovery and tool calls reuse the same loopback
+ * MCP session instead of paying an initialize/connect round trip per outer
+ * Connector session.
+ */
+export async function readCanonicalRuntimeToolSchema(
+  ctx: MultiRepositoryMcpToolContext,
+  sharedProxy?: CanonicalRuntimeProxy,
+): Promise<CanonicalRuntimeToolSchema> {
+  const proxy = sharedProxy ?? createCanonicalRuntimeProxy(ctx);
+  try {
+    const response = await proxy.listTools();
+    const definitions = response.tools.map(mcpToolDefinitionFromSdk);
+    const toolNames = definitions
+      .map((tool) => tool.name)
+      .filter((name): name is string => typeof name === 'string' && name.length > 0)
+      .sort();
+    return {
+      definitions,
+      toolNames,
+      fingerprint: forgeToolSurfaceFingerprint(definitions),
+    };
+  } finally {
+    if (!sharedProxy) await proxy.close();
+  }
+}
+
+type GatewayProxyTiming = Pick<McpTimingTrace,
+  | 'gatewayProxyResolveMs'
+  | 'gatewayProxyConnectMs'
+  | 'gatewayProxyCallMs'
+  | 'gatewayProxyCanonicalDispatchLagMs'
+  | 'gatewayProxyCanonicalDurationMs'
+  | 'gatewayProxyReturnMs'
+  | 'gatewayProxyConnectionState'
+>;
+
+type MutableGatewayProxyTiming = Partial<GatewayProxyTiming>;
+
+function roundedMs(value: number): number {
+  return Math.round(Math.max(0, value) * 100) / 100;
+}
+
+export function deriveCanonicalForwardingTiming(input: {
+  gatewayCallStartedAtMs: number;
+  gatewayCallDurationMs: number;
+  response: CallToolResult;
+}): Pick<GatewayProxyTiming, 'gatewayProxyCanonicalDispatchLagMs' | 'gatewayProxyCanonicalDurationMs' | 'gatewayProxyReturnMs'> {
+  const structured = input.response.structuredContent;
+  const responseMeta = structured && typeof structured === 'object' && !Array.isArray(structured)
+    && (structured as Record<string, unknown>).responseMeta
+    && typeof (structured as Record<string, unknown>).responseMeta === 'object'
+    ? (structured as Record<string, unknown>).responseMeta as Record<string, unknown>
+    : undefined;
+  const canonicalStartedAt = typeof responseMeta?.serverStartedAt === 'string'
+    ? Date.parse(responseMeta.serverStartedAt)
+    : Number.NaN;
+  const canonicalDurationMs = typeof responseMeta?.serverDurationMs === 'number'
+    ? responseMeta.serverDurationMs
+    : Number.NaN;
+  if (!Number.isFinite(canonicalStartedAt) || !Number.isFinite(canonicalDurationMs)) return {};
+  const dispatchLagMs = roundedMs(canonicalStartedAt - input.gatewayCallStartedAtMs);
+  const boundedCanonicalDurationMs = roundedMs(canonicalDurationMs);
+  return {
+    gatewayProxyCanonicalDispatchLagMs: dispatchLagMs,
+    gatewayProxyCanonicalDurationMs: boundedCanonicalDurationMs,
+    gatewayProxyReturnMs: roundedMs(input.gatewayCallDurationMs - dispatchLagMs - boundedCanonicalDurationMs),
+  };
+}
+
+export const DEFAULT_CANONICAL_RUNTIME_PROXY_LANES = 8;
+export const MAX_CANONICAL_RUNTIME_PROXY_LANES = 16;
+export const DEFAULT_CANONICAL_RUNTIME_PROXY_IDLE_TTL_MS = 30_000;
+
+export function canonicalRuntimeProxyLaneLimit(raw = process.env.FORGE_CANONICAL_RUNTIME_PROXY_LANES): number {
+  const parsed = Number(raw);
+  if (!Number.isInteger(parsed) || parsed < 1) return DEFAULT_CANONICAL_RUNTIME_PROXY_LANES;
+  return Math.min(parsed, MAX_CANONICAL_RUNTIME_PROXY_LANES);
+}
+
+export type CanonicalRuntimeLaneClass = 'interactive' | 'wait';
+
+export interface CanonicalRuntimeLaneScheduler {
+  acquire(laneClass?: CanonicalRuntimeLaneClass): Promise<number>;
+  release(laneId: number): void;
+  close(): void;
+  size(): number;
+}
+
+export const RESERVED_INTERACTIVE_CANONICAL_RUNTIME_PROXY_LANES = 2;
+
+/**
+ * Lease one inner Runtime MCP session exclusively to each in-flight request.
+ * Long process_wait calls are capped below total capacity so they can never
+ * consume every lane. Interactive waiters are handed released lanes first.
+ */
+export function createCanonicalRuntimeLaneScheduler(maxLanes = canonicalRuntimeProxyLaneLimit()): CanonicalRuntimeLaneScheduler {
+  const boundedMax = Math.max(1, Math.min(Math.trunc(maxLanes), MAX_CANONICAL_RUNTIME_PROXY_LANES));
+  const maxWaitLanes = boundedMax <= RESERVED_INTERACTIVE_CANONICAL_RUNTIME_PROXY_LANES
+    ? Math.max(1, boundedMax - 1)
+    : boundedMax - RESERVED_INTERACTIVE_CANONICAL_RUNTIME_PROXY_LANES;
+  const idle: number[] = [];
+  const waiters: Array<{ laneClass: CanonicalRuntimeLaneClass; resolve: (laneId: number) => void; reject: (error: Error) => void }> = [];
+  const activeClasses = new Map<number, CanonicalRuntimeLaneClass>();
+  let waitLaneCount = 0;
+  let laneCount = 0;
+  let closed = false;
+
+  const activate = (laneId: number, laneClass: CanonicalRuntimeLaneClass): number => {
+    activeClasses.set(laneId, laneClass);
+    if (laneClass === 'wait') waitLaneCount += 1;
+    return laneId;
+  };
+
+  const nextEligibleWaiterIndex = (): number => {
+    const interactiveIndex = waiters.findIndex((waiter) => waiter.laneClass === 'interactive');
+    if (interactiveIndex >= 0) return interactiveIndex;
+    return waitLaneCount < maxWaitLanes ? waiters.findIndex((waiter) => waiter.laneClass === 'wait') : -1;
+  };
+
+  return {
+    async acquire(laneClass: CanonicalRuntimeLaneClass = 'interactive'): Promise<number> {
+      if (closed) throw new Error('CANONICAL_RUNTIME_PROXY_CLOSED');
+      if (laneClass === 'wait' && waitLaneCount >= maxWaitLanes) {
+        return await new Promise<number>((resolve, reject) => waiters.push({ laneClass, resolve, reject }));
+      }
+      const available = idle.shift();
+      if (available !== undefined) return activate(available, laneClass);
+      if (laneCount < boundedMax) return activate(laneCount++, laneClass);
+      return await new Promise<number>((resolve, reject) => waiters.push({ laneClass, resolve, reject }));
+    },
+    release(laneId: number): void {
+      if (closed) return;
+      const previousClass = activeClasses.get(laneId);
+      activeClasses.delete(laneId);
+      if (previousClass === 'wait') waitLaneCount = Math.max(0, waitLaneCount - 1);
+      const waiterIndex = nextEligibleWaiterIndex();
+      if (waiterIndex >= 0) {
+        const [waiter] = waiters.splice(waiterIndex, 1);
+        waiter.resolve(activate(laneId, waiter.laneClass));
+        return;
+      }
+      idle.push(laneId);
+    },
+    close(): void {
+      if (closed) return;
+      closed = true;
+      const error = new Error('CANONICAL_RUNTIME_PROXY_CLOSED');
+      for (const waiter of waiters.splice(0)) waiter.reject(error);
+      activeClasses.clear();
+      idle.length = 0;
+    },
+    size(): number {
+      return laneCount;
+    },
+  };
+}
+
+interface CanonicalRuntimeProxyConnection {
+  identity: CanonicalRuntimeProxyIdentity;
+  client: Client;
+  transport: StreamableHTTPClientTransport;
+}
+
+interface CanonicalRuntimeProxyLane {
+  current?: CanonicalRuntimeProxyConnection;
+  connecting?: Promise<CanonicalRuntimeProxyConnection>;
+  idleCloseTimer?: ReturnType<typeof setTimeout>;
+  leased: boolean;
+  leaseGeneration: number;
+}
+
+export interface CanonicalRuntimeProxyOptions {
+  /** Keep burst-created inner MCP sessions hot briefly, then release their retained Runtime graph. */
+  idleTtlMs?: number;
+}
+
+export function createCanonicalRuntimeProxy(
+  ctx: MultiRepositoryMcpToolContext,
+  options: CanonicalRuntimeProxyOptions = {},
+): CanonicalRuntimeProxy {
+  const scheduler = createCanonicalRuntimeLaneScheduler();
+  const lanes = new Map<number, CanonicalRuntimeProxyLane>();
+  const idleTtlMs = Math.max(1, Math.trunc(options.idleTtlMs ?? DEFAULT_CANONICAL_RUNTIME_PROXY_IDLE_TTL_MS));
+  let closed = false;
+
+  const laneState = (laneId: number): CanonicalRuntimeProxyLane => {
+    const existing = lanes.get(laneId);
+    if (existing) return existing;
+    const created: CanonicalRuntimeProxyLane = { leased: false, leaseGeneration: 0 };
+    lanes.set(laneId, created);
+    return created;
+  };
+
+  const clearLaneIdleTimer = (lane: CanonicalRuntimeProxyLane): void => {
+    if (!lane.idleCloseTimer) return;
+    clearTimeout(lane.idleCloseTimer);
+    lane.idleCloseTimer = undefined;
+  };
+
+  const closeLane = async (laneId: number, expectedClient?: Client): Promise<void> => {
+    const lane = lanes.get(laneId);
+    if (!lane) return;
+    if (lane.connecting) {
+      await lane.connecting.catch(() => undefined);
+      lane.connecting = undefined;
+    }
+    if (!lane.current || (expectedClient && lane.current.client !== expectedClient)) return;
+    clearLaneIdleTimer(lane);
+    const closing = lane.current;
+    lane.current = undefined;
+    // Client.close() only tears down the local SDK client. Explicitly terminate
+    // the Streamable HTTP session so the canonical Runtime drops its Server and
+    // transport graph instead of retaining an orphaned server-side session.
+    await closing.transport.terminateSession().catch(() => undefined);
+    await closing.client.close().catch(() => undefined);
+  };
+
+  const leaseLane = (laneId: number): void => {
+    const lane = laneState(laneId);
+    clearLaneIdleTimer(lane);
+    lane.leased = true;
+    lane.leaseGeneration += 1;
+  };
+
+  const releaseLane = (laneId: number): void => {
+    const lane = laneState(laneId);
+    lane.leased = false;
+    const releasedGeneration = lane.leaseGeneration;
+    scheduler.release(laneId);
+    if (closed || !lane.current) return;
+    clearLaneIdleTimer(lane);
+    lane.idleCloseTimer = setTimeout(() => {
+      const current = lanes.get(laneId);
+      if (!current || closed || current.leased || current.leaseGeneration !== releasedGeneration) return;
+      current.idleCloseTimer = undefined;
+      void closeLane(laneId);
+    }, idleTtlMs);
+    lane.idleCloseTimer.unref?.();
+  };
+
+  const closeAllLanes = async (): Promise<void> => {
+    await Promise.all(Array.from(lanes.keys()).map(async (laneId) => await closeLane(laneId)));
+  };
+
+  const connect = async (identity: CanonicalRuntimeProxyIdentity): Promise<CanonicalRuntimeProxyConnection> => {
+    const connectOnce = async (): Promise<CanonicalRuntimeProxyConnection> => {
+      const abort = new AbortController();
+      const timeout = setTimeout(() => abort.abort(new Error('CANONICAL_RUNTIME_TIMEOUT')), CANONICAL_RUNTIME_CONNECT_TIMEOUT_MS);
+      const headers: Record<string, string> = {
+        Authorization: `Bearer ${identity.token}`,
+      };
+      const transport = new StreamableHTTPClientTransport(identity.endpoint, {
+        requestInit: { headers, signal: abort.signal },
+      });
+      const client = new Client({ name: 'forge-public-gateway-proxy', version: '1.0.0' });
+      try {
+        await client.connect(transport);
+        return { identity, client, transport };
+      } catch (error) {
+        await client.close().catch(() => undefined);
+        throw error;
+      } finally {
+        clearTimeout(timeout);
+      }
+    };
+    return await retryCanonicalRuntimeConnectDuringHandoff(
+      connectOnce,
+      () => canonicalRuntimeReleaseHandoffActive(ctx.controllerHome),
+    );
+  };
+
+  const clientForCurrentRuntime = async (laneId: number, timing: MutableGatewayProxyTiming): Promise<Client> => {
+    const lane = laneState(laneId);
+    const resolveStarted = performance.now();
+    const identity = canonicalRuntimeProxyIdentity(ctx);
+    timing.gatewayProxyResolveMs = roundedMs(performance.now() - resolveStarted);
+    if (lane.current && sameCanonicalRuntimeProxyIdentity(lane.current.identity, identity)) {
+      timing.gatewayProxyConnectionState = 'reused';
+      timing.gatewayProxyConnectMs = 0;
+      return lane.current.client;
+    }
+
+    if (lane.connecting) {
+      const connectStarted = performance.now();
+      const pending = await lane.connecting;
+      timing.gatewayProxyConnectMs = roundedMs(performance.now() - connectStarted);
+      if (sameCanonicalRuntimeProxyIdentity(pending.identity, identity)) {
+        timing.gatewayProxyConnectionState = 'coalesced_connect';
+        return pending.client;
+      }
+    }
+
+    const reconnect = Boolean(lane.current);
+    await closeLane(laneId);
+    const connectStarted = performance.now();
+    lane.connecting = connect(identity);
+    try {
+      lane.current = await lane.connecting;
+      timing.gatewayProxyConnectMs = roundedMs(performance.now() - connectStarted);
+      timing.gatewayProxyConnectionState = reconnect ? 'identity_reconnect' : 'cold_connect';
+      return lane.current.client;
+    } finally {
+      lane.connecting = undefined;
+    }
+  };
+
+  const prepareLane = async (
+    timing: MutableGatewayProxyTiming,
+    laneClass: CanonicalRuntimeLaneClass = 'interactive',
+  ): Promise<{ laneId: number; client: Client }> => {
+    if (closed) throw new Error('CANONICAL_RUNTIME_PROXY_CLOSED');
+    const handoff = await waitForCanonicalRuntimeReleaseHandoff(
+      () => canonicalRuntimeReleaseHandoffActive(ctx.controllerHome),
+    );
+    if (handoff.waited) await closeAllLanes();
+    const laneId = await scheduler.acquire(laneClass);
+    leaseLane(laneId);
+    try {
+      return { laneId, client: await clientForCurrentRuntime(laneId, timing) };
+    } catch (error) {
+      releaseLane(laneId);
+      throw error;
+    }
+  };
+
+  return {
+    async listTools() {
+      const { laneId, client } = await prepareLane({});
+      try {
+        return await client.listTools() as { tools: Tool[] };
+      } catch (error) {
+        await closeLane(laneId, client);
+        throw error;
+      } finally {
+        releaseLane(laneId);
+      }
+    },
+    async callTool(callerContext, name, args, timing = {}) {
+      const prepared = await prepareLane(timing, name === 'process_wait' ? 'wait' : 'interactive');
+      const laneId = prepared.laneId;
+      let activeClient = prepared.client;
+      const gatewayCallStartedAtMs = Date.now();
+      const callStarted = performance.now();
+      const forwardedRequest = {
+        name,
+        arguments: args,
+        _meta: canonicalRuntimeForwardingMeta(callerContext),
+      } as Parameters<Client['callTool']>[0];
+      try {
+        // The outer Connector session is already bound to the Canonical Runtime
+        // schema and fenced by the Runtime-published fingerprint before dispatch.
+        // Caller identity stays request-scoped while the inner MCP lane is reused.
+        // Only keyed Work-attributed Process Runtime commands may replay after a
+        // transient inner disconnect: the stable request_id resolves to the
+        // original Process instead of spawning the side effect twice.
+        const response = await callCanonicalRuntimeToolWithReplay({
+          name,
+          args,
+          call: async () => await activeClient.callTool(
+            forwardedRequest,
+            { timeout: CANONICAL_RUNTIME_TOOL_CALL_TIMEOUT_MS },
+          ),
+          reconnect: async () => {
+            await closeLane(laneId, activeClient);
+            activeClient = await clientForCurrentRuntime(laneId, timing);
+          },
+        });
+        const result = response as unknown as CallToolResult;
+        timing.gatewayProxyCallMs = roundedMs(performance.now() - callStarted);
+        Object.assign(timing, deriveCanonicalForwardingTiming({
+          gatewayCallStartedAtMs,
+          gatewayCallDurationMs: timing.gatewayProxyCallMs,
+          response: result,
+        }));
+        return result;
+      } catch (error) {
+        timing.gatewayProxyCallMs = roundedMs(performance.now() - callStarted);
+        // Poison only the failed lane. Other hot lanes remain available so one
+        // Runtime/session failure cannot collapse the whole Gateway proxy pool.
+        await closeLane(laneId, activeClient);
+        throw error;
+      } finally {
+        releaseLane(laneId);
+      }
+    },
+    async close() {
+      if (closed) return;
+      closed = true;
+      scheduler.close();
+      await closeAllLanes();
+    },
+  };
+}
+
+function canonicalRuntimeSchemaMatchesGateway(ctx: MultiRepositoryMcpToolContext): boolean {
+  const runtimeFingerprint = observeRuntimeStatus(ctx.controllerHome).snapshot?.toolSurfaceFingerprint;
+  // Older Runtime status records do not have a schema fence. Preserve a
+  // diagnosable compatibility path until the Runtime has been upgraded.
+  return !runtimeFingerprint || runtimeFingerprint === controllerExposureSnapshot(ctx).fingerprint;
+}
+
+function toolSurfaceMismatchResult(): CallToolResult {
+  const value = {
+    error: {
+      code: 'MCP_TOOL_SURFACE_MISMATCH',
+      message: 'Gateway and Canonical Runtime tool surfaces differ. The session must reinitialize against a matching Runtime release.',
+      recoverable: true,
+      action: 'reinitialize',
+    },
+  };
+  return { content: [{ type: 'text', text: JSON.stringify(value) }], structuredContent: value, isError: true };
+}
+
+export function createForgeMcpServerFromContext(
+  baseContext: ServerToolContext,
+  runtimeSchema?: CanonicalRuntimeToolSchema,
+  sharedRuntimeProxy?: CanonicalRuntimeProxy,
+): Server {
+  const server = new Server(
+    { name: 'forge-mcp', version: FORGE_VERSION },
+    { capabilities: { tools: { listChanged: true } }, instructions: mcpServerInstructions(baseContext.policy.profile) },
+  );
+  const ownsRuntimeProxy = isMultiRepositoryContext(baseContext) && !getRuntimeWriteClaim() && !sharedRuntimeProxy;
+  const runtimeProxy = isMultiRepositoryContext(baseContext) && !getRuntimeWriteClaim()
+    ? sharedRuntimeProxy ?? createCanonicalRuntimeProxy(baseContext)
+    : undefined;
+  server.onclose = () => {
+    if (ownsRuntimeProxy) void runtimeProxy?.close();
+  };
+  // The connector process can outlive the canonical Runtime release it proxies.
+  // Refresh tools after every MCP session initialization so clients do not keep
+  // a stale schema when the bounded Runtime surface changes behind the gateway.
+  server.oninitialized = () => {
+    void server.sendToolListChanged().catch(() => undefined);
+  };
+  server.setRequestHandler('tools/list', async () => {
+    if (isMultiRepositoryContext(baseContext) && !runtimeSchema && !getRuntimeWriteClaim() && !canonicalRuntimeSchemaMatchesGateway(baseContext)) {
+      throw new Error('MCP_TOOL_SURFACE_MISMATCH: Gateway source does not match the Canonical Runtime schema.');
+    }
+    const definitions = runtimeSchema?.definitions ?? (isMultiRepositoryContext(baseContext)
+      ? controllerExposureSnapshot(baseContext).definitions
+      : buildMcpToolDefinitions(baseContext.policy, { enableChatgptBrowser: baseContext.enableChatgptBrowser === true }));
+    return { tools: definitions.map(mcpToolDefinitionToSdk) };
+  });
+  server.setRequestHandler('tools/call', async (request, handlerContext) => {
+    const name = request.params.name;
+    const args = (request.params.arguments ?? {}) as Record<string, unknown>;
+    const forwardedBaseContext = canonicalRuntimeRequestContext(
+      baseContext,
+      (request.params as { _meta?: unknown })._meta,
+    );
+    const ctx: ServerToolContext = { ...forwardedBaseContext, signal: handlerContext.mcpReq.signal };
+    if (isMultiRepositoryContext(ctx)) {
+      const rpcId = handlerContext.mcpReq.id;
+      return traceControllerMcpRequest(ctx, name, args, typeof rpcId === 'string' || typeof rpcId === 'number' ? rpcId : undefined, async (requestId, _traceId, phaseTimings) => {
+        // The public Gateway only exposes its stable facade schema. It may
+        // proxy execution, but it never discovers one Runtime schema and
+        // validates calls against another source checkout.
+        if (!getRuntimeWriteClaim() && !runtimeSchema && !canonicalRuntimeSchemaMatchesGateway(ctx)) {
+          return toolSurfaceMismatchResult();
+        }
+        if (runtimeSchema && !runtimeSchema.toolNames.includes(name)) {
+          return toolSurfaceMismatchResult();
+        }
+        if (!runtimeSchema && !isControllerToolExposed(ctx, name)) {
+          const value = {
+            error: {
+              code: 'TOOL_NOT_FOUND',
+              message: `${name} is not registered by this forge build. Tool availability is independent of Request vs Full Access.`,
+            },
+          };
+          return { content: [{ type: 'text', text: JSON.stringify(value, null, 2) }], structuredContent: value, isError: true };
+        }
+        // The public OAuth Gateway is intentionally thin. If this process does
+        // not own the Canonical Runtime write claim, forward the complete tool
+        // call to the loopback Runtime MCP endpoint. This prevents source-checkout
+        // Gateway processes from acquiring Process Runtime leases or evaluating
+        // Runtime source coherence against their own checkout.
+        if (!getRuntimeWriteClaim()) {
+          // Keep trace correlation separate from Tool Contract idempotency. Only
+          // a caller-supplied request_id may authorize semantic replay/dedup.
+          const forwardedArgs = args;
+          if (runtimeProxy && runtimeSchema) return runtimeProxy.callTool(ctx, name, forwardedArgs, phaseTimings);
+          if (runtimeProxy && observeRuntimeStatus(ctx.controllerHome).ready) return runtimeProxy.callTool(ctx, name, forwardedArgs, phaseTimings);
+        }
+        const executeLocalTool = async (): Promise<CallToolResult> => {
+          const accessResult = callAccessTool(ctx, name, args);
+          if (accessResult) return accessResult;
+          const executionResult = await callExecutionTool(ctx, name, args);
+          if (executionResult) return executionResult;
+          const processResult = await callProcessTool(ctx, name, args);
+          if (processResult) return processResult;
+          if (isGatewayIsolatedTool(name)) {
+            const isolatedResult = await routeDurableMcpCall(ctx, name, args, { allowReadOnly: true, forceDurable: true });
+            if (isolatedResult) return isolatedResult;
+          }
+          const runtimeResult = await callRuntimeTool(ctx, name, args);
+          if (runtimeResult) return runtimeResult;
+          const durableResult = await routeDurableMcpCall(ctx, name, args);
+          if (durableResult) return durableResult;
+          const repositoryResult = await callRepositoryToolWithPostFinalizeAttribution(ctx.controllerHome, name, args, ctx);
+          if (repositoryResult) return repositoryResult;
+          return await callMultiRepositoryTool(ctx, name, args);
+        };
+        const localResult = await executeLocalTool();
+        return settleCognitionAfterTool(ctx, name, args, localResult) ?? localResult;
+      });
+    }
+    return callMcpTool(ctx, name, args);
+  });
+  return server;
+}
+
+export function createForgeMcpServer(opts: McpServerOptions): Server {
+  return createForgeMcpServerFromContext(createMcpToolContext(opts));
+}