improvement(mship): mcp persistence in chat

Author: Sg312

apps/sim/lib/copilot/chat/payload.ts

@@ -34,7 +34,10 @@ interface BuildPayloadParams {
   model: string
   provider?: string
   contexts?: Array<{ type: string; content: string; tag?: string; path?: string }>
-  /** MCP servers explicitly tagged on this turn. Untagged servers stay unavailable. */
+  /**
+   * MCP servers enabled for this chat — every server tagged on this or any
+   * earlier turn. Servers never tagged in the chat stay unavailable.
+   */
   mcpServerIds?: string[]
   fileAttachments?: Array<{ id: string; key: string; size: number; [key: string]: unknown }>
   commands?: string[]

apps/sim/lib/copilot/chat/post.test.ts

@@ -289,6 +289,81 @@ describe('handleUnifiedChatPost', () => {
     )
   })
 
+  it('keeps MCP servers tagged on earlier turns enabled for the rest of the chat', async () => {
+    resolveOrCreateChat.mockResolvedValue({
+      chatId: 'chat-1',
+      chat: { id: 'chat-1' },
+      conversationHistory: [
+        {
+          id: 'msg-1',
+          role: 'user',
+          content: '/Docs search auth',
+          contexts: [{ kind: 'mcp', serverId: 'mcp-server-1', label: 'Docs' }],
+        },
+        { id: 'msg-2', role: 'assistant', content: 'here you go' },
+      ],
+      isNew: false,
+    })
+
+    const response = await handleUnifiedChatPost(
+      new NextRequest('http://localhost/api/copilot/chat', {
+        method: 'POST',
+        body: JSON.stringify({
+          message: 'now search billing',
+          workspaceId: 'ws-1',
+          chatId: 'chat-1',
+        }),
+      })
+    )
+
+    expect(response.status).toBe(200)
+    expect(buildCopilotRequestPayload).toHaveBeenCalledWith(
+      expect.objectContaining({ mcpServerIds: ['mcp-server-1'] }),
+      { selectedModel: '' }
+    )
+    // The tools ride the tool array every turn, so re-expanding the listing for
+    // an inherited server would only duplicate what the model already sees.
+    const expandedContexts = processContextsServer.mock.calls[0]?.[0] ?? []
+    expect(expandedContexts).not.toContainEqual(expect.objectContaining({ kind: 'mcp' }))
+  })
+
+  it('unions MCP servers across turns without duplicating a re-tagged server', async () => {
+    resolveOrCreateChat.mockResolvedValue({
+      chatId: 'chat-1',
+      chat: { id: 'chat-1' },
+      conversationHistory: [
+        {
+          id: 'msg-1',
+          role: 'user',
+          content: '/Docs search auth',
+          contexts: [{ kind: 'mcp', serverId: 'mcp-server-1', label: 'Docs' }],
+        },
+      ],
+      isNew: false,
+    })
+
+    const response = await handleUnifiedChatPost(
+      new NextRequest('http://localhost/api/copilot/chat', {
+        method: 'POST',
+        body: JSON.stringify({
+          message: '/Docs /Issues cross-reference',
+          workspaceId: 'ws-1',
+          chatId: 'chat-1',
+          contexts: [
+            { kind: 'mcp', serverId: 'mcp-server-1', label: 'Docs' },
+            { kind: 'mcp', serverId: 'mcp-server-2', label: 'Issues' },
+          ],
+        }),
+      })
+    )
+
+    expect(response.status).toBe(200)
+    expect(buildCopilotRequestPayload).toHaveBeenCalledWith(
+      expect.objectContaining({ mcpServerIds: ['mcp-server-1', 'mcp-server-2'] }),
+      { selectedModel: '' }
+    )
+  })
+
   it('persists cancelled partial responses from the server lifecycle', async () => {
     await handleUnifiedChatPost(
       new NextRequest('http://localhost/api/copilot/chat', {

apps/sim/lib/copilot/chat/post.ts

@@ -246,6 +246,44 @@ function normalizeContexts(contexts: UnifiedChatRequest['contexts']) {
   })
 }
 
+/**
+ * An MCP server tagged with `/name` stays enabled for the rest of the chat, not
+ * just the turn it was tagged on. Persisted user messages already carry their
+ * `mcp` contexts, so the transcript is the source of truth — enablement survives
+ * reloads and reopened chats with no extra state to keep in sync. There is
+ * deliberately no off switch: history is append-only.
+ *
+ * Only the ids travel forward, not the contexts themselves. The tools ride the
+ * tool array on every turn, so the model always sees their names and schemas;
+ * re-expanding the prompt listing each turn would just duplicate that. Keeping
+ * inherited servers out of the persisted contexts also keeps the `/name` chips
+ * on a sent message showing only what the user actually typed that turn.
+ */
+function collectChatMcpServerIds(
+  conversationHistory: unknown[],
+  currentContexts: UnifiedChatRequest['contexts']
+): string[] {
+  const serverIds = new Set<string>()
+
+  const collect = (contexts: unknown) => {
+    if (!Array.isArray(contexts)) return
+    for (const ctx of contexts) {
+      if (!ctx || typeof ctx !== 'object') continue
+      const { kind, serverId } = ctx as { kind?: unknown; serverId?: unknown }
+      if (kind === 'mcp' && typeof serverId === 'string' && serverId) {
+        serverIds.add(serverId)
+      }
+    }
+  }
+
+  for (const message of conversationHistory) {
+    collect((message as { contexts?: unknown } | null)?.contexts)
+  }
+  collect(currentContexts)
+
+  return Array.from(serverIds)
+}
+
 async function resolveAgentContexts(params: {
   contexts?: UnifiedChatRequest['contexts']
   resourceAttachments?: UnifiedChatRequest['resourceAttachments']
@@ -1003,9 +1041,7 @@ export async function handleUnifiedChatPost(req: NextRequest) {
           [TraceAttr.CopilotContextsCount]: normalizedContexts.length,
         },
         () => {
-          const mcpServerIds = normalizedContexts.flatMap((context) =>
-            context.kind === 'mcp' && context.serverId ? [context.serverId] : []
-          )
+          const mcpServerIds = collectChatMcpServerIds(conversationHistory, normalizedContexts)
           return branch.kind === 'workflow'
             ? branch.buildPayload({
                 message: body.message,

apps/sim/lib/copilot/chat/process-contents.ts

@@ -94,9 +94,9 @@ export async function processContextsServer(
           type: 'mcp',
           tag: ctx.label ? `/${ctx.label}` : '/',
           content: [
-            `The user explicitly enabled the MCP server "${ctx.label || ctx.serverId}" for this turn.`,
-            'Its request-scoped tools are listed below. Load a tool with load_custom_tool({ type: "mcp", name: "<exact name>" }) before calling it.',
-            'Do not narrate discovery, loading, tool-name selection, or retries. Call the tool first, then respond once with the result. Never claim the server works before a successful tool result. Do not automatically retry a timed-out or abandoned MCP call.',
+            `The user explicitly enabled the MCP server "${ctx.label || ctx.serverId}". It stays enabled for the rest of this chat, and its tools remain callable on every later turn.`,
+            'Its tools are listed below and are callable directly by the exact name shown — there is no loading step.',
+            'Do not narrate discovery, tool-name selection, or retries. Call the tool first, then respond once with the result. Never claim the server works before a successful tool result. Do not automatically retry a timed-out or abandoned MCP call.',
             ...toolLines,
           ].join('\n'),
         }

apps/sim/lib/copilot/mcp-tools.test.ts

@@ -37,7 +37,8 @@ describe('mothership MCP tool schemas', () => {
     expect(tools).toEqual([
       expect.objectContaining({
         name: 'mcp-server-1-search',
-        defer_loading: true,
+        // Explicitly enabled by the user, so it is callable without an unlock step.
+        defer_loading: false,
         executeLocally: false,
         params: expect.objectContaining({
           mothershipToolKind: 'mcp',

apps/sim/lib/copilot/mcp-tools.ts

@@ -21,7 +21,12 @@ function toMothershipMcpTool(tool: {
     description:
       tool.description || `MCP tool ${tool.name} from ${tool.serverName || tool.serverId}`,
     input_schema: tool.inputSchema,
-    defer_loading: true,
+    // Not deferred: deferral keeps the 200+ unrequested integration schemas out
+    // of the tool array, but an MCP server is explicitly enabled by the user and
+    // stays enabled for the chat. Deferring it means every turn after the first
+    // needs a load_custom_tool round-trip the model has no listing to prompt it
+    // for, and a direct call is rejected as unavailable.
+    defer_loading: false,
     executeLocally: false,
     service: `mcp:${tool.serverId}`,
     params: {