fix(realtime): tear down chat streams inside the request's workspace scope

apps/web/src/routes/api/chat/__tests__/stream.test.ts

@@ -473,6 +473,42 @@ describe('GET /api/chat/stream - assistant activity snapshot replay', () => {
   })
 })
 
+describe('GET /api/chat/stream - teardown keeps the workspace scope', () => {
+  it('clears presence inside the request workspace scope when the client disconnects', async () => {
+    const { createWorkspaceScope, getWorkspaceScope, runWithWorkspaceScope } =
+      await import('@/lib/server/workspaces/workspace-context')
+    const scope = createWorkspaceScope({
+      workspace: { workspaceKey: 'inst_stream' },
+      db: {},
+      sql: {},
+      origin: 'request',
+      secrets: { secretKey: 'd'.repeat(64), storage: null, storageProblem: 'not read here' },
+    } as never)
+    tokenPrincipal('member')
+    let scopeSeenByClear: string | null | undefined
+    mockClearPresence.mockImplementation(async () => {
+      scopeSeenByClear = getWorkspaceScope()?.workspace.workspaceKey ?? null
+      return false
+    })
+
+    // Opened inside a workspace scope, exactly as the middleware would run it…
+    const controller = new AbortController()
+    const request = new Request('http://test/api/chat/stream?scope=inbox', {
+      signal: controller.signal,
+    })
+    const res = await runWithWorkspaceScope(scope, () => GET({ request }))
+    expect(res.status).toBe(200)
+    await vi.waitFor(() => expect(mockMarkPresent).toHaveBeenCalled())
+
+    // …and aborted from outside it, which is where the runtime fires the
+    // signal. Teardown must not see an empty scope.
+    expect(getWorkspaceScope()).toBeNull()
+    controller.abort()
+    await vi.waitFor(() => expect(mockClearPresence).toHaveBeenCalled())
+    expect(scopeSeenByClear).toBe('inst_stream')
+  })
+})
+
 describe('GET /api/chat/stream - abandoned heartbeat timeout', () => {
   it('stops polling presence and unsubscribes when pings go unconsumed', async () => {
     vi.useFakeTimers()

apps/web/src/routes/api/chat/stream.ts

@@ -40,6 +40,7 @@ import { createSseStream, SSE_RESPONSE_HEADERS } from '@/lib/server/utils/sse'
 import { streamLimiter } from '@/lib/server/realtime/stream-connection-limit'
 import { startStreamHeartbeat } from '@/lib/server/realtime/stream-heartbeat'
 import { getClientIp } from '@/lib/server/domains/api/rate-limit'
+import { getWorkspaceScope, runWithWorkspaceScope } from '@/lib/server/workspaces/workspace-context'
 import { logger } from '@/lib/server/logger'
 
 const log = logger.child({ component: 'chat-stream' })
@@ -224,7 +225,13 @@ export const Route = createFileRoute('/api/chat/stream')({
         let heartbeat: { stop: () => void } | null = null
         let unsubscribe: (() => Promise<void>) | null = null
 
-        cleanup = async () => {
+        // The abort listener and the stream's cancel hook are invoked by the
+        // runtime, outside the request's AsyncLocalStorage context, so under
+        // pooled tenancy `clearPresence` (and the requeue behind it) would run
+        // with no workspace scope and throw. Capture the scope now and re-enter
+        // it for teardown — the same pattern the auth stash sweeps use.
+        const workspaceScope = getWorkspaceScope()
+        const teardown = async () => {
           if (cleanedUp) return
           cleanedUp = true
           heartbeat?.stop()
@@ -251,6 +258,8 @@ export const Route = createFileRoute('/api/chat/stream')({
           }
           slot.release()
         }
+        cleanup = () =>
+          workspaceScope ? runWithWorkspaceScope(workspaceScope, teardown) : teardown()
 
         const run = async () => {
           try {