Files
supabase/apps/studio/state/ai-assistant-state.notebook-cache-invalidation.test.ts
Charis de3a8799d6 fix(studio): invalidate notebook caches after assistant create/update (#49415)
## Summary
- The assistant's `create_notebook`/`update_notebook` tools run entirely
server-side, so an open notebook tab's React Query cache and Valtio
store never learn a write happened — the tab keeps showing stale content
until a manual reload.
- Adds `collectNotebookCacheEffects`/`applyNotebookCacheEffects`
(`apps/studio/lib/ai/notebook-cache-invalidation.ts`), which scan
finished assistant messages for completed
`create_notebook`/`update_notebook` tool calls and evict the affected
notebook via `evictNotebookFromCaches`
(`apps/studio/data/content/notebooks/notebook-cache.ts`), plus
invalidate the nav list.
- Wired into `createChatInstance`'s `onFinish` in
`state/ai-assistant-state.tsx`, with per-chat dedupe so replayed history
isn't reprocessed.
- Removes the cache entry outright rather than invalidating it, since a
remounting `useNotebookQuery` would otherwise read the stale cached
value synchronously before its refetch lands.
- Explicitly skips eviction when the open tab has unsaved local edits,
so an assistant write can't silently discard them.

Related: [FE-4235](https://linear.app/supabase/issue/FE-4235)

**Out of scope:** this only protects the client-side cache/store from
being clobbered after the fact. Preventing the assistant's
`update_notebook` tool call itself from overwriting a user's unsaved
edits (a data-layer conflict, not a cache-freshness one) is tracked
separately in [FE-4255](https://linear.app/supabase/issue/FE-4255).

## Test plan
- [x] `pnpm test:studio -- notebook-cache notebook-cache-invalidation
ai-assistant-state.notebook-cache-invalidation
ExplorerNotebookTab.assistant-cache-invalidation
ExplorerNotebookTabCoordinator` — all passing
- [x] Reproduction-first component test
(`ExplorerNotebookTab.assistant-cache-invalidation.test.tsx`) — verified
it fails without the fix (stale content persists) and passes with it
- [x] Regression test for the dirty-notebook guard (an edited, unsaved
notebook is left untouched by an assistant write)
- [x] `pnpm typecheck --filter=studio` / `pnpm lint --filter=studio`
clean


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Bug Fixes**
* Notebook changes made through the AI assistant now appear correctly in
open notebook tabs and after reopening them.
* Saved notebook caches are refreshed after completed create or update
actions, preventing stale content from being displayed.
  * Unsaved notebook changes are preserved during cache cleanup.
  * Closing a notebook tab now consistently removes its cached content.

* **Tests**
* Added coverage for assistant-driven updates, remounts, duplicate
actions, project context changes, and cache behavior.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
2026-08-25 14:15:27 +08:00

182 lines
6.2 KiB
TypeScript

import type { UIMessageChunk } from 'ai'
import { convertArrayToReadableStream } from 'ai/test'
import { afterEach, describe, expect, it, vi } from 'vitest'
import { contentKeys } from '@/data/content/keys'
import { getQueryClient } from '@/data/query-client'
import { notebooksState } from '@/state/notebooks/notebooks-state'
import type { Notebook } from '@/state/notebooks/types'
const testContext = vi.hoisted(() => ({
queuedStreams: [] as Array<Array<UIMessageChunk>>,
onSend: undefined as (() => void) | undefined,
}))
vi.mock('ai', async (importOriginal) => {
const actual = await importOriginal<typeof import('ai')>()
return {
...actual,
DefaultChatTransport: class {
constructor(private options: any = {}) {}
async sendMessages(opts: any) {
await this.options.prepareSendMessagesRequest?.({
api: this.options.api,
id: opts.chatId,
messages: opts.messages,
body: { ...this.options.body, ...opts.body },
headers: {},
credentials: undefined,
requestMetadata: opts.metadata,
trigger: opts.trigger,
messageId: opts.messageId,
})
testContext.onSend?.()
const chunks = testContext.queuedStreams.shift()
if (!chunks) throw new Error('No queued stream for this sendMessages call')
return convertArrayToReadableStream(chunks)
}
async reconnectToStream() {
return null
}
},
}
})
const { createAiAssistantState } = await import('./ai-assistant-state')
const PROJECT_REF = 'default'
const NOTEBOOK_ID = 'notebook-onfinish-test'
const seedNotebook = () => {
delete notebooksState.notebooks[NOTEBOOK_ID]
const notebook: Notebook = {
id: NOTEBOOK_ID,
type: 'notebook',
name: 'Test notebook',
visibility: 'project',
favorite: false,
owner_id: 1,
project_id: 1,
content: {
schema_version: 1,
cells: [{ _tag: 'markdown_cell', _id: 'cell-1', text: 'Original' }],
},
}
notebooksState.setNotebook({ projectRef: PROJECT_REF, notebook })
}
afterEach(() => {
delete notebooksState.notebooks[NOTEBOOK_ID]
notebooksState.needsSaving.clear()
testContext.queuedStreams = []
testContext.onSend = undefined
})
describe('createChatInstance onFinish — notebook cache invalidation via a real Chat/stream', () => {
it('evicts the notebook from caches once update_notebook reaches output-available, after an approval round trip', async () => {
seedNotebook()
const queryClient = getQueryClient()
queryClient.setQueryData(contentKeys.resource(PROJECT_REF, NOTEBOOK_ID), { id: NOTEBOOK_ID })
const state = createAiAssistantState()
state.setContext({ projectRef: PROJECT_REF })
const chatId = state.createChat({ name: 'Delete a cell' })
const chatInstance = state.chatInstances[chatId]
// First stream: assistant proposes update_notebook, which needs approval.
testContext.queuedStreams.push([
{ type: 'start' },
{
type: 'tool-input-available',
toolCallId: 'call-1',
toolName: 'update_notebook',
input: {
id: NOTEBOOK_ID,
expected_updated_at: '2024-01-01T00:00:00.000Z',
operations: [{ _tag: 'delete_cell', cell_id: 'cell-1' }],
},
},
{ type: 'tool-approval-request', approvalId: 'approval-1', toolCallId: 'call-1' },
{ type: 'finish' },
])
await chatInstance.sendMessage({ text: 'Delete the first cell' })
// Still pending approval — must not evict yet.
expect(notebooksState.notebooks[NOTEBOOK_ID]).toBeDefined()
// Second stream: after approval, the tool actually executes and returns its result.
testContext.queuedStreams.push([
{ type: 'start' },
{
type: 'tool-output-available',
toolCallId: 'call-1',
output: { id: NOTEBOOK_ID, name: 'Test notebook' },
},
{ type: 'finish' },
])
await chatInstance.addToolApprovalResponse({ id: 'approval-1', approved: true })
await vi.waitFor(() => {
expect(notebooksState.notebooks[NOTEBOOK_ID]).toBeUndefined()
})
expect(queryClient.getQueryData(contentKeys.resource(PROJECT_REF, NOTEBOOK_ID))).toBeUndefined()
})
it('binds cache effects to the project the request was sent under, not whatever project is active when the stream finishes', async () => {
seedNotebook()
const OTHER_PROJECT_REF = 'other-project'
const queryClient = getQueryClient()
queryClient.setQueryData(contentKeys.resource(PROJECT_REF, NOTEBOOK_ID), { id: NOTEBOOK_ID })
const state = createAiAssistantState()
state.setContext({ projectRef: PROJECT_REF })
const chatId = state.createChat({ name: 'Delete a cell' })
const chatInstance = state.chatInstances[chatId]
testContext.queuedStreams.push([
{ type: 'start' },
{
type: 'tool-input-available',
toolCallId: 'call-1',
toolName: 'update_notebook',
input: {
id: NOTEBOOK_ID,
expected_updated_at: '2024-01-01T00:00:00.000Z',
operations: [{ _tag: 'delete_cell', cell_id: 'cell-1' }],
},
},
{ type: 'tool-approval-request', approvalId: 'approval-1', toolCallId: 'call-1' },
{ type: 'finish' },
])
await chatInstance.sendMessage({ text: 'Delete the first cell' })
testContext.queuedStreams.push([
{ type: 'start' },
{
type: 'tool-output-available',
toolCallId: 'call-1',
output: { id: NOTEBOOK_ID, name: 'Test notebook' },
},
{ type: 'finish' },
])
// Simulate navigating to a different project while the approval round-trip is in
// flight: after the request was sent (with the origin project ref in its body), but
// before its stream resolves and onFinish runs.
testContext.onSend = () => state.setContext({ projectRef: OTHER_PROJECT_REF })
await chatInstance.addToolApprovalResponse({ id: 'approval-1', approved: true })
await vi.waitFor(() => {
expect(notebooksState.notebooks[NOTEBOOK_ID]).toBeUndefined()
})
// Must evict the origin project's cache entry — the one the write actually happened
// in — not whichever project happened to be active once the stream finished.
expect(queryClient.getQueryData(contentKeys.resource(PROJECT_REF, NOTEBOOK_ID))).toBeUndefined()
})
})