AI Persistence/build Cloudflare Adapter

Cloudflare Chat Persistence

The deliverable is one file in the Worker — src/lib/chat-persistence.ts — exporting a factory that builds a ChatPersistence from the request's D1 binding, plus (when the app needs coordination) a Durable Object lock store. Tables go into the app's existing migrations/ directory and are applied with wrangler d1 migrations apply.

Do not create a package or a migration runner. Wrangler already tracks applied migrations; a second bookkeeping table only creates drift.

Read the Store Reference (docs/persistence/store-reference.md) for the store contracts, and ai-persistence/stores for the shape rules. This skill covers only the Cloudflare-specific parts.

1. Read the app before writing anything

2. Two independent pieces

D1 database      -> messages, runs, interrupts, metadata   (AIPersistence.stores)
Durable Object   -> LockStore                              (withLocks — NOT a store)

These do not compose into one object. AIPersistence.stores accepts exactly four keys. Putting locks in the map throws Unknown AIPersistence store key: locks; putting it in a composePersistence override throws Unknown AIPersistence override key: locks. Both also fail to type-check. Return the state persistence from one factory and the lock store from another, then wire them as two middlewares.

Most apps need only the first piece. Add the Durable Object when other middleware genuinely needs mutual exclusion across isolates — InMemoryLockStore gives none, because a Worker runs on many isolates at once.

3. Bindings are per-request

This is the one rule that separates Cloudflare from every other backend. A D1 binding does not exist at module scope, so chat-persistence.ts must export a factory, not a const:

import { defineAIPersistence } from '@tanstack/ai-persistence'
import type { ChatPersistence } from '@tanstack/ai-persistence'

/** Call inside a request handler — `env` is not available at module scope. */
export function chatPersistence(d1: D1Database): ChatPersistence {
  return defineAIPersistence({
    stores: {
      messages: createMessageStore(d1),
      runs: createRunStore(d1),
      interrupts: createInterruptStore(d1),
      metadata: createMetadataStore(d1),
    },
  })
}

Annotate ChatPersistence — bare AIPersistence is the all-optional bag and withPersistence rejects it. Building it per request is cheap: the stores hold no state beyond the binding.

4. The stores

Two routes, same invariants:

• Drizzle over D1 — if the app already runs Drizzle, wrap the binding with drizzle(env.DB, { schema }) and follow ai-persistence/build-drizzle-adapter verbatim (its "if db is per-request" section is exactly this case). Stop reading here.
• Raw D1 — implement the four stores against d1.prepare(sql).bind(...): .first() for get, .all() for list*, .run() for writes. D1 speaks SQLite, so this mirrors the node:sqlite walkthrough in the guide one-for-one; everything is already async, so no Promise.resolve wrapping.

The invariants are the whole game, whichever route you take:

On runs, findActiveRun is required; listByThread and listReclaimable are optional, so implement those two only if the app needs them. withPersistence calls none of the three — the consumers are reconstruct.ts (findActiveRun, for rejoin-by-thread) and @tanstack/ai-sandbox's reapDetachedRuns (listReclaimable, without which the store cannot be reaped); nothing in the framework calls listByThread. Consumers of the two OPTIONAL methods feature-detect with store.method?.(...) and degrade to "not supported" when one is absent. The conformance testkit does not: either of those you leave out must be listed in skipMethods or the suite fails, so declare them and it reports the omission as a skip.

RunRecord.error is a structured RunError ({ message: string, code?: string }), so the table gets two columns rather than one JSON blob: error for the provider's prose and error_code for the stable classification an operator filters and groups by. Write both together in update, so a later code-less failure can never leave a stale error_code from an earlier one behind, and omit code from the mapped record when the column is null: ...(row.error != null ? { error: { message: row.error, ...(row.error_code != null ? { code: row.error_code } : {}) } } : {}). Other row mappers omit absent optionals the same way (...(row.sandbox_key != null ? { sandboxKey: row.sandbox_key } : {})) so records compare cleanly against the reference in-memory backend. JSON columns are text: JSON.parse on read, JSON.stringify on write. Timestamps are integer epoch ms.

sandbox_key, detached_since, cancel_requested, and driver_epoch are the durable-agent-runs columns. In update, check 'field' in patch for all four — never patch.field !== undefined — because a reattach clears detachedSince by passing it explicitly as undefined, which must bind NULL into the SET clause rather than being filtered out of it (a filtered clear leaves the stale value and the run looks permanently detached to the reaper). The same applies to cancelRequested: false written explicitly is a real, meaningful value distinct from "never set", not something to coerce away. See examples/ts-react-chat/src/lib/sqlite-persistence.ts (D1 speaks the same SQLite dialect) for the worked update body.

5. The migration

Write the tables into the app's migrations/ directory as a new numbered file:

CREATE TABLE IF NOT EXISTS chat_threads (
  thread_id text PRIMARY KEY NOT NULL,
  messages_json text NOT NULL,
  updated_at integer NOT NULL
);
CREATE TABLE IF NOT EXISTS chat_runs (
  run_id text PRIMARY KEY NOT NULL,
  thread_id text NOT NULL,
  status text NOT NULL,
  started_at integer NOT NULL,
  finished_at integer,
  error text,
  error_code text,
  usage_json text,
  sandbox_key text,
  detached_since integer,
  cancel_requested integer,
  driver_epoch integer
);
CREATE INDEX IF NOT EXISTS chat_runs_thread_status ON chat_runs (thread_id, status);
CREATE INDEX IF NOT EXISTS chat_runs_thread_started ON chat_runs (thread_id, started_at);
-- Powers listReclaimable: status = 'running' AND detached_since <= cutoff.
CREATE INDEX IF NOT EXISTS chat_runs_status_detached ON chat_runs (status, detached_since);
CREATE TABLE IF NOT EXISTS chat_interrupts (
  interrupt_id text PRIMARY KEY NOT NULL,
  run_id text NOT NULL,
  thread_id text NOT NULL,
  status text NOT NULL,
  requested_at integer NOT NULL,
  resolved_at integer,
  payload_json text NOT NULL,
  response_json text
);
CREATE INDEX IF NOT EXISTS chat_interrupts_thread ON chat_interrupts (thread_id, requested_at);
CREATE TABLE IF NOT EXISTS chat_metadata (
  namespace text NOT NULL,
  key text NOT NULL,
  value_json text NOT NULL,
  PRIMARY KEY (namespace, key)
);

Apply with wrangler d1 migrations apply <database-name> (--local first, then --remote). If the app also uses Drizzle, generate this file with drizzle-kit generate instead of hand-writing it — the SQL and the Drizzle table definitions must agree, so let one of them own the other.

6. Wire it into the chat route

import {
  chat,
  chatParamsFromRequest,
  toServerSentEventsResponse,
} from '@tanstack/ai'
import { openaiText } from '@tanstack/ai-openai'
import { withPersistence } from '@tanstack/ai-persistence'
import { chatPersistence } from './lib/chat-persistence'

export default {
  async fetch(request: Request, env: Env) {
    const params = await chatParamsFromRequest(request)
    const stream = chat({
      adapter: openaiText('gpt-5.5'),
      messages: params.messages,
      threadId: params.threadId,
      runId: params.runId,
      ...(params.resume ? { resume: params.resume } : {}),
      middleware: [withPersistence(chatPersistence(env.DB))],
    })
    return toServerSentEventsResponse(stream)
  },
}

threadId is a bare string to the stores. Authorize thread access at the route — derive the user from the session, never trust a client-supplied id.

7. Durable Object lock store (only if needed)

Implement LockStore from @tanstack/ai/locks. withLock(key, fn) routes each key to a Durable Object instance (idFromName(key)) that serializes owners. Use leases so a crashed owner cannot block forever: the DO grants a lease with an expiry, an alarm reclaims it, and the lock passes the callback an AbortSignal that fires when ownership can no longer be guaranteed. Callbacks must stop starting external mutations once the signal aborts.

Export the DO class from the Worker entry so wrangler can bind it:

export { ChatLockDurableObject } from './locks'

Then wire both middlewares:

import { withLocks } from '@tanstack/ai/locks'
import { withPersistence } from '@tanstack/ai-persistence'

const middleware = [
  withPersistence(chatPersistence(env.AI_STATE)),
  withLocks(createDurableObjectLockStore(env.AI_LOCKS)),
]

wrangler bindings

{
  "d1_databases": [
    {
      "binding": "AI_STATE",
      "database_name": "tanstack-ai-state",
      "database_id": "<id>",
      "migrations_dir": "migrations",
    },
  ],
  "durable_objects": {
    "bindings": [{ "name": "AI_LOCKS", "class_name": "ChatLockDurableObject" }],
  },
  "migrations": [
    { "tag": "v1", "new_sqlite_classes": ["ChatLockDurableObject"] },
  ],
}

Durable Object locks do not use the D1 table migration set; their state is configured through the migration tags above.

Verify

import { runPersistenceConformance } from '@tanstack/ai-persistence/testkit'
import { env } from 'cloudflare:test'
import { chatPersistence } from '../src/lib/chat-persistence'

runPersistenceConformance('app-d1', () => chatPersistence(env.AI_STATE), {
  skip: ['generationRuns', 'artifacts', 'blobs'],
})

Run it against a Miniflare D1 binding with the migration applied, reset between runs. All four state stores are provided; the suite also covers the three generation stores, so a chat-only adapter declares them skipped (drop the skip once you add the R2-backed set from ai-persistence/build-cloudflare-artifact-store). skip never accepts 'locks', which is not a store.

If your recipe leaves an optional runs method (listByThread/listReclaimable) unimplemented, declare it with skipMethods, e.g. { skipMethods: ['runs.listByThread'] }. An omitted method that is not declared fails the suite instead of silently passing.

The lock store needs its own tests, because nothing in the conformance suite touches it. Cover at minimum: two concurrent withLock calls on the same key serialize; different keys do not block each other; a lease that expires aborts the signal handed to the critical section; and a callback that throws still releases the lock.

Only if you are publishing this as a package

For a reusable npm adapter rather than a file in the app: peer-dep @cloudflare/workers-types >=4.x, and prepend /// <reference types="@cloudflare/workers-types" /> to the generated index.d.ts so consumers get the D1/DurableObject types. Emit the table SQL into the consumer's migrations/ directory rather than shipping a runner, and if you offer both raw-D1 and Drizzle paths, guard with a test that the emitted SQL and the Drizzle tables describe the same schema.