Cloudflare

Cloudflare stores thread history and accepted work in Durable Object SQLite. Each Object owns its database, separate from other Objects. Records survive Object eviction and reconstruction. While the Object is live, the adapters share bounded, write-through memory for thread headers, submissions, recovery state, and the journal tail. Native reads reuse those committed views; cache misses reload validated rows. Canonical row retention shares one eight-MiB serialized-byte budget across the isolate, rather than allocating that allowance to each Object. Decoded row views are released with their retained raw rows. Memory stores over the same owner share its transaction gate and document, receipt, and usage views. Failed transactions discard cached state.

Run durable agents

Use @yielded/agent-platform-cloudflare for the complete host. ThreadObject.layer and ThreadObject.make assemble storage, registrations, admission, RPC, and alarm-driven recovery. The Cloudflare setup includes the agent, provider, Object class, and binding configuration. No separate storage Layer is needed.

Use an existing Object

If your application already owns a SQLite Durable Object, follow custom host integration to share its database and alarm slot with the runtime. Reuse the host’s Effect SQL client for application queries.

@yielded/agent-storage-cloudflare also exposes DoThreadStore and DoSubmissionLedger for custom assemblies. Their convenience Layers accept ctx.storage and supply the SQL client and Crypto. Both stores must use the same database so ownership claims fence the same records. These adapters provide storage ports; acquiring them alone does not drive agent execution or recovery.

Use the adapters for ordinary writes. If maintenance writes private adapter tables directly, call DoThreadStore.invalidate(ctx.storage) afterward, before serving port reads. Keep port operations quiesced during both steps and enroll the write with the host’s mutation gate. This invalidates the adapters’ shared views; it does not change durable data. Canonical records remain append-only.

Background worker starts validate authority and reserve capacity at the source, then admit and initialize the child through one destination RPC. Custom routed assemblies must install PortRouting.routedWorkerAdmissionLayer alongside the routed ledger and thread store. The owner-side port handler requires WakeScheduler and DurableRuntimeFailpoint and must run mutations through the host’s maintenance gate. ThreadObject.make and ThreadObject.layerInHost provide this wiring, including enrollment of affected maintenance lanes. Upgrade the framework, storage and host packages together; stored records and retry receipts need no reset.

Optional shared memory

Thread history and memory shared across threads have different owners. For shared documents, recall, and routed memory operations, see Cloudflare shared memory.