# v0.0.1 limitations celld v0.0.1 is an alpha of the standalone engine, not a drop-in replacement for the entire Cloudflare Workers platform. - A fleet runs one application deployment. There is no public multi-tenant scheduler, account service, managed ingress, or global placement layer. - celld's peer HTTP does not terminate TLS. Use a private network or encrypted overlay and terminate public TLS in a separate ingress proxy. - The fleet bucket is administrative authority. celld does not make shared or broadly scoped object-store credentials safe. - `celld deploy` requires `esbuild` on `PATH` for Worker code, accepts `wrangler.json` or `wrangler.jsonc`, and deliberately rejects unsupported configuration instead of silently ignoring it. Static assets are supported, including asset-only projects; symlinks and special files are refused, and `.assetsignore` still requires Wrangler. `wrangler.toml`, routes, and vars remain outside the native deploy path. - The implemented Worker surface focuses on modules, Durable Objects, SQLite storage, alarms, service bindings, fetch, and hibernatable WebSockets. It does not promise every Web API, Node API, Cloudflare binding, compatibility flag, or arbitrary untrusted tenant program. Which Cloudflare services are supported, planned, or out of scope is [cloudflare-compat](cloudflare-compat.md). - WebSocket upgrades and messages route across nodes through the signed peer tunnel, so any node can serve as ingress for any cell. Close-code propagation and reconnect edge cases across nodes have thinner coverage than same-node WebSockets; latency-sensitive deployments still benefit from owner-sticky ingress. - Outbound WebSockets work from both Workers and Durable Objects. A Worker socket lives and dies with its request, as it does on Cloudflare: when the request and its `waitUntil` work finish, the socket is closed. A Durable Object socket outlives individual events and keeps its cell resident for the life of the connection; it does not hibernate or preserve a live transport across node migration, so applications needing failover must persist connection intent and reconnect after activation. A node bounds how much of its residency outbound sockets may pin. - R2 Worker bindings are not implemented. The S3-compatible fleet bucket is celld infrastructure, not an application R2 binding. - Linux is the production target. See [support](support.md) for the macOS replication limitation and exact prebuilt targets. - Pressure shedding is opt-in until release measurements establish safe defaults. A new node receives released cells through ordinary traffic; there is no central placement controller. - Updates are explicit. There is no unattended update agent. The installer retains immutable releases and switches one `current` pointer, so installing an exact previous SHA performs a runtime-pair rollback. If an unsupported deployment key or runtime API is encountered, celld should fail loudly. Silent compatibility gaps are bugs.