Security and hardening
Binding isolation, operator token, egress policy, deploy confinement, and credential hygiene over upstream denoland/celld.
Patches that shrink what a Worker — or a stolen credential, or a malicious project — can reach. Opt-in hardening stays opt-in; isolation that fails closed is on for everyone. Each entry names its fork patch number.
Binding and socket isolation (fail closed)
0040— Bindings resolve from the script’s declared resources: cell dispatch and id derivation refuse names the calling script never bound (D1, KV, queues, workflows, Durable Object namespaces, reserved cells), runtime cells answer an RPC allowlist, and cell-scoped ops (storage, SQL, alarms, containers, facets, socket registry) answer only the executing cell. Stubs received over RPC or storage keep working through host-recorded grants: the host checks each lifted stub’s brand and the sender’s rights and holds the stub table itself, writing the envelope tag in Rust. Stored-stub rows written before this patch decode as plain data on first read; storing them again writes the new envelope. Upgrade every node together: stubs sent from an old node revive as plain data on a new one.0041— The operator-only KV and Queue fetch branches demand the same host-set operator-ingress marker the D1 export introduced: every KV fetch op and the Queue info/purge/pause/resume/peek/redrive ops refusestub.fetch(), while the bindings (RPC/send) keep working. The tunnel markers are protocol-negotiated — a peer that sends no marker flag has its markers ignored, so a Worker on an older node cannot forge them. Queue DLQ and redrive need the fleet on this patch: mixed versions fail closed both ways, and the alarm replays the durable DLQ and redrive marks once the fleet converges.0042— Socket ids answer only their owner: a WebSocket id is usable only by the cell scope or request context that holds it, and TCP socket ids likewise by their event. Ids are unguessable (f64-safe random, nonzero); joining another cell’s upgrade presents a single-use host-minted token.0119—cello deployrefuses Durable Object bindings with a foreignscript_name— previously counted as a class of the deploying script, boot-looping the node — and namesscript_namein the refusal. A binding naming its own script stays local; a non-stringscript_nameis refused the same way (upstream #312).0037— Shared D1, KV, and Queue cells run in a dedicated runtime isolate pool with no tenant bundle, bindings, vars, services, loaders, or schedule (see Workers compatibility).
Operator token and network posture
0013—CELLO_OPERATOR_TOKENguards the internal listener: when set, every internal route requiresAuthorization: Bearer <token>(constant-time compare, 401 without it, no exempt routes — probes use the public health endpoint) and the node refuses to start on an empty value or one with surrounding whitespace. Unset is today’s open listener. Every in-tree internal client sends the token from its environment. Set it only once every node runs this release: older nodes send no token, so a mixed fleet fails internal calls with 401.0021— Tenant egress policy:CELLO_EGRESS=publicrefuses Workerfetch()/connect()/ outbound-WebSocket dials to non-public addresses (loopback, private, link-local, CGNAT, ULA, unspecified, multicast, reserved, IPv4-mapped and transition forms, the metadata address), withCELLO_EGRESS_ALLOWCIDR/host:portexceptions matched on every hop of the resolved address. Tenant clients ignore proxies; Cello’s own traffic and published container endpoints are exempt; refused fetches reject withTypeError. Defaultopenis unchanged.0014— Container fence admits replies to a node in a container netns:CELLO_FENCE_NODE_NETNS=auto(defaultoff, byte-identical) cuts anestablished,relatedexception for the node’s own addresses, discovered on Linux or pinned withCELLO_FENCE_NODE_ADDRS. Replies only — a container still cannot open a connection to the node.
Confined deploys
0050—cello deployandcello d1read the project tree only inside the project’s own directory: every project path is canonicalized and refused with an error naming its config key when it resolves elsewhere — the config,main, assets, Dockerfiles and build contexts,migrations_dir(entries must be regular files, never links), Python sources. Unlike Wrangler, the deploy may run with more privilege than the project’s author, so on Linux esbuild runs under Landlock with reads confined to the project (it must be the self-contained release binary, not a wrapper script); every run also writes a metafile, and the build fails unless every file esbuild read resolves inside the project. A monorepo that hoistsnode_modulesabove the config directory no longer bundles: install the packages inside it.0051—cello deploy --read-root DIR(repeatable) names directories beyond the project that build inputs may come from: esbuild’s inputs andcontainers[].imageDockerfiles with their build contexts. Only the operator names a read root, on the command line, so a project cannot widen its own bounds. A platform that deploys a built subdirectory of a checkout names the checkout.
Credential hygiene
0011—CELLO_MEMORY_SCOPE=selfmeasures the process’s own cgroup v2 leaf instead of the container root, so nodes sharing one container each shed against their ownmemory.max(which an unsetCELLO_MAX_RSS_MBdefaults to a share of); cgroup v1 and an unreadable leaf fall back to the root with a warning.0052—CELLO_DEPLOY_AWS_ACCESS_KEY_ID,CELLO_DEPLOY_AWS_SECRET_ACCESS_KEY, andCELLO_DEPLOY_AWS_SESSION_TOKENgive deployment-artifact reads their own credential: with the trio set, a node servesdeploy/anddeploy-blobs/through a second S3 client while every other key keepsAWS_*. A platform can then mint a read-only grant over the deployment prefixes and a read-write grant over the state prefixes, so a credential copied out of the node’s environment can read a deployment but never publish one. Partial settings are a startup error; serving nodes refuse the trio ongs:///az://buckets; control-plane nodes (managed or BYO bucket) must not set it at all. Unset keeps one credential for everything, exactly as before.0053— A serving process is non-dumpable:maincallsprctl(PR_SET_DUMPABLE, 0)when the invocation is a node or thecello devsupervisor, so a same-uid sibling — a tenant build running as the fleet uid — can no longer read/proc/<pid>/environ,mem, ormaps, orptracethe node. Operator commands stay dumpable. Linux-only; no core dumps from a node, and a non-root debugger cannot attach (root can). Children a serving process spawns start with an emptied environment (PATH,HOME,TMPDIRonly).0131— TheCELLO_NODEbinding name is reserved at deploy time while.enventries are skipped with a warning (see Workers compatibility).