Skip to content

12_worker_backend.md

Latest commit

 

History

History
306 lines (252 loc) · 12.5 KB

12_worker_backend.md

File metadata and controls

306 lines (252 loc) · 12.5 KB
Raw

12. Worker backend

Note

This doc reflects shipped code in packages/computer/src/backends/worker-shell/. The example deployment lives at examples/worker-shell/.

The worker backend is the second WorkspaceBackend shape the package ships. It pairs a Workspace with a just-bash shell running in a Dynamic Worker minted through env.LOADER. The shell reaches the host workspace over Workers RPC, so there is no second store and no sync round trip; the host Durable Object's SQLite is the only authoritative state.

Import via the sub-path so the bundled just-bash payload tree-shakes out of consumers that don't use it:

import { WorkerShellBackend } from "@cloudflare/computer/backends/worker-shell";

When to reach for it

The container backend (@cloudflare/computer/backends/container) gives you a real Linux environment with arbitrary binaries on $PATH, network, and a full POSIX filesystem. It costs a container per session and a real roundtrip on every filesystem op.

The worker backend trades the real environment for a Workers isolate that boots instantly, scales out cheaply, and has no container lifecycle. The shell is the just-bash interpreter; the supported command set is broad (cat, grep, awk, sed, jq, sort) but not the full Linux userland. JavaScript modules run through the worker-javascript backend, not through just-bash's Node-only language commands. Filesystem operations forward into the same SQLite store as the container backend, so the storage shape, mount rules, and read-only enforcement are unchanged.

Reach for the worker backend when:

  • The agent only needs textual shell tooling — read, search, transform, summarize. Not when it needs to compile, install packages, or run a browser.
  • You want isolation without a container. A runaway Bash run in one workspace can OOM its Dynamic Worker isolate without affecting the host DO.
  • You're sensitive to the per-session container cost.

Reach for the container backend when the agent runs npm, a real language runtime, native binaries, or anything else that needs a real kernel. Prefer the worker backend's host-forwarded git command for supported Git operations; use the container only for Git behavior that requires its native environment.

Wire shape

agent code
  │  Workers RPC
  ▼
host DO ─── Workspace ─── WorkerShellBackend
                              │
                              │  env.LOADER.get(loaderId, codeCallback)
                              │    .getEntrypoint("ShellWorker")
                              ▼
                       Dynamic Worker (ShellWorker)
                       WorkerEntrypoint with exec / getExec /
                       killExec RPC methods
                              │
                              │  env.HOST.getWorkspace() per exec
                              ▼
                       new Bash({ fs: WorkspaceFsAdapter(ws.fs) })
                              │
                              │  every fs RPC the shell makes lands
                              │  in the host DO's own request context
                              ▼
                       back to host DO's SQLite

The Worker Loader caches the Dynamic Worker isolate by id. The default id is workspace-shell:${workspace.id} — one isolate per workspace, so concurrent execs in the same workspace share a warm isolate, and a runaway Bash run in one workspace can't touch another workspace's shell.

globalOutbound: null on the Dynamic Worker blocks fetch() and connect() from inside the shell. The only path out of the isolate is back through the host DO over env.HOST.

Built-in custom commands

The worker backend registers two just-bash custom commands on every exec:

  • git ... forwards to the host workspace's workspace.git.cli(...).
  • assets publish <path> [<expiry>] forwards to the host workspace's configured assets publisher and prints the share URL to stdout.

assets publish accepts an absolute path or a path relative to the current working directory. The optional expiry defaults to one hour; a bare number is milliseconds, and ms, s, m, and h suffixes are accepted (30000, 30s, 5m, 2h). If the Workspace was not constructed with an assets client, the command exits 1 with a clear message.

The Dynamic Worker never receives the R2 bucket binding or signing secrets. The host Durable Object configures the Workspace with an assets client, and the command reaches that host-side capability over the same env.HOST.getWorkspace() loopback as filesystem and git calls.

Why a loopback proxy

The natural impulse is to hand the Dynamic Worker the host DO's DurableObjectNamespace directly in its env. That fails: values passed through the Worker Loader's env go through structured clone, and a raw DurableObjectNamespace doesn't survive that.

WorkspaceServiceProxy (in @cloudflare/computer) is a tiny WorkerEntrypoint whose getWorkspace() method does the namespace lookup on the host side. The backend mints a stub through ctx.exports.WorkspaceServiceProxy({ props: { binding, id } }); the resulting Fetcher does survive structured clone because the runtime serializes it as a binding reference rather than a value. The shell calls env.HOST.getWorkspace() per exec and reaches back into the host DO through a normal Workers RPC hop.

A per-call lookup also matters for I/O context: every filesystem RPC the shell makes lands inside the host DO's own request context, where the DO's storage handles are valid. An earlier attempt at handing the shell a long-lived WorkspaceFilesystemStub through env hit workerd's cross-request I/O guard immediately — the shell's request context and the DO's request context aren't the same.

Push and pull

BackendHandle.sync is "none". With a single authoritative store there's nothing to ship or fetch; Workspace.push and Workspace.pull short-circuit on the bit and the reconcile pass on connect is skipped. The exec sync bracket still calls them so the surface stays uniform — the pushed, pulled, and skipped counts on the runtime result are empty.

Event stream framing

ShellWorker.exec returns { id, events: ReadableStream<Uint8Array> }. The byte stream is newline- delimited JSON, one ExecEvent per line. Workers RPC carries byte streams natively; the JSON-framed log is the lowest-friction shape across the isolate hop.

The backend decodes the frames into structured ExecEvent values and re-encodes string payloads (stdout / stderr) into Uint8Array so the runtime's utf8 decoder transforms, which accumulate the result from raw events, see the shape they already handle.

Fetcher factory escape hatch

WorkerShellBackend is source-agnostic. The common case takes { loader, workspace, ctx } and builds the loader callback itself. For deployments that need a different Fetcher source — a Workers service binding, a Workers-for-Platforms dispatch namespace, a fake in tests — pass fetcher: () => unknown | Promise<unknown> instead. The factory is consulted once on connect(); the resolved Fetcher is held for the life of the handle.

Known fidelity gaps

  • Single-chunk stdout/stderr. just-bash returns the full stdout and stderr at the end of a run rather than streaming chunks. ShellWorker.exec emits one stdout event (when non-empty), one stderr event (when non-empty), and one exit event. Live streaming would need a fork of just-bash or a wrapper writer.
  • No hard links, no utimes. The adapter throws ENOSYS on link (the store has no hard-link model) and no-ops on utimes (no atime column). chmod, symlink, readlink, and lstat all work end-to-end against the DO's store.
  • No cross-request reattach. ShellWorker.getExec always returns ENOENT; killExec is a no-op. Each exec is scoped to its own call. The previous in-isolate event log shape didn't survive the move to per-call workspace stubs (each exec fetches its own stub through env.HOST.getWorkspace()).
  • PATH-walk diagnostics in wrangler dev. just-bash probes every directory on $PATH for every command name. Each miss resolves a rejected RPC promise that workerd's rpc layer flags as "Uncaught" in the dev log, even though just-bash catches the rejection locally. The noise is cosmetic; exec still returns the correct result.

Custom commands

The shell isolate registers built-in git and artifacts commands that forward through the WorkspaceStub returned by getWorkspace(): git calls workspace.git.cli(...), and artifacts calls workspace.artifacts.cli(...). The artifacts command also bridges the two: its create shorthand registers a git remote, so the command hands the artifacts CLI a remoteAdd closure backed by the same workspace.git.cli(...). The artifacts package owns no git of its own; the bridge lives at the backend wiring layer. ShellWorker still exposes an extraCommands(ws) hook for layering project-specific commands onto the same Bash instance. The hook runs once per exec with the live host stub the shell already reached, so a command shares that stub's lifetime without refetching.

A host durable object wires the Artifacts command by passing the binding to Workspace:

export class MyAgent extends DurableObject<Env> {
  #workspace: Workspace;

  constructor(ctx: DurableObjectState, env: Env) {
    super(ctx, env);
    this.#workspace = new Workspace({
      storage: ctx.storage,
      sessionId: ctx.id.toString(),
      artifacts: { binding: env.ARTIFACTS },
      backends: [new WorkerShellBackend(/* ... */)],
    });
  }
}

With the binding declared in Wrangler:

{
  "artifacts": [{ "binding": "ARTIFACTS", "namespace": "default" }]
}

artifacts repo list inside bash.exec then forwards to the client's cli(...). The shell isolate has no network of its own; the binding call lands host-side, the same way network-bound git subcommands do. See docs/15_artifacts_interface.md.

Example

examples/worker-shell/ is a single wrangler project that mirrors examples/container/ beat for beat:

  • One wrangler.jsonc with the Durable Object, an R2 mount at /workspace/r2, and a worker_loaders binding named LOADER.
  • src/index.ts holds the DO and the HTTP surface (the /c/<name>/file/... and /c/<name>/exec routes the container example also exposes).
  • No Dockerfile, no build script. The shell ships with @cloudflare/computer/backends/worker-shell as feature groups: an always-on core (SHELL_CORE_MODULES) plus one optional group per command at @cloudflare/computer/shell/<feature>. The backend assembles core with whatever groups you opt into and hands the result to the Loader callback itself.

The DO's backend wiring:

import curlModules from "@cloudflare/computer/shell/curl";
import sqliteModules from "@cloudflare/computer/shell/sqlite";

new WorkerShellBackend({
  loader: env.LOADER,
  workspace: { binding: "ContainerExample", id: ctx.id.toString() },
  ctx,
  commands: [curlModules, sqliteModules],
})

Run with npm run dev --workspace @example/computer-worker. The same curl recipes from the container example work once curlModules is passed to commands.

Optional shell commands

Core carries the always-on command set (cat, ls, grep, sed, awk, sort, …). The heavier commands are split into optional groups that are opt-in by import: import a group from @cloudflare/computer/shell/<feature> and pass it to the commands option, and only then does its code enter your bundle.

import curlModules from "@cloudflare/computer/shell/curl";
import htmlToMarkdownModules from "@cloudflare/computer/shell/html-to-markdown";

// commands: [curlModules, htmlToMarkdownModules]

A group you never import is unreachable in your module graph, so the bundler drops it — there is no build-time flag to set and no default-on cost to opt out of. The full set of optional groups is curl, html-to-markdown, python, sqlite, js-exec, yq, file, xan, and jq.

curl runs on a SecureFetch adapter over the isolate's global fetch — undici is redirected to a throwing stub at build time and never ships. Egress stays governed by the Dynamic Worker's globalOutbound (left null, i.e. closed), not by the shell, so enabling curl does not by itself open the network.

Consumers that build the Loader callback by hand (the fetcher path) assemble the modules table themselves with assembleShellModules([...groups]) from the same package.