Running commands

Every exec call runs your command in just-bash, a bash interpreter written in JavaScript. It runs inside a Dynamic Worker that the durable object starts for the call. There is no Linux machine, no container, and no process table.

What works

• Pipes, redirects, variables, loops, functions, && and ||, globs and here-documents, the same way they work in bash.
• Text tools: cat, ls, find, grep, sed, awk, head, tail, sort, uniq, wc, cut, tr, diff, xargs and the rest of the usual set.
• jq and curl. This example turns both on. They are optional command groups that the host chooses.
• A built-in git for clone, status, diff, log and commits. It works directly on the workspace files.

What doesn't

• No npm, node, python, compilers, package managers or other binaries. If a command isn't built in, it is "not found", and installing it isn't possible.
• No direct network access. curl is there, but requests fail unless the host sets up an egress policy, and this example doesn't.
• No background jobs that outlive the call. When the command finishes, the Worker is gone.

Files

• cwd defaults to /workspace.
• Every file operation goes straight to the durable object's storage. There's no copy to sync, so a write is visible to the next call and to the HTTP file routes right away.
• /workspace/r2 is a read-only copy of an R2 bucket.
• /workspace/.agents/skills (where this file lives) is a read-only copy of files shipped with the Worker.
• A write under either read-only folder doesn't just fail that one command. It ends the whole call with exit code 1 and a "read-only mount" error, and output printed earlier in the call is lost. Write somewhere else under /workspace instead.

Tips

• Starting a call is fast, so many small commands are fine.
• For anything that needs a real runtime, such as building, testing or installing, use a workspace with a container backend.