Skip to content
mikan

Remote task executor

Run parallel, throwaway commands in a remote sandbox through the remote_task tool.

remote_task gives the agent a task executor: one command, run in a fresh remote sandbox, output returned. It is not a sandbox mode, and it is deliberately not the agent’s computer.

The distinction is the subject of ADR 0002: a sandbox runtime holds a workspace projection across turns, so files written in one turn are still there in the next. Remote execution cannot do that — sharing a POSIX filesystem over a WAN has no working answer — so it is offered as a tool instead of as --sandbox=cloudflare:*, which is no longer accepted.

Setup

Section titled “Setup”
Terminal window
export CLOUDFLARE_SANDBOX_URL="https://your-bridge.workers.dev"
export CLOUDFLARE_SANDBOX_TOKEN="replace-me" # optional
export CLOUDFLARE_SANDBOX_CWD="/workspace"   # optional, default /workspace

The tool appears in the agent’s tool list only when CLOUDFLARE_SANDBOX_URL is set, so a host without a bridge never advertises a capability it cannot fulfil. You deploy the bridge Worker and its container image yourself — see the bridge example.

What each call gets

Section titled “What each call gets”
  • Its own sandbox. Every call uses a fresh mikan-task-<uuid> id, so parallel calls cannot see or disturb each other. That is the point: fan-out work and dynamic workflows.
  • No workspace. The conversation’s files are not mounted. Input arrives in the command; results come back on stdout.
  • No credentials. The conversation vault belongs to the sandbox runtime. A throwaway remote task gets only what the command itself carries — do not expect GH_TOKEN and friends to be present.
  • Nothing afterwards. The sandbox is discarded when the command finishes. Anything written there is gone.

A command that exceeds the 50 KB output limit is truncated from the tail, and there is nowhere to spill the rest — a task executor has no workspace to write to — so have the command summarise or filter its own output.

Migrating from --sandbox=cloudflare:*

Section titled “Migrating from --sandbox=cloudflare:*”

That mode is rejected at startup. Pick a sandbox mode that actually holds the workspace — image:<image> or gondolin:default — and keep CLOUDFLARE_SANDBOX_URL set if you want the agent to retain remote fan-out through remote_task.

Nothing is lost in the move: the mode never mounted the workspace either. ActorExecutionResolver resolved mounts and the Cloudflare branch discarded them, the bridge payload has no mount concept, and vault file projection was unsupported. Sessions and MEMORY.md were always on the mikan host, never on the remote side.