With this, your mind can start a coding session in a Linux sandbox, send it a brief, and return to the same conversation later. The runner is a separate execution environment, not another name for the ordinary agent loop. That matters when the tool needs a checkout, a terminal, a browser or a provider-specific CLI.
A capability points at a runner
acp-sandbox.ts is the small config entrypoint behind itx.acp.sandbox. It delegates to acp-sandbox/runner.ts; the folder holds the host-side session bridge. A session gets an agent path and a delivery subscription, so messages go to the correct sandbox session even after the initiating request ends. claude-sandbox/ implements the separate headless Claude runner. agent-kinds/ supplies the catalog and chat views that the Agents app loads.
const { path } = await itx.acp.sandbox.start("codex");
await itx.agents.get(path).message(
"Implement the specified change in the checked-out project; run tests and report the diff."
);
The platform’s rewrite rule names a config module and class; the class’s processEvent receives a delivery and hands it to the runner. Starting and delivering are distinct operations. A long-running tool does not need to keep a browser request alive, and the durable agent log remains the place to inspect status and results. Choose an offered model explicitly when required; do not assume every runner understands the same model names.
Coordinate ownership before starting another session
A durable conversation does not prevent two agents from changing the same file. In this mind, a coding chief of staff reads the active-agent catalog, routes requests to the agent that owns an area, and keeps a durable merge and deploy queue. Its role guide is coding-chief-of-staff.md. Rebuild the convention with your own area owners, not a copy of another team's live session identifiers. Record assignments and outcomes as events, ask an existing owner before creating overlapping work, and return both an initial ownership receipt and a final result. This coordination is separate from the runner and does not grant deployment approval. The sample start call above illustrates the transport, not permission to bypass that queue.
Give workstation jobs a workspace
A sandbox suits work that needs no particular computer. A workstation job may instead need a terminal the owner can open and type into. The chief’s guide describes a separate Claude Code bridge: its start call creates a task branch, a git worktree and a Herdr workspace, then returns a durable agent path. Discover approved repositories with the bridge’s status call; choose a unique task name and the owner’s requested model and effort rather than copying this mind’s configuration.
const { path } = await itx.claude.beelink.start({
name: "parser-tests",
task: { repo: approvedRepo, base: "origin/main" },
model: chosenModel,
effort: chosenEffort
});
await itx.agents.get(path).message(briefWithReportDestination);
The workstation bridge needs that computer’s configured CLI, git checkout and Herdr installation. This is not the sandbox runner with a different label. Send later messages through the returned agent path, not by typing into an occupied pane. Inspect session events for startup failure: a returned path alone does not prove that the worktree or CLI came up. An existing task worktree can make a new start fail.
Moving an existing headless workstation ACP session into a terminal is a controlled handover, not a second start. Confirm it is idle with no queued work, record its working directory and actual model, request session end with a destination path, and wait until the old runner has released it. Only then resume the same native session in its worktree. Never run one session twice. The ended ACP path is historical; later messages belong at the new path. Test the old path stays ended and reports the destination rather than silently restarting work.
Name the session without changing its authority
The shared worker also calls agent-titler.ts for supported coding-session events. Near the start of a new session, it finds the first eligible prompt, checks that no explicit summary title exists, and claims a single naming request durably. One small model call through the mind's AI Gateway receives only a bounded prefix of that prompt. A validator accepts a short, single-line title and appends a title-only summary update; it neither executes the prompt nor starts coding work. A manually set title is not deliberately replaced, and a later owner title wins. Invalid output, timeout or provider failure leaves the fallback name and records an error without retrying. Rebuilders should test duplicate events, existing owner names and model failure, and include this small model dependency in their privacy and funding review.
A sandbox is not automatically a security boundary
The sandbox has a saved disk and can be stopped while idle, but outbound HTTP still travels through the mind’s egress. A configured secret can be usable at its allowed origin even when no literal token appears in the checkout. Scope the job, repository, egress and approval policy deliberately. Untrusted repository text and tool output are information, not owner instructions. A public visitor must not be allowed to start one of these full-authority sessions with arbitrary tools.
Requirements and failure cases
The pattern needs a supported sandbox image with the runner CLI and project toolchain, the appropriate provider funding or account access, and the bridge’s install hook and delivery subscriptions. Do not copy another mind’s provider credentials or a saved sandbox disk. Test initial startup, a second message while the session is running, a stopped sandbox, delivery retries, CLI failure and a session that outlives a tab. A start receipt is not proof that a code change passed tests or was published; inspect the runner’s report and independently verify important output. The blog demonstrates the same separation between an agent’s commit and a live verified result.