The reply disappeared. Did your AI helper finish the job?

What changed after the setup announcement

Claude Code introduced guided auto-mode setup on 19 October and followed it with a provider-request fix on 20 October. The 22 October v2.1.309 notes report further repairs to sandbox boundaries and helpers inheriting tool restrictions. They also say an interrupted tool call is shown as having an unknown outcome when a self-hosted session moves to a new runner. A runner is the computer environment carrying out the work. These are software changes to follow alongside model releases.

Give the session a small world

Imagine a folder containing your own public reading links. You want an assistant to make a simple local reading page with a title, subject and original link for each entry. Start with five invented entries. Write down the output before asking for code: a page you can open locally, a clear way to edit an entry and no account requirement. The assistant now has a finished job to aim at.

A workspace is the set of files, tools and settings available during that job. For this project, it can be one folder. Explain which files are inputs, which output may be created, and where to put an unfinished attempt. This makes it possible to stop and resume without rebuilding the plan from a long conversation. You are giving the assistant a small, comprehensible environment rather than asking it to manage your entire computer.

Make the first result tangible

Ask the assistant to sketch the reading page first. A title should open its original link. A subject should be readable without opening another screen. Try one missing title and one unusually long link before polishing the colours. These two awkward entries help you see whether the page is useful with ordinary messy input.

You can finish this exercise with a specification or a local file. Neither needs a public domain, a database subscription or an always-running agent. Save the input list and the result together. If another assistant takes over tomorrow, it should be able to see what was intended, what was produced and which detail still needs work. That is the practical benefit of organising a working session.

Give every helper the same small boundary

For the fictional reading-page project, make a simple agreement: the assistant may read the five-link input and create one page in the practice folder. It may not alter the original list or publish the page. If a second helper checks the result, give it read access to that same practice folder. A helper needs the material required for its part of the job; it does not need access to unrelated personal files.

A sandbox is a technical boundary around what a program can reach. Think of a practice tray that holds the materials for this job. The analogy stops at the tray edge: a software boundary depends on its actual settings and implementation. A written instruction expresses what you want; permissions determine what tools can do. For a first rehearsal, use invented inputs and retain a copy before any editing.

An unanswered message has three possible endings

Imagine that the assistant was creating your reading page when the connection disappeared. There are at least three possibilities: it created nothing, it saved part of the page, or it finished the page but its reply never reached you. The empty chat window alone cannot distinguish them. Asking it to start over immediately treats all three as the first case.

Inspect the practice folder first. If the page exists, open it and compare its entries with the original list. Suppose it contains all five titles but only four working links. Keep that useful partial result. Record which link is missing and ask for that specific repair. If the page is complete, the next job is checking it, not generating another copy. If it does not exist, record that observation before trying again.

The same reasoning matters when an assistant eventually sends messages or changes a shared system. Repeating an action can create a second copy even though the first reply was lost. This exercise stays local: you can learn the distinction using an invented file and a written interruption scenario. A known failure, an unfinished result and an unknown outcome deserve different next steps.

Leave a handoff that another assistant can use

Save a short note beside the page: the original five-link input, the output filename, what you inspected, what passed and what remains. Include the actual tool version if you ran one. For the four-working-links example, identify the missing entry and say that publication has not been requested. Another helper can now continue from observable work rather than reconstructing a conversation.

Before calling the page finished, check every title and destination, the missing-title example and one long link. Keep your original input available while checking. If you only completed a paper rehearsal, label it a plan. If you actually built the page, record the checks you really performed. The useful end of this small project is a readable page with a clear state, including any unfinished part.