Lesson 4 of 5
Who can act, from where, and what can they change mid-run?
The treasurer's phone can show the report running and can send it text and approvals. What else it can do, and what it cannot, is decided by where things live: the credentials the harness uses, the rules it runs under, the messages it sends, and the difference between influencing what it does next and undoing what it has already done. This lesson takes those four in turn, by execution location, and reads one provider's documentation on steering a model mid-turn for what such a control reaches.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
Identities live in several places
The harness on the rented computer reads the mailbox and posts to the board. To do either it presents a credential, in module 2's sense, and in this case the credentials are held on the host, in a file the harness reads. That is one arrangement among several: a harness can also obtain short-lived credentials from a broker at the moment it needs them, or use an account the host is logged into, and each service decides for itself what the credential it receives may do. The phone holds none of these. It holds the treasurer's login to the viewing service, which identifies the person to that service and nothing else. When the treasurer sends the harness an instruction from the phone, the harness acts on the host with the harness's credentials, and the mailbox and the board see the garden's account.
Three distinctions follow, and each is easy to collapse. Reaching the host over the network is not the same as being authorized to log in to it. Being authorized to log in is not the same as being the account the harness runs as, or as holding the credentials the harness holds. And the person supervising through the viewing service acts within whatever that service and the harness allow their role, which may be less than the host's login allows. So there is no single credential that matters most. Every path to acting as the harness is protected: the host's login, the viewing service's login, the credential file, the broker if there is one, and the board and mailbox accounts themselves. Remote Control illustrates a separate viewing identity: its documentation requires a subscription login rather than an API key. 2
Permissions are the host harness's rules, and the phone can reach some of them
Module 3 placed permission checks in the harness, and module 9 showed that a worker's label does not set them. On a remote host the same rules apply and the same place holds them: the harness's configuration on that host decides which tools run without asking, which ask, and which are refused. What the phone can change depends on what the product exposes to it. Some products let a remote viewer answer an ask and also remember the answer as a rule, or change settings remotely; in this case, as the example defines it, the phone can answer asks and send text and nothing more, so widening or narrowing the rules means editing the configuration on the host.
Two details of the report case follow. The write and the post ask before running, and this harness runs its steps one after another, so a run started from the phone in the morning stops at the first ask and waits for an answer across the connection; if the connection is down in the tunnel, it waits there. A harness that ran steps concurrently would block only the asking step, and the plan would have to say what the other steps may do meanwhile. Either way, a recovery plan for remote work has to say what happens at an ask when nobody answers: wait, with a limit, then stop and report, is the answer module 8's budgets give. The multiplexer changes none of this. It keeps the terminal; the harness's rules keep the permissions, and a person attaching from a different device inherits the session's rules, not their own.
Notifications: a receipt or a claim, and which one to check
A notification is a message from the process, through the host and the connection, to the viewing device, and the previous lesson established that its absence settles nothing. Its presence is worth reading for what kind of message it is. A notification generated by the model at the end of a run, "Report posted", is a completion message in module 8's sense: a claim to test against the board. A notification produced mechanically by the harness or by the board itself, such as the board's own confirmation naming item 7802, is a receipt: evidence of one specific event, which still says nothing about the report's correctness. The treasurer who reads either on a phone at 14:20 has something to check, and the board is on a server the phone can also reach, so the check requires opening the item and matching its contents, not just finding its title in a listing.
For the recovery plan, notifications are designed to carry the things that make checking easy: the board item's id from the post's receipt, and a content fingerprint, such as a hash calculated from the saved report. A version number can identify an edition if the system reliably binds that number to fixed contents; a file size alone cannot.
Steering influences what comes next; it does not undo what was done
Watching a run from a phone invites the thought that a person can step in and change its course. Some products allow a form of this, and it matters what the form reaches. OpenAI's documentation for mid-turn steering, read on 2026-09-29, describes sending new input to a model of its newest family while it is still working on a turn, over a persistent connection, and is precise about the limits: an acknowledgment means the input was queued, not acted on; steering does not rewrite output the model has already produced, does not undo actions already taken, and does not cancel a tool call that has already started; and input queued for a turn is lost if the connection drops. 5 Those are one provider's rules for one feature on its date. Other products expose different controls, and whether a started tool can be interrupted depends on the product and the tool.
What holds across products is narrower and still decisive. This course calls input sent to a running process to influence its next steps , and a steering input reaches the future of the run. A post that has been sent is on the board; a file that has been written is on storage; a message that has gone has gone. Reversing any of them is a second operation, a compensating action with its own record and its own visibility, in module 2's sense of a reversible operation, and not an undo. A person watching from a phone can stop a run, answer its asks, and send it text that may change what it does next. What has to be reversible is decided when the workflow is designed, by keeping the before-state and choosing operations that can be compensated, not by the presence of a steering control. One more product-specific note belongs here: a message sent to a session whose process has crashed may, in some products, be what triggers the session to be served again, so sending "stop" to a dead run is not always inert.
Carry this forward
- Identities live in several places: the harness's credentials on the host or from a broker, the person's login to the viewing service, and each service's own account. Reaching a host, being authorized to log in, and being the harness's account are different things, and every path is protected.
- Permissions are the harness's rules on the host. What the phone can change is what the product exposes to it; in this case, asks and text only. A plan for remote work says what happens at an ask when nobody answers.
- A model-generated completion notification is a claim. A receipt tied to a recorded event is evidence of that event; being produced by the harness alone does not make every notification a receipt. Carry the item id and a content fingerprint, not a file size.
- Steering input reaches only the future of a run. Reversing a completed effect is a compensating action, not an undo, and what can be compensated is designed in before the run.
Lesson check
Choose an answer for each question.
This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.
The drawing, the multiplexer, the four cases and the limits of control together make a plan. The last lesson writes it down, asks how it can be checked without touching the real run, and reads two guides to one product that disagree.
Sources · Reviewed 2026-09-29
- 2. Anthropic, Claude Code documentation: Continue local sessions from any device with Remote Control
- 5. OpenAI documentation: Mid-turn steering
Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.