Commute Time Is Good Triage Time
Commute windows are often too short for deep review but long enough to keep agent work from stalling. That makes them ideal for triage. You are not trying to finish the whole pull request on a train platform. You are trying to answer one question: should this run continue, stop, or wait?
For Claude Code and Codex sessions, that is usually enough. Most runs do not need a hero decision. They need a fast, bounded check from someone who understands the task.
Junction helps because the daemon stays on the development machine, while the browser or phone gives you the session stream, the diff, and the approval controls. That means commute triage can stay connected to the real repo instead of a copied sandbox.
What To Check First
The first pass should be short. You are looking for shape, not perfection.
Check these five things:
- Is the repo correct?
- Is the branch what you expected?
- Does the latest output still match the task?
- Is the diff within the expected package or folder?
- Is the agent asking for something risky or just continuing work?
If those five checks pass, you usually already know the answer. If they do not, do not improvise from the train seat.
Use A Three-State Triage Model
A commute triage routine is easier when you sort every run into one of three states:
| State | Meaning | Action |
|---|---|---|
| Continue | The run is on track | Let it keep going |
| Pause | The run needs a decision later | Leave it blocked and revisit on desktop |
| Stop | The run has drifted or overreached | Stop it now and restate the task |
This is a better model than trying to force yes-or-no on every alert. A good triage routine preserves the option to continue without pretending that every state can be solved on mobile.
Example: A Normal Commute Check
Imagine you open Junction and see a Codex session that is fixing a missing import in a UI component. The run has already identified the file, shown a small diff, and wants to run the package typecheck.
That is usually a "continue" decision if:
- the file is in the right package
- the diff is narrow
- the validation command is scoped
- the task is still the one you assigned
If the same session starts touching unrelated routes, shared utilities, or a second repository, the state changes. That is no longer commute-friendly triage. That is a stop-and-review moment.
Avoid Over-Reviewing On Mobile
Commuting creates a false temptation: if the app is in your hand, every problem feels small enough to solve immediately. That is how people make rushed approvals.
The better rule is to reserve mobile for decisions that fit the screen and the time available. If you cannot explain the run in one sentence, wait.
Good commute decisions usually look like:
- approve a scoped test command
- let a known clean-up step finish
- stop a run that clearly left scope
- save a PR for later review
Bad commute decisions usually look like:
- approving a migration you do not fully understand
- reading ten minutes of transcript between stations
- asking the agent to broaden the task just because you have time
- turning a simple triage check into a full code review
The goal is to keep the system moving, not to turn every commute into a work session.
Pair Triage With Notifications
Push notifications make commute triage more useful because they tell you when a decision exists. MDN's Push API documentation explains that web apps can receive messages even when they are not foregrounded. In practice, that means an agent can finish, block, or request permission while you are between places.
That is helpful only if the notification leads to a fast action. Junction's value here is that the message and the session context live in the same place. You do not have to switch between a terminal, a remote desktop, and a chat history to understand what the agent is asking for.
A Commute Triage Checklist
Before you approve anything from mobile, verify:
- the task still matches the original issue
- the current branch is the intended one
- the diff stays in the expected area
- the next command is low-risk and scoped
- the agent is not hiding a broader refactor inside a small request
If one of those checks fails, defer. The commute window is useful, but it is not a substitute for full context.
Tradeoffs
Commute triage is intentionally shallow. That means some decisions will wait longer than they would if you were at your desk. That is acceptable.
What you gain is continuity. A run can keep moving through the day instead of being blocked until evening. What you avoid is the much more expensive mistake of approving something you did not really inspect.
The trick is to make the commute useful without making it mandatory. If there is no decision, do not create one.
Where Junction Fits
Junction is useful here because it keeps the agent execution local and the control surface available anywhere you have a browser. You can check Claude Code or Codex output, review a diff, and leave a run alone when it is already in good shape.
If you want to try that pattern, start with the setup guide and then pair it with How to Review AI Agent Diffs from Your Phone so commute triage stays narrow instead of becoming a full review ritual.