AI coding agents work better when someone does a small amount of operational housekeeping every week.
That sounds mundane because it is. It is also the difference between a clean local workflow and a pile of stale sessions, half-reviewed diffs, noisy notifications, and branches nobody wants to own.
This checklist is for teams using Claude Code, Codex, or both inside Junction.
The weekly goal
The goal is not perfection. The goal is to keep the system understandable.
You want to know:
- what is still active
- what is blocked
- what can be archived
- what needs approval
- what changed on each branch
- whether any daemon or auth state needs attention
If you can answer those questions quickly, the rest of the workflow stays lighter during the week.
1. Clear finished runs
Start by archiving or closing runs that are already done.
Completed sessions are easy to ignore because they are no longer urgent. But if they stay active in your mental stack, they keep adding noise. A clean history is easier to search, and a smaller active queue is easier to supervise.
This also helps when you revisit a task later. The history of what happened is more useful when the finished work is not mixed in with the live work.
2. Review blocked work
Next, inspect anything that is waiting on permission, missing context, or stalled on a decision.
Blocked runs are where drift often starts. If a run has been waiting too long, it might need:
- a narrower prompt
- a clearer acceptance check
- a different owner
- a stop and restart
Do not let blocked work accumulate silently. It tends to become harder to interpret the longer it sits.
3. Check diffs before they accumulate
Look through the current diffs and ask whether each one still matches the original task.
If a diff has grown beyond the intended scope, split it now. Weekly review is a good time to catch work that became too broad but still looks "close enough" to keep going.
That is especially important for runs that touched shared state, because a small unresolved change can create bigger cleanup later.
Junction's built-in diff review is valuable here because you can inspect the change from the same place where you see the run history and approvals.
4. Triage notifications
Notification hygiene matters more than people expect.
If every alert feels urgent, none of them are.
Use the weekly sweep to decide which notification patterns are useful and which are just noise. If a run only needs attention when it finishes, keep that signal lightweight. If it needs a human to approve a destructive or shared-state action, that signal deserves a higher priority.
The point is to keep interruption reserved for the events that actually matter.
5. Reconcile branches and worktrees
A weekly review should include branch cleanup.
Look for:
- branches that are ready to merge
- branches that should be abandoned
- worktrees that are no longer needed
- branch names that no longer match the session they came from
This matters because the cost of a messy branch state compounds. The more agent work you run, the more you benefit from a steady cleanup habit.
If your team uses isolated worktrees for agent runs, this step keeps the workspace from drifting into clutter.
6. Check daemon health and auth
If you have more than one daemon, confirm that each machine still looks healthy.
That means verifying:
- the machine is still reachable
- the provider auth is still valid
- the correct repo or branch is connected
- the daemon is still associated with the right work
This is not glamorous work, but it prevents a lot of avoidable friction later. Junction's multi-daemon model is useful only if the routing state stays trustworthy.
7. Revisit what should be automated next
Once the active queue is under control, decide whether any repeated task has become a better automation candidate.
The weekly sweep is a good time to ask:
- did we see the same kind of work multiple times?
- did one of those runs go smoothly enough to standardize?
- was there a manual task that now looks stable enough for Switchboard?
That keeps the automation program tied to real behavior instead of guesswork.
A practical weekly order
You do not need to spend all morning on this. A focused pass is usually enough.
One simple order is:
- archive finished runs
- inspect blocked work
- review active diffs
- clean up branches and worktrees
- check notifications
- confirm daemon and auth state
- pick one automation candidate for the following week
That sequence keeps the operational cost low and the active queue understandable.
Why this works
The weekly checklist works because it breaks maintenance into small decisions.
You are not trying to solve every workflow problem in one sitting. You are just making sure the local agent system is still legible. That is what keeps the rest of the week moving.
It is also a good fit for a browser-based control surface. You can do the sweep without sitting in the terminal, which matters when the goal is to supervise local agent work from wherever you happen to be.
A final practical rule
If a weekly review reveals work you do not understand, stop and reclassify it before you leave it alone for another week.
That one habit prevents a lot of future cleanup.
If you want to make the checklist easier to run, start with the setup guide and then compare pricing. For a related workflow, read What an AI Coding Agent Dashboard Should Actually Do.