I plan in Claude Code and hand the implementation to other models through a small dispatcher of my own. An order runs anywhere from 2 minutes to 2.5 hours. My way of knowing whether one had finished was to remember to ask.

Claude Code now has mods, small plugins that can draw a pane beside the chat. So I wrote one that shows the queue: dispatch-board, 1,330 lines of TypeScript, MIT.

In flight 2/4 · queued 1 · needs you 3 · updated 12:01:22 Needs you (3) 20260107-101833-fix-i… · billing-ui · gpt-mini · check_failed · r2 · 41m 20260107-091204-add… · atlas-api · gpt-large · needs_review · r1 · 2h14m 20260106-221540-migrat… · docs-site · local-qwen · timeout · r1 · 13h00m In flight (3) 20260107-112250-table-pa… · billing-ui · gpt-mini · reviewing · r2 · 27m 20260107-113902-split-auth… · atlas-api · gpt-large · running · r1 · 12m 20260107-114417-changelog-l… · docs-site · local-qwen · queued · r1 · 4m Lane budget, this month gpt-large · $72.60 / $70.00 · 103% gpt-mini · $12.90 / $15.00 · 86% claude-review · $6.10 / $30.00 · 20% local-qwen · $0.42 / no cap Finished in the last 24h: 7 

That is the built-in demo data, not my orders. The real pane told me two things in its first minute that I did not know: 9 orders were waiting for a decision from me, some for more than two weeks, and one lane was already over its monthly cap.

Why not use an existing mod

I started from lot-map, which draws a live map of what a session is doing. But lot-map holds to a principle I respect: no network, no files, no processes. It draws the session's own events and nothing else. Work I hand to another model leaves the session the moment the submit command returns, so it is exactly what lot-map will never show. That needed a second pane, one that reads something outside.

The status command was not read-only

The plan was ten lines: run status --json every 10 seconds, map the rows, draw them.

Before pointing a 10-second timer at a command, I read what the command does. Mine reconciles. Every status call, in the list form too, checks each running order's runner process, and if the process is gone it rewrites the order's state file, and for an order under review it signals the review worker's process group first. That is the right thing for a command I type by hand. Polled every 10 seconds from four open sessions, it makes the dashboard a second writer next to the runner.

I had believed the list form was safe and only status <id> reconciled. The source said otherwise in one line.

The list had a second problem. It prints the newest 20 orders, and my newest 20 were all finished. The 9 that needed me were older and fell off the bottom.

So the board does not run the command. It reads the state files: one directory listing, then the JSON of each order. The dispatcher writes them with a temp file and an atomic rename, so a read sees a whole old file or a whole new one, and the adapter keeps the last good row if a file ever fails to parse.

What it costs

Measured on my machine against 299 real orders, 289 finished and 10 open:

Time
The status command I did not use 554 ms
First poll of a session, every state file once 37 ms
Each later poll, one listing plus the open orders 1 to 4 ms
Drawing the pane 2 to 7 ms

A finished order cannot change, so after the first poll the board only re-reads the open ones. The cost of a poll follows the 10 open orders, not the 299.

It polls every 10 seconds while something is in flight and once a minute when nothing is, and not at all until I open the pane. The pane is redrawn only when what it shows has changed, which is why durations are shown to the minute. With several sessions open, they share one snapshot through the plugin's store, so four panes fetch about as often as one. When a real order moved from review to waiting on me, the pane showed it 10.5 seconds after the state file changed.

Monthly spend per lane is summed from the cost of each attempt in the same state files. I checked it against the dispatcher's own stats command and it matched for every lane to four decimals.

What it will not show

A mod runs with your permissions and can see every tool call, so I wrote down what this one may do before writing it. It hooks three events: session start, its own slash command, and the drawing of its own pane. It adds nothing to the conversation, calls no model and makes no network request. A row holds an id, a directory name, a lane, a state and numbers. Titles are off by default and cut to 40 characters when on.

31 tests hold it to that. The four I would not ship without: a command that fails, hangs or prints half a JSON document leaves the session's own events untouched; nothing but those fields reaches a label, whatever the input carries; a poll that changes nothing redraws nothing; a thousand finished orders draw as many elements as five.

I have looked at it in a terminal and in the VS Code extension, both on macOS. The test kit accepts the same tree for the desktop app and mobile, which is not the same as having seen it there.

If your dispatcher is not mine

It almost certainly is not. The board takes any command that prints this:

{ "orders": [ { "id": "…", "project": "atlas-api", "lane": "gpt-large", "state": "running", "round": 1, "startedAt": "2026-01-07T11:39:02Z", "updatedAt": "2026-01-07T11:50:41Z" } ], "lanes": [{ "name": "gpt-large", "costMonth": 72.6, "budgetMonth": 70 }], "capacity": { "running": 2, "max": 4 } } 

The command is run as an argument list with no shell and killed after 5 seconds. Before you point it at yours, do the check I nearly skipped: read what your status command does when a worker has died. If it repairs anything, it is not a command to poll.

Install, at the prompt of a terminal session:

/plugin install dispatch-board --marketplace lizhuojunx86/dispatch-board 

Version 0.1 only displays. Accepting or redoing an order still happens in the dispatcher, and I have not decided whether a button in a pane should be allowed to merge code.

This follows two earlier threads of mine. The token audits started with a missing model line that was half my agent overspend, and the routing one found that my routing policy and my traces disagreed 96 times. Those were about measuring after the fact. This one is the same numbers while the work is still running.

If you write an adapter for another orchestrator, or find a state your dispatcher has that the board sorts wrongly, tell me. I'll add it.

Zhuojun Li