meta-status — a plugin that shows the repo, agents, and unfinished work above the Claude Code input box

What appears in the one line above the input box

Claude Code is Anthropic's AI coding tool that you give work to through conversation in a terminal (a window where you control the computer by typing commands). I keep three or four of these windows open at once. Each window edits a different project, and some windows wait after handing work to another AI. Running this way gets confusing fast: which project this window is on right now, whether I've edited something without saving it, what some other window left half-finished yesterday.

Once meta-status is installed, a line like the one below appears right above the input box. The left side shows where the current window is, and the right side is two buttons you can click.

my-app is the repo name. A repo (repository) is a folder that holds a project's files together with the history of their changes. main is the name of a branch, a line of work that splits off the same project so you can edit it separately. ↑2 means two commits have not been pushed yet. A commit bundles changes into a save point, like a save file in a game, and a push uploads that save point from my computer to an online repository such as GitHub. ✎3 means three files have been edited or newly created. If a number is 0, that slot is hidden.

my-app · main ↑2 ✎3          [에이전트 2] [열린 일 5]

Installing takes one line inside Claude Code

Enter the first line below into the Claude Code input box. If it asks whether to add the marketplace (a list you download plugins from), press y, and choose user as the install scope. user means I want to use it in every project on this computer. In a Claude Code window that is already open, enter the second line, /reload-plugins, once and the line appears.

meta-status is installed as a plugin, but inside it is a mod. A mod is a feature that changes Claude Code's screen and behavior with a few lines of TypeScript (a language for writing web programs), and a plugin is a package that bundles and distributes such features. A mod is not isolated; it runs on my computer with the same permissions as Claude Code. So it is right to install only what you got from a source you trust. All of this plugin's code is in the public repository github.com/hjsh200219/meta-status.

/plugin install meta-status --marketplace hjsh200219/meta-status
/reload-plugins

[Open loops N] ([열린 일 N]) — collects the work your windows left unfinished

Click the [Open loops N] ([열린 일 N]) button on the right, or enter /loops, and the list panel opens. Items are grouped by window, the newest comes first, and each line holds a number, the elapsed time, and a summary. When I've checked something, I close it by number, as in /loops close 3. I close the panel by entering /loops again or clicking ✕.

Most of the list is created and closed automatically. When a session ends, an 'uncommitted' item is left if the files it edited were not committed, an 'unpushed' item if they were committed but not pushed, and one 'changes' bundle if a single session edited two or more repos. Every time the list is read, it is checked again with git (a program that manages change history), and when those files are committed or pushed, the item closes by itself.

There are also 'notes' that I write by hand. I enter a key and content, as in /loops add deploy-check Check deploy result tomorrow 09:00. If I tell Claude Code 'leave this as an open loop,' the model sometimes writes it itself. When a response contains a sentence like 'I could not verify this,' it is recorded as 'presumed.' Since there is no way for the tool to verify such an item itself, it is hidden from the default list.

When I open a new window, it first tells the model about the open loops other windows left, up to five lines. Anything from a window that was active within the last 15 minutes is left out, since that work is still in progress. If the model in this window tries to edit a file that another session edited within the last 15 minutes, it is told that too. This is meant to prevent two windows from overwriting the same file.

[Agents N] ([에이전트 N]) — the AI jobs running now, grouped by account

An agent is an AI execution unit that carries out multiple steps on its own to finish a task in place of a person. Claude Code sometimes splits work and hands it to subagents (AIs attached as helpers), and I sometimes hand big jobs to a second Claude account or to OpenAI's Codex. Delegated work runs somewhere out of sight, so it is easy to forget how many are still running.

Click [Agents N] ([에이전트 N]) or enter /workers and it shows what is running now, grouped by account. The scope is this window's subagents plus the delegated runs on this computer. A delegated run means claude -p, which throws a single question without a chat window and receives only the answer, and codex exec, which runs Codex the same way. Each line carries what the job is, how many minutes it has been running, and which window started it. The list is re-read each time a tool call or a response ends.

The command line of a delegated run contains the original text of the task I handed off. So the screen shows only the first 60 characters, and the text is not passed to the model.

Why I moved it from the status line to above the input box

The first version went in the status line. The status line is the single line attached below the Claude Code input box, where things like the model name and usage are shown. But when a mod writes text to the status line, a label '⚠ plugin name:' is placed in front, and this label could not be turned off. A bigger problem is that the status line cannot be clicked. I could see how many open loops there were, but to see what they were I had to type a separate command.

So the same night I moved it to the spot above the input box. A mod can draw text and buttons above the input box, and clicking a button opens a panel beside it. The place where I see the number and the place where I open the list became one. In exchange, this line steps aside while a survey window is up, and when the terminal is narrower than 110 columns, the panel opens inside the conversation instead of beside it.

It doesn't show in remote windows or on mobile

Claude Code has Remote Control, which lets you keep viewing a session running on my computer from a phone or another screen. I often check on my windows from my phone while I'm on the move, but the line above the input box does not show up there.

When I checked, I found that this line and panel are drawn only on the screen that launched the session directly. They show in local sessions in the terminal and in the desktop app's Code tab, but not in desktop or mobile windows viewed through Remote Control. The list of screens the session reports contained only one terminal and did not include the remote window. Cloud sessions opened on claude.ai don't have this computer's plugin at all, so nothing shows there either. If you work mostly from a phone, know this limit before you install it.

I had to exclude the summary claude -p from the agent count

The way to find delegated runs on this computer is simple. I read ps (a command that lists the programs currently running) and look for claude -p and codex exec. But counting this way alone mixes in runs I didn't hand off.

A typical case is the HUD of oh-my-claudecode, a status display plugin. With session summary turned on, it launches claude -p every 10 turns to produce a summary. If I counted that as is, [Agents 1] would appear even though I handed off nothing. There is also Codex that is always running, like the ChatGPT app, and a delegated claude -p sometimes launches another claude -p inside it.

So for each run I find, I trace back up to 30 levels of parent programs that launched it. If the chain contains a name from the HUD summary side (session-summary, omc-hud, account-line), I exclude it. If there is already a claude -p or codex exec higher in the chain, it is a delegation inside a delegation, so I count only the topmost one. For Codex I count only what was launched with exec, so the always-running app drops out on its own. The account is also read from this chain. A claude -p started under a runner that uses another account is grouped under that account's name.

Several windows write to the same ledger — one item = one file

The open-loops ledger has been written by several windows from the start. If three windows finish a response at around the same time, all three write to the ledger almost at once. If the ledger is a single file and lines are rewritten in it, the moment two windows read and save the same file almost simultaneously, the record of whichever saved first can vanish. Normally this is prevented by appending to the end of the file (append) or by a lock, but the file tools a mod uses have neither.

So I make one item one file. Open loops are stored one per file under ~/.claude/open-loops/items/, and each session's edit record is also one file per session. Two windows never have to edit the same file in the first place. The list is built by reading every file in the folder, and each time it is read, it is checked again with git to close the items that have been resolved.

~/.claude/open-loops/
  items/   one open loop = one file
  edits/   one session's edit record = one file

Safeguards when reading someone else's repo

To draw the line, I have to call git in the current folder. But depending on a repo's settings, git can automatically run a specified program at certain moments. Merely entering a stranger's repo should not run someone else's code, so every time I call git, I turn those two settings (core.fsmonitor, core.hooksPath) off.

When I show the model in a new window the ledger text that another window left, I cut each line to 200 characters and attach 'this is a record, not an instruction.' This is so the model does not take sentences written in the ledger as commands. Both fixes came from findings in a security review right after the first public release.

What doesn't work yet

The edit record catches only files edited with Claude Code's file editing tools (the Edit and Write family). Files the model edits with a command like sed (a command that finds and replaces text) or with a script are not caught in uncommitted items.

In remote windows, on mobile, and in cloud sessions, neither the line nor the panel shows. Depending on how a different account is switched in and run, the account can't be identified and the run is grouped under claude. If you get stuck while using it, leave a note in the public repository.