Waitroom

STILL GUESSING?
Bet while your agent thinks → LANDED → review the diff. Four Claude Code rooms for dead time.

   

Features  ·  Install  ·  Demo  ·  Contributing

Demo

A real session, not a mockup — Claude Code 2.1.304 with both plugins, sped up only while Claude works; this is a 15-second cut of it. Bet on the turn while it runs, read the diff as it lands, flag a hunk with a reason, and "Fix what I flagged" is all the next message needs. Full-length video.

Don't fill the wait with your phone — spend it on the diff and your call on the turn (local Claude Code plugins, no telemetry).

Features

The wait is the one moment you are already holding the problem in your head and have nothing to do with it. The rooms spend it on the diff the agent just wrote, on how well you called the turn, on the commit message and PR description that will ship it, and — quietly, off by default — on a long turn's toll on you.

• 🔍 wait-review — The diff Claude already wrote, a hunk at a time. What you flag goes back with your next message, with your reason. A pane. On.
• 🎯 bet-on-your-agent — Your call on how the turn will go, scored afterwards: a Brier score, and a calibration table that shows where your confidence runs hot or cold. No currency and no winnings. The row. On.
• 🚢 wait-ship — A drafted commit message, PR description, and changelog entry, to review the way wait-review lets you review the code. The row. On.
• 🌬️ wait-breathe — One quiet line during a long turn: a posture check, an eye break, a breath, a stretch. The row. Off until /breathe.
• 📋 wait-changed — The files this turn changed, listed once the turn has ended. Read-only. Not shown during the wait, and not shown when nothing changed. Not in the four-room install. The row. On.
• 🔔 wait-ping — The handoff. A local notification when a long turn has ended, or when a permission confirm is waiting. The OS. On. Not in the four-room install.

One family, four rooms

wait-review keeps a pane beside the transcript. The others take turns in the one row above the prompt. wait-changed is not a fifth way to spend the wait: it uses that same row only after a turn ends, and only to list the files that turn changed. wait-ping draws nothing in the session. When a turn runs long, or a permission confirm is waiting, it sends a notification through the operating system on this machine.

	Plugin	Spends the wait on	Where	Starts
	wait-review	The diff Claude already wrote, a hunk at a time. What you flag goes back with your next message, with your reason.	A pane	On
	bet-on-your-agent	Your call on how the turn will go, scored afterwards: a Brier score, and a calibration table that shows where your confidence runs hot or cold.	The row	On
	wait-ship	The words that ship the change: a drafted commit message, PR description and changelog entry, to review the way wait-review lets you review the code.	The row	On
	wait-breathe	You. One quiet line during a long turn: a posture check, an eye break, a breath, a stretch.	The row	Off until /breathe
	wait-changed	The files this turn changed, listed once the turn has ended. Read-only. Not shown during the wait, and not shown when nothing changed.	The row	On
	wait-ping	The handoff. A local notification when a long turn has ended, or when a permission confirm is waiting.	The OS	On

Repository

wait-review/                  pane
bet-on-your-agent/            row above the prompt
wait-ship/                    row above the prompt
wait-breathe/                 row above the prompt · off until /breathe
wait-changed/                 same row, after the turn · separate install
waitroom/                     four-room install
docs/demo.gif                 session capture used as the README demo
docs/demo.mp4                 full-length cut of that session
docs/brand/                   pixel README lockup
ads/assets/                   pixel promo placeholders
wait-ping/                    local notification · separate install
wait-review-opencode/         unpublished OpenCode prototype
bet-on-your-agent-opencode/   unpublished OpenCode prototype

Install

claude plugin marketplace add coindef/waitroom && claude plugin install waitroom@waitroom

Star it on GitHub if it earned it.

That installs wait-review, bet-on-your-agent, wait-ship, and wait-breathe.

Optional — turn-end file list (not in the four-room install):

claude plugin install wait-changed@waitroom

After a turn that changed files, lists them above the prompt. Read-only. Skip if you only want the four rooms. What it shows.

Optional — local notification when a long turn ends (not in the four-room install):

claude plugin install wait-ping@waitroom

A turn under 30 seconds ends quietly. Sound stays off until you turn it on. What it sends.

Then start Claude Code (restart it, if it was already running). Each room runs on its defaults — no key, no service, nothing written to your repository. wait-breathe stays off until you send /breathe. The installer's userConfig options not yet set lines are fine: every option has a working default (more).

Already at the Claude Code prompt? Send two slash commands, each as its own message — /plugin marketplace add coindef/waitroom, then /plugin install waitroom@waitroom — and save the install screen as it is.

Both halves are safe to run again: adding a marketplace that is already there, or installing a plugin that already is, changes nothing. With the four rooms in, /plugin reads 4 mods active · wait-review, bet-on-your-agent, wait-ship, wait-breathe.

Or install one room

claude plugin install wait-review@waitroom
claude plugin install bet-on-your-agent@waitroom
claude plugin install wait-ship@waitroom
claude plugin install wait-breathe@waitroom

The marketplace add in the line above still comes first. At the prompt, each of those is its own message: /plugin install wait-review@waitroom, and the same shape for the other three.

Every plugin here needs Claude Code 2.1.291 or later — claude --version says which you have, claude update moves you on. An older build installs it without complaint and then shows it as failed to load; see If a plugin shows an error.

The one-liner adds the marketplace from its repository path, coindef/waitroom. From then on the marketplace goes by its name, waitroom, the part after the @. waitroom@waitroom is this four-room install.

You may also see /plugin install waitroom --marketplace coindef/waitroom. It works once. Sent again, it answers Plugin "waitroom" not found in any marketplace, because by then --marketplace wants the name waitroom, not the repository path. The one-liner above has no such trap.

Updating

Run these at the prompt, then restart Claude Code — an update is written to disk straight away but only loaded by the next session:

/plugin marketplace update waitroom
/plugin update waitroom@waitroom
/plugin update wait-review@waitroom
/plugin update bet-on-your-agent@waitroom
/plugin update wait-ship@waitroom
/plugin update wait-breathe@waitroom
/plugin update wait-changed@waitroom
/plugin update wait-ping@waitroom

/plugin update waitroom@waitroom picks up a room newly added to that install. Each room still updates on its own line. Each reports updated from the old version to the new one, or already at the latest version for a mod that has not changed. What changed, and in which mod, is in CHANGELOG.md; if one it lists as changed still says it is at the latest version, the marketplace was not refreshed yet, so run the first line again.

Coming from the old name

The repository was called Plugin, and the marketplace wait-mods (in the very first release, wait-review). If you installed under one of those, the stale marketplace is still registered and has to go before the new one can take its place. /plugin marketplace list shows which you have. Removing a marketplace uninstalls the plugins that came from it, so remove the old one, then install again:

/plugin marketplace remove wait-mods
/plugin marketplace add coindef/waitroom
/plugin install waitroom@waitroom

(From the first release, the first line is /plugin marketplace remove wait-review instead.) Leave the old one in place and one of the two copies reports holds the same plugin name and comes first — see If a plugin shows an error.

bet-on-your-agent's record does not come with it: each install keeps its own, so your calibration starts fresh under the new name.

Screenshots

Real captures from this build, not pixel mockups. The walkthrough is in The first minute.

Requirements

• Claude Code 2.1.291 or later, run interactively — the current stable release, 2.1.304, is fine. Checked live on 2.1.291, 2.1.292, 2.1.300, 2.1.302 and 2.1.304. Older builds do not run them, and claude update is the fix:
  • before 2.1.242, the plugins show failed to load (below);
  • 2.1.242 to 2.1.280 install them but reject their hooks module, which uses state helpers those builds predate (claude plugin validate says so on 2.1.242, 2.1.260 and 2.1.280);
  • 2.1.281 to 2.1.290 accept the module but were not run live.
• A git repository for wait-review and wait-ship. Both read the diff, so outside one they have nothing to show. bet-on-your-agent and wait-breathe work anywhere.
• Plugins that hook Claude Code's own events are early access. See If a plugin shows an error.

The first minute

Each room shows up by itself the first time it has something to say. Every image below is a real capture of this build: Claude Code 2.1.311 in a scratch repository, cut from the terminal as it was, never redrawn.

wait-review · diff review hunk by hunk

wait-review needs no setup. Ask Claude for a change. When the turn ends, the pane starts offering the diff a hunk at a time. Here it is paging through a hunk longer than the pane, with the one before it already flagged:

Give the pane the keys first: /wait-review or ctrl+x tab (or click it). Its letters only work while it holds the keyboard — press o without that and you type an o into your prompt. The pane reminds you at the bottom, and esc hands the keys back when you are done. /wait-review always works; ctrl+x tab gives the keys to the row above the prompt instead while that row has buttons up, as it does when bet-on-your-agent asks "Did it land?".

Key	Does
o	Looks right — next hunk
f	Flag it; it goes to Claude with your next message
n	After f: say why, in a line
l	Later — comes back once the rest are read
x	Skip the rest of this file
u	Undo the last call
j / k	Read on / back through a hunk longer than the pane

A new file is a single hunk as long as the file, so j matters more than it looks: the pane shows lines 1–30 of 200 and pages through the rest.

Running /wait-review opens the pane already holding the keyboard, so the letters answer at once. The pane that opens by itself at the start of a session deliberately does not take the keys: it would eat the first keystrokes of a prompt you have not sent yet. The first time that happens it says so, in bold, Focus this pane: /wait-review · then o / f work, and that line does not come back after you have run /wait-review or pressed o or f with the pane focused.

If the pane never appears, your terminal is under 144 columns — a pane nobody asked for needs that much room. Run /wait-review and it opens at any width, even while a turn is still going. The status line tells you so:

wait-review: 3 to read · terminal too narrow — /wait-review opens anyway

On Claude Code 2.1.304 or later, Claude Code has a diff panel of its own, and it opens beside the transcript the first time Claude edits — in front of wait-review's pane. Send /diff once to hide it. Claude Code remembers that, so the side is wait-review's from then on, in later sessions too; /diff again brings the built-in panel back whenever you want it. If you run /wait-review and see nothing, its reply says Claude’s /diff is covering this side — send /diff once to show wait-review. The first time a turn writes a diff while that panel is covering the pane, the transcript says once: Claude's Diff is covering Waitroom. Send /diff once to show wait-review — Claude remembers.

bet-on-your-agent · forecast agent turns

bet-on-your-agent scores your calibration, not a wager. It keeps no currency and no winnings, just whether you called it. While a turn runs, it puts a question above the prompt:

Type 1 to 5 into the empty composer and it answers: no pane, no focus, no click. 0: Skip is the exception. A bare 0 stays in the composer (seen live on 2.1.311), so give the row the keys first with ctrl+x tab, or click it. /bet shows what it has learned about you. In your first week, the forecast and the file count are all it asks. The extras (a bonus question, a daily challenge, a badge wall) wait in /bet-recap until then.

wait-ship · draft ship artifacts

wait-ship needs no setup either. When a turn ends having changed anything, the same row offers the kit it drafted: a commit message, a PR description, and, where CHANGELOG.md exists, a changelog entry:

Its keys are letters, so give the row the keys first (ctrl+x tab, or a click, as the row's last line says). Then a takes every draft and s skips. m puts each draft on a key of its own (c commit message, p PR body, l changelog, e drafts for editing). A letter typed into the composer stays in the composer; only bet-on-your-agent's digits 1 to 5 answer from there. Accepted drafts land under .waitroom/ship, a folder that ignores itself in git, with the exact command to run printed in the transcript — nothing here ever runs on your behalf. A turn that changed nothing gets no row, even with earlier changes still uncommitted.

wait-breathe · long-turn wait

wait-breathe is the one plugin that starts off. Send /breathe to turn it on (see Install above); it answers wait-breathe: on · one quiet nudge once a turn runs 90s · /breathe again turns it off. Once it is on, a single turn running past longTurnSeconds (90s by default) gets one quiet line, never more than one every ten minutes:

No health claim and no "you should": just the action, rotating so it is never the same one twice running. Its tag stays grey, the quietest thing on the screen. /breathe turns it on, and off again; the same switch is Enable the breathing nudges under /plugin → wait-breathe.

wait-changed · what this turn changed

wait-changed needs no setup. When a turn ends having changed files, the row above the prompt lists them under What changed this turn, with Read-only local git — nothing uploaded under the title. Up to eight files, then N more files…. No hunks, no commit, no PR. Close clears this turn. Don't show again turns the card off; /wait-changed turns it back on. A clean turn, a turn still running, and a folder that is not a git repository all stay silent.

How they share the screen

They compose rather than compete for room. wait-review takes the pane, and the others share the row just above the prompt, one at a time: a breathe nudge first, then — once a turn has ended and it changed files — wait-changed's list, then a ship offer, then the bet's question. Two sets of buttons never stack there; the one behind waits its turn. Close on the list is what gives the ship offer the row. While a turn runs, the bet's forecast gives way only to a breathe nudge, since the window to call the turn closes when it lands, and the diff card is not drawn yet.

The idea behind all four

The cost of waiting on a coding agent is not boredom, it is context collapse. A 90-second wait filled with something entertaining costs minutes of refocus on the way back, and any game with a win condition quietly teaches you to want the agent to be slow. You also cannot win on entertainment: the person already has a phone.

So none of these compete with a phone. Each spends the wait on something that is only available during the wait, needs no context switch to start, and can be abandoned mid-thought:

• wait-review spends it on the diff, because the agent now writes code faster than anyone reads it and a session otherwise ends with a large diff nobody reviewed.
• bet-on-your-agent spends it on a forecast, because knowing how far to trust the agent is the skill that matters most and is currently learned by vibes.
• wait-ship spends it on the words around the diff, because a commit message and a PR description are worth getting right too, and writing them well after the fact competes with whatever comes next instead of riding the one moment the change is already fresh in mind.
• wait-breathe, alone of the four, spends it on you rather than the work — the same wait that is free for the others to fill is the only unscheduled moment in a long session. It is off by default because not everyone wants a plugin making that particular claim on their attention.

bet-on-your-agent asks what you expect from what Claude is writing, wait-ship hands you the words for what it just wrote, and wait-breathe — once you turn it on — asks something of you instead of Claude.

If a plugin shows an error

failed to load, or ✗ 1 error on its row in /plugin, most likely means Claude Code is too old. The error itself reads like this:

Hook load failed: [ { "expected": "record", "code": "invalid_type",
  "path": [ "hooks" ], "message": "Invalid input: expected record, received undefined" } ]

Builds before 2.1.242 cannot read the hooks module these plugins are built on, so every plugin built that way fails to load the same way — including ones from other marketplaces. Run claude update, restart, and they load. (Reproduced on 2.1.200 and 2.1.241; 2.1.242 is the first build that reads it.)

Otherwise, check whether they are running at all. Open /plugin: when all four are loaded, the line under the tabs reads

4 mods active · wait-review, bet-on-your-agent, wait-ship, wait-breathe

If one is missing from it, the Errors tab in the same menu names the reason — so does claude --debug, on a line starting hooks module wait-review@waitroom not loaded:. What it can say:

It says	What it means	What to do
Its hooks module does not load. "wait-review@…" holds the same plugin name and comes first	Two copies are installed — almost always one from the old wait-mods marketplace beside the new one	/plugin marketplace remove wait-mods removes the old copy with its marketplace. See Coming from the old name
hooks modules are turned off for installed plugins in this process: …	The early-access switch, below	Depends on the rest of the line
only managed plugins and built-in plugins run (allowManagedHooksOnly / disableAllHooks)	Your organization's managed settings allow only its own plugins' hooks	Ask whoever manages Claude Code for you
… load no hooks module in this mode (--bare)	Claude Code was started with --bare	Start it without --bare

No error anywhere, and no mods active line at all, means disableAllHooks is on in your settings. It turns off every hook, these included.

The same "holds the same plugin name" error appears if you run claude --plugin-dir ./wait-review while the installed copy is enabled. Disable the installed one while you work on a clone.

The early-access switch

Plugins that hook the engine's own events are early access, gated by a rollout switch. The end of the line says which state it is in:

• this wake's flag fetch has not answered the rollout switch is the harmless case. The process started before the answer arrived, and the mods load at the next turn start once it does — the debug log then shows … loaded (worker, environment 4, tier user). A short claude -p run can end in that gap; an interactive session never does.
• the rollout switch was saved off by an earlier session and is not refreshed yet is a stale answer. Start claude once with network access and it refreshes.
• the rollout switch served off means the switch is off for this build or account. Nothing in these mods can turn it on, and they load by themselves once it reads on.

Either way the switch is not about these mods in particular: it holds back every installed plugin that carries a hooks module, whoever made it.

What the installer says on the way in

"4 userConfig options not yet set" is informational: every option has a working default and the plugin runs on them — the debug log puts it as every option is its default. Configure them to change something, not to get started.

Contributing

Each mod is a self-contained plugin folder. From the repository root:

claude plugin validate ./wait-review
claude plugin test     ./wait-review          # 101 tests

claude plugin validate ./bet-on-your-agent
claude plugin test     ./bet-on-your-agent    # 130 tests

claude plugin validate ./wait-ship
claude plugin test     ./wait-ship            # 70 tests

claude plugin validate ./wait-breathe
claude plugin test     ./wait-breathe         # 28 tests

claude plugin validate ./wait-changed
claude plugin test     ./wait-changed         # 32 tests

claude plugin validate ./wait-ping
claude plugin test     ./wait-ping            # 31 tests

claude plugin validate ./waitroom
claude plugin test     ./waitroom             # 6 tests

Neither of those runs a real type checker, so also run tsc from each mod's folder once it has loaded at least once (that's what writes .claude-plugin/types/, which the mod's tsconfig.json points at):

npx -y -p typescript@5.7 tsc -p .

(typescript@latest currently ships a rewritten tsc whose CLI rejects -p outright — pin a 5.x release rather than chasing that down again.) A clean run prints nothing. This caught a real bug once already (two call sites where a status literal silently widened to string inside a .map, erasing the protection HunkStatus was there to give) that 84 passing tests and a clean validate both missed — run it after any change to any mod's hooks/, not just when something looks wrong.

Run a working copy without installing it, or install from this clone so edits reload in place:

claude --plugin-dir ./wait-review
# or
claude plugin marketplace add .     # then /plugin install <mod>@waitroom

Each mod keeps the logic that needs no engine in one plain module (hooks/diff.ts, hooks/bet.ts, hooks/ship.ts, hooks/breathe.ts, hooks/changed.ts, hooks/ping.ts, hooks/tip.ts) so it can be tested directly, with the hooks module beside it and a types/index.d.ts state contract that claude plugin validate holds it to. Each mod's own tests/flow.test.* or tests/band.test.tsx drives the real loop through the engine, git and all.

.claude-plugin/types/ inside a mod is written by the engine on each load and is not checked in.

userConfig conventions

Two option names recur across mods; a new one should follow the same meaning rather than pick a fresh one:

• autoOpen governs whether a mod's pane opens by itself at session start — not whether the mod itself runs. Its default follows whether the pane is that mod's main surface or a secondary record: true for wait-review, whose pane is where it does its work; false for bet-on-your-agent, whose main surface is the row above the prompt, so the pane is only worth the screen space once you ask for it with /bet.
• enabled is a master switch, present only on mods with a core behavior worth turning off entirely — wait-ship, wait-breathe, wait-changed, and wait-ping. wait-review and bet-on-your-agent have no enabled key at all, because neither has a mode where it does nothing; they are always on. Where enabled exists, default it to whatever keeps the mod's main claim on the user modest: true for wait-ship (drafting costs nothing until you act on a draft) and for wait-changed (the list is read-only, and Don't show again turns it off), false for wait-breathe (its claim is on your attention, not your screen). wait-ping's enabled defaults to true for the same reason as wait-changed: it is its own install, not part of the four rooms. Its sound defaults to false, and thresholdSeconds defaults to 30, so a shorter turn does not notify when it ends.

OpenCode prototypes

wait-review-opencode and bet-on-your-agent-opencode are unpublished prototypes of the same two mods for OpenCode's own plugin API. Each reuses its Claude Code counterpart's engine-free core (hooks/diff.ts, hooks/bet.ts) unmodified; each README explains what's ported and what's blocked on OpenCode's API rather than merely unported. bet-on-your-agent-opencode has since been run as a real plugin inside a real, pty-driven OpenCode 1.19.3 TUI session — installed the same way a user would, through the TUI's own "Install plugin" dialog — which is how a real api.kv restart across two processes and a narrow-terminal layout bug (since fixed) were found. wait-review-opencode is still verified by type-checking against the real OpenCode/OpenTUI packages rather than a live session, since exercising it needs a configured model provider this environment doesn't have. Either way: adapter-only, not installable through this marketplace, not versioned, not part of a release — see each README's own "How this was verified" section for the full detail.

License

MIT — see LICENSE. What changed in each release is in CHANGELOG.md.