modz

Orchestrated delivery: the session model plans and reviews, five tiered worker agents implement. Pairs with savvy-progress.

1 skill · 5 agents

Themes for the transcript: tool rows, reply gutters and spinner words, and living ASCII scenes above the prompt. /skin switches them live. It never cuts output short.

1 skill

Holds a shell command that would lose work (rm -rf, find -delete, git reset/checkout/clean/stash drop/branch -D, force pushes, writes over files, docker, terraform, kubectl and database drops) and shows exactly what it would lose. A command that loses nothing runs; one nobody can answer is refused. Also holds an edit to a file another chat just changed.

Shows what a cold send would cost, after a break or a model change, and can keep the one-hour prompt cache warm. Keepwarm starts only on request, stops after 12 pings, and uses tokens.

A silent git snapshot before each turn's first write, from a file tool or a Bash command that may write. /checkpoints lists the turns and undoes one; each session in each worktree sees only its own. Git-ignored files are not covered.

A notification when a long turn finishes or Claude is waiting on you, with optional sound packs. On the desktop it stays quiet unless tests failed or the turn ended on an error.

Shows each turn's effort and model. In auto mode, light prompts go at low effort; a short follow-up goes at your own effort, and a lowered turn has a re-send button.

A file tree of what Claude touches: what it is doing now, the files touched this turn and their git status. Does nothing until its pane is opened.

Saves working notes before every compaction and offers them back after with one press; near auto-compaction it asks once: compact now, or hand off to a fresh chat.

One quiet line above the prompt: the 5-hour limit and its reset, 7-day use, context and session cost; alerts only on the desktop by default (/hud band). The Usage pane shows what fills the context and 7 or 30 days of usage history.

Visual QA for the MODZ kit: /kit-gallery opens a pane drawing every kit component in several states

The MODZ panel. Opens beside the chat with a live line from every installed mod, one click to each mod's own pane.

The plan's steps as Claude works through them, from the task list, an approved plan or a plan file. The plan comes back after /clear and compaction.

A card in the conversation for a PR Claude opens or the branch already has: files with the largest changes on the desktop, checks, review and merge state in the terminal. No poller of its own.

Every project in one place: percent, plans, where it stopped, tokens and agents; /projects and a full-screen board

Spots corrections in your prompts worth keeping and saves them to the CLAUDE.md you choose (this project by default); /reflect-undo takes the last one back. /mine turns friction that repeats across past chats into ideas.

Replays the last turn's edits step by step, each tied to its plan step, with a coloured diff; earlier turns and past sessions too. Opens from /replay.

Every agent gets a face: progress, cost and time for each subagent, live, in the Agents pane and a band above the prompt. Other chats waiting on you show at the top.

Stops keys and tokens before they land: scans writes, edits, text a Bash command writes, git add and commit, and blocks committing .env. If the guard fails, the write is blocked. Pasted keys reach the model as placeholders; /rec masks the screen.

A spend budget per session, as a share of the 5-hour limit (or the weekly one without it) or in dollars: a notice at 80%, a stop at 100%, and a question before new subagents when the window runs low.

A browser inside the terminal: preview sites and HTML files and let the agent drive it. Terminal only; needs the external terminal-browser CLI.

After you allow it once per folder, runs tests and lint in the package you edited and shows the result as a badge; a failed run is offered to your next prompt. Watches your PR's GitHub CI too.

Beta. Chat and zero-token agent questions between your own computers; team rooms across accounts are experimental.

Your year with Claude Code as a story and a video, from your own transcripts: persona, which agents did the work, month by month. /wrapped; Hide names masks project and file names before you share. Uses no model tokens.