Mission Control Claude Code mod: watch Claude Code subagents and tools live

Mission Control is a community mod by Hamza Zafar (hamzafer), part of his claude-code-mods collection. It shows what Claude Code is doing in one place: the main agent, its subagents and every tool call they make. Run /mission and a pane opens next to the chat with a tree. The main loop is at the top, each Claude Code subagent hangs under it, and the tool calls sit under the agent that made them. Runs of the same tool are folded into one line, such as read: README.md → src/login.js ×2, and each line shows its duration.

While Claude works, and for 30 seconds after, a single line above the prompt sums it up: ◈ mission: 2 tools, with agent and changed-file counts when there are any. A second view, /mission code, draws a map of the files Claude read or edited, with arrows for the relative imports between them, a glow on the file being touched now, and a one-line note on each change. The author's README asks for Claude Code 2.1.287 or later for the collection; we tested it on 2.1.296.

The recording starts Claude Code with the mod loaded in the small demo project and asks Claude to read README.md and src/login.js and say in one sentence what they do. Claude reads both files, and while it works the band line above the prompt counts the tools. After the answer, the clip types /mission. The pane opens above the prompt with the Who, Code and Close buttons, the summary 0 agents (0 running) · 2 tool calls, the green main line with the prompt text, and the folded read: demo-app/README.md → src/login.js ×2 line under it. The code map is not shown in the clip.

Use any of the methods in the install block; they install straight from the author's own marketplace. Then run /mission in a session. After installing into a session that is already open, run /reload-plugins first.

The mod keeps its tree, file list and view in host-held $.state atoms, so a hot reload keeps the picture. Every tool call is wrapped in a tool.call hook that adds a node, lets the call run, then marks it done or failed. From hooks/register.tsx:

let failed = true
let r: Awaited<ReturnType<typeof next>>
try {
  r = await next(e)
  failed = r.deny !== undefined || r.isError === true
} finally {
  await update($, nodes, list => list.map(n => (n.id === id ? { ...n, status: failed ? 'failed' : 'done', end: Date.now() } : n)))
}

agent.spawn adds subagents, turn.start resets the tree, and turn.complete closes whatever is still marked running. Calls to Agent and Claude Code's own bookkeeping tools (ToolSearch, TaskCreate and similar) are left out. A one-second clock only updates the display while something is running; the tree is capped at 300 nodes. For the code map, hooks/map.ts builds an SVG: it reads each touched source file through $.fs.read, finds relative imports with a regex, and lays the files out in columns by import depth. The SVG is turned into a PNG by headless Chrome and shown with the pane's Image element.

There are no options in /config. The only controls are /mission, /mission who and /mission code, and the pane buttons w, c and q. To change behavior, edit the constants at the top of register.tsx: MAX_NODES (tree size), MAX_FILES (files on the map), BAND_AFTER_MS (how long the band stays), INTERNAL (tools kept out of the tree) and CODE_FILE (file types that appear on the map).

We read every file the plugin can execute before running it: .claude-plugin/plugin.json, hooks/hooks.json, hooks/register.tsx, hooks/map.ts, hooks/tree.ts, the type declarations and the tests. There is no package.json, install script, binary or bundled code, and we found no hidden Unicode, eval or remote code.

• Network: none of its own. No fetch, sockets or hosts.
• Model call: explain() in register.tsx calls $.model.complete with model: 'haiku' after a main turn that edited or wrote files. The prompt holds up to 300 characters of each old_string and new_string (400 of a written file), and the reply becomes a note of at most 80 characters on the map. The reply is parsed as JSON, cut short and escaped before it enters the SVG. It uses your session's quota.
• Processes: one sh -c script through $.process.run that starts headless Chrome with a throwaway profile, waits for the screenshot, then kills that Chrome. File and folder names are passed as arguments, not spliced into the shell string. Chrome is only started while the pane is open on the code view.
• File writes: $TMPDIR/mission-control/ only: map.svg, two map-N.png files and a Chrome profile folder. Nothing under ~/.claude, settings, shell files or LaunchAgents. The mod reads the files Claude touches, to find imports, and keeps edit snippets in memory.
• Daemons: none. The one-second clock stops with the session, and Chrome is killed after each screenshot.
• Secrets and environment: it reads TMPDIR only.
• Telemetry: none.
• Prompt influence: it adds no additionalContext and does not rewrite prompts or tool results.
• Gating hooks: tool.call and agent.spawn are wrapped, and claude plugin validate lists them as gating hooks without .catch. They only observe; tool.call returns the tool's result unchanged and does not change arguments.
• External requirement: the code map needs Google Chrome at /Applications/Google Chrome.app (macOS path); the Who view does not.
• License: MIT, with a LICENSE file.

claude plugin validate reported no errors or warnings, and the author's 7 tests passed on Claude Code 2.1.296.

We checked it with Claude Code 2.1.296 on macOS (validate, tests and the recording). If something is off:

• /mission is an unknown command. Check the plugin is loaded in /plugin, and run /reload-plugins after a fresh install.
• The code map says it needs Google Chrome. Install Chrome at /Applications/Google Chrome.app. The mod looks only at that path.
• "Nothing yet". The tree fills as a turn runs. Start a task and the lines appear.
• No map outside the terminal. In other surfaces the pane says the code map draws in the terminal.
• Missing band line. It shows only while Claude works and for 30 seconds after, and not while a survey is open.