All shelves / Mods / Workflow
agent-watch
By ice-lfernandes
agent-watch is a Claude Code workflow mod by ice-lfernandes.
Draws each subagent's tokens and a stall toast, and the last tool calls, with a summary of the heaviest agent when they finish.
Draws
Install
The marketplace name is lfernandes-mods.
/plugin install agent-watch --marketplace ice-lfernandes/claude-code-modsIt runs inside the Claude Code process. The side panel is only what the author's README names. Read the repo before you install, or run claude plugin validate on the plugin folder.
- Version
- 0.3.1
- License
- MIT
- Repo stars
- 7
- Last change
- 2026-10-09
What it can reach
- Draws in the session
- Reads usage or context figures
- Sees tool calls
The repository README says these mods require Claude Code 2.1.287 or later. While subagents run, a status line shows how many are running, their tokens, and a stall mark, and it clears when they finish. Tokens come from each model request: input that was not cached, cache read, cache written, and output. The lead row is the main conversation. Tokens from loops that are not agents show as other loops. A running agent with no model request and no tool call for 5 minutes, by default, gets one toast that says what it is stuck on: thinking, a tool call, or nothing. The alert re-arms when the agent does something again. Teammates between turns, and agents held on their own background work, never count as stalled. When the last active agent ends, a toast names the agents, the tokens, the wall time, and the heaviest agent with its share. On a subscription it also says how much of the 5-hour window that wave used. /watch opens the pane. Press an agent's name and the row shows its last 5 tool calls. A stalled agent's investigate button puts a request about it in the prompt. The README says nothing stops the agent. /watch demo adds three fake agents. Language defaults to auto: Portuguese when the system language is Portuguese, English otherwise. From a clone, one session is claude --plugin-dir ./agent-watch. The README says reading each response's usage adds a small cost to each request, the main conversation included. Its reach table marks network, processes, files, model calls, and sending data anywhere as no.
The install name is agent-watch. This page adds the author's name. The install line has no @suffix. The marketplace name is lfernandes-mods. The marketplace also holds limits-meter, allowlist-coach, test-hud, launchpad, and plain-view. The status line and the pane of the subagents are why this page files it under workflow. agent-radar draws one line per running subagent, with the time and the tool count. review-watch draws one line per running code review. This one is the tokens, the stall, and the last tool calls. investigate fills the prompt. Demo agents are fake.
Demo
What it hooks into
session.startWhen a session startsagent.spawnWhen a subagent startsturn.stepAt each step of a turntool.callAround a tool call, where it can hold oneturn.completeWhen a turn endscommand.runWhen you run its slash commandui.scrollA Claude Code eventPaneui.renderDraws into the sessionPane
Read from the on() calls in its hook modules. Open its source under Files to check what each one does.
Files
The author's README and source, 4 files. They load when you open them.
Facts
- Author
- ice-lfernandes
- Category
- Workflow
- Marketplace
- lfernandes-mods
- Folder
- agent-watch
- Version
- 0.3.1
- License
- MIT
- Repo stars
- 7
- Last change
- 2026-10-09
- Language
- TypeScript
- Files
- 13
