pr-watch by heidiks: Claude Code mod code

What it does

Draws the status line, toasts
Adds /pr-watch
Watches Bash calls
Hooks config.set, session.start
Keeps its own state between turns

Good points

No network calls
Written in TypeScript
Handles errors
Writes no files

Code to start from

A starter ModsCode writes from the drawing it read, and the mod's own lines that draw, where there are any.

Starter (register.tsx)

import type { Register } from 'claude-code'

// Drawn like pr-watch by heidiks (pane, band, status, toast), as ModsCode read it from the mod's code:
// https://modscode.com/claude-mods/mods/heidiks-pr-watch/
// Written by ModsCode from that reading, not copied from the mod: use it as you like.
// Each value below stands for one the mod works out at run time; put in your own.
export const register: Register = on => {
  on('session.start', async ($, e, next) => {
    const parts = '…'
    const heading = '…'
    const describe = '…'
    const url = '…'
    await $.ui.status(`PRs ${parts}`)
    await $.ui.toast(`${heading} · ${describe}`)
    await $.ui.toast(`Could not open ${url}`)
    return next(e)
  })
}

The mod's own code

No drawing lines were cut from this mod's code; its source is linked.

heidiks/agent-kit

How to build this: Pane (panel), Deny a tool call

What its code calls

runs programsreads filesdrives Claudetool callsremembers

What its code shows

Read from the source at commit 4f106cbb9c42 on 2026-10-11. The first group decides whether a mod is listed; the rest is for you to weigh. A later commit is not read until the next crawl.

Decides the listing: code you can check
Entry read	Every module hooks/hooks.json names was read.plugins/pr-watch/hooks/register.tsx
Parses	Every file parses as TypeScript or JavaScript.7 files parsed
All its code is in the mod	Every import is a file of the mod or the engine's own module; no packages, no require.all its code is in the mod
Readable	No minified lines, no eval or new Function, no import of a computed path, no encoded blobs.no minified code, no eval, no encoded blobs
Calls in the open	Network and programs are reached as $.http.fetch and $.process.run in place, never passed around.every call outside goes through $ in the open
For you to weigh before installing
Where it draws	The terminal, the desktop app, both, or nothing (hooks only: a guard or a prompt rewriter), as its code says. Shown, never a reason to leave a mod out.draws in the terminal and the desktop app, the same in both
Install line	Whether its repository's own .claude-plugin/marketplace.json lists it, so it installs from that marketplace with the lines on its page.listed in its repository's .claude-plugin/marketplace.json: installs from there
Programs on Windows	Whether every program it starts through $.process exists on Windows (tail, date, open, osascript, /usr/… do not).runs no program a Windows PC lacks
Network	Whether every fetch names its https host in the code.no network
Programs	Whether every program it starts is named in the code and is not a shell, an interpreter or a network tool.plugins/pr-watch/hooks/register.tsx:49 a program named at run time; plugins/pr-watch/hooks/register.tsx:309 a program named at run time
Reads and sends	Whether it both reads something of yours (files, environment variables, the transcript, settings) and reaches outside (the network, an MCP server, another session, telemetry).reads files, but reaches nothing outside
Keys in the source	Whether the source holds what looks like an API key or a private key.no keys in the source
Licence	An open licence (MIT, Apache-2.0, BSD, ISC, Unlicense, 0BSD, CC0) with its text lets ModsCode show its code; otherwise the code is linked.MIT
Settings hooks	Hooks declared in hooks/hooks.json or plugin.json beside the module: a command (a shell when it has no args), an HTTP request, an MCP tool or a model prompt at an event. They run outside $.no settings hooks: its function hooks alone
Servers it ships	MCP servers (a program, or a URL; a headersHelper is a shell command; a .mcpb bundle is unpacked at run time), LSP servers, monitors (a shell command for the whole session) and bin/ (on the PATH of the Bash tool).ships no MCP or LSP server, no monitor, no bin/
Skills, commands, agents	What its Markdown declares beyond text: allowed-tools granted for the turn, hooks added when a skill is invoked, a command run before Claude reads the skill.1 skill: text only
Tools and MCP	Whether every tool it calls through $.tool.call and $.mcp.call is named in the code, and none is a shell (Bash through the tool runs a shell without $.process).calls no tool, MCP server or command
Asks of the model	Completions it runs ($.model.complete, fork, classify), subagents it spawns, prompts it submits as you, and the tools, commands and agents it adds.prompt.submit (as the person), model.complete, adds command pr-watch
Events it hooks	By what a hook there can do: gates (refuse a tool call, a prompt, a setting, another plugin), shapes (what the model reads: the system prompt, tool descriptions, each request), reads (your input, the events of the settings hooks, every event with *), and the rest.gates: command.run (pr-watch), tool.call (Bash, mcp__azure-devops__repo…), config.set · also: session.start, ui.render (AbovePrompt, Pane) · timers: clock.after, clock.every
Reads, by name	The environment variables and files it names, and whether it reads the transcript, settings.json, the prompt box or the mouse selection.files: …/…, …/spec.md, …/docs/prd
Writes, by name	The files and environment variables it names, and whether it changes settings, the prompt box, the clipboard, the conversation (a row the model reads) or another session.the clipboard

This is a reading of the code text, not a security review: it does not run the mod, and it cannot see what a program the mod starts does. The check before installing is yours. How mods are chosen

Install it

Check it yourself before you install it. Mods run unsandboxed, with your permissions, and ModsCode does not vouch for what this code does. Read its code (or run claude plugin validate . in its folder), weigh what it reaches, and install the commit you read.

From its marketplace

Its repository is its marketplace. Add that marketplace (in the terminal with these commands, or in the desktop app's plugin settings) and install the mod from it:

/plugin marketplace add heidiks/agent-kit

/plugin install pr-watch@agent-kit

/reload-plugins

This installs the marketplace's latest commit, which may not be the one read here (4f106cbb9c42). Compare them before you install. /reload-plugins loads it into a session that is already open.

Try the commit ModsCode read

No marketplace needed. In a terminal, clone the repository, check out the commit read, validate the mod, and load it for one session:

git clone https://github.com/heidiks/agent-kit

git -C agent-kit checkout 4f106cbb9c424e1dccf3148f84618fe5602501d7

claude plugin validate agent-kit/plugins/pr-watch

claude --plugin-dir agent-kit/plugins/pr-watch

The validate line lists the events the mod hooks and what it asks Claude Code to do, without running it. The last line loads the mod for this session only and writes nothing to your settings. If the author has rewritten history, the commit may no longer exist.

Ask Claude about it

With the ModsCode connector added, Claude can read this mod's code and what it reaches, and build on them. Ask it like this:

Read the ModsCode mod "heidiks-pr-watch" with its code and make me one like it

Search them from Claude: the MCP connector

Similar mods

Badge for your README

If this is your mod, paste this into your README. It links to this page.

README.md

[![ModsCode](https://modscode.com/badge/heidiks-pr-watch.svg)](https://modscode.com/claude-mods/mods/heidiks-pr-watch/)