SLOPSHOPPER

fork-explorer

Ask "what if..." in a side pane: a side question over the current session returns 2-3 alternative approaches side by side, without adding anything to your…

newpanecommandtoastmodel
v0.1.0no licenseupdated 2026-10-03harshitmywork17/claude-mods/plugins/fork-explorer
A shopper browsing a rack in a slop shop
Preview · a replayed session in a sandbox
claude · ~/work/app · fork-explorer
│ ┃ Fork Explorer ✕ › fix the failing auth test and add an audit log call │ ┃ What if…: we used Redis instead of Postgres │ ┃ Ask a "what if…" question. The answer stays ⏺ Read(src/auth.ts) │ ┃ here and never enters your conversation. ⎿ Read 6 lines │ ⏺ Update(src/auth.ts) │ ⎿ Added 2 lines, removed 1 line │ ⏺ Bash(bun test) │ ⎿ 3 pass, 1 fail │ │ ● Done. refresh now rejects expired claims and logs an audit event. │ │ ✻ Worked for 42s · done 4:20 PM │ │ › /fork │ │ ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts

Draws

Pane · Fork Explorer
What if…: we used Redis instead of Postgres for sessions? ⏎ Ask a "what if…" question. The answer stays here and never enters your conversation.
README

harshit-mods

A Claude Code plugin marketplace, and the single home for Claude customizations that stay the same on every device: ten mods, the Claritymaxx skill, the sync-home skill, and the karpathy-guidelines and python-standards coding skills. To add something, ask Claude to put it "on all my devices". The sync-home skill and CLAUDE.md tell it how.

PluginCommandWhat it does
usage-forecast/forecast, /forecast hide, /forecast showA band above the prompt. 5-hour and Weekly plan limits as coloured bars (green, then yellow from 70%, red from 90%) with the percent used, a reset countdown and the reset time (↻ resets in 1h 12m · 15:40), and ⚠ at this pace, out in 40m when you would run out before the reset. Below it, the context window as weather with a 12-turn sparkline and turns left. Replaces token-weather.
blast-radius/blast-radius, /blast-radius <command> (dry run)Catches rm -rf, git reset --hard, force pushes, git clean -f, git checkout ., git branch -D, git stash drop/clear, DROP/TRUNCATE, mkfs, dd of=/dev/.... It shows what the command would touch in a pane and forces a permission prompt. A deny from your settings is never loosened.
replay-theater/replay, /replay 2Records each turn's Edit and Write calls. In the pane, n/p step through the diffs, o/w change turn, and Esc closes it.
changed-files/changed-filesA live sidebar of every file touched this session, with +/- counts, edit counts and a new tag. It opens by itself on the first edit.
session-meter/meterThe status line shows elapsed time, tool calls and files edited. A toast appears when a turn takes 60 s or more; change it with /plugin configure session-meter@harshit-mods.
claritymaxx/claritymaxx:explain <topic>, or ask "explain..."Third-party skill from v60samurai/claritymaxx. Builds a mental model first, then explains it as text, a diagram or a small HTML page.
fork-explorer/fork <what if...>, or type in the paneRuns a side question over the current session and shows 2-3 alternative approaches as cards side by side, with pros, cons and effort. "use this" puts the choice in your prompt box to edit and send. Nothing is added to your conversation.
mission-control/missionA live pane of what Claude is executing: each tool call with its input, status badge (✓ ✗ ⊘ ●), duration and output; subagents in their own lanes with tool and token counts; a network-tab style waterfall; time spent in the model between tools. v switches between the timeline and the outputs. Observe-only; it never changes a call.
diff-minimapautomaticA thin strip beside each Edit and Write row in the transcript. Green, red and yellow ticks show where in the file the change sits, with the line range and the file length.
codebase-atlas/atlas, /atlas changes, /atlas components <file>, /atlas calls <name>, /atlas decisions, /atlas rescanOne pane with six tabs. Map: folders as boxes in dependency layers (entry points on top, foundations at the bottom), lit yellow when Claude edits and cyan when it reads; select one for what it imports and what uses it. Changes: every uncommitted change mapped to the functions and classes it added, modified or removed, with call-site counts, the layers crossed and the files that import what changed. Components: a file's classes, functions and methods with size bars, what each calls, and which changed. Calls: callers and callees of any function; type a name or select one in the transcript, then walk the graph. Decisions: a timeline of decisions a small model pulls from each turn (turn off with extractDecisions). Explain: side-question explanations of a folder, a function or the session's changes that never enter your conversation. Works with or without git: in a git repo, changes are measured against the last commit; in a plain folder, against the files as Atlas first read them that session (dependencies, virtual environments and build output are skipped). Reads Python and JS/TS.
workbench/wb, /wb changes [file], /wb preview <file>, /wb artifacts, /wb map, /wb usage, /wb calls <name>, /wb forecast, /wb workbench, /wb hud offOne workspace. A one-line band above the prompt with two slides, switched with ◀ ▶: Workbench (session time, tool calls, files changed with +/−, new files, plan limits and context on wide terminals, then the running tool or the last edit and its functions) and Forecast (5-hour and weekly limits with bars, reset times and pace, plus the context forecast; it shows the usage-forecast mod's band when that mod is on). A pane with six tabs on one row: Now, the turn's tool calls with status and timing, failures with their error, and subagents nested under the call that started them; Changes, every file with a change bar and the functions it touched, then a file's changed functions and highlighted diff since Claude first touched it, then each edit (replay with n/p); Preview, a live render of the file being edited (Markdown, CSV table, JSON tree, HTML screenshot, PNG, code) with recent files one press away; Artifacts, files created this session by Claude ✦ or its commands, with preview and copy path; Map, architecture layers with folder sizes and imports, a file's components, and the call graph; Usage, limits with reset times and pace, context, and cost per turn. Works with or without git. Claude Code's own diff panel sits above mod panes: run /diff to hide it if it covers the workbench.
crew/crew, /crew auto off, /crew closeAn Agents side pane, like a mission board for subagents. Three cards on top total the session's cost, tokens and time. Each subagent is a little pixel crew member whose hat shows its effort (heavy, careful, medium, light), with its model and effort level, context used with a bar, tokens, its share of the cost and how long it ran; it walks while it works and shows the tool it is running. Sections: Running, Completed (✓ done, ✗ failed, ⊘ stopped; folds) and Planned, the tasks from Claude's task or todo list with what each still waits for ("after 1, 3"). Click an agent for its task, tool count and answer. Opens by itself when the first subagent starts; /crew auto off stops that.
explainer-page/explainer-page:explainer-page, or ask for a visual or interactive explainer pageA skill for textbook-style HTML pages that teach one concept: each idea in words, maths, a figure and code, with colour-coded notation (inputs blue, outputs orange, parameters green, functions violet, probabilities pink) kept identical everywhere; tables and figures (3D bars, node diagrams) drawn from one data source and highlighted together on hover; sliders where a parameter is the point; code cells with their real output; margin notes, callouts and check-yourself questions; light and dark themes and a contents sidebar. Publishes as a claude.ai artifact when that tool is available, otherwise writes an .html file. Triggers only on an explicit request for such a page.
savvy-flow/savvy-flow:savvy-flow <task>A copy of savvy-flow from johnnyvizz/claude-kit with one change: the top tier, savvy-fable, runs on Opus at max effort instead of the Fable model, so it works without Fable access. The session model plans the task, delegates pieces to five tiered worker agents (savvy-fable, savvy-heavy, savvy-careful, savvy-medium, savvy-light) and reviews what they return. tools/sync-savvy-flow.sh refreshes the copy from upstream and reapplies the change; the Sync savvy-flow GitHub Action runs it daily and commits when upstream changed.
savvy-progress/agents-infoFrom the same repo, following its main branch the same way. A progress bar above the prompt driven by savvy-flow, and an agents panel with model, effort, step progress, context, estimated cost and time. Animated crabs draw in the desktop app; the terminal shows text rows (crew draws its sprites in both).
sync-home(automatic) or ask "add X to all my devices"A skill that tells Claude this repository is where plugins, mods, skills and other customizations go so they sync to every device. Claude adds them under plugins/, validates, pushes and tells you what to turn on.
karpathy-guidelines/karpathy-guidelines:clean-code, or automatic on any coding taskGuidelines from Andrej Karpathy's notes on LLM coding mistakes: think before coding, simplicity first, surgical changes, goal-driven execution, with worked examples in EXAMPLES.md.
python-standards/python-standards:python-standards, or automatic on Python workPython standards: SOLID, DRY, KISS and YAGNI; FastAPI project layout; naming; OWASP-aligned secure coding; linting and testing setup.

Use them on every device

Add this marketplace to your claude.ai account once:

  1. On claude.ai, open Customize, then Plugins, then Add, then Add marketplace.
  2. Enter harshitmywork17/claude-mods. If the repository is private, your GitHub account must be connected to Claude.
  3. Turn on each plugin. Turn on sync-home too, so every session knows to put new synced customizations here.

Plugins on your account sync at session start, on every machine where you sign in to Claude Code with that account. Mods need Claude Code v2.1.287 or later (claude --version, then claude update). After you push a change here, press Check for updates on the marketplace in claude.ai, and turn on any new plugin. Then start a new session.

If a command such as /wb is missing on a machine:

  1. claude --version must be 2.1.287 or later.
  2. claude plugin list should show workbench@synced. If it doesn't, the sync hasn't picked it up: check for updates in claude.ai, or install on that machine as below.
  3. claude plugin test run in an empty folder should say no hooks module to load. Any other message means mods are turned off there.

Mods run in Claude Code only: the terminal, the desktop app's Code tab, IDE extensions and cloud sessions set up as below. Claude chat on the web, desktop and mobile shows no mod UI. Claritymaxx is a skill, so it also works in Claude chat and Cowork.

Use them in cloud sessions

Cloud sessions at claude.ai/code don't load account plugins or the plugins a repository lists in .claude/settings.json. They do load plugin folders named in CLAUDE_CODE_PLUGIN_DIRS. Set it once on the cloud environment:

  1. Open the environment menu in a session's title bar, then Edit.
  2. Under Environment variables, add: `` CLAUDE_CODE_PLUGIN_DIRS=/home/user/claude-mods/plugins ``
  3. When you start a cloud session, select harshitmywork17/claude-mods alongside the repository you're working on. It is cloned to /home/user/claude-mods, and every plugin in plugins/ loads, including new ones.

A cloud session can only clone the repositories selected for it, because this repository is private. In a session without it, the variable points at a missing folder, which Claude Code skips.

Install on one machine

claude plugin marketplace add harshitmywork17/claude-mods
claude plugin install workbench@harshit-mods   # repeat for each plugin

Or run every plugin from a local clone, picking up changes with git pull: add "env": { "CLAUDE_CODE_PLUGIN_DIRS": "<path to clone>/plugins" } to ~/.claude/settings.json. Don't combine this with account sync on the same machine: when two copies share a name, only the first one loads.

Develop

Each mod has tests: claude plugin test plugins/<mod>. Check a mod with claude plugin validate plugins/<mod>. Increase version in a mod's plugin.json when you change it, because installed copies are cached by version.

Source 3 files
hooks/register.tsx 163 lines
1import { atom, read, update } from 'claude-code'
2import type { EngineInterface, Register } from 'claude-code'
3
4import type { Fork } from '../types'
5import { forkPrompt, parseApproaches, promptFor } from './approaches'
6
7const PANE = 'fork-explorer'
8const MAX_FORKS = 5
9const WIDE_ENOUGH = 100
10const forks = atom({ plugin: 'fork-explorer', key: 'forks' } as const, [])
11const viewing = atom({ plugin: 'fork-explorer', key: 'viewing' } as const, 0)
12
13const REASON_TEXT: Record<string, string> = {
14  'nothing-to-fork': 'There is no conversation to explore yet. Ask Claude something first.',
15  'api-error': 'The side question failed (API error). Try again.',
16  'empty-reply': 'The side question came back empty. Try rephrasing.',
17  aborted: 'The side question was cancelled.',
18}
19
20async function setFork($: EngineInterface, id: string, change: Partial<Fork>): Promise<void> {
21  await update($, forks, list => list.map(one => (one.id === id ? { ...one, ...change } : one)))
22}
23
24/** Runs one side question over the session and records its approaches. Never touches the conversation. */
25async function explore($: EngineInterface, question: string): Promise<Fork | undefined> {
26  const id = `fork-${await $.clock.now()}`
27  const fork: Fork = { id, question, status: 'thinking', approaches: [] }
28  await update($, forks, list => [fork, ...list].slice(0, MAX_FORKS))
29  await update($, viewing, () => 0)
30
31  const reply = await $.model.fork({ prompt: forkPrompt(question) })
32  if (!reply.isAnswered) {
33    await setFork($, id, { status: 'error', error: REASON_TEXT[reply.reason] ?? `No answer (${reply.reason}).` })
34  } else {
35    const approaches = parseApproaches(reply.text)
36    await setFork(
37      $,
38      id,
39      approaches.length === 0
40        ? { status: 'error', error: 'The answer could not be read as approaches. Try again.' }
41        : { status: 'done', approaches },
42    )
43  }
44  return (await read($, forks)).find(one => one.id === id)
45}
46
47export const register: Register = on => {
48  on('session.start', async ($, e, next) => {
49    await $.command.register({
50      name: 'fork',
51      description: 'Fork Explorer: "what if..." side question with 2-3 alternatives, kept out of the conversation',
52      argumentHint: '[what if...]',
53      immediate: true,
54    })
55
56    return next(e)
57  })
58
59  on('command.run', { command: 'fork' }, async ($, e) => {
60    await $.ui.open({ id: PANE, title: 'Fork Explorer', focus: true, closeOnEscape: true })
61    const question = e.args.trim()
62    if (question !== '') await explore($, question)
63
64    return {}
65  })
66
67  on('session.end', async ($, e, next) => {
68    if (e.reason === 'clear') await update($, forks, () => [])
69
70    return next(e)
71  })
72
73  on('ui.render', { component: 'Pane', requestId: PANE }, async ($, e) => {
74    const { Box, Button, Text } = $.ui.resolve(e)
75    const list = await read($, forks)
76    const at = Math.min(await read($, viewing), Math.max(0, list.length - 1))
77    const fork = list[at]
78    const isWide = e.props.bodyColumns >= WIDE_ENOUGH
79    const table = $.ui.resolve(e)
80    const ask =
81      'Input' in table ? (
82        <table.Input
83          key="ask"
84          label="What if…"
85          placeholder="we used Redis instead of Postgres for sessions?"
86          submitLabel="explore"
87          autoFocus
88          onSubmit={value => (value.trim() === '' ? undefined : void explore($, value.trim()))}
89        />
90      ) : null
91
92    return (
93      <Box flexDirection="column">
94        {ask}
95        {fork === undefined && (
96          <Text dimColor>Ask a "what if…" question. The answer stays here and never enters your conversation.</Text>
97        )}
98        {fork !== undefined && (
99          <Box flexDirection="column" marginTop={1}>
100            <Box flexDirection="row" gap={1}>
101              <Text bold wrap="truncate-end">
102                What if {fork.question}
103              </Text>
104              {list.length > 1 && (
105                <Text dimColor>
106                  ({at + 1}/{list.length})
107                </Text>
108              )}
109            </Box>
110            {fork.status === 'thinking' && <Text color="cyan">Exploring alternatives…</Text>}
111            {fork.status === 'error' && <Text color="red">{fork.error}</Text>}
112            {fork.status === 'done' && (
113              <Box flexDirection={isWide ? 'row' : 'column'} gap={1} marginTop={1}>
114                {fork.approaches.map((approach, index) => (
115                  <Box
116                    key={`card-${index}`}
117                    flexDirection="column"
118                    borderStyle="round"
119                    borderColor="cyan"
120                    paddingX={1}
121                    width={isWide ? `${Math.floor(100 / fork.approaches.length)}%` : undefined}
122                    flexGrow={1}
123                  >
124                    <Text bold color="cyan">
125                      {index + 1}. {approach.title}
126                    </Text>
127                    <Text dimColor>effort {approach.effort}</Text>
128                    <Text>{approach.idea}</Text>
129                    {approach.pros.map(pro => (
130                      <Text color="green">+ {pro}</Text>
131                    ))}
132                    {approach.cons.map(con => (
133                      <Text color="red">− {con}</Text>
134                    ))}
135                    <Button
136                      key={`use-${index}`}
137                      label="use this"
138                      hotkey={String(index + 1)}
139                      onPress={async () => {
140                        await $.prompt.fill({ text: promptFor(fork.question, approach), mode: 'replace' })
141                        $.ui.toast('Fork Explorer: put in your prompt. Edit it, then send.')
142                      }}
143                    />
144                  </Box>
145                ))}
146              </Box>
147            )}
148            <Box flexDirection="row" gap={1} marginTop={1}>
149              {at < list.length - 1 && (
150                <Button key="older" label="older" hotkey="o" onPress={() => update($, viewing, n => n + 1)} />
151              )}
152              {at > 0 && <Button key="newer" label="newer" hotkey="w" onPress={() => update($, viewing, n => n - 1)} />}
153              {fork.status !== 'thinking' && (
154                <Button key="again" label="ask again" hotkey="a" onPress={() => void explore($, fork.question)} />
155              )}
156            </Box>
157          </Box>
158        )}
159      </Box>
160    )
161  })
162}
163
hooks/approaches.ts 57 lines
1import type { Approach } from '../types'
2
3const MAX_APPROACHES = 3
4const MAX_TEXT = 400
5const MAX_ITEMS = 4
6
7export function forkPrompt(question: string): string {
8  return [
9    'Side question from the user. Answer it on its own: do not use tools and do not continue or change the current task.',
10    `What if: ${question}`,
11    'Give 2 or 3 genuinely different approaches, grounded in this conversation and codebase.',
12    'Reply with JSON only, no prose and no code fence, in exactly this shape:',
13    '{"approaches":[{"title":"short name","idea":"one or two sentences","pros":["..."],"cons":["..."],"effort":"S, M or L"}]}',
14  ].join('\n')
15}
16
17function text(value: unknown, limit = MAX_TEXT): string {
18  return typeof value === 'string' ? value.trim().slice(0, limit) : ''
19}
20
21function list(value: unknown): string[] {
22  return Array.isArray(value) ? value.map(item => text(item, 160)).filter(item => item !== '').slice(0, MAX_ITEMS) : []
23}
24
25/** Reads the approaches out of a reply, tolerating prose or a code fence around the JSON. */
26export function parseApproaches(reply: string): Approach[] {
27  const start = reply.indexOf('{')
28  const end = reply.lastIndexOf('}')
29  if (start < 0 || end <= start) return []
30
31  let parsed: unknown
32  try {
33    parsed = JSON.parse(reply.slice(start, end + 1))
34  } catch {
35    return []
36  }
37
38  const raw = (parsed as { approaches?: unknown }).approaches
39  if (!Array.isArray(raw)) return []
40
41  return raw
42    .map((item: Record<string, unknown>) => ({
43      title: text(item.title, 80),
44      idea: text(item.idea),
45      pros: list(item.pros),
46      cons: list(item.cons),
47      effort: text(item.effort, 8) || '?',
48    }))
49    .filter(approach => approach.title !== '' && approach.idea !== '')
50    .slice(0, MAX_APPROACHES)
51}
52
53/** The text "use this" puts in the prompt box for the person to edit and send. */
54export function promptFor(question: string, approach: Approach): string {
55  return `Let's go with "${approach.title}" for: ${question}\n${approach.idea}`
56}
57
types/index.d.ts 27 lines
1export type Approach = {
2  title: string
3  idea: string
4  pros: string[]
5  cons: string[]
6  effort: string
7}
8
9export type Fork = {
10  id: string
11  question: string
12  status: 'thinking' | 'done' | 'error'
13  approaches: Approach[]
14  error?: string
15}
16
17declare module 'claude-code' {
18  interface PluginState {
19    'fork-explorer': {
20      /** Newest first. */
21      forks: Fork[]
22      /** Index into forks of the one on screen. */
23      viewing: number
24    }
25  }
26}
27