SLOPSHOPPER

cc-graph-subagents-probe

Recording probe for the subagents session, throwaway

guardagents
★ 1v0.0.0MITupdated 2026-10-12kuehn-lars/cc-graph/fixtures/sessions/subagents-probe
A shopper browsing a rack in a slop shop
README

<img src=".github/assets/header.svg" alt="cc-graph: see what Claude knows" width="100%">

Version Claude Code Surface TypeScript CI Network License

<samp>install · the pane · privacy · roadmap · changelog</samp>

<samp>After a long turn, nobody can say without scrolling back what Claude actually has in context. cc-graph is a Claude Code mod that draws it in a live pane beside the conversation: every file Claude read, searched or changed, how those files import each other, and what compaction took away.</samp>

<a name="install"></a> <img src=".github/assets/h-install.svg" alt="install" width="100%">

/plugin install cc-graph --marketplace kuehn-lars/cc-graph

<samp>Then type /graph. Needs Claude Code v2.1.287 or later, in the terminal.</samp>

<a name="the-pane"></a> <img src=".github/assets/h-the-pane.svg" alt="the pane" width="100%">

<img src=".github/assets/showcase.svg" alt="The cc-graph pane in the terminal, annotated: the context window meter, partly read and compacted files, import edges, folder frames, size bars, the focus, search-only files, an edit and a new file, the now line and the legend" width="100%">

<samp>control</samp><samp>what it does</samp>
<kbd>g</kbd><samp>graph view</samp>
<kbd>l</kbd><samp>list view</samp>
<kbd>1</kbd> … <kbd>9</kbd><samp>pick a row in the list</samp>
maxNodes = 60<samp>most files in the graph, in /config</samp>
importEdges = on<samp>show file imports, in /config</samp>
blindSpots = on<samp>look for blind spots with git, in /config</samp>
captureThinking = on<samp>keep thinking excerpts, in /config</samp>

<samp>Files a shell command created, deleted or changed show too, named with the command that did it, and so do files Claude read with cat, head or sed -n. Files that import a change but were never opened are flagged as blind spots while the turn is still running.</samp>

<samp>The list also shows Claude's newest thinking: a short excerpt per step, with its turn and the files that step read or changed. Claude Code passes thinking text to mods only when "showThinkingSummaries": true is in your settings (or with --thinking-display summarized). This holds for every 5.5 model: Opus, Sonnet and Haiku. Without it, the footer says thinking is hidden.</samp>

<a name="privacy"></a> <img src=".github/assets/h-privacy.svg" alt="privacy" width="100%">

<samp>cc-graph only watches. Its hooks pass every tool result on untouched and record what happened in session memory.</samp>

  • <samp>Nothing leaves your machine. No network requests, no model calls, no files written.</samp>
  • <samp>Paths, sizes and counts, never file contents. The one exception is Claude's newest 20 thinking excerpts, at most 400 characters each, and only while captureThinking is on. All of it is gone when the conversation ends or on /clear.</samp>
  • <samp>The only process it starts is read-only git, by argument list, never through a shell.</samp>
  • <samp>Check it yourself: claude plugin validate lists every event cc-graph hooks and every API method it calls. CI fails if the code goes beyond the allowed list.</samp>

<a name="roadmap"></a> <img src=".github/assets/h-roadmap.svg" alt="roadmap" width="100%">

<img src=".github/assets/roadmap.svg" alt="Roadmap: M0 Foundations, v0.1 What Claude knows, v0.2 The graph and v0.3 Blind spots shipped; next v0.4 Why it did that, then v0.5 Desktop graph and pointer, v0.6 Decisions and timeline, v1.0 Hardening and listing" width="100%">

<samp>Features, exit gates and spikes per milestone are in ROADMAP.md.</samp>

<samp>The project is spec-driven: specs/ is the source of truth, and each feature goes from spec to failing tests to code. Instructions for Claude Code working on this repository are in CLAUDE.md.</samp>

claude --plugin-dir .          # run with the mod loaded (use a copy of fixtures/sample-app)
npx tsc -p . && claude plugin test && node scripts/check-calls.mjs
node scripts/trace.mjs --check # every approved acceptance criterion has a test

<samp><a href="LICENSE">MIT</a> © 2026 Lars Kühn · built with and for <a href="https://claude.com/claude-code">Claude Code</a></samp>

Source 1 files
hooks/register.tsx 27 lines
1// Throwaway recording probe for fixtures/sessions/subagents.ts (see ../../README.md). Not part of the mod.
2// It logs tool.call (with the next(e) result), agent.spawn, turn.start and session.start as the hooks see them, plus the rows
3// of $.agent.list() after each spawn and at each subagent's first tool.call.
4import type { Register } from 'claude-code'
5
6const OUT = '/tmp/cc-graph-subagents/log-' + Date.now() + '.json'
7const rows: any[] = []
8const seen = new Set<string>()
9const open = (event: string, input: any) => { const row: any = { event, input }; rows.push(row); return row }
10const dump = ($: any) => $.fs.write(OUT, JSON.stringify(rows, null, 1))
11const log = async (event: string, $: any, e: any, next: any) => {
12  const row = open(event, e)
13  if (event === 'tool.call' && e.agentId && !seen.has(e.agentId)) { seen.add(e.agentId); row.agentListAtFirstCall = await $.agent.list() }
14  const out = await next(e)
15  row.result = out
16  if (event === 'agent.spawn') row.agentListAfterSpawn = await $.agent.list()
17  await dump($)
18  return out
19}
20
21export const register: Register = on => {
22  on('session.start', ($, e: any, next: any) => log('session.start', $, e, next)).catch(($, e, next) => next(e))
23  on('turn.start', ($, e: any, next: any) => log('turn.start', $, e, next)).catch(($, e, next) => next(e))
24  on('agent.spawn', ($, e: any, next: any) => log('agent.spawn', $, e, next)).catch(($, e, next) => next(e))
25  on('tool.call', ($, e: any, next: any) => log('tool.call', $, e, next)).catch(($, e, next) => next(e))
26}
27