Map, audit, and refactor a 1000-file codebase the way a senior engineer would — but with 17 specialist reviewers running in parallel.
A Claude Code plugin that builds a dependency + call graph, segments the repo by integration / domain, dispatches per-segment specialist reviewers in waves of ≤5, synthesizes findings into a P0–P3 refactor backlog, and optionally applies surgical fixes behind a confirmation gate.
Inside Claude Code:
/plugin install cartographer
Cartographer is published in the official Claude Code plugin marketplace — no marketplace add step needed.
/cartographer:audit # full pipeline — graph → segments → review → backlog → fixes
/cartographer:map # graph + diagrams only, no review (cheap, no LLM cost)
Or just say it in plain English — the skill triggers on phrases like "audit this project", "map out my codebase", "find duplication across the repo".
The plugin walks you through scope → clarifying questions → segment confirmation → cost estimate → review → backlog → fixes, asking before any expensive step.
- Dependency + call graph + class hierarchy (Python AST-precise, JS/TS/Go regex-best-effort)
- Pipeline traces from entry points (API handlers,
main(), workers) — one Mermaid flowchart per pipeline - OpenAPI extraction + per-endpoint deep call trace with cross-endpoint reuse map
- 17 specialist reviewers: auth-security, supabase, db-schema, telegram-bot, ai-pipeline, backend-api, data-pipeline, queue-worker, webhook-integration, frontend-ui, frontend-designer, mobile, cli-tool, test-suite, file-storage, realtime-streaming, devops-config, caching, generalist
- Cross-cutting synthesis — duplication, naming drift, centralization candidates
- Refactor backlog ranked P0–P3
- Fix application (opt-in, behind a gate) — branch isolation, surgical edits, test runner
Phase 0 Scope & opt-in (interactive)
Phase 1 Build dep + call + class graph (script)
Phase 1.5 Trace pipelines from entry points (script)
Phase 1.6 OpenAPI + per-endpoint deep call trace (script)
Phase 2 Detect & label functional segments (script)
Phase 3 Plan review waves (script)
Phase 3.5 Match a specialist to each segment (script)
Phase 4 Dispatch review subagents (≤5 per wave) (Claude)
Phase 5 Synthesize cross-cutting findings (script + Claude)
Phase 6 Final report + refactor backlog (Claude)
Phase 7 Apply backlog fixes (opt-in) (Claude + script)
Phases 1 → 3.5 are deterministic Python (~30s for 300 files, no LLM cost). Only Phase 4 and 7 spend subagent invocations. A 1000-file repo with 20 segments costs ~20 subagent calls, not 1000 file reads.
For CI or to just get the diagrams:
pip install pathspec networkx rapidfuzz
bash run_pipeline.sh /path/to/your/repo # writes to <repo>/.cartographer/
bash run_pipeline.sh /path/to/your/repo --readonly # writes to ~/.cartographer/<hash>/- Python — AST-precise (full call graph + class hierarchy with bases/methods/fields)
- JavaScript / TypeScript — regex-based, best-effort (TS path aliases resolved)
- Go — regex-based, best-effort
Tree-sitter for non-Python is on the roadmap.
MIT — see LICENSE.
PRs welcome. Open an issue first for non-trivial changes — the issue templates cover bug reports and feature requests. New specialist reviewers are the easiest contribution: drop a markdown file under agents/ with YAML frontmatter and the matcher auto-discovers it.