claude-reimagined

A one-command bootstrap that turns a fresh machine into a fully wired Claude Code workstation — CLI, plugins, hooks, MCP servers, statusline, and a 40+ skill library, all configured to work together out of the box.

If you've ever spent an afternoon stitching together Claude Code, RTK, context-mode, code-review-graph, caveman, custom hooks, and a skill router — this repo is that afternoon, automated.

Table of Contents

• Install
• Repository Structure
• What You Get
• code-review-graph Per-Repo Setup
• Hooks Deep-Dive
• Settings
• Skills Library
• Where Files Land
• Rollback
• Requirements
• Troubleshooting
• Contributing
• License

Install

Three steps. macOS or Linux.

git clone https://github.com/kunaltulsidasani/claude-reimagined.git
cd claude-reimagined
./bootstrap.sh

The bootstrapper will:

1. Detect your OS and check for missing dependencies (curl, git, python3, npm, jq, pipx).
2. Walk through each component, explain what it does, and ask before installing — default is yes, just hit Enter to accept. Type n to skip.
3. Print a verification summary at the end.

Other modes

./bootstrap.sh --dry-run                              # preview only — no changes
./bootstrap.sh --yes                                  # auto-approve every prompt (no confirmations)
./bootstrap.sh --force                                # reinstall components already present
./bootstrap.sh --clean --yes                          # wipe ~/.claude first, then fresh install
./bootstrap.sh --skip rtk,caveman                     # skip specific components
./bootstrap.sh --skills-only python-pro,golang-pro    # install only chosen skills

Repository Structure

claude-reimagined/
├── bootstrap.sh                  # entry point — orchestrates every install script
├── statusline.sh                 # Claude Code statusline renderer
├── Makefile                      # `make install`, `make test`, `make verify`
├── CLAUDE.md                     # project-level Claude Code instructions
├── LICENSE                       # MIT
│
├── configs/
│   └── settings.json             # reference settings.json (user-managed copy)
│
├── hooks/
│   ├── pre-compact.sh            # PreCompact hook — survives auto-compaction
│   └── subagent-model-router.sh  # PreToolUse[Agent] hook — routes models per subagent
│
├── lib/
│   └── common.sh                 # shared bash helpers (logging, prompts, results)
│
├── scripts/
│   ├── install_deps.sh             # curl, git, python3, npm, jq, pipx
│   ├── install_claude_code.sh      # Claude Code CLI
│   ├── install_rtk.sh              # RTK token-killer proxy
│   ├── install_context_mode.sh     # context-mode MCP plugin
│   ├── install_code_review_graph.sh# code-review-graph MCP server
│   ├── install_caveman.sh          # caveman terse-mode plugin
│   ├── install_statusline.sh       # statusline registration
│   ├── install_subagent_router.sh  # subagent model-routing hook
│   ├── install_pre_compact.sh      # pre-compact hook
│   ├── install_skills.sh           # 39 skills from registry.yaml via sparse clone
│   └── verify_installation.sh      # post-install verification
│
├── skills/
│   └── registry.yaml             # curated skill registry (id → repo:path, tier)
│
├── tests/
│   ├── run_tests.sh              # zero-dep bash test runner
│   ├── lib/helpers.sh            # test fixtures and assertions
│   ├── unit/                     # unit tests — common.sh, hooks, routers
│   └── integration/              # E2E tests for every install script
│
└── logs/                         # ignored — populated at runtime

What You Get

Each component solves a specific pain point with vanilla Claude Code. You can skip any of them with --skip <id>.

Statusline Preview

Shows model, repo, branch, context %, cost, session time, 5h/7d burn rate, and active mode tag.

code-review-graph Per-Repo Setup

The bootstrapper installs the code-review-graph binary globally, but the MCP integration is per-repo — you opt in for each project where you want it.

cd /path/to/your/repo
code-review-graph install

That command:

1. Adds an MCP server entry to the project's .mcp.json (or creates one).
2. Sets up file-watcher hooks (via Claude Code's PostToolUse hook system) so the graph updates incrementally on every Edit, Write, or MultiEdit.
3. Builds the initial graph by parsing the repo with Tree-sitter (one-time, takes seconds to a minute depending on repo size).

Verify it worked

code-review-graph status      # shows nodes/edges/files counts
ls .code-review-graph/         # graph DB lives here, gitignored
cat .mcp.json                  # MCP server entry should be present

In Claude Code, the graph tools (detect_changes, get_review_context, query_graph, semantic_search_nodes, etc.) will appear in the tool list once you restart the session. See [CLAUDE.md](https://github.com/kunaltulsidasani/claude-reimagined/blob/main/CLAUDE.md) in this repo for an example of instructing Claude to prefer graph tools over Grep.

Update or rebuild

code-review-graph build       # incremental rebuild (called automatically by hooks)
code-review-graph build --full # full rebuild from scratch

Add .code-review-graph/ to gitignore

echo '.code-review-graph/' >> .gitignore

The graph database is local-machine state — don't commit it.

Hooks Deep-Dive

pre-compact — survive context compaction

• File: hooks/pre-compact.sh → ~/.claude/hooks/pre-compact.sh
• Fires: immediately before Claude auto-compacts (triggered at 80% context via CLAUDE_AUTOCOMPACT_PCT_OVERRIDE).
• Does: detects your stack, schema files, API directories, and git state, then writes preservation instructions to stdout. Claude's compaction summarizer reads these and decides what to keep.

Stack detection covers: JS/TS (Next.js, NestJS, React, Node), Go, Python (FastAPI, Django, Flask), Rust, Flutter, Java/Maven, Kotlin/Gradle, Ruby/Rails, Swift, C#/.NET. Each stack has its own preservation rules — test command, build command, schema file, API dirs.

The compacted summary is guaranteed to contain: active task and next step, exact error messages, every file touched, non-obvious decisions and their reasons, pending/blocked work, discovered env constraints.

subagent-model-router — auto-pick the cheapest capable model

• File: hooks/subagent-model-router.sh → ~/.claude/hooks/subagent-model-router.sh
• Fires: before every Agent tool call.
• Does: inspects subagent_type and prompt complexity, rewrites the tool input via updatedInput to pin a specific model.

Settings

~/.claude/settings.json is user-managed. The bootstrapper verifies it exists at the end. A reference copy lives in [configs/settings.json](https://github.com/kunaltulsidasani/claude-reimagined/blob/main/configs/settings.json).

Hooks registered in settings.json:

• PreCompact → pre-compact.sh
• PreToolUse[Agent] → subagent-model-router.sh

Skills Library

39 skills from community repos, sparse-cloned into ~/.claude/skills/<id>/. Use --skills-only <ids> to pick a subset.

Each entry in [skills/registry.yaml](https://github.com/kunaltulsidasani/claude-reimagined/blob/main/skills/registry.yaml) is tagged with a tier:

Gold tier (full anatomy with executable helpers)

Silver tier (helpers inline)

Bronze tier (knowledge skills)

Upstream sources (full repo list):

• anthropics/skills
• obra/superpowers
• awesome-skills/code-review-skill
• Jeffallan/claude-skills
• Mindrally/skills
• alirezarezvani/claude-skills
• seokheejang/claude-ops-skills
• SpillwaveSolutions/document-specialist-skill
• clear-solutions/unit-tests-skills

Where Files Land

Rollback

Every file modified by a bootstrap script is backed up first:

~/.claude/settings.json.bak.20240101_120000

Format: <original-path>.bak.YYYYMMDD_HHMMSS. To restore, copy the .bak.* file back over the original.

For a clean reinstall: ./bootstrap.sh --clean --yes wipes ~/.claude first.

Install results land in ~/.claude-bootstrap/logs/results.tsv with status OK, FAILED, SKIPPED, or DECLINED.

Requirements

• macOS 12+ or Linux (x86_64 or arm64)
• bash 4.0+
• Internet access for downloads

The bootstrapper installs missing deps automatically (Homebrew on macOS; apt-get/dnf/yum on Linux).

• Hard deps: curl, git, python3, npm
• Soft deps: jq, pipx, uv (for code-review-graph)

Troubleshooting

**[FAIL] CLAUDE.md or [FAIL] settings** — these files are user-managed. Create them at ~/.claude/CLAUDE.md and ~/.claude/settings.json. The bootstrapper only verifies their existence; it does not install them.

**rtk gain fails** — you may have the wrong rtk binary on your PATH (name collision with reachingforthejack/rtk Rust Type Kit). Run which rtk to check, then uninstall the wrong one (cargo uninstall rtk or brew uninstall rtk).

Component fails mid-bootstrap — non-critical components don't abort the run. Check ~/.claude-bootstrap/logs/results.tsv and the per-script logs in the same directory.

Skill missing scripts/ or reference/ after install — fixed in the latest install_skills.sh. Re-run with --force --skills-only <skill-id>.

Contributing

See CONTRIBUTING.md for development setup, testing, and how to add new components or skills.

Quick reference:

make install     # run bootstrap.sh interactively
make test        # run the full test suite (~140 tests)
make verify      # run verify_installation.sh

License

MIT