Claude Code Plugin

The fastest way to get repowise — Claude handles everything.

Table of contents

  1. Overview
  2. Installation
  3. Slash commands
    1. /repowise:init
    2. /repowise:status
    3. /repowise:update
    4. /repowise:search
    5. /repowise:ask
    6. /repowise:context
    7. /repowise:symbol
    8. /repowise:reindex
    9. /repowise:health
    10. /repowise:risk
    11. /repowise:security
    12. /repowise:coverage
    13. /repowise:impacted-tests
    14. /repowise:dead-code
    15. /repowise:export
    16. /repowise:decision
    17. /repowise:why
    18. /repowise:doctor
  4. Automatic behaviors
    1. Codebase exploration
    2. Pre-modification checks
    3. Change review
    4. Code health
    5. Architectural decision queries
    6. Dead code awareness
  5. How skills work
  6. Requirements

Overview

The Claude Code plugin integrates repowise directly into Claude Code. It handles installation, API key setup, MCP server registration, and teaches Claude to use repowise tools proactively — without manual configuration.

Installation

Open Claude Code and run:

/plugin marketplace add repowise-dev/repowise
/plugin install repowise@repowise

That’s the entire installation. The plugin:

  • Installs repowise via pip if not already installed
  • Registers the MCP server with Claude Code
  • Loads the slash commands
  • Configures Claude to use repowise tools automatically

Slash commands

/repowise:init

Interactive setup and indexing for the current repository.

Claude will guide you through:

  1. Mode selection — choose between:
    • Full — complete wiki generation with LLM docs (requires API key)
    • Index-only — graph + git + dead code + code health, no LLM (free)
    • Advanced — manual control over provider, concurrency, exclude patterns

  2. Provider selection — Anthropic, OpenAI, Gemini, Codex CLI, OpenCode, or local Ollama

  3. API key entry — saved to .repowise/.env (gitignored) when a hosted provider needs one

  4. Indexing — runs in the background with live progress updates

When done, Claude confirms the MCP server is active and the codebase is queryable.

/repowise:status

Show the current state of the repowise index.

Output includes:

  • Last sync commit and timestamp
  • Total pages, symbols, decisions indexed
  • Provider and model in use
  • Pages marked stale (need regeneration)
  • MCP server connection status

/repowise:update

Incrementally sync the wiki after code changes.

Claude detects which files have changed since the last indexed commit, regenerates only the affected pages, updates CLAUDE.md, and confirms when done.

/repowise:search

Search the indexed wiki from within Claude Code.

Claude will ask for your query and search mode (fulltext, semantic, or symbol), then display results inline with links to relevant pages.

/repowise:ask

Ask a codebase question and get a cited, synthesised answer (repowise ask). This is the CLI adapter over the get_answer MCP tool — hybrid retrieval plus an LLM answer — so it costs a model call where /repowise:search does not.

Use it when you want an answer with confidence and citations, not a hit list. --scope restricts retrieval to a path prefix; --format json / --full match the other tool-adapter commands.

/repowise:context

Triage card for files, modules, or symbols (repowise context). CLI adapter over get_context: architectural layer, hotspot / fix history, doc freshness, and optional relationship blocks (--include callers, skeleton, …). Batch targets in one call. No source bytes by default — use /repowise:symbol for exact bodies.

/repowise:symbol

Read one symbol body with live-verified line bounds (repowise symbol). CLI adapter over get_symbol. Accepts path/to/file.py::Name, a live range (path:start-end), or a repowise#<hex> distill omission ref. --context-lines adds surrounding lines; a truncated body returns a continuation you can pass straight back.

/repowise:reindex

Rebuild the vector search index without making LLM calls.

Use this after switching embedding providers, or if semantic search results seem off.

/repowise:health

Code-health KPIs, lowest-scoring files, refactoring targets, or trends (repowise health).

/repowise:risk

Repo-relative review percentile/classification with a supporting diff-shape score (repowise risk).

/repowise:security

Full-history secret scan (repowise security scan --history). Working-tree scanning already runs during init/update; without --history the CLI only prints a hint. Default history mode is secrets-only; --all-patterns also reports code-smell patterns.

/repowise:coverage

Ingest or inspect coverage reports (repowise coverage add / status). Lights up untested-hotspot markers and builds the per-test map when contexts are present.

/repowise:impacted-tests

Tests whose coverage intersects a change (repowise impacted-tests).

/repowise:dead-code

Unreachable files, unused exports, and zombie packages by confidence.

/repowise:export

Export wiki pages or the architecture model (repowise export). Formats: markdown (default), html, json, and structurizr. JSON --full keeps tombstones and adds decisions / dead code / hotspots. Structurizr can emit a model fragment or a --standalone workspace (--components, --no-externals, --force).

/repowise:decision

List, inspect, add, or confirm architectural decisions.

/repowise:why

Why the code is shaped this way (repowise why). CLI adapter over get_why: decision search by question, governing decisions + origin story for a path, target-anchored search (--target), or the decision-health dashboard with no args. Falls back to git archaeology when no decision is recorded. /repowise:decision manages records; this command queries them.

/repowise:doctor

Diagnose (and optionally repair) setup, keys, and index drift.

Automatic behaviors

Beyond the slash commands, the plugin teaches Claude skills it uses automatically — without being asked.

Codebase exploration

Before reading raw source files, Claude calls:

  • get_overview() at the start of new tasks to orient itself
  • get_answer(question) for direct how/where/why questions
  • search_codebase(query) to locate code instead of using grep
  • get_context(targets) to get docs and ownership before opening files

Pre-modification checks

Before editing any file, Claude calls get_risk(targets) to assess:

  • Bug-fix history (defect_profile / bug_magnet) when present
  • Whether the file is a hotspot (high churn)
  • How many other files depend on it
  • Whether there are co-change patterns to be aware of

If the risk is high, Claude surfaces this before making changes.

Change review

For a PR / branch / working-tree diff, Claude combines get_change_risk (whole-change score) with get_risk’s per-file directive block.

Code health

Quality / complexity / “what to refactor” questions go through get_health.

Architectural decision queries

When facing “why is this structured this way” questions, Claude calls get_why(query) to check decision records and git archaeology before suggesting changes that might conflict with existing decisions.

Dead code awareness

During refactoring or cleanup tasks, Claude calls get_dead_code() to find confirmed unused code rather than guessing.

How skills work

Skills in Claude Code are prompt instructions that modify Claude’s behavior. The repowise plugin registers six skills that Claude loads when working in an indexed repo.

You don’t need to trigger them manually. When Claude detects it’s working in a repo with a connected repowise MCP server, the skills activate automatically.

The CLAUDE.md generator reinforces these skills by writing the mandatory MCP tool workflow directly into your project’s context file — so even without the plugin, any Claude session that reads your CLAUDE.md will follow the same workflow.

Requirements

  • Claude Code (desktop app, CLI, or IDE extension)
  • Python 3.11+
  • An LLM API key (for full mode — not needed for index-only)