Claude Code Plugins

Community-maintained marketplace

Feedback

analyzing-codebases

@devame/llm-context-tools
1
0

Generates LLM-optimized code context with function call graphs, side effect detection, and incremental updates. Processes JavaScript/TypeScript codebases to create compact semantic representations including multi-level summaries, entry point identification, and hash-based change tracking. Provides 74-97% token reduction compared to reading raw source files. Useful for understanding code architecture, debugging complex systems, reviewing pull requests, and onboarding to unfamiliar projects.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name analyzing-codebases
description Generates LLM-optimized code context with function call graphs, side effect detection, and incremental updates. Processes JavaScript/TypeScript codebases to create compact semantic representations including multi-level summaries, entry point identification, and hash-based change tracking. Provides 74-97% token reduction compared to reading raw source files. Useful for understanding code architecture, debugging complex systems, reviewing pull requests, and onboarding to unfamiliar projects.

LLM Context Tools

Generate compact, semantically-rich code context for LLM consumption with 99%+ faster incremental updates.

What This Skill Does

Transforms raw source code into LLM-optimized context:

  • Function call graphs with side effect detection
  • Multi-level summaries (System → Domain → Module)
  • Incremental updates (only re-analyze changed files)
  • Hash-based change tracking
  • Query interface for instant lookups

Token Efficiency: 74-97% reduction vs reading raw files

When To Use

✅ User asks to "analyze this codebase" ✅ User wants to understand code architecture ✅ User needs help debugging or refactoring ✅ You need efficient context about a large codebase ✅ User wants LLM-friendly code documentation

Quick Start

# Check if tool is available
llm-context version

# If not available, user needs to install
# See setup.md for installation

# Run analysis
llm-context analyze

# Query results
llm-context stats

How It Works

Progressive Disclosure Strategy

Read in this order for maximum token efficiency:

  1. L0 (200 tokens) → .llm-context/summaries/L0-system.md

    • Architecture overview, entry points, statistics
  2. L1 (50-100 tokens/domain) → .llm-context/summaries/L1-domains.json

    • Domain boundaries, module lists
  3. L2 (20-50 tokens/module) → .llm-context/summaries/L2-modules.json

    • File-level exports, entry points
  4. Graph (variable) → .llm-context/graph.jsonl

    • Function details, call relationships, side effects
  5. Source (as needed) → Read targeted files only

Never read raw source files first! Use summaries and graph for context.

Common Commands

# Analysis
llm-context analyze              # Auto-detect full/incremental
llm-context check-changes        # Preview what changed

# Queries
llm-context stats                # Show statistics
llm-context entry-points         # Find entry points
llm-context side-effects         # Functions with side effects
llm-context query calls-to func  # Who calls this?
llm-context query trace func     # Call tree

Usage Patterns

Pattern 1: First-Time Codebase Understanding

# 1. Run analysis
llm-context analyze

# 2. Read L0 for overview
cat .llm-context/summaries/L0-system.md

# 3. Get statistics
llm-context stats

# 4. Find entry points
llm-context entry-points

Response template:

I've analyzed the codebase:

[L0 content - architecture, components, entry points]

Statistics: X functions, Y files, Z calls

Would you like me to:
1. Explain a specific domain?
2. Trace a function's call path?
3. Review the architecture?

Pattern 2: After Code Changes

# Incremental analysis (only changed files)
llm-context analyze

# See what changed
cat .llm-context/manifest.json

Response: Highlight new/modified functions and their impact

Pattern 3: Debugging

# Find function
llm-context query find-function buggyFunc

# Trace calls
llm-context query trace buggyFunc

# Check side effects
llm-context side-effects | grep buggy

Response: Explain call path and identify potential issues based on side effects

Detailed Guides

Setup & Installation: See setup.md Usage Examples: See examples.md Command Reference: See reference.md Common Workflows: See workflows.md

Side Effect Types

When analyzing functions, these effects are detected:

  • file_io - Reads/writes files
  • network - HTTP, fetch, API calls
  • database - DB queries, ORM
  • logging - Console, logger
  • dom - Browser DOM manipulation

Graph Format

Each function in graph.jsonl:

{
  "id": "functionName",
  "file": "path/file.js",
  "line": 42,
  "calls": ["foo", "bar"],
  "effects": ["database", "network"]
}

Best Practices

✅ DO

  • Read L0 → L1 → L2 → Graph → Source (in order)
  • Use queries before reading files
  • Run incremental analysis after edits
  • Mention detected side effects when debugging
  • Check manifest age before using cached data

❌ DON'T

  • Read raw source files first
  • Grep through files manually
  • Re-read entire codebase on changes
  • Skip summaries and go straight to source

Token Efficiency

Traditional approach:

  • Read 10 files = 10,000 tokens
  • Missing: call graphs, side effects

LLM-context approach:

  • L0 + L1 + Graph = 500-2,000 tokens
  • Includes: complete context + relationships

Savings: 80-95%

Performance

Initial Analysis

  • 100 files: 2-5s
  • 1,000 files: 30s-2min
  • 10,000 files: 5-15min

Incremental Updates

  • 1 file: 30-50ms
  • 10 files: 200-500ms
  • 50 files: 1-2s

Key: Incremental is 99%+ faster at scale

Troubleshooting

"No manifest found" → Run llm-context analyze first

"Cannot find module" → User needs to install: See setup.md

"Graph is empty" → No JavaScript files found. Check directory.

Success Criteria

This skill is working when:

  1. ✅ Analysis completes without errors
  2. .llm-context/ exists with all files
  3. llm-context stats shows functions
  4. ✅ You use summaries before reading source
  5. ✅ Token usage is 50-95% less than raw reading

Summary

Transform from:

  • ❌ Reading thousands of lines token-by-token
  • ❌ Missing global context
  • ❌ Slow re-analysis

To:

  • ✅ Compact semantic representations
  • ✅ Call graphs + side effects
  • ✅ 99%+ faster incremental updates
  • ✅ 50-95% token savings