Skip to content

CLAUDE_CODE_INTEGRATION.md

Latest commit

 

History

History
162 lines (129 loc) · 4.07 KB

CLAUDE_CODE_INTEGRATION.md

File metadata and controls

162 lines (129 loc) · 4.07 KB

Claude Code Architecture Integration

Research Summary

What I Learned from Anthropic GitHub Repos

1. Agent SDK Pattern

Claude Code is built on the Agent SDK which provides:

  • async query() for one-off tasks
  • ClaudeSDKClient for interactive sessions
  • Tools,.hooks, subagents, MCP servers as first-class configurations

2. Core Tools

Claude Code includes these built-in tools:

Tool Purpose
Read Read any file
Write Create new files
Edit Precise edits with search/replace
Bash Run terminal commands
Glob Find files by pattern
Grep Search content with regex
WebSearch Search the web
WebFetch Fetch page content
AskUserQuestion Clarifying questions
Task Spawn subagents

3. Skills System

Agent Skills are folders with SKILL.md files:

  • Auto-discovered from .claude/skills/ or ~/.claude/skills/
  • YAML frontmatter with metadata
  • Markdown body with instructions
  • Model invokes based on description matching

Example SKILL.md:

---
name: code-review
description: Security-focused code review
---
When reviewing code, check:
1. OWASP Top 10 vulnerabilities
2. Error handling completeness
3. Performance implications

4. Permission Modes

  • ask - Always ask for permission (default)
  • accept_edits - Auto-approve file edits, ask for other actions
  • auto - Auto-approve safe operations (dangerous)
  • never - Dry-run mode

5. Hooks System

Lifecycle hooks for custom behaviors:

  • PreToolUse / PostToolUse
  • SessionStart / SessionEnd
  • UserPromptSubmit

6. Model Context Protocol (MCP)

Open standard for connecting AI to external systems:

  • Primitives: Tools, Resources, Prompts
  • Transport: Stdio (local) or HTTP (remote)
  • Servers: Filesystem, Git, Memory, etc.

NVIDIA CLI v6.0 Integration

Tools Implemented (11 total)

Tool Claude Code NVIDIA CLI v6
read_file ✅ ✅
write_file ✅ ✅
edit_file ✅ ✅ (new)
execute_command ✅ ✅
glob_search ✅ ✅
grep_search ✅ ✅
list_directory ✅ ✅
web_search ✅ ✅
web_fetch ✅ ✅
git ✅ ✅
ask_user ✅ ✅

Skills System

Built-in Skills (4):

  • code-review - Security-focused reviews
  • refactor - Best practice refactoring
  • test - AAA pattern testing
  • debug - Scientific debugging

Auto-activation: Skills activate when user input matches regex patterns

Permission Modes

Full implementation with 4 modes:

  • ask - Always confirm
  • accept_edits - Auto-approve file writes
  • auto - Auto-approve safe operations
  • never - Dry-run

ReAct Agent Loop

Implemented the core ReAct pattern:

  1. Plan - User input + system prompt
  2. Act - Model generates tool calls
  3. Observe - Tools execute and return results
  4. Iterate - Model processes results, may call more tools

Hooks System

Framework implemented (extensible):

  • session_start
  • session_end
  • user_prompt
  • pre_tool_use
  • post_tool_use

New Slash Commands (v6.0)

Command Purpose
/init Analyze project
/add Add files to context
/drop Remove files from context
/glob Search files by pattern
/grep Search file contents
/skill Toggle skills
/mode Change permission mode
/status Show session metrics
/compact Compress history
/undo Remove last exchange

Files Created

  • nv.py - Main v6.0 with all features
  • nv_v6.py - Clean minimal version
  • CLAUDE_CODE_INTEGRATION.md - This documentation

About "Ralph Wiggum"

No public references found in Anthropic repos. It may be:

  • An internal codename
  • A community/third-party project
  • Not yet publicly documented

The NVIDIA CLI uses functional naming conventions instead of character-based naming.

Usage

# Install
pip install -e .

# Chat with ReAct agent
nv chat

# With specific mode
nv chat --mode accept_edits

# One-shot query
nv ask "explain this code"