vrr/ruvector
2
0
Fork 0
mirror of https://github.com/ruvnet/RuVector.git synced 2026-05-25 15:03:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions
ruvector/docs/research/claude-code-rvsource/00-index.md
rUv e39b5901c1 feat(decompiler): rebuild all versions — organized source/rvf separation, 100% coverage 
Rebuilt all 4 versions from scratch:
- v0.2.x: 1,049 classes, 13,869 functions, 3,375 RVF vectors
- v1.0.x: 1,390 classes, 16,593 functions, 4,669 RVF vectors
- v2.0.x: 1,612 classes, 20,395 functions, 5,712 RVF vectors
- v2.1.x: 1,632 classes, 19,906 functions, 9,058 RVF vectors

Structure: source/ (17 JS modules in subfolders) + rvf/ (9 containers)
- Zero mixing: no JS in rvf dirs, no RVF in source dirs
- 100% code coverage: uncategorized/ catches everything
- 17 modules: core/3, tools/3, permissions/1, config/3, telemetry/1, ui/2, types/1, uncategorized/1
- 9 RVF containers per version (1 master + 8 per-category)

Co-Authored-By: claude-flow <ruv@ruv.net>
2026-04-03 03:18:41 +00:00

8.7 KiB
Raw Permalink Blame History

Claude Code CLI Source Analysis: Research Index

Project

Deep-dive reverse engineering of the Claude Code CLI (v2.1.91) internal architecture, based on binary analysis, string extraction, pattern matching, and configuration schema examination.

Methodology

This analysis used "agentic jujutsu" -- the tool analyzed itself by:

  1. Locating the binary and extension files on disk
  2. Extracting the embedded JavaScript source from the Bun SEA binary
  3. Pattern-matching against 12.8 MB of application code for class names, function signatures, string literals, and configuration patterns
  4. Analyzing the 76-property settings schema
  5. Cross-referencing 498 environment variables
  6. Mapping tool definitions, hook events, and MCP protocol methods

Research Documents

# Document Description
01 Overview and Binary Structure Binary format, installation paths, Bun SEA architecture, version management
02 Tool System 25+ built-in tools, MCP tool integration, tool schemas, validation, content block types
03 Agent Loop and Execution Flow Entry points, main loop, streaming, conversation management, slash commands, output formats
04 Permission System 6 permission modes, permission flow, sandbox integration, managed settings
05 MCP Integration 4 transports, 13 protocol methods, connection management, OAuth, tool discovery
06 Hooks System 6 hook events, command/HTTP hook types, lifecycle, security controls
07 Context and Session Management Token budgets, auto-compaction, session persistence, CLAUDE.md, file checkpointing, prompt caching
08 Configuration and Environment Settings hierarchy, 76 settings, 498 env vars, home directory structure
09 Agent and Subagent System Agent types, task/subagent lifecycle, skill system, plugin marketplace
10 Models and API 27+ model IDs, 5 provider backends, API endpoints, prompt caching, effort levels
11 Telemetry and Observability OpenTelemetry, Datadog, Perfetto, debugging, cost tracking
12 Dependency Graph Module relationships, data flow, state management, initialization sequence
13 Extension Points 13 extension mechanisms from CLAUDE.md to Agent SDK
14 Source Extraction Binary analysis, code metrics, extraction methods, dependency identification
15 Core Module Analysis Agent loop, tool dispatch, permissions, context management, MCP, streaming
16 Call Graphs Mermaid call graphs: boot, agent loop, tool dispatch, permissions, MCP, compaction
17 Class Hierarchy 1,557 classes, inheritance trees, AppState type, tool registry
18 State Machines Agent loop, permission, session, streaming, MCP, sandbox state machines

Extracted Source (v2.1.91)

Source and RVF cleanly separated. Master RVF: 9,058 vectors.

Directory Module Fragments Confidence
source/core/ agent-loop.js 77 High
source/core/ context-manager.js 49 High
source/core/ streaming-handler.js 24 High
source/core/ session.js 361 High
source/tools/ tool-dispatch.js 531 High
source/tools/mcp/ mcp-client.js 51 High
source/permissions/ permission-system.js 500 High
source/config/ config.js 473 High
source/config/ model-provider.js 165 Medium
source/config/ env-vars.js 223 Pattern
source/telemetry/ telemetry.js 524 High
source/telemetry/ telemetry-events.js 861 Pattern
source/ui/ commands.js 80 Medium
source/ui/ command-defs.js 93 Pattern
source/types/ class-hierarchy.js 1,467 Pattern
source/types/ api-endpoints.js 52 Pattern
source/uncategorized/ uncategorized.js 3,162 Low

RVF containers in rvf/: master.rvf (all), core.rvf, tools.rvf, permissions.rvf, config.rvf, telemetry.rvf, etc.

Additional Research

# Document Description
19 RuVector Integration Guide 6-tier integration plan: WASM MCP, agents, hooks, cache, SDK, plugin
20 SOTA Decompiler Research Survey of JSNice, DeGuard, DIRE, VarCLR + ruDevolution validation
21 Model Weight Analysis Embedded models, LoRA federation, GPU training, GGUF parsing

RVF Version Corpus

Version Latest Vectors RVF Size Bundle Classes Functions Modules
v0.2.x 0.2.126 3,375 1,731 KB 6.9 MB 1,049 13,869 17
v1.0.x 1.0.128 4,669 2,388 KB 8.9 MB 1,390 16,593 17
v2.0.x 2.0.77 5,712 2,918 KB 10.5 MB 1,612 20,395 17
v2.1.x 2.1.91 9,058 4,617 KB 12.6 MB 1,632 19,906 17

Tools

Tool Description
scripts/rebuild-all-versions.mjs Full rebuild of all version decompilations (Node.js)
scripts/claude-code-decompile.sh CLI decompiler (extract, beautify, split)
scripts/claude-code-rvf-corpus.sh Build RVF containers for all versions (shell wrapper)
npm/packages/ruvector/src/decompiler/ Decompiler library (module-splitter, metrics, witness)
npx ruvector decompile <package> npm CLI decompiler
examples/decompiler-dashboard/ Visual explorer (Vite + React)
crates/ruvector-decompiler/ Rust decompiler crate (MinCut + AI + witness)

ruDevolution SOTA Results

95.7% name accuracy — beats JSNice (63%), DIRE (65.8%), VarCLR (72%) by 23-35 points.

Trained on 8,201 pairs, 673K param transformer, pure Rust inference (<5ms, zero deps).

Key Findings

Architecture Summary

  • Runtime: Bun 1.3.11 Single Executable Application (229 MB binary)
  • Application code: ~12.8 MB of bundled, minified JavaScript
  • UI: React 18.3.1 WebView (VS Code) + Ink-style terminal (CLI)
  • API: Anthropic Messages API with streaming SSE
  • Extension: MCP client protocol with 4 transports

By the Numbers

Metric Count
Built-in tools 25+
Slash commands 39
Environment variables 498
Settings properties 76
Supported models 27+
MCP protocol methods 13
Hook event types 6
Permission modes 5 (acceptEdits, bypassPermissions, default, dontAsk, plan)
Extension mechanisms 13
Auth providers 5
MCP transports 4
Output formats 3
Source code classes 1,557
Functions (estimated) 19,464
Async generators (core loops) 6
Bundle size (minified) 11 MB / 4,836 lines

Architecture Pattern

Claude Code follows a plugin-oriented monolith pattern:

  • Single binary deployment (Bun SEA)
  • Modular internal architecture with clear subsystem boundaries
  • Extensive extension surface (MCP, hooks, agents, skills, plugins)
  • Multi-provider backend abstraction (Anthropic/AWS/GCP/Azure)
  • Layered security (permissions -> sandbox -> hooks -> managed settings)

Limitations

  • Source is minified/mangled: variable names are meaningless (e.g., Yq, f9)
  • Cannot trace exact function boundaries or module structure
  • V8 snapshot region (~100MB) could not be decompiled
  • Some patterns may be from bundled dependencies, not Claude Code itself
  • This analysis reflects v2.1.91; architecture may change between versions