Skip to content

COMPARISON.md

Latest commit

 

History

History
157 lines (135 loc) · 8.79 KB

COMPARISON.md

File metadata and controls

157 lines (135 loc) · 8.79 KB
Raw

LegnaCode vs Claude Code Feature Comparison

🌐 中文文档

LegnaCode is built on top of the Claude Code upstream project, deeply enhanced while maintaining full compatibility. Below is a detailed side-by-side comparison.

Core Capabilities

Feature Claude Code LegnaCode
Terminal AI coding assistant ✅ ✅
45+ built-in tools ✅ ✅
MCP protocol support ✅ ✅
Multi-agent collaboration ✅ ✅
Plugin & skill system ✅ ✅
Git workflow management ✅ ✅

Models & Backends

Feature Claude Code LegnaCode
Anthropic API ✅ ✅
AWS Bedrock ✅ ✅
GCP Vertex ✅ ✅
Azure OpenAI ❌ ✅
MiniMax deep native integration ❌ ✅ 6 multimodal tools auto-registered
OpenAI-compatible bridge ❌ ✅ Full streaming bridge with SSE→Anthropic event translation
Auto API format detection ❌ ✅ URL-based: /anthropic suffix → Anthropic SDK, else → OpenAI fetch
apiFormat setting ❌ ✅ Force Anthropic or OpenAI per profile, or auto-detect
DeepSeek / Qwen / GLM / Kimi / MiMo ❌ ✅ 7 dedicated adapters, dual-endpoint (Anthropic + OpenAI)
reasoning_content passback ❌ ✅ Auto-extracts thinking blocks for DeepSeek/Kimi/MiMo multi-turn
MiniMax reasoning_details ❌ ✅ Array format thinking content in OpenAI streaming
Provider-specific finish_reason ❌ ✅ sensitive (GLM), repetition_truncation (MiMo), content_filter
Intelligent model routing ❌ ✅ Auto-selects model tier by prompt complexity
Model Adapter architecture Partial ✅ 7 adapters + OpenAI-compat bridge, per-adapter apiFormat
JSON repair for weak models ❌ ✅ Fixes markdown fences, trailing commas, unbalanced brackets

Multimodal (MiniMax-exclusive)

Feature Claude Code LegnaCode
AI image generation ❌ ✅ MiniMaxImageGenerate
AI video generation ❌ ✅ MiniMaxVideoGenerate
AI speech synthesis ❌ ✅ MiniMaxSpeechSynthesize
AI music generation ❌ ✅ MiniMaxMusicGenerate
AI image understanding ❌ ✅ MiniMaxVisionDescribe
AI web search ❌ ✅ MiniMaxWebSearch
Multimodal workflow orchestration ❌ ✅ Auto-chains multiple tools for complex tasks
Multimodal skill packs ❌ ✅ 5 built-in skills to guide AI orchestration

Memory System

Feature Claude Code LegnaCode
MEMORY.md persistent memory ✅ ~8K tokens/turn ✅ Compatible & preserved
4-layer memory stack ❌ ✅ L0+L1 only ~800 tokens/turn (88% savings)
Vector semantic search ❌ ✅ TF-IDF + cosine similarity, <5ms
DrawerStore persistence ❌ ✅ SQLite + WAL audit log
Temporal knowledge graph ❌ ✅ Entity-relation + time-validity queries
Room auto-classification ❌ ✅ 6 categories (facts/decisions/events/...)
PreCompact memory saving ❌ ✅ Auto-extracts high-value exchanges before compaction
Exchange-pair chunking ❌ ✅ Q+A pairing + 5-category scored tagging
Cross-session search ❌ ✅ /recall command
Memory Provider plugins ❌ ✅ Pluggable memory backends
Content-hash deduplication ❌ ✅ sha256 + 30s window prevents duplicate observations
Token economics tracking ❌ ✅ discoveryTokens + readTokens per drawer
Relevance feedback ❌ ✅ Frequently recalled memories get up to +100% boost
90-day time decay ❌ ✅ Old memories fade but never fully disappear
Privacy tag filtering ❌ ✅ <private> content auto-redacted before extraction
L0/L1/L2 content tiering ❌ ✅ Budget-driven degradation (full → summary → one-liner)
Cross-session knowledge ❌ ✅ Auto-writes .legna/knowledge.md on session end
Project-local memory ❌ Global ~/.claude/ ✅ <cwd>/.legna/memory/ with auto-migration

Context Management

Feature Claude Code LegnaCode
Auto Compact ✅ ✅
Micro Compact ✅ ✅
History Snip ✅ ✅
Tool output pre-pruning ❌ ✅ Auto-trims large tool_result payloads
Budget pressure injection ❌ ✅ Nudges model to wrap up when context >80%
Worker thread pool ❌ ✅ Offloads large file operations

Agent Enhancements

Feature Claude Code LegnaCode
Sub-agent spawning ✅ ✅
Team collaboration ✅ ✅
RPC subprocess tool execution ❌ ✅ UDS RPC, compresses multi-step ops into one inference
Autonomous skill detection ❌ ✅ Detects repeated patterns, auto-creates SKILL.md
Self-evolution loop (Hermes) ❌ ✅ Auto-learns corrections, preferences, patterns to memory
Background Review Agent ❌ ✅ Post-session experience extraction to .legna/memory/
Nudge system ❌ ✅ Counter-driven session learning summary
Tool schema export ❌ ✅ Anthropic-compatible format
Parallel file edit mode ❌ ✅ One sub-agent per file + sibling skeletons
Code Graph (symbol index) ❌ ✅ Regex-based, TS/JS/Python/Go/Rust, incremental mtime
Blast radius analysis ❌ ✅ blastRadius() — files affected if a file changes
Caller tracing ❌ ✅ traceCallers() — who calls this symbol
Tool call loop detection ❌ ✅ Same (tool, args) 3+ times → blocks
Negative feedback detection ❌ ✅ Detects frustration, injects strategy-shift hint (EN/ZH/JA)
Error file pre-injection ❌ ✅ Bash fail → auto-reads files from stderr/compiler output
First-read full file ❌ ✅ Forces full read on first encounter, prevents fragmented reads
Compound engineering ❌ ✅ Auto-writes docs/solutions/, prefetch searches past learnings

User Experience

Feature Claude Code LegnaCode
Verbose mode Off by default ✅ On by default
Token/Timer display Shows after 30s ✅ Shows from second 1
Compact status indicator ❌ Silent execution ✅ Displays "Compacting..."
Interruption reason display ❌ Only shows it stopped ✅ Shows specific interruption reason
Output retry indicator ❌ Silent retry ✅ Shows retry progress
Tool execution log ❌ ✅ Tool name + queue depth
Apple Terminal notifications ❌ Inverted logic bug ✅ Fixed
Pangu CJK spacing ❌ ✅ Auto-inserts spaces between CJK and ASCII in markdown
/undo command ❌ ✅ Reverts last file edit (Edit/Write), max 20 entries
Workflow engine ❌ Raw markdown only ✅ Structured steps with checks, retry, dependencies
Project-local plans ❌ Global ~/.claude/plans/ ✅ <cwd>/.legna/plans/

Configuration & Deployment

Feature Claude Code LegnaCode
Global config directory ~/.claude/ ~/.legna/ (auto-migrated)
WebUI admin panel ❌ ✅ legna admin
WebUI profile inline editing ❌ ✅ Edit any profile directly from card, per-file read/write API
WebUI preset templates ❌ ✅ 7 provider presets (DeepSeek/Kimi/GLM/Qwen/MiniMax/MiMo/Anthropic)
WebUI chat viewer ❌ ✅ Session replay with thinking/tool call visualization
WebUI live chat ❌ ✅ SSE streaming chat for API connectivity testing (single-turn only)
Kiro Gateway optimization ❌ ✅ Client-side history compression (thinking/tool_result/schema)
Config migration tool ❌ ✅ legna migrate
MiniMax authentication ❌ ✅ /auth-minimax
Pure TS syntax highlighting ❌ Requires native modules ✅ Zero native dependencies
Official registry install — ✅ --registry=https://registry.npmjs.org
Pixel office visualization ❌ ✅ LegnaCode Office — VS Code extension + Admin panel
Office real-time agent status ❌ ✅ Canvas 2D characters animate based on tool calls
Office conversation sidebar ❌ ✅ Collapsible per-agent conversation flow
Office join-key remote access ❌ ✅ 8-char shareable key for remote CLI instances
Office notification sounds ❌ ✅ Web Audio oscillator tones for state changes
Office layout persistence ❌ ✅ Custom office layout saved to ~/.legna-office/layout.json

Platform Support

Platform Claude Code LegnaCode
macOS arm64 ✅ ✅
macOS x64 ✅ ✅
Linux x64 ✅ ✅
Linux arm64 ✅ ✅
Windows x64 ✅ ✅

LegnaCode maintains full compatibility with the Claude Code upstream. All original features work as expected. Enhancements are non-invasive by design and do not affect existing workflows.