wulun811/LiuHe

**LLM-Native Code Operations Toolkit** — Code tooling reinvented for the LLM that has no hands, no eyes, and no memory.
**English** | 简体中文 | Docs | GitHub
**Malong LiuHe** is a toolkit built for LLMs rather than humans. It ships two components:
- **malong/** — an MCP toolset (44 tools): symbol read/write, indexing, impact analysis, reference tracing, call graphs, dead-code detection, code review, security scanning, atomic batch editing, test orchestration, pipeline verification — served to the LLM over MCP (JSON-RPC over stdio).
- **malong-parse/** — a Rust parsing service: a tree-sitter-based multi-language symbol extraction engine (JavaScript/TypeScript/TSX/Python/Go/Rust/C/C++/Java/Bash), talking to the toolset over a Unix socket, with an LRU tree cache and batched parallel extraction.
Core Idea
Traditional tools (git, sed, IDE) assume a user with hands, eyes, and memory. An LLM has none of these. Malong LiuHe is redesigned around that fact:
| Missing | Design Consequence |
|---------|--------------------|
| **No hands** | Operations must be atomic, undoable, retryable (edit_transaction with rollback and undo journal) |
| **No eyes** | Output must be structured and self-explanatory (JSON for direct consumption; errors carry suggestion/next_action) |
| **No memory** | Every call is self-contained (takes workspace_dir); optimistic concurrency with version anchors (read_symbol → write_symbol(base_version) conflict state machine) |
Performance (Measured)
| Metric | Value | Note |
|--------|-------|------|
| Small-file read P95 | **1ms** | 50 runs, warmed |
| Small-file write P95 | **7ms** | 30 runs, incl. post-write re-index |
| repo_map | **98ms** | down from tens of seconds (SQLite index + Rust parsing) |
| Memory | **RSS 134MB / 26%** | under a real docker --memory=512m cgroup limit |
| Concurrency | **128 concurrent / 256 in-flight requests, zero OOM** | 32-way mixed read+write on one hot file, no tearing, DB integrity_check PASS |
| Throughput | **~588 calls/s** | single-threaded ceiling — 60–600× real LLM demand |
| Token savings | **↓65.3%** | same task 7673 → 2662 est. (primitives) |
| Call reduction | **↓50.0%** | 6 calls (legacy) → 3 calls (primitives) |
| Tool-description compression | **44 tools ≈ 1.33k tokens** | tiered: core kept, low-freq ≤70 chars, verbose ≤230; detail flows via next_step hints |
| Index throughput | **538 files / 7s** | scoped reindex, measured |
| dry_run fidelity | **47/47** | golden-hash comparison across 3 languages, 100% |
| Concurrent writes to same symbol | **16/16** | exactly one winner + one conflict, zero silent overwrites |
Language Support
| Language | Symbol Extraction | Reference Extraction |
|----------|:---:|:---:|
| JavaScript / TypeScript / TSX / JSX / MTS / CTS | ✓ | ✓ |
| Python | ✓ | ✓ |
| Go | ✓ | ✓ |
| Rust (impl blocks / trait / enum semantics) | ✓ | ✓ |
| C / C++ (incl. headers) | ✓ | ✓ |
| Java | ✓ | ✓ |
| Bash | ✓ | ✓ |
Tool Overview
| Category | Tools |
|----------|-------|
| **I/O primitives** | read_symbol (symbol read + version), write_symbol / write_symbols (safe write: conflict state machine + atomic write + undo) |
| **Index & search** | reindex, symbol_search, code_search (keyword + regex intent patterns — NOT semantic search), repo_map (98ms file map; skeleton pagination: page/page_size/prefix — truncated output always carries the full top-level skeleton with page numbers)
| **Analysis** | impact_analysis (blast radius + risk), call_chain, references, dep_graph, inspect (outline + refs + call chain in one), trace_symbol (constant tracing + hardcoded copies) |
| **Editing** | edit_batch, edit_transaction (atomic txn + rollback + cross-session lock), edit_collision_guard (write conflict detection), edit_sandbox (pre-write validation), diff_facts (symbol changes + test sync), rename_symbol (atomic cross-file rename), git_worktree (isolated branch + verify + merge) |
| **Quality gates** | code_review, security_review (regex security patterns), code_quality (5-dim probe), style_sniffer, guard_patterns, naming_consistency, exception_guard, config_drift, dependency_gatekeeper, fix_imports, sweep_dead_code, mock_sync — all pure regex/AST, zero LLM calls, deterministic |