lean-ctx
GitHub通过压缩LLM上下文（最高99%）和提供81个MCP工具，优化AI代理的文件读取、代码搜索及Shell执行效率。支持树状AST分析和锚点编辑，旨在减少Token消耗并提升开发工作流性能。
Trigger Scenarios
Install
npx skills add AjayIrkal23/agentic-mercy-10x --skill lean-ctx -g -y
SKILL.md
Frontmatter
{
"name": "lean-ctx",
"description": "Context Engineering for AI Agents — 81 MCP tools, 10 read modes, 95+ shell patterns, tree-sitter AST for 27 languages. Compresses LLM context by up to 99%. Use when reading files, running shell commands, searching code, or exploring directories. Auto-installs if not present."
}
lean-ctx — Context Engineering for AI Agents
Setup
which lean-ctx || curl -fsSL https://raw.githubusercontent.com/yvgude/lean-ctx/main/skills/lean-ctx/scripts/install.sh | bash
lean-ctx setup
Core Tools (10 always visible)
| Tool | Purpose |
|---|---|
ctx_read(path, mode) |
Read file with compression and caching |
ctx_search(pattern, path) |
Search code with compressed results |
ctx_shell(command) |
Run shell with compressed output |
ctx_tree(path, depth) |
Directory listing |
ctx_patch(path, ops) |
Anchored editing (line+hash, no old-text echo) |
ctx_session(action) |
Session state and persistence |
ctx_knowledge(action) |
Project knowledge across sessions |
ctx_overview(task) |
Task-relevant project map |
ctx_graph(action) |
Code relationships and impact |
ctx_call(name, args) |
Invoke any tool by name |
Shell Hook (use instead of raw exec)
lean-ctx -c "git status"
lean-ctx -c "cargo test"
lean-ctx -c "npm install"
lean-ctx ls src/
ctx_read Modes
| Mode | When |
|---|---|
anchored |
Files you will edit (full text + N:hh| anchors for ctx_patch) |
full |
Verbatim cached read |
map |
Context-only (deps + exports) |
signatures |
API surface only |
diff |
After edits (changed lines) |
aggressive |
Large files, syntax-stripped; JSON arrays row-deduped (lossless) |
entropy |
Shannon filtering |
task |
Task-relevant lines |
lines:N-M |
Specific range |
auto |
System selects optimal |
Re-reads cost ~13 tokens. fresh=true bypasses cache.
Redundant JSON (arrays of like objects) is crushed losslessly into a compact
_defaults + per-row form; if a slice was dropped, recover it with
ctx_expand(id, json_path=… | search=…).
File Editing
Anchored editing saves output tokens: ctx_read(mode="anchored") → ctx_patch(path, op, line, hash, new_text).
Never reproduce old text byte-for-byte; batch via ops:[…]; op=create writes new files.
Stale anchor → CONFLICT with fresh anchors (retry once). Native Edit/StrReplace stay fine;
ctx_edit (str_replace) is the legacy fallback via ctx_call/power profile.
More Tools (via ctx_call or ctx_load_tools)
Architecture: ctx_symbol, ctx_callgraph, ctx_impact, ctx_architecture, ctx_routes, ctx_smells, ctx_quality ↳ "What breaks if I change this file/class/type?" → ctx_impact (file-level blast radius; resolves same-package/namespace type usage with no import for C#, Java, Go and Kotlin). "Who calls this function?" → ctx_callgraph (symbol-level). "How navigable / how much is complexity costing me?" → ctx_quality (navigability score + token quality tax). Multi-agent: ctx_agent, ctx_share, ctx_task, ctx_handoff, ctx_workflow Verify: ctx_benchmark, ctx_verify, ctx_proof, ctx_review Batch: ctx_fill, ctx_execute, ctx_expand, ctx_pack
Full docs: https://leanctx.com/docs
Version History
- 581d130 Current 2026-07-19 09:12
