update test

Author: shibing624

README.md

@@ -16,7 +16,16 @@
 [![GitHub issues](https://img.shields.io/github/issues/shibing624/agentica.svg)](https://github.com/shibing624/agentica/issues)
 [![Wechat Group](https://img.shields.io/badge/wechat-group-green.svg?logo=wechat)](#社区与支持)
 
-**Agentica** 是一个轻量级 Python 框架，用于构建 AI 智能体。Async-First 架构，支持工具调用、RAG、多智能体团队、工作流编排和 MCP 协议。
+**Agentica** 不是套一层 LLM API 的聊天壳，而是一个 Async-First 的 agent harness。
+它让 Agent 能真正跑起来: 调工具、跑长任务、做多智能体协作、跨会话保留记忆，并通过 Skill system 接入可演进的 self-learn 工作流。
+
+| 能力 | 说明 |
+|------|------|
+| **Long-running Agent Loop** | `Runner` 驱动的 LLM ↔ Tool 循环，内置压缩、重试、成本预算、死循环防护 |
+| **Works Beyond Chat** | 文件、执行、搜索、浏览器、MCP、多智能体、Workflow，不依附单一 IDE 场景 |
+| **Memory That Survives Sessions** | Workspace 记忆按条目存储、相关性召回，并可把确认过的偏好同步到 `~/.agentica/AGENTS.md` |
+| **Skill-Based Self-Learn** | SkillTool 可加载外部技能；内置 Agent 持续学习策略 |
+| **Open, Composable Harness** | 模型、工具、记忆、Skill、Guardrails、MCP 都是可替换部件，而不是封闭 SaaS 黑盒 |
 
 ## 架构
 
@@ -76,13 +85,13 @@ export DEEPSEEK_API_KEY="your-api-key"      # DeepSeek
 - **多智能体** — Team（动态委派）、Swarm（并行/自治）和 Workflow（确定性编排）
 - **安全守卫** — 输入/输出/工具级 Guardrails，流式实时检测
 - **MCP / ACP** — Model Context Protocol 和 Agent Communication Protocol 支持
-- **Skill 系统** — 基于 Markdown 的技能注入，模型无关
+- **Skill 系统** — 基于 Markdown 的技能注入，支持项目级、用户级和外部托管 skill 目录
 - **多模态** — 文本、图像、音频、视频理解
-- **持久化记忆** — 索引/内容分离、相关性召回、四类型分类、drift 防御
+- **持久化记忆** — 索引/内容分离、相关性召回、四类型分类、drift 防御，并可同步长期偏好到全局 `AGENTS.md`
 
 ## Workspace 记忆
 
-Workspace 提供跨会话的持久化记忆，采用索引/召回设计：
+Workspace 提供跨会话的持久化记忆，采用索引/召回设计；需要时还可以把确认过的用户/反馈记忆编译进全局 `~/.agentica/AGENTS.md`，让新 session 自动继承：
 
 ```python
 from agentica import Workspace
@@ -96,6 +105,7 @@ await workspace.write_memory_entry(
     content="User prefers concise, typed Python.",
     memory_type="feedback",              # user|feedback|project|reference
     description="python coding style",   # 相关性匹配关键词
+    sync_to_global_agent_md=True,        # 同步到 ~/.agentica/AGENTS.md 的 Learned Preferences 区块
 )
 
 # 相关性召回（根据当前 query 返回最相关的 ≤5 条）
@@ -105,23 +115,49 @@ memory = await workspace.get_relevant_memories(query="how to write python")
 Agent 自动根据当前 query 召回最相关记忆，而非全量注入：
 
 ```python
-from agentica import Agent, Workspace
+from agentica import DeepAgent, Workspace
 from agentica.agent.config import WorkspaceMemoryConfig
 
-agent = Agent(
+agent = DeepAgent(
     workspace=Workspace("./workspace"),
     long_term_memory_config=WorkspaceMemoryConfig(
         max_memory_entries=5,  # 最多注入 5 条相关记忆
+        sync_memories_to_global_agent_md=True,
     ),
 )
 ```
 
+`DeepAgent` 默认启用 `SkillTool(auto_load=True)`，会自动发现 `~/.agentica/skills/` 和 `.agentica/skills/` 目录下的 skill；同时默认开启 `tool_config.auto_load_mcp=True`，启动时会自动读取工作目录里的 `mcp_config.json/yaml/yml`。这样 DeepAgent 开箱就是带 skills + MCP + memory 的一键完全体。
+
 ## CLI
 
 ```bash
 agentica --model_provider zhipuai --model_name glm-4.7-flash
 ```
 
+安装外部 skill 集合：
+
+```bash
+agentica extensions install https://github.com/obra/superpowers
+```
+
+如果你已经进入交互式 CLI，也可以直接在会话里安装并立刻刷新当前 skills：
+
+```text
+> /extensions install https://github.com/obra/superpowers
+> /extensions list
+> /extensions remove learn-from-experience
+> /extensions reload
+```
+
+也支持安装本地目录或指定目标目录：
+
+```bash
+agentica extensions install /path/to/skill-repo --target-dir ~/.agentica/skills
+```
+
+如果你安装到自定义目录而不是标准搜索路径，记得把这个目录加入 `AGENTICA_EXTRA_SKILL_PATH`，这样 `DeepAgent` 和 CLI 才会自动发现它。
+
 <img src="https://github.com/shibing624/agentica/blob/main/docs/assets/cli_snap.png" width="800" />
 
 ## Web UI

README_EN.md

@@ -16,7 +16,16 @@
 [![GitHub issues](https://img.shields.io/github/issues/shibing624/agentica.svg)](https://github.com/shibing624/agentica/issues)
 [![Wechat Group](https://img.shields.io/badge/wechat-group-green.svg?logo=wechat)](#community--support)
 
-**Agentica** is a lightweight Python framework for building AI agents. Async-First architecture with support for tool calling, RAG, multi-agent teams, workflow orchestration, and MCP protocol.
+**Agentica** is not just a chat wrapper around an LLM API. It is an Async-First agent harness for running real agents:
+tool calling, long-running task loops, multi-agent orchestration, cross-session memory, and skill-driven self-learning workflows.
+
+| Capability | What it means |
+|------------|---------------|
+| **Long-running Agent Loop** | `Runner` manages the LLM ↔ tool loop with compression, retries, cost budgets, and loop safety |
+| **Works Beyond Chat** | Files, execution, search, browser, MCP, multi-agent teams, and workflows instead of a single chat endpoint |
+| **Memory That Survives Sessions** | Workspace memory is stored as indexed entries with relevance recall, and confirmed preferences can sync into `~/.agentica/AGENTS.md` |
+| **Skill-Based Self-Learn** | SkillTool can load external skills, built-in agent self-learning strategy |
+| **Open Composable Harness** | Models, tools, memory, skills, guardrails, and MCP are replaceable building blocks instead of a closed hosted platform |
 
 ## Architecture
 
@@ -76,13 +85,13 @@ export DEEPSEEK_API_KEY="your-api-key"      # DeepSeek
 - **Multi-Agent** — Team (dynamic delegation), Swarm (parallel / autonomous), and Workflow (deterministic orchestration)
 - **Guardrails** — Input / output / tool-level guardrails, streaming real-time detection
 - **MCP / ACP** — Model Context Protocol and Agent Communication Protocol support
-- **Skill System** — Markdown-based skill injection, model-agnostic
+- **Skill System** — Markdown-based skill injection with project, user, and managed external skill directories
 - **Multi-Modal** — Text, image, audio, video understanding
-- **Persistent Memory** — Index/content separation, relevance-based recall, four-type classification, drift defense
+- **Persistent Memory** — Index/content separation, relevance-based recall, four-type classification, drift defense, and optional sync into global `AGENTS.md`
 
 ## Workspace Memory
 
-Workspace provides persistent cross-session memory with index/recall design:
+Workspace provides persistent cross-session memory with index/recall design. When needed, confirmed user and feedback memories can also be compiled into global `~/.agentica/AGENTS.md` so new sessions inherit them automatically:
 
 ```python
 from agentica import Workspace
@@ -96,6 +105,7 @@ await workspace.write_memory_entry(
     content="User prefers concise, typed Python.",
     memory_type="feedback",              # user|feedback|project|reference
     description="python coding style",   # keywords for relevance scoring
+    sync_to_global_agent_md=True,        # sync into ~/.agentica/AGENTS.md
 )
 
 # Relevance-based recall (returns top-k most relevant entries for the query)
@@ -105,23 +115,49 @@ memory = await workspace.get_relevant_memories(query="how to write python")
 Agents automatically recall the most relevant memories for the current query, rather than dumping all memory:
 
 ```python
-from agentica import Agent, Workspace
+from agentica import DeepAgent, Workspace
 from agentica.agent.config import WorkspaceMemoryConfig
 
-agent = Agent(
+agent = DeepAgent(
     workspace=Workspace("./workspace"),
     long_term_memory_config=WorkspaceMemoryConfig(
         max_memory_entries=5,  # inject at most 5 relevant memories
+        sync_memories_to_global_agent_md=True,
     ),
 )
 ```
 
+`DeepAgent` enables `SkillTool(auto_load=True)` by default, so it automatically discovers skills from `~/.agentica/skills/` and `.agentica/skills/`; it also turns on `tool_config.auto_load_mcp=True`, which auto-loads `mcp_config.json/yaml/yml` from the working directory when present. In practice, `DeepAgent` now boots as a one-command runtime with skills + MCP + memory already wired in.
+
 ## CLI
 
 ```bash
 agentica --model_provider zhipuai --model_name glm-4.7-flash
 ```
 
+Install an external skill pack:
+
+```bash
+agentica extensions install https://github.com/obra/superpowers
+```
+
+If you are already inside the interactive CLI, you can install and refresh skills in-place:
+
+```text
+> /extensions install https://github.com/obra/superpowers
+> /extensions list
+> /extensions remove learn-from-experience
+> /extensions reload
+```
+
+Local directories and custom targets are also supported:
+
+```bash
+agentica extensions install /path/to/skill-repo --target-dir ~/.agentica/skills
+```
+
+If you install into a custom directory instead of a standard search path, add that directory to `AGENTICA_EXTRA_SKILL_PATH` so `DeepAgent` and the CLI can auto-discover it.
+
 <img src="https://github.com/shibing624/agentica/blob/main/docs/assets/cli_snap.png" width="800" />
 
 ## Web UI

agentica/agent/base.py

@@ -310,6 +310,9 @@ def _post_init(self):
                     tool.set_parent_agent(self)
                 elif isinstance(tool, BuiltinMemoryTool):
                     tool.set_workspace(self.workspace)
+                    tool.set_sync_global_agent_md(
+                        self.long_term_memory_config.sync_memories_to_global_agent_md
+                    )
 
         # Register BuiltinMemoryTool when memory=True and workspace exists
         if self.memory and self.workspace is not None:
@@ -318,6 +321,9 @@ def _post_init(self):
             if not has_memory_tool:
                 memory_tool = BuiltinMemoryTool()
                 memory_tool.set_workspace(self.workspace)
+                memory_tool.set_sync_global_agent_md(
+                    self.long_term_memory_config.sync_memories_to_global_agent_md
+                )
                 if self.tools is None:
                     self.tools = [memory_tool]
                 else:
@@ -353,7 +359,13 @@ def _post_init(self):
             if self.long_term_memory_config.auto_archive:
                 auto_hooks.append(ConversationArchiveHooks())
             if self.long_term_memory_config.auto_extract_memory:
-                auto_hooks.append(MemoryExtractHooks())
+                auto_hooks.append(
+                    MemoryExtractHooks(
+                        sync_memories_to_global_agent_md=(
+                            self.long_term_memory_config.sync_memories_to_global_agent_md
+                        )
+                    )
+                )
             if auto_hooks:
                 self._default_run_hooks = _CompositeRunHooks(auto_hooks)
 

agentica/agent/config.py

@@ -111,6 +111,9 @@ class WorkspaceMemoryConfig:
     # Only fires when the LLM didn't call save_memory during the run.
     # Costs one extra LLM request per run (mirrors CC's extractMemories service).
     auto_extract_memory: bool = False
+    # Recompile confirmed user/feedback memories into ~/.agentica/AGENTS.md so
+    # future sessions automatically inherit long-lived preferences.
+    sync_memories_to_global_agent_md: bool = False
 
 
 @dataclass

agentica/agent/deep.py

@@ -74,6 +74,7 @@ class DeepAgent(Agent):
     - 5-stage compression pipeline (compress_tool_results=True)
     - Context overflow handling at 80% (context_overflow_threshold=0.8)
     - Repeated tool-call detection at 3 (max_repeated_tool_calls=3)
+    - MCP auto-loading from local mcp_config.json/yaml when available
     - Workspace memory with relevance recall (max_memory_entries=10)
     - Conversation auto-archive (auto_archive=True)
     - Agentic prompt with datetime and agent name
@@ -160,6 +161,7 @@ def __init__(
 
         if tool_config is None:
             tool_config = ToolConfig(
+                auto_load_mcp=True,
                 compress_tool_results=True,
                 context_overflow_threshold=0.8,
                 max_repeated_tool_calls=3,
@@ -171,6 +173,7 @@ def __init__(
                 load_workspace_context=True,
                 load_workspace_memory=True,
                 max_memory_entries=10,
+                sync_memories_to_global_agent_md=True,
             )
 
         super().__init__(

agentica/cli/config.py

@@ -134,11 +134,66 @@ def _get_tool_import_path(tool_name: str) -> str:
 
 
 def parse_args():
-    parser = argparse.ArgumentParser(description='CLI for agentica')
-    
     # Check if running in ACP mode (special handling)
     if len(sys.argv) > 1 and sys.argv[1] == 'acp':
         return None  # Signal to run in ACP mode
+
+    if len(sys.argv) > 1 and sys.argv[1] == "extensions":
+        parser = argparse.ArgumentParser(description="Manage Agentica skill extensions")
+        subparsers = parser.add_subparsers(dest="extensions_command", required=True)
+
+        install_parser = subparsers.add_parser(
+            "install",
+            help="Install skills from a git repository URL or local directory",
+        )
+        install_parser.add_argument("source", help="Git repository URL or local path")
+        install_parser.add_argument(
+            "--target-dir",
+            default=None,
+            help="Install target directory (default: ~/.agentica/skills)",
+        )
+        install_parser.add_argument(
+            "--force",
+            action="store_true",
+            help="Replace already installed skills with the same name",
+        )
+
+        list_parser = subparsers.add_parser(
+            "list",
+            help="List installed skills from the target directory",
+        )
+        list_parser.add_argument(
+            "--target-dir",
+            default=None,
+            help="Skill directory to inspect (default: ~/.agentica/skills)",
+        )
+
+        remove_parser = subparsers.add_parser(
+            "remove",
+            help="Remove an installed skill by name",
+        )
+        remove_parser.add_argument("skill_name", help="Installed skill directory name")
+        remove_parser.add_argument(
+            "--target-dir",
+            default=None,
+            help="Skill directory to modify (default: ~/.agentica/skills)",
+        )
+
+        reload_parser = subparsers.add_parser(
+            "reload",
+            help="Reload skills from disk and print the current registry count",
+        )
+        reload_parser.add_argument(
+            "--target-dir",
+            default=None,
+            help="Skill directory to inspect (default: ~/.agentica/skills)",
+        )
+
+        args = parser.parse_args(sys.argv[2:])
+        args.command = "extensions"
+        return args
+
+    parser = argparse.ArgumentParser(description='CLI for agentica')
 
     parser.add_argument('--query', type=str, help='Question to ask the LLM', default=None)
     parser.add_argument('--model_provider', type=str,

agentica/cli/display.py

@@ -204,6 +204,8 @@ def show_help():
     execute, web_search, fetch_url, write_todos, task
   - Use --tools to add extra tools, e.g.: --tools calculator shell wikipedia
   - Use --enable-skills to load skills, then /skills to list them
+  - Install skill packs before launch: agentica extensions install <git-url>
+  - Install skill packs in-session: /extensions install <git-url>
   - Say "remember this" or "save this" to trigger memory saving
 """
     console.print(help_text, style="yellow")