Feature/advanced memory - #317

- 问题：hy3要求开思考的时候，后面的对话要带上思考内容，如果不带上，会导致思维链断裂

- Add real agent demo (Tool/Skill/MCP Tool/CodeExecutor) with allow/review/deny scenarios
- Add register_rule() and custom_rules support to ToolScriptSafetyScanner
- Expand dangerous command coverage in _rules.py and _policy.py
- Add quick-start usage examples and running instructions to README
- Update tests to cover new policy fields and custom rule registration

- 问题：之前遗漏了重试这块上报，模型调用重试在trace上无显示,同时发现模型调用失败的信息不会放在对应的span里,一起修复了

Unknown language segments may contain valid non-Python input. Continue Bash scanning, but merge Python findings only when AST parsing succeeds so strict review mode does not block safe Bash scripts.

Updates #90

RELEASE NOTES: NONE

1. Restore HAS_MAGIC as backward-compatible public alias (updated lazily
   when magic is successfully imported on non-win32)
2. Add threading.Lock to _import_docker() for concurrent safety
3. Remove dead code: unused `import sys as _sys` in test method
4. All yapf + flake8 clean, 47 passed, 0 regressions

….update

Address AI review Critical: eliminate # noqa: F821 and NameError risk
by declaring docker/Container/consume_socket_output/demux_adaptor/
frames_iter as module-level None placeholders, populated by
_import_docker() via direct assignment instead of globals().update.

- Remove TYPE_CHECKING guard (placeholders serve the same purpose)
- Remove all # noqa: F821 comments
- _exec_run_with_stdin now references real module variables, not
  dynamically injected names
- Tests @patch works naturally without _import_docker() side effects
- yapf + flake8 clean, 47 passed, 0 regressions

…ests

Address latest AI review:
1. Change except ImportError to except Exception for magic import
   failure, catching OSError and other non-Import failures safely
2. Include HAS_MAGIC in test mock save/restore to prevent state leakage
3. Add exc_info=True to the debug log for better diagnostics

Address AI review Critical: if _import_docker() fails silently and docker
remains None, the except docker.errors.DockerException line would throw
AttributeError: NoneType has no attribute errors, masking the real error.
Now explicitly checks docker is not None after import and raises a clear
RuntimeError with install instructions.

…me semantics

1. _import_docker(): catch ImportError so docker stays None when SDK
   is not installed, making the `if docker is None` guard in
   _init_docker_client() actually reachable
2. _files.py: probe magic at module import time on non-win32 to set
   HAS_MAGIC immediately, preserving the original semantics where
   external code can check HAS_MAGIC right after import

1. test_container_cli.py: replace module-level _import_docker() call
   with autouse module-scoped fixture, removing import-time global side
   effects and fixing import ordering (all imports at top)
2. _files.py: add threading.Lock for magic lazy import to prevent
   concurrent race condition on _magic_module/_magic_checked state
3. Consolidate _magic_checked=True inside lock for atomic state update

…owing

Fix AI review Critical: module-level docker=None/Container=None shadowed
the TYPE_CHECKING type aliases, breaking external
`from ..._container_cli import Container`.

- Use TYPE_CHECKING guard for docker/Container type annotations (static only)
- Store runtime symbols in private _docker_mod/_docker_container_cls/
  _docker_consume_socket_output/_docker_demux_adaptor/_docker_frames_iter
- Update all runtime references to use private names
- Update test @patch targets to _docker_mod
- No public module-level names are shadowed

1. _files.py: remove module-level import magic on non-win32 entirely
   (was still triggering at import time). Now truly deferred to first
   detect_content_type() call. HAS_MAGIC defaults False, updated lazily.
2. Remove unused _has_magic() function (dead code per review)
3. _container_cli.py: remove unused _docker_container_cls
4. _container_cli.py: widen except ImportError to except Exception
   ensuring _docker_imported=True always set, preventing retry storms

Address latest AI review warnings (no Critical this round):
1. Add logger.debug(exc_info=True) for _import_docker except path
   to aid debugging when docker import fails for non-obvious reasons
2. Move import threading to standard library group at top of file
3. Remove duplicate import threading below the third-party imports

问题：使用langfuse插件上报时,用户在span上设置的attribute没有携带到最终上报数据中,导致用户无法观测自定义埋点
解决方案：langfuse上报时保留用户自定义的attribute,带langfuse前缀的原样透传,其他自定义字段收进metadata便于观测

1. invocation 初始化异常时 span 缺失 runner 业务属性且显示未知：
   - trace_runner() 的 invocation_context 改为 Optional，初始化异常时传入
     None（此时 InvocationContext 尚未构造），依赖它的 runner.name 在 None
     时跳过，其余业务属性照常写入
   - runner 初始化失败分支统一调用 trace_runner() 上报错误状态

2. call_llm 模型调用错误被记录为成功：
   - 模型异常被 retry 层转成 LlmResponse(error_code=...) 后，trace 层未识别
     error_code，导致 span 误标成功；现补充检查 llm_response.error_code

3. 外部取消(asyncio.CancelledError)时根 invocation span 仍显示成功且缺 partial：
   - CancelledError 继承 BaseException 而非 Exception，漏过 except Exception；
   现单独捕获 CancelledError，标记 span 为错误
   - 累积已流式输出的 partial 文本，防止取消时已生成内容丢失

4. GeneratorExit 时 agent_run 无部分输出：
   - 生成器关闭(aclose/break)触发的 GeneratorExit 分支现会累积并回填
     partial 流式文本到 agent_action 中，避免已输出内容丢失

问题: 部署成AGUI服务时,如果Agent产生Event带error,则会立即终止,但有的Agent并不是终止(比如GraphAgent),它会接着执行下一个Node,之前产生Event带error时,将会返回RunErrorEvent事件,导致连接被断开

解决方案: AGUI服务端,在收到Event时,不立即返回RunErrorEvent,如果服务是最后一个Event包含错误(Agent此时已经结束),则再发送RunErrorEvent

问题：LLM调用流式输出被中断（如网络中断、模型业务错误、多轮中工具调用后再次调用被中断）时，上游 agent_run 和 runner 上报的 span 状态误报为成功、output 为空，已输出到一半的文本也随之丢失，监控侧只能下钻到 call_llm 才能看到 error_code。

解决方案：错误响应不再清空已流式的文本，span 标记为失败并保留中断前的输出内容（带 [INTERRUPTED] 标记）；多轮场景在已有内容后追加中断；call_llm 的 span 也回填已流式内容。

同步 main 与 r0.1 分支版本号，携带 1.1.18 已包含的修复

feature: 增加advanced memory的示例

Update .env

feature: 第一版可用advanced memory

Preserve default Runner post-turn behavior

feature: 优化session memory实现逻辑，加入清理机制

增加coordination的test脚本

feature: 增加记忆新鲜度机制

feature: 增加环境变量设置，用户可以在.env中设置模型窗口大小

feature: 为advanced_memory增加preload功能