synapse-cli-agent 0.1.16__tar.gz → 0.1.18__tar.gz

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/.gitignore

@@ -18,6 +18,7 @@ wheels/
 .env
 .coding-agent/
 .cmd-agent/
+.tmp/
 .synapse/
 .pytest_cache/
 .ruff_cache/
@@ -40,4 +41,4 @@ site/
 
 # OS
 Thumbs.db
-.DS_Store
+.DS_Store.tmp/

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/AGENTS.md

@@ -1,27 +1,5 @@
 # Synapse 项目协作指南
 
-## 用户信息
-- 姓名：肖前程
-- 年龄：35 岁
-- 职业：Java 程序员
-- 当前学习方向：Agent 应用开发
-- 主语言背景：Java → 正在学习 Python/LangChain Agent 生态
-
-## 用户信息
-- 姓名：肖前程
-- 年龄：35 岁
-- 职业：Java 程序员
-- 当前学习方向：Agent 应用开发
-- 主语言背景：Java → 正在学习 Python/LangChain Agent 生态
-
-## 约定
-- Python 依赖使用 `uv` 管理。
-- 优先小步、可测试的改动。
-- CLI 输出简洁，不使用 emoji。
-- 默认关闭人工审批（自动放行）；除非用户要求，不要自行加审批门槛。
-- Backend 仅使用 `LocalShellBackend`，无远程 sandbox。
-- 对用户回复与模型思考过程使用中文；代码标识符/路径/命令可保留原文。
-
 ## 当前仓库结构
 
 | 路径                               | 主要职责                                                 |
@@ -153,3 +131,55 @@ powershell -ExecutionPolicy Bypass -File scripts/release.ps1
 ```
 
 8. 脚本会创建并推送 `v{version}` tag；`.github/workflows/release.yml` 会从 `CHANGELOG.md` 提取同名段落、执行 `uv build` 并创建 GitHub Release。
+
+## 10. cdp_take_screenshot 工具使用说明
+
+### 10.1 已知限制
+- `filePath` 参数无法写入工作区目录（浏览器沙箱限制），所有指定 `filePath` 的尝试均会报 `Access denied`。
+- 不指定 `filePath` 时，截图以文本格式（含 base64）自动存入 `/large_tool_results/call_xx_xxx`。
+
+### 10.2 标准流程：截图 → 解码 → 识图
+
+执行完 `cdp_take_screenshot`（不指定 `filePath`）后，按以下步骤操作：
+
+#### Step 1 — 编写/复用解码脚本
+如果工作区不存在 `/decode_screenshot.py`，创建它：
+```python
+import base64, re, sys, glob, os
+
+files = glob.glob('large_tool_results/call_*')
+if not files:
+    print("No screenshot files found")
+    sys.exit(1)
+
+latest = max(files, key=os.path.getctime)
+print(f"Processing: {latest}")
+
+content = open(latest, 'r').read()
+m = re.search(r"data='([^']+)'", content)
+if not m:
+    print("No base64 data found")
+    sys.exit(1)
+
+os.makedirs('.tmp', exist_ok=True)
+out = f'.tmp/screenshot_{os.path.getmtime(latest)}.png'
+open(out, 'wb').write(base64.b64decode(m.group(1)))
+print(f"OK -> {out}")
+```
+
+#### Step 2 — 执行解码
+```bash
+python decode_screenshot.py
+# 输出示例: OK -> .tmp/screenshot_1712345678.1234567.png
+```
+
+#### Step 3 — 调用识图
+```bash
+# 使用 Step 2 返回的文件名
+describe_image(image_path="/.tmp/screenshot_1712345678.1234567.png", ...)
+```
+
+### 10.3 注意事项
+- 每次截图都会在 `/large_tool_results/` 生成一个新文件，解码脚本自动选最新的。
+- 输出到 `/.tmp/` 目录，该目录需加入 `.gitignore`。
+- 不要在对话中手动执行多步解码，应直接调用解码脚本。

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/CHANGELOG.md

@@ -7,6 +7,43 @@ The release workflow automatically extracts the matching section as release note
 
 ---
 
+## v0.1.18
+
+### 新功能
+
+- 新增 LLM Debug Inspector（`F11`），实时监控模型通信、工具调用和 token 消耗。
+- Inspector 支持采集开关、跟随最新、按类型筛选（异常/工具/慢调用）、回合折叠和调用详情查看。
+- Inspector 概览栏显示失败率（基于工具级错误检测），工具标签页展示失败工具及原因。
+
+### 修复
+
+- TUI：`F10` 恢复删除 session 弹框入口；修复鼠标选中与点击复制的冲突，拖选后自动复制。
+
+### 工程改进
+
+- `DebugCaptureRecord` 增加工具级错误检测（LangChain `ToolMessage.status` + 内容模式）。
+- `_tool_pairs` 返回 `error` 字段，区分 "等待中"（result null）与 "真失败"（有错误内容）。
+- Inspector 前端：失败率仅统计真正失败的工具，"待响应"不计入。
+
+---
+
+## v0.1.17
+
+### 新功能
+
+- `find_files` / `search_files` 工具新增 `context_lines`、`case_insensitive`、`head_limit`、`offset` 参数。
+- `search_files` 支持忽略大小写（`case_insensitive`，由 `synapse-search-core` 原生引擎实现）。
+- 支持分页查询（`head_limit` + `offset`），Agent 可按需翻页而非一次性获取全部结果。
+
+### 工程改进
+
+- 工具的 Pydantic schema 不再定义 `intent` 字段，改由 `build_intent_schema_middleware` 中间件统一管理。
+- 新增 Synapse 自有文件搜索工具 `find_files` / `search_files`，排除 deepagents 内置 `ls`/`glob`/`grep` 工具。
+- 系统提示词中的 `glob`/`grep` 工具名修正为 `find_files` / `search_files`。
+- `synapse-search-core` 升级至 0.1.1（新增 `case_insensitive` 参数）。
+
+---
+
 ## v0.1.16
 
 ### 新功能

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: synapse-cli-agent
-Version: 0.1.16
+Version: 0.1.18
 Summary: Local coding agent built on LangChain Deep Agents (LocalShell, no sandbox)
 Project-URL: Homepage, https://github.com/alex8224/synapse-agent
 Project-URL: Repository, https://github.com/alex8224/synapse-agent
@@ -287,7 +287,7 @@ OpenAI Responses API 同时支持 HTTP/SSE 与普通 LLM WebSocket。profile 中
 }
 ```
 
-`vision_model.model`、`base_url`、`api_key_env` 均可更换；`api_key` 也支持，但建议使用环境变量。`think` 是识图模型独立的思考开关，开启后发送 `thinking: {"type": "enabled"}`；它不影响主模型。默认只处理本地/内存图片；将 `allow_remote_urls` 设为 `true` 才会把消息中的 HTTP(S) 图片 URL 转发给识图服务。该服务只用于把图片转换成文字，配置后图片内容会发送到指定服务。多模态主模型会跳过该 middleware，不调用识图服务。
+`vision_model.model`、`base_url`、`api_key_env` 均可更换；`api_key` 也支持，但建议使用环境变量。`think` 是识图模型独立的思考开关，开启后发送 `thinking: {"type": "enabled"}`；它不影响主模型。默认只处理本地/内存图片；将 `allow_remote_urls` 设为 `true` 才会把消息中的 HTTP(S) 图片 URL 转发给识图服务。该服务只用于把图片转换成文字，配置后图片内容会发送到指定服务。主模型 profile 显式配置 `image_input=false` 时，Synapse 还会自动注入 `describe_image` 工具，让 Agent 主动读取并识别工作区内的图片；未配置 `vision_model` 时工具会明确返回不可用。多模态主模型会跳过 middleware，且不会注入该工具。
 
 ### MCP 服务器配置
 

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/README.md

@@ -251,7 +251,7 @@ OpenAI Responses API 同时支持 HTTP/SSE 与普通 LLM WebSocket。profile 中
 }
 ```
 
-`vision_model.model`、`base_url`、`api_key_env` 均可更换；`api_key` 也支持，但建议使用环境变量。`think` 是识图模型独立的思考开关，开启后发送 `thinking: {"type": "enabled"}`；它不影响主模型。默认只处理本地/内存图片；将 `allow_remote_urls` 设为 `true` 才会把消息中的 HTTP(S) 图片 URL 转发给识图服务。该服务只用于把图片转换成文字，配置后图片内容会发送到指定服务。多模态主模型会跳过该 middleware，不调用识图服务。
+`vision_model.model`、`base_url`、`api_key_env` 均可更换；`api_key` 也支持，但建议使用环境变量。`think` 是识图模型独立的思考开关，开启后发送 `thinking: {"type": "enabled"}`；它不影响主模型。默认只处理本地/内存图片；将 `allow_remote_urls` 设为 `true` 才会把消息中的 HTTP(S) 图片 URL 转发给识图服务。该服务只用于把图片转换成文字，配置后图片内容会发送到指定服务。主模型 profile 显式配置 `image_input=false` 时，Synapse 还会自动注入 `describe_image` 工具，让 Agent 主动读取并识别工作区内的图片；未配置 `vision_model` 时工具会明确返回不可用。多模态主模型会跳过 middleware，且不会注入该工具。
 
 ### MCP 服务器配置
 

synapse_cli_agent-0.1.18/decode_screenshot.py

@@ -0,0 +1,26 @@
+import base64
+import glob
+import os
+import re
+import sys
+
+# 找到最新的 large_tool_results 文件
+files = glob.glob('large_tool_results/call_*')
+if not files:
+    print("No screenshot files found")
+    sys.exit(1)
+
+latest = max(files, key=os.path.getctime)
+print(f"Processing: {latest}")
+
+content = open(latest).read()
+m = re.search(r"data='([^']+)'", content)
+if not m:
+    print("No base64 data found")
+    sys.exit(1)
+
+os.makedirs('.tmp', exist_ok=True)
+ts = str(os.path.getmtime(latest))
+out = f'.tmp/screenshot_{ts}.png'
+open(out, 'wb').write(base64.b64decode(m.group(1)))
+print(f"OK -> {out}")

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synapse-cli-agent"
-version = "0.1.16"
+version = "0.1.18"
 description = "Local coding agent built on LangChain Deep Agents (LocalShell, no sandbox)"
 readme = "README.md"
 authors = [

synapse_cli_agent-0.1.18/release_body.md

@@ -0,0 +1,19 @@
+## v0.1.18
+
+### 新功能
+
+- 新增 LLM Debug Inspector（`F11`），实时监控模型通信、工具调用和 token 消耗。
+- Inspector 支持采集开关、跟随最新、按类型筛选（异常/工具/慢调用）、回合折叠和调用详情查看。
+- Inspector 概览栏显示失败率（基于工具级错误检测），工具标签页展示失败工具及原因。
+
+### 修复
+
+- TUI：`F10` 恢复删除 session 弹框入口；修复鼠标选中与点击复制的冲突，拖选后自动复制。
+
+### 工程改进
+
+- `DebugCaptureRecord` 增加工具级错误检测（LangChain `ToolMessage.status` + 内容模式）。
+- `_tool_pairs` 返回 `error` 字段，区分 "等待中"（result null）与 "真失败"（有错误内容）。
+- Inspector 前端：失败率仅统计真正失败的工具，"待响应"不计入。
+
+---

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/rust/synapse-search-core/Cargo.lock

@@ -391,7 +391,7 @@ dependencies = [
 
 [[package]]
 name = "synapse-search-core"
-version = "0.1.0"
+version = "0.1.1"
 dependencies = [
  "globset",
  "grep-regex",

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/rust/synapse-search-core/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "synapse-search-core"
-version = "0.1.0"
+version = "0.1.1"
 edition = "2021"
 rust-version = "1.80"
 license = "Apache-2.0"

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/rust/synapse-search-core/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "synapse-search-core"
-version = "0.1.0"
+version = "0.1.1"
 description = "Native workspace search algorithms for Synapse"
 requires-python = ">=3.12"
 license = { text = "Apache-2.0" }

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/rust/synapse-search-core/src/lib.rs

@@ -3,7 +3,7 @@ use std::path::{Path, PathBuf};
 use std::time::UNIX_EPOCH;
 
 use globset::{Glob, GlobSet, GlobSetBuilder};
-use grep_regex::RegexMatcher;
+use grep_regex::{RegexMatcher, RegexMatcherBuilder};
 use grep_searcher::{BinaryDetection, Searcher, SearcherBuilder, Sink, SinkContext, SinkMatch};
 use ignore::WalkBuilder;
 use pyo3::exceptions::PyValueError;
@@ -121,7 +121,7 @@ fn search_file(
     Ok(sink.result)
 }
 
-#[pyfunction(signature = (base_path, pattern, include_glob=None, max_results=1000, context_lines=0))]
+#[pyfunction(signature = (base_path, pattern, include_glob=None, max_results=1000, context_lines=0, case_insensitive=false))]
 fn grep(
     py: Python<'_>,
     base_path: &str,
@@ -129,6 +129,7 @@ fn grep(
     include_glob: Option<&str>,
     max_results: usize,
     context_lines: usize,
+    case_insensitive: bool,
 ) -> PyResult<Py<PyDict>> {
     if pattern.is_empty() {
         return Err(PyValueError::new_err("pattern must not be empty"));
@@ -143,7 +144,9 @@ fn grep(
             base_path.display()
         )));
     }
-    let matcher = RegexMatcher::new(pattern)
+    let matcher = RegexMatcherBuilder::new()
+        .case_insensitive(case_insensitive)
+        .build(pattern)
         .map_err(|error| PyValueError::new_err(format!("invalid regex pattern: {error}")))?;
     let include = include_glob
         .map(compile_glob)

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/src/synapse/app/agent.py

@@ -35,6 +35,7 @@ from synapse.runtime.middleware import (
     build_strip_redundant_prompt_blocks,
     build_task_namespace_middleware,
     build_tool_error_recovery_middleware,
+    build_tool_exclusion_middleware,
 )
 from synapse.runtime.model_request_compression_middleware import (
     build_model_request_compression_middleware,
@@ -48,7 +49,11 @@ from synapse.settings import Settings
 from synapse.tool_output.pipeline import ToolOutputTransformPipeline
 from synapse.tool_output.repository import ToolOutputRepository
 from synapse.tool_output.transformers import load_transformer_plugins
-from synapse.tools import build_session_tools
+from synapse.tools import (
+    build_describe_image_tools,
+    build_filesystem_search_tools,
+    build_session_tools,
+)
 
 
 def _build_checkpointer(settings: Settings):
@@ -228,6 +233,12 @@ def build_coding_agent(
         readonly=settings.readonly,
         excluded_tools=settings.excluded_tools,
     )
+    # Keep deepagents' built-in ``ls``, ``glob``, and ``grep`` out of model
+    # requests. Synapse registers the non-conflicting ``find_files`` and
+    # ``search_files`` tools explicitly below.
+    model_request_excluded_tools = set(settings.excluded_tools) | {"ls", "glob", "grep"}
+    if settings.readonly:
+        model_request_excluded_tools.update({"execute", "write_file", "edit_file"})
 
     interrupt_on = build_interrupt_on(require_approval=settings.require_approval)
     with span("checkpointer"):
@@ -279,6 +290,7 @@ def build_coding_agent(
     tools: list[Any] = []
     if extra_tools:
         tools.extend(extra_tools)
+    tools.extend(build_filesystem_search_tools(backend))
     # 跨会话查阅工具
     try:
         session_tools = build_session_tools(
@@ -382,6 +394,13 @@ def build_coding_agent(
         getattr(selected_profile, "base_url", None) or getattr(settings, "openai_base_url", None),
     )
     vision_config = VisionModelConfig.from_registry(registry, settings)
+    tools.extend(
+        build_describe_image_tools(
+            image_input=getattr(selected_profile, "image_input", None),
+            backend=backend,
+            config=vision_config,
+        )
+    )
 
     output_repository = ToolOutputRepository(settings.resolved_tool_output_db_path())
     middleware: list[Any] = [
@@ -392,6 +411,7 @@ def build_coding_agent(
         ),
         build_model_retry_middleware(),
         build_task_namespace_middleware(),
+        build_tool_exclusion_middleware(model_request_excluded_tools),
     ]
     transform_enabled = bool(getattr(settings, "enable_tool_output_transform", True))
     try:
@@ -449,6 +469,13 @@ def build_coding_agent(
 
         middleware.append(build_openai_oauth_compat_middleware())
 
+    # Debug capture middleware — innermost, records final provider-ready
+    # request/response pairs for the Debug Inspector (F11 / /debug).
+    from synapse.observability.llm_debug import get_debug_store
+    from synapse.runtime.debug_capture_middleware import build_debug_capture_middleware
+
+    middleware.append(build_debug_capture_middleware(get_debug_store()))
+
     if progress is not None:
         progress("compiling agent graph")
     with span("create_deep_agent"):

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/src/synapse/commands/mcp.py

@@ -1,4 +1,5 @@
 """MCP slash-command handler and presentation helpers."""
+
 from __future__ import annotations
 
 import json
@@ -102,6 +103,7 @@ def _md_tool_list(names: list[str], warnings: list[str] | None = None) -> str:
         lines.append(f"\nwarn: {w}")
     return "\n".join(lines)
 
+
 def handle_mcp(
     args: list[str],
     *,
@@ -216,8 +218,7 @@ def handle_mcp(
         md_lines.append("|---|---|")
         for transport, names in sorted(by_transport.items()):
             md_lines.append(
-                f"| {markdown_escape(transport)} | "
-                f"{markdown_escape(', '.join(names))} |"
+                f"| {markdown_escape(transport)} | {markdown_escape(', '.join(names))} |"
             )
         if result.tools:
             md_lines.append("")
@@ -291,9 +292,7 @@ def handle_mcp(
         loaded = getattr(build_coding_agent, "last_mcp_servers", []) or []
         warnings_list = getattr(build_coding_agent, "last_mcp_warnings", []) or []
         tools = getattr(build_coding_agent, "last_mcp_tool_names", []) or []
-        notice = (
-            f"mcp '{target}' {'enabled' if changed.enabled else 'disabled'} 路 tools={len(tools)}"
-        )
+        notice = f"mcp '{target}' {'enabled' if changed.enabled else 'disabled'} tools={len(tools)}"
         lines = [
             f"mcp server '{target}' {'enabled' if changed.enabled else 'disabled'}; agent rebuilt",
             f"loaded servers: {', '.join(loaded) or '(none)'}",
@@ -325,7 +324,7 @@ def handle_mcp(
         return SlashResult(
             handled=True,
             lines=["mcp enabled (run /mcp reload to rebuild agent)"],
-            notice="mcp enabled 路 reload pending",
+            notice="mcp enabled reload pending",
             settings_changed=True,
         )
 
@@ -348,7 +347,7 @@ def handle_mcp(
         return SlashResult(
             handled=True,
             lines=["mcp disabled; agent rebuilt without MCP tools"],
-            notice="mcp disabled 路 tools=0",
+            notice="mcp disabled tools=0",
             agent=new_agent,
             settings_changed=True,
         )
@@ -372,9 +371,9 @@ def handle_mcp(
         warnings = getattr(build_coding_agent, "last_mcp_warnings", []) or []
         tools = getattr(build_coding_agent, "last_mcp_tool_names", []) or []
         enabled_flag = bool(getattr(settings, "enable_mcp", True))
-        notice = f"mcp reloaded 路 servers={len(loaded)} tools={len(tools)}"
+        notice = f"mcp reloaded servers={len(loaded)} tools={len(tools)}"
         if not enabled_flag:
-            notice = "mcp reloaded 路 disabled"
+            notice = "mcp reloaded disabled"
         lines = [
             f"agent rebuilt; mcp enabled={getattr(settings, 'enable_mcp', True)}",
             f"loaded servers: {', '.join(loaded) or '(none)'}",

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/src/synapse/content/prompts.py

@@ -46,7 +46,7 @@ Never scan the entire repository without a specific reason.
 
 ## Virtual filesystem
 
-File tools such as `ls`, `glob`, `grep`, `read_file`, `edit_file`, and `write_file`
+File tools such as `find_files`, `search_files`, `read_file`, `edit_file`, and `write_file`
 operate on a virtual filesystem rooted at `/`.
 
 Valid paths:
@@ -109,7 +109,7 @@ Do not use generic intent values such as `run tool` or `read_file`.
 Search only when required by a clear task.
 Keep searches targeted and avoid unnecessary full-file output.
 
-For `glob` and `grep` tools, automatically exclude common build artifacts and caches:
+For the `search_files` tool, automatically exclude common build artifacts and caches via the `glob` parameter:
 * `target/`
 * `.venv/`
 * `.node_modules/`
@@ -133,7 +133,7 @@ Run independent tool calls in parallel within the same turn.
 Parallelize when arguments are already known and results do not depend on each other, including:
 
 * Multiple file reads
-* `glob` and related file reads
+* `find_files` and related file reads
 * `execute` and related reads
 * Edits to different files
 * Independent test or lint commands

{synapse_cli_agent-0.1.16 → synapse_cli_agent-0.1.18}/src/synapse/integrations/mcp_client.py

@@ -13,11 +13,13 @@ import asyncio
 import atexit
 import json
 import logging
+import os
+import sys
 import threading
 from concurrent.futures import Future
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Any
+from typing import Any, TextIO
 
 logger = logging.getLogger(__name__)
 
@@ -71,6 +73,31 @@ def _expand_mapping(raw: dict[str, Any] | None) -> dict[str, str]:
     return out
 
 
+def _valid_subprocess_stream(stream: TextIO | None) -> TextIO | None:
+    """Return a stream only when its OS descriptor is valid for inheritance."""
+    if stream is None or getattr(stream, "closed", False):
+        return None
+    try:
+        fd = stream.fileno()
+        os.fstat(fd)
+    except (AttributeError, OSError, ValueError):
+        return None
+    return stream
+
+
+def _stdio_errlog() -> tuple[TextIO, bool]:
+    """Choose a valid stderr target and report whether the caller owns it.
+
+    Textual, redirected Windows launchers, and GUI executables can leave
+    ``sys.stderr`` present while its underlying descriptor is closed. The MCP
+    SDK passes this object to subprocess creation, which then raises EBADF.
+    """
+    stream = _valid_subprocess_stream(sys.stderr)
+    if stream is not None:
+        return stream, False
+    return open(os.devnull, "w", encoding="utf-8"), True  # noqa: SIM115
+
+
 def _parse_server(raw: dict[str, Any]) -> McpServerConfig:
     name = str(raw.get("name") or "").strip()
     if not name:
@@ -337,10 +364,11 @@ class _LoopThread:
 class _LiveServer:
     config: McpServerConfig
     session: Any
-    # Keep transport context managers open for process lifetime.
+    # Keep transport context managers and owned streams open for process lifetime.
     transport_cm: Any
     session_cm: Any
     streams: Any = None
+    errlog: TextIO | None = None
 
 
 class McpSessionPool:
@@ -385,6 +413,8 @@ class McpSessionPool:
                 await live.transport_cm.__aexit__(None, None, None)
             except Exception:  # noqa: BLE001
                 pass
+            if live.errlog is not None:
+                live.errlog.close()
         self._servers.clear()
 
     async def _open_stdio(self, server: McpServerConfig) -> _LiveServer:
@@ -398,16 +428,38 @@ class McpSessionPool:
             args=server.args,
             env=server.env or None,
         )
-        transport_cm = stdio_client(params)
-        read, write = await transport_cm.__aenter__()
-        session_cm = ClientSession(read, write)
-        session = await session_cm.__aenter__()
-        await session.initialize()
+        errlog, owns_errlog = _stdio_errlog()
+        transport_cm = stdio_client(params, errlog=errlog)
+        transport_entered = False
+        session_cm = None
+        session_entered = False
+        try:
+            read, write = await transport_cm.__aenter__()
+            transport_entered = True
+            session_cm = ClientSession(read, write)
+            session = await session_cm.__aenter__()
+            session_entered = True
+            await session.initialize()
+        except BaseException:
+            if session_entered and session_cm is not None:
+                try:
+                    await session_cm.__aexit__(None, None, None)
+                except BaseException:
+                    pass
+            if transport_entered:
+                try:
+                    await transport_cm.__aexit__(None, None, None)
+                except BaseException:
+                    pass
+            if owns_errlog:
+                errlog.close()
+            raise
         return _LiveServer(
             config=server,
             session=session,
             transport_cm=transport_cm,
             session_cm=session_cm,
+            errlog=errlog if owns_errlog else None,
         )
 
     async def _open_http(self, server: McpServerConfig) -> _LiveServer:

synapse_cli_agent-0.1.18/src/synapse/observability/debug_inspector.html

@@ -0,0 +1,19 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width,initial-scale=1">
+<title>Synapse Inspector</title>
+<style>
+:root{color-scheme:dark;--bg:#0d131b;--panel:#121b25;--raised:#182431;--hover:#1d2b3a;--line:#29394b;--text:#e7edf4;--muted:#93a4b7;--blue:#68b5f8;--red:#ff8f8f;--amber:#efc36b;--green:#74c69d}*{box-sizing:border-box}body{margin:0;height:100vh;overflow:hidden;background:var(--bg);color:var(--text);font:13px/1.55 system-ui,-apple-system,"Segoe UI",sans-serif}button,input{font:inherit}button{border:0;color:inherit;background:transparent;cursor:pointer}button:focus-visible,input:focus-visible{outline:2px solid var(--blue);outline-offset:2px}.top{height:52px;display:flex;align-items:center;gap:16px;padding:0 16px;border-bottom:1px solid var(--line);background:#101822}.brand{font-size:15px;font-weight:700}.brand span{color:var(--blue)}.capture{display:flex;align-items:center;gap:7px;color:var(--muted)}.switch{width:31px;height:18px;padding:2px;border-radius:10px;background:#465466}.switch i{display:block;width:14px;height:14px;border-radius:50%;background:#edf3f8;transition:.15s}.switch.on{background:var(--green)}.switch.on i{transform:translateX(13px)}.actions{display:flex;align-items:center;gap:4px;margin-left:auto}.btn{min-height:28px;padding:4px 9px;border:1px solid transparent;border-radius:5px;color:var(--muted)}.btn:hover{border-color:var(--line);background:var(--raised);color:var(--text)}.danger:hover{color:var(--red)}#app{display:grid;grid-template-columns:300px minmax(0,1fr);height:calc(100vh - 52px)}#side{display:flex;min-width:0;min-height:0;flex-direction:column;border-right:1px solid var(--line);background:var(--panel)}.side-head{padding:12px;border-bottom:1px solid var(--line)}.side-title{display:flex;justify-content:space-between;color:var(--muted);font-size:11px;font-weight:700}.search{width:100%;height:32px;margin-top:9px;padding:0 9px;border:1px solid var(--line);border-radius:5px;background:#0d151e;color:var(--text)}.filters{display:flex;gap:5px;margin-top:8px}.filter{padding:3px 7px;border:1px solid var(--line);border-radius:4px;color:var(--muted);font-size:11px}.filter.active,.filter:hover{border-color:#3b6d98;background:#1b3044;color:var(--blue)}#records{flex:1;min-height:0;overflow-y:scroll;overflow-x:hidden;padding:6px 0;scrollbar-color:#40546a #111923;scrollbar-width:thin}#records::-webkit-scrollbar{width:9px}#records::-webkit-scrollbar-track{background:#111923}#records::-webkit-scrollbar-thumb{border:2px solid #111923;border-radius:8px;background:#40546a}.turn-group{margin:0}.turn{display:flex;width:100%;align-items:center;gap:7px;padding:8px 12px;color:var(--muted);font-size:10px;font-weight:700;text-align:left;text-transform:uppercase}.turn:hover{background:var(--hover);color:var(--text)}.turn-arrow{width:10px;color:var(--blue);transition:transform .15s}.turn-group.collapsed .turn-arrow{transform:rotate(-90deg)}.turn-count{margin-left:auto;font-weight:500}.turn-group.collapsed .turn-records{display:none}.record{display:block;width:100%;padding:8px 12px;border-left:3px solid transparent;text-align:left}.record:hover{background:var(--hover)}.record.selected{border-left-color:var(--blue);background:#1d3043}.r-top,.r-bottom{display:flex;align-items:center;gap:7px}.r-id{font-weight:700}.r-model{flex:1;overflow:hidden;color:var(--muted);text-overflow:ellipsis;white-space:nowrap}.r-bottom{justify-content:space-between;margin-top:3px;color:var(--muted);font:11px ui-monospace,Consolas,monospace}.bad{color:var(--red)}.warn{color:var(--amber)}.tag{padding:0 5px;border-radius:8px;background:#26384a;color:#b9c8d6}#main{min-width:0;overflow:auto;padding:20px 24px 56px}.empty{display:grid;height:100%;place-content:center;gap:5px;text-align:center;color:var(--muted)}.empty strong{color:var(--text);font-size:16px}.call-head{display:flex;align-items:flex-start;justify-content:space-between;gap:18px}.eyebrow{color:var(--blue);font-size:10px;font-weight:700;text-transform:uppercase}h1{margin:3px 0 2px;font-size:20px}.sub{color:var(--muted);font-size:12px}.metrics{display:grid;grid-template-columns:repeat(6,minmax(80px,1fr));margin:16px 0;border:1px solid var(--line);border-radius:6px;background:var(--panel)}.metric{padding:9px 11px;border-right:1px solid var(--line)}.metric:last-child{border:0}.metric label{display:block;color:var(--muted);font-size:10px}.metric strong{display:block;margin-top:1px;font:600 14px ui-monospace,Consolas,monospace}.tabs{display:flex;gap:2px;border-bottom:1px solid var(--line)}.tab{padding:9px 12px;border-bottom:2px solid transparent;color:var(--muted)}.tab.active{border-bottom-color:var(--blue);color:var(--text)}#view{padding-top:14px}.section{margin-bottom:12px}.section-head{display:flex;align-items:center;margin-bottom:7px}.section-head h2{margin:0;font-size:14px}.section-head .meta{margin-left:8px;color:var(--muted);font-size:11px}.section-head .btn{margin-left:auto}.message,.tool-row,.raw-box{border:1px solid var(--line);border-radius:6px;background:var(--panel)}.message{margin-bottom:8px}.message-head{display:flex;align-items:center;padding:7px 10px;border-bottom:1px solid var(--line)}.role{font-size:11px;font-weight:700;text-transform:uppercase}.message-head .meta{margin-left:8px;color:var(--muted);font-size:11px}.message-head .btn{margin-left:auto}.content,pre{margin:0;overflow:auto;padding:10px;white-space:pre-wrap;overflow-wrap:anywhere;color:#dce6ef;font:12px/1.6 ui-monospace,SFMono-Regular,Consolas,monospace}.content{max-height:340px}.tool-summary{padding:9px 11px;border:1px solid var(--line);border-radius:6px;background:var(--panel);color:var(--muted)}.tool-row{margin-top:7px;overflow:hidden}.tool-toggle{display:grid;width:100%;grid-template-columns:62px 140px minmax(0,1fr);align-items:center;gap:8px;padding:9px 11px;text-align:left}.tool-toggle:hover{background:var(--hover)}.status{font-size:11px}.status.ok{color:var(--green)}.status.missing{color:var(--amber)}.tool-name{overflow:hidden;color:var(--text);font-weight:650;text-overflow:ellipsis;white-space:nowrap}.tool-intent{overflow:hidden;color:var(--muted);text-overflow:ellipsis;white-space:nowrap}.tool-detail{display:none;grid-template-columns:1fr 1fr;gap:1px;border-top:1px solid var(--line);background:var(--line)}.tool-row.open .tool-detail{display:grid}.tool-side{min-width:0;background:#101821}.tool-label{padding:6px 10px;border-bottom:1px solid var(--line);color:var(--muted);font-size:10px;font-weight:700}.tool-side pre{max-height:300px}.raw-box pre{max-height:65vh}.notice{margin-bottom:10px;padding:9px 11px;border-left:3px solid var(--amber);background:#2a251a;color:#dbcda9}.new-records{display:none;color:var(--blue)}@media(max-width:850px){#app{grid-template-columns:250px minmax(0,1fr)}.metrics{grid-template-columns:repeat(3,1fr)}.metric:nth-child(3),.metric:nth-child(6){border-right:0}.tool-detail{grid-template-columns:1fr}}@media(max-width:650px){#app{display:block}#side{display:none}#main{padding:16px}.metrics{grid-template-columns:repeat(2,1fr)}.metric:nth-child(even){border-right:0}}
+</style>
+</head>
+<body>
+<header class="top"><div class="brand"><span>//</span> Synapse Inspector</div><button class="capture" id="capture" title="开启或暂停模型通信采集">采集 <span class="switch"><i></i></span><strong id="state">关闭</strong></button><div class="actions"><button class="btn new-records" id="newRecords"></button><label class="btn"><input id="follow" type="checkbox" checked> 跟随最新</label><button class="btn" id="refresh">刷新</button><button class="btn danger" id="clear">清空</button></div></header>
+<div id="app"><aside id="side"><div class="side-head"><div class="side-title"><span>调用记录</span><span id="count">0 条</span></div><input class="search" id="search" type="search" placeholder="筛选模型、工具、回合或错误"><div class="filters"><button class="filter active" data-filter="all">全部</button><button class="filter" data-filter="error">异常</button><button class="filter" data-filter="tool">工具</button><button class="filter" data-filter="slow">慢调用</button></div></div><div id="records"></div></aside><main id="main"><div class="empty"><strong>等待通信记录</strong><span>开启采集并运行一个回合后，在这里检查模型调用。</span></div></main></div>
+<script>
+const $=id=>document.getElementById(id);let records=[],selected=-1,loaded=null,filter='all',tab='overview',timer,newCount=0;const collapsedTurns=new Set;const esc=v=>String(v??'').replace(/[&<>"']/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));const dump=v=>JSON.stringify(v,null,2);const ms=v=>v>=1000?(v/1000).toFixed(2)+'s':Math.round(v||0)+'ms';const tok=v=>Number(v||0).toLocaleString();async function api(url,options){try{const response=await fetch(url,options);return response.ok?response.json():null}catch(_){return null}}function copy(text,button){navigator.clipboard?.writeText(text).then(()=>{const old=button.textContent;button.textContent='已复制';setTimeout(()=>button.textContent=old,900)}).catch(()=>{})}function visibleRows(){const query=$('search').value.trim().toLowerCase();return records.filter(r=>(filter==='all'||filter==='error'&&r.error||filter==='tool'&&r.has_tools||filter==='slow'&&r.duration_ms>=10000)&&(!query||`${r.turn_index} ${r.model_call_index} ${r.model_name} ${r.provider} ${r.error||''} ${r.tool_names||''}`.toLowerCase().includes(query)))}function drawList(){const out=$('records'),rows=visibleRows(),groups=new Map,scrollTop=out.scrollTop;out.innerHTML='';$('count').textContent=`${rows.length} / ${records.length} 条`;for(const record of rows){if(!groups.has(record.turn_index))groups.set(record.turn_index,[]);groups.get(record.turn_index).push(record)}for(const [turn,items] of groups){const group=document.createElement('section'),collapsed=collapsedTurns.has(turn);group.className='turn-group'+(collapsed?' collapsed':'');group.innerHTML=`<button class="turn" aria-expanded="${!collapsed}"><span class="turn-arrow">▼</span><span>回合 ${turn}</span><span class="turn-count">${items.length} 次调用</span></button><div class="turn-records"></div>`;const container=group.querySelector('.turn-records');for(const r of items){const button=document.createElement('button'),tools=r.tool_count?`<span class="tag">${r.tool_count} 工具</span>`:'';button.className='record'+(r.index===selected?' selected':'');button.innerHTML=`<div class="r-top"><span class="r-id ${r.error?'bad':''}">#${r.model_call_index}</span><span class="r-model">${esc(r.model_name)}</span>${r.error?'<span class="bad">失败</span>':''}</div><div class="r-bottom"><span class="${r.duration_ms>=10000?'warn':''}">${ms(r.duration_ms)}</span><span>${tok(r.usage?.input_tokens)} → ${tok(r.usage?.output_tokens)}</span>${tools}</div>`;button.onclick=()=>select(r.index);container.appendChild(button)}group.querySelector('.turn').onclick=()=>{if(collapsedTurns.has(turn))collapsedTurns.delete(turn);else collapsedTurns.add(turn);drawList()};out.appendChild(group)}out.scrollTop=scrollTop}function messageCard(message){const text=message.content_full||message.content_preview||'',hasCalls=(message.tool_calls||[]).length;let body=text&&!looksLikeRawCalls(text)?`<div class="content">${esc(text)}</div>`:'';if(!body&&hasCalls)body=`<div class="tool-summary">模型发起 ${hasCalls} 个结构化工具调用。详细参数请在“工具”标签查看。</div>`;if(!body)body='<div class="tool-summary">此消息没有可展示的文本内容。</div>';return `<article class="message"><div class="message-head"><span class="role">${esc(message.role||'unknown')}</span><span class="meta">${message.name?' / '+esc(message.name):''} ${tok(message.estimated_tokens)} tokens${hasCalls?' · '+hasCalls+' 工具':''}</span><button class="btn copy-message" data-text="${encodeURIComponent(text)}">复制文本</button></div>${body}</article>`}function looksLikeRawCalls(text){return text.includes("'type': 'function_call'")||text.includes('"type":"function_call"')}function section(title,meta,body,raw){return `<section class="section"><div class="section-head"><h2>${title}</h2><span class="meta">${meta}</span>${raw?'<button class="btn copy-section">复制 JSON</button>':''}</div>${body}</section>`}function intentOf(pair){try{const value=typeof pair.args==='string'?JSON.parse(pair.args):pair.args;return value?.intent||value?.command||value?.file_path||value?.pattern||''}catch(_){const match=String(pair.args||'').match(/["']intent["']\s*:\s*["']([^"']+)/);return match?.[1]||''}}function toolsView(pairs){if(!pairs.length)return '<div class="empty"><strong>本次调用没有关联工具</strong><span>这里只展示本次请求新增的工具结果和本次响应发起的工具调用。</span></div>';const failed=pairs.filter(p=>p.error===true);let failHtml='';if(failed.length)failHtml=`<div class="notice bad">${failed.length} 个工具失败：${failed.map(p=>esc(p.name)+' — '+esc((p.result||'').slice(0,100))).join('<br>')}</div>`;return `${failHtml}<div class="notice">默认折叠参数和结果；缺失响应的工具会优先展开。</div>${pairs.map((pair,index)=>`<article class="tool-row ${pair.result==null?'open':''}"><button class="tool-toggle"><span class="status ${pair.error===true?'bad':pair.result==null?'missing':'ok'}">${pair.error===true?'失败':pair.result==null?'待响应':'已完成'}</span><span class="tool-name">${esc(pair.name)}</span><span class="tool-intent">${esc(intentOf(pair)||pair.id)}</span></button><div class="tool-detail"><div class="tool-side"><div class="tool-label">调用参数</div><pre>${esc(pair.args||'{}')}</pre></div><div class="tool-side"><div class="tool-label">工具响应</div><pre>${esc(pair.result??'尚未捕获对应工具响应')}</pre></div></div></article>`).join('')}`}
+function renderView(record){const req=record.request_messages||[],delta=req.slice(record.request_delta_start||0),resp=record.response_messages||[],pairs=record.tool_pairs||[];let html;if(tab==='overview'){const directCalls=resp.reduce((n,m)=>n+(m.tool_calls||[]).length,0),results=delta.filter(m=>m.role==='tool').length;html=section('本次请求新增',`${delta.length} 条 · 约 ${tok(delta.reduce((n,m)=>n+(m.estimated_tokens||0),0))} tokens`,delta.length?delta.map(messageCard).join(''):'<div class="tool-summary">本次调用沿用同回合上下文，没有新增请求消息。</div>',dump(delta))+section('模型输出',`${resp.length} 条消息 · ${directCalls} 个工具调用`,resp.length?resp.map(messageCard).join(''):'<div class="tool-summary">模型没有返回消息。</div>',dump(resp))+section('执行摘要',`${results} 条工具结果输入 · ${directCalls} 个新工具调用`,`<div class="tool-summary">本次调用接收了 ${results} 条工具结果，并发起 ${directCalls} 个后续工具调用。切换到“工具”查看结构化参数和结果。</div>`)}else if(tab==='request'){html=section('请求增量',`${delta.length} 新增 / ${req.length} 总消息`,delta.map(messageCard).join('')||'<div class="tool-summary">没有新增消息。</div>',dump(delta))+section('完整请求',`${req.length} 条消息`,req.map(messageCard).join(''),dump(req))}else if(tab==='response'){html=section('响应消息',`${resp.length} 条 · ${tok(record.usage?.output_tokens)} output tokens`,resp.map(messageCard).join('')||'<div class="tool-summary">模型没有返回消息。</div>',dump(resp))}else if(tab==='tools'){html=toolsView(pairs)}else{html='<div class="raw-box"><div class="tool-summary">正在加载捕获快照…</div></div>';loadRaw(record)}$('view').innerHTML=html;bindView(record)}function bindView(record){$('view').querySelectorAll('.tool-toggle').forEach(button=>button.onclick=()=>button.parentElement.classList.toggle('open'));$('view').querySelectorAll('.copy-message').forEach(button=>button.onclick=()=>copy(decodeURIComponent(button.dataset.text),button));$('view').querySelectorAll('.copy-section').forEach(button=>button.onclick=()=>{const title=button.parentElement.querySelector('h2').textContent;let value=title.includes('完整')?record.request_messages:title.includes('响应')?record.response_messages:(record.request_messages||[]).slice(record.request_delta_start||0);copy(dump(value),button)})}async function loadRaw(record){const raw=await api('/api/records/'+selected+'/raw');if(tab!=='raw'||loaded!==record)return;const box=$('view');box.innerHTML=`<section class="section"><div class="section-head"><h2>捕获快照</h2><span class="meta">按需加载</span><button class="btn" id="copyRaw">复制 JSON</button></div><div class="raw-box"><pre>${esc(dump(raw))}</pre></div></section>`;$('copyRaw').onclick=()=>copy(dump(raw),$('copyRaw'))}function drawDetail(record){if(!record){$('main').innerHTML='<div class="empty"><strong>等待通信记录</strong><span>开启采集并运行一个回合后，在这里检查模型调用。</span></div>';return}const pairs=record.tool_pairs||[],delta=(record.request_messages||[]).slice(record.request_delta_start||0),newTokens=delta.reduce((n,m)=>n+(m.estimated_tokens||0),0),tools=pairs.length,failed=tools?pairs.filter(p=>p.error===true).length:0,failRate=tools?Math.round(failed/tools*100)+'%':'—',status=record.error?'失败':'已完成';$('main').innerHTML=`<div class="call-head"><div><div class="eyebrow">模型通信 / 回合 ${record.turn_index}</div><h1>调用 #${record.model_call_index} <span class="${record.error?'bad':''}">${status}</span></h1><div class="sub">${esc(record.provider)} · ${esc(record.model_name)} · ${new Date((record.started_at||0)*1000).toLocaleString()}</div></div><button class="btn" id="copyRawTop">复制原始记录</button></div><div class="metrics"><div class="metric"><label>耗时</label><strong>${ms(record.duration_ms)}</strong></div><div class="metric"><label>总输入</label><strong>${tok(record.usage?.input_tokens)}</strong></div><div class="metric"><label>本次新增</label><strong>+${tok(newTokens)}</strong></div><div class="metric"><label>输出</label><strong>${tok(record.usage?.output_tokens)}</strong></div><div class="metric"><label>关联工具</label><strong>${tools}</strong></div><div class="metric"><label>失败率</label><strong>${failRate}</strong></div></div>${record.error?`<div class="notice bad">${esc(record.error)}</div>`:''}<nav class="tabs">${[['overview','概览'],['request','请求'],['response','响应'],['tools',`工具 ${tools}`],['raw','原始']].map(([id,label])=>`<button class="tab ${tab===id?'active':''}" data-tab="${id}">${label}</button>`).join('')}</nav><div id="view"></div>`;$('main').querySelectorAll('.tab').forEach(button=>button.onclick=()=>{tab=button.dataset.tab;drawDetail(record)});$('copyRawTop').onclick=async()=>{const raw=await api('/api/records/'+selected+'/raw');if(raw)copy(dump(raw),$('copyRawTop'))};renderView(record)}async function select(index){selected=index;drawList();$('main').innerHTML='<div class="empty"><strong>正在加载调用详情</strong></div>';const record=await api('/api/records/'+index);if(selected===index&&record){loaded=record;drawDetail(record)}}async function refresh(forceFollow=false){const [status,next]=await Promise.all([api('/api/status'),api('/api/records')]);if(status){$('capture').querySelector('.switch').classList.toggle('on',!!status.enabled);$('state').textContent=status.enabled?'开启':'关闭'}if(!next)return;const previousCount=records.length,hadSelection=selected>=0;records=next;drawList();if(records.length>previousCount&&hadSelection&&!$('follow').checked){newCount+=records.length-previousCount;$('newRecords').style.display='block';$('newRecords').textContent=`${newCount} 条新记录`}if((!hadSelection||forceFollow||$('follow').checked&&records.length>previousCount)&&records.length)select(records.at(-1).index);else if(selected>=records.length){selected=-1;loaded=null;drawDetail(null)}}$('capture').onclick=async()=>{await api('/api/toggle',{method:'POST'});refresh()};$('refresh').onclick=()=>refresh(false);$('newRecords').onclick=()=>{newCount=0;$('newRecords').style.display='none';refresh(true)};$('clear').onclick=async()=>{if(records.length&&confirm('清空全部已采集的通信记录？')){await api('/api/clear',{method:'POST'});records=[];selected=-1;loaded=null;drawList();drawDetail(null)}};$('search').oninput=drawList;document.querySelectorAll('.filter').forEach(button=>button.onclick=()=>{filter=button.dataset.filter;document.querySelectorAll('.filter').forEach(item=>item.classList.toggle('active',item===button));drawList()});document.addEventListener('visibilitychange',()=>{clearInterval(timer);if(!document.hidden)timer=setInterval(refresh,1500)});refresh();timer=setInterval(refresh,1500);
+</script>
+</body>
+</html>