synapse-cli-agent 0.1.38__tar.gz → 0.1.39__tar.gz

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/.github/workflows/release.yml

@@ -102,6 +102,7 @@ jobs:
             --collect-all termaid `
             --collect-all texicode `
             --collect-all synapse_tool_compress_core `
+            --collect-data synapse `
             src/synapse/entry.py
 
       - name: Rename executable for artifact

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/AGENTS.md

@@ -4,7 +4,8 @@
 
 | Path | Responsibility |
 | --- | --- |
-| `src/synapse/app/` | Agent assembly and `AGENTS.md` injection |
+| `src/synapse/app/` | Agent assembly (composition root in `agent.py`, contracts in `agent_assembly.py`) and `AGENTS.md` injection |
+| `src/synapse/acp/` | Official Agent Client Protocol (ACP) adapter: run Synapse as an ACP agent over stdio; session registry, permission coordinator, event bridge |
 | `src/synapse/commands/` | Slash command parsing, completion, and results |
 | `src/synapse/content/` | Prompts, skills, input history, multimodal content |
 | `src/synapse/goals/` | Goal accounting, steering, and persistence |
@@ -29,7 +30,7 @@
 
 ## Architecture & compatibility
 
-- `src/synapse/app/agent.py` is the assembly layer; keep domain algorithms in their own packages.
+- `src/synapse/app/agent.py` is the assembly composition root; the reusable assembly contracts (middleware build, typed resource wiring) live in `src/synapse/app/agent_assembly.py`. Keep domain algorithms in their own packages.
 - New features belong in the matching domain package; cross-domain wiring goes in `app/` or an explicit runtime middleware.
 - `src/synapse/config.py` is a compatibility re-export layer. Import from `synapse.settings` in new code; keep legacy import paths working.
 - `__init__.py` exports are public API. Keep necessary re-exports when moving implementations.

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/CHANGELOG.md

@@ -8,6 +8,28 @@ All entries are written in English.
 
 ---
 
+## v0.1.39
+
+### New Features
+
+- Stamped session-affinity headers on model requests in both the Python runtime and the native Rust OpenAI client, for consistent session binding at upstream gateways.
+- Added a `session-cache-analysis` skill for computing per-session prompt cache hit rates and diagnosing cache eviction.
+
+### Bug Fixes
+
+- Hashed tools by their OpenAI wire form instead of the raw `StructuredTool` dump, keeping `tools_hash`/`schema_hash` stable across agent rebuilds (previously embedded the function object's memory address).
+- Inserted pasted text at the cursor position in the TUI prompt instead of appending.
+- Persisted the model binding before agent rebuild and on exit.
+- Loaded the debug inspector HTML via `importlib.resources` for frozen (PyInstaller) builds.
+
+### Engineering
+
+- Removed the tool-response truncation middleware and its settings.
+- Deprecated the reversible compression middleware and reader.
+- Kept the AGENTS.md repository-layout table in sync with the current tree.
+
+---
+
 ## v0.1.38
 
 ### New Features

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: synapse-cli-agent
-Version: 0.1.38
+Version: 0.1.39
 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

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/README.zh-CN.md

@@ -415,7 +415,6 @@ uv run synapse run "修复 calculator.sub 的 bug，使 tests 全部通过" -w t
 | `AGENT_MCP_CONFIG` | - | MCP servers JSON |
 | `CHECKPOINT_BACKEND` | `sqlite` | `sqlite` 或 `memory` |
 | `AGENT_SESSION_PREWARM_ENABLED` | `false` | 恢复大上下文会话后在后台预热 provider 缓存（首次请求仍按输入 token 计费；详见 `docs/config.md`） |
-| `AGENT_ENABLE_TOOL_RESPONSE_TRUNCATE` | `false` | 启用工具响应折叠（keep 窗口外的大工具输出压成摘要+引用，大上下文会话可省 20-30%；详见 `docs/config.md`） |
 | `INHERIT_ENV` | `true` | shell 继承主机环境 |
 | `VIRTUAL_MODE` | `true` | 文件路径虚拟根 |
 | `AGENT_SHOW_REASONING_PLACEHOLDERS` | `true` | 是否显示网关未暴露推理文本时的占位思考节点；Codex 加密推理可设为 `false` 隐藏 |

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/architecture-flow.md

@@ -467,7 +467,7 @@ load_settings(**overrides):
 | 工作区 | workspace, shell_timeout(120s), max_output_bytes(100KB) | Path.cwd() |
 | 安全 | require_approval(False), safety_profile("dev-autopass") | -- |
 | 检查点 | checkpoint_backend("sqlite") | project/.synapse/ |
-| 工具输出 | enable_tool_output_transform(True), threshold_bytes(512) | -- |
+| 工具输出 | enable_tool_output_transform(False), threshold_bytes(512) | -- |
 | 子代理 | enable_subagents(True) | -- |
 | MCP | enable_mcp(True), mcp_eager(False) | -- |
 | UI | theme("cursor-dark"), history_tail_turns(20) | -- |

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/config.md

@@ -77,13 +77,6 @@ Synapse 使用 **Pydantic Settings** 实现分层配置系统。
 | `SESSION_SUMMARY_MAX_CHARS` | `600` | 本地会话摘要的最大字符数（含多轮条目，超出时从最旧条目开始裁剪） |
 | `AGENT_HISTORY_TAIL_TURNS` | `20` | TUI 从轻量 transcript 投影初始读取/渲染的最近可见会话轮数；滚动到顶部后按 turn 游标加载更早历史，最多挂载 5 页 |
 | `AGENT_SESSION_PREWARM_ENABLED` | `false` | 恢复大上下文会话后，在后台对该会话历史发起一次最小模型请求，让 provider 预填充并缓存历史前缀，用户第一条消息可命中缓存、大幅缩短首 token 等待。注意：预热本身会按输入 token 计费一次，仅在需要时开启 |
-| `AGENT_ENABLE_TOOL_RESPONSE_TRUNCATE` | `false` | 启用工具响应裁剪：keep 窗口（默认最近 40K tokens）外的 `execute`/`read_file`/`search_files` 输出折叠为摘要，减小每轮请求体积（大上下文会话可省 20-30%）。请求层变换，不修改历史 |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_KEEP_TOKENS` | `40000` | 工具响应裁剪的 keep 窗口（token），窗口内消息不裁剪 |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_MAX_HEAD_CHARS` | `2000` | 关闭 fold 模式时，超长输出保留的头部字符数 |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_MAX_TAIL_CHARS` | `0` | 关闭 fold 模式时，超长输出额外保留的尾部字符数（>0 表示 head+tail） |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_FOLD_ENABLED` | `true` | fold 模式：窗口外输出一律折叠（有 `tool-output://` 引用的可经 `read_tool_result` 找回原文，无引用的保留头部锚点） |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_FOLD_HEAD_CHARS` | `300` | fold 模式下无引用输出保留的头部锚点字符数 |
-| `AGENT_TOOL_RESPONSE_TRUNCATE_TOOLS` | `["execute","read_file","search_files"]` | 参与折叠/裁剪的工具名单（JSON 数组） |
 
 ### 长程目标（Goals）
 

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/design.md

@@ -250,7 +250,9 @@ py-agent/
 │   └── design.md                 # 本方案
 ├── AGENTS.md                     # 项目级记忆/约定（给 agent 读）
 ├── skills/                       # 可选 skills（渐进披露）
-│   └── python-testing/
+│   ├── session-crash-repair/
+│   │   └── SKILL.md
+│   └── session-cache-analysis/
 │       └── SKILL.md
 ├── src/
 │   └── coding_agent/

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/skills.md

@@ -6,9 +6,9 @@ Skills 是 Synapse 的可扩展插件系统，位于项目 `skills/` 目录下
 
 ```
 skills/
-├── python-testing/
+├── session-cache-analysis/
 │   └── SKILL.md
-└── python-memory-diagnostics/
+└── session-crash-repair/
     └── SKILL.md
 ```
 
@@ -18,16 +18,16 @@ skills/
 
 ```markdown
 ---
-name: python-testing
-description: How to run and write Python tests in this project with uv and pytest.
-license: MIT
-compatibility: Requires uv and pytest in a Python 3.12+ project.
-allowed_tools: execute, run_tests, read_file, write_file, edit_file
+name: session-cache-analysis
+description: Analyze the prompt cache hit rate of a Synapse session (thread).
+license: Apache-2.0
+compatibility: Requires a synapse source checkout with uv and Python 3.12+.
+allowed_tools: execute, read_file, write_file, search_files, find_files, search_session, read_session
 ---
 
-# Python Testing Skill
+# Session Cache Hit-Rate Analysis
 
-## Running Tests
+## 数据来源
 ...
 ```
 
@@ -47,8 +47,8 @@ Synapse 自带两个 Skills：
 
 | Skill | 说明 |
 |---|---|
-| `python-testing` | 使用 uv + pytest 运行和编写测试 |
-| `python-memory-diagnostics` | 诊断 Python 进程内存问题 |
+| `session-crash-repair` | 检测并修复异常退出后 checkpoint 不一致的会话（悬挂工具调用、挂起图任务、过期转录投影） |
+| `session-cache-analysis` | 分析会话的 prompt 缓存命中率，区分增量 miss 与全量逐出并定位根因 |
 
 ## 如何编写自定义 Skill
 

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/tool-output-transform.md

@@ -18,7 +18,7 @@ All transforms are deterministic. Synapse uses a low-cost `512` byte prefilter,
 
 ```json
 {
-  "enable_tool_output_transform": true,
+  "enable_tool_output_transform": false,
   "tool_output_transform_threshold_bytes": 512,
   "tool_output_disabled_types": ["code"],
   "tool_output_transform_plugins": ["my_package.transforms:my_transformer"],
@@ -26,10 +26,12 @@ All transforms are deterministic. Synapse uses a low-cost `512` byte prefilter,
 }
 ```
 
+Tool-output transformation is **off by default** (`enable_tool_output_transform: false`); set it to `true` to enable.
+
 Environment equivalents:
 
 ```text
-AGENT_ENABLE_TOOL_OUTPUT_TRANSFORM=true
+AGENT_ENABLE_TOOL_OUTPUT_TRANSFORM=false
 AGENT_TOOL_OUTPUT_TRANSFORM_THRESHOLD_BYTES=512
 AGENT_TOOL_OUTPUT_DISABLED_TYPES=["code"]
 AGENT_TOOL_OUTPUT_TRANSFORM_PLUGINS=["my_package.transforms:my_transformer"]

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/docs/tutorial.md

@@ -328,9 +328,9 @@ synapse-agent/                       # 项目根目录
 │   └── fixtures/                    # 测试 fixtures
 │
 ├── skills/                          # Agent Skills
-│   ├── python-testing/              # Python 测试技能
+│   ├── session-crash-repair/        # 会话崩溃修复技能
 │   │   └── SKILL.md                 # 技能定义文件
-│   └── python-memory-diagnostics/   # 内存诊断技能
+│   └── session-cache-analysis/      # 会话缓存命中率分析技能
 │       └── SKILL.md
 │
 ├── docs/                            # 文档
@@ -1073,20 +1073,20 @@ Skills 是给 Agent 的可加载"专业说明手册"。
 ### 10.1 Skill 文件格式
 
 ```markdown
-<!-- skills/python-testing/SKILL.md -->
+<!-- skills/session-cache-analysis/SKILL.md -->
 
 ---
-name: python-testing
-description: How to run and write Python tests in this project with uv and pytest.
-license: MIT
-allowed-tools: execute run_tests read_file write_file edit_file
+name: session-cache-analysis
+description: Analyze the prompt cache hit rate of a Synapse session (thread).
+license: Apache-2.0
+allowed-tools: execute read_file write_file search_files find_files
 ---
 
-# Python testing skill
+# Session cache hit-rate analysis
 
-## Run tests
+## Data source
 Prefer:
-`uv run pytest`
+`load_messages_from_sqlite_file(checkpoints.sqlite, thread_id)`
 
 ...
 ```
@@ -1114,8 +1114,8 @@ Agent 启动时会扫描 `skills/` 目录，将所有 Skill 的描述注入系
 
 ### 10.3 现有 Skills
 
-- **python-testing**: 告诉 Agent 如何正确地运行和编写 Python 测试
-- **python-memory-diagnostics**: 告诉 Agent 如何诊断 Python 进程的内存问题
+- **session-crash-repair**: 告诉 Agent 如何检测并修复异常退出后 checkpoint 不一致的会话
+- **session-cache-analysis**: 告诉 Agent 如何分析会话的 prompt 缓存命中率并定位逐出根因
 
 ### 10.4 如何创建新 Skill
 

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synapse-cli-agent"
-version = "0.1.38"
+version = "0.1.39"
 description = "Local coding agent built on LangChain Deep Agents (LocalShell, no sandbox)"
 readme = "README.md"
 license = { file = "LICENSE" }

synapse_cli_agent-0.1.39/release_body.md

@@ -0,0 +1,21 @@
+## v0.1.39
+
+### New Features
+
+- Stamped session-affinity headers on model requests in both the Python runtime and the native Rust OpenAI client, for consistent session binding at upstream gateways.
+- Added a `session-cache-analysis` skill for computing per-session prompt cache hit rates and diagnosing cache eviction.
+
+### Bug Fixes
+
+- Hashed tools by their OpenAI wire form instead of the raw `StructuredTool` dump, keeping `tools_hash`/`schema_hash` stable across agent rebuilds (previously embedded the function object's memory address).
+- Inserted pasted text at the cursor position in the TUI prompt instead of appending.
+- Persisted the model binding before agent rebuild and on exit.
+- Loaded the debug inspector HTML via `importlib.resources` for frozen (PyInstaller) builds.
+
+### Engineering
+
+- Removed the tool-response truncation middleware and its settings.
+- Deprecated the reversible compression middleware and reader.
+- Kept the AGENTS.md repository-layout table in sync with the current tree.
+
+---

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/app/agent_assembly.py

@@ -29,11 +29,6 @@ from synapse.runtime.model_request_compression_middleware import (
     build_model_request_compression_middleware,
 )
 from synapse.runtime.steer import SteerQueue, build_steer_middleware
-from synapse.runtime.tool_output_middleware import build_tool_output_transform_middleware
-from synapse.runtime.tool_output_usage_middleware import build_tool_output_usage_middleware
-from synapse.runtime.tool_response_truncate_middleware import (
-    build_tool_response_truncate_middleware,
-)
 from synapse.tool_output.pipeline import ToolOutputTransformPipeline
 from synapse.tool_output.repository import ToolOutputRepository
 from synapse.tool_output.transformers import load_transformer_plugins
@@ -105,7 +100,11 @@ class MiddlewareContext:
 
 
 def build_tool_output_pipeline(settings: Any) -> ToolOutputTransformPipeline:
-    """Build the reversible tool-output pipeline with plugin fallback."""
+    """Deprecated: build the reversible tool-output pipeline with plugin fallback.
+
+    The transform middleware is no longer registered; this helper is kept for
+    compatibility and for anyone reviving the feature.
+    """
     try:
         return ToolOutputTransformPipeline(
             transformers=load_transformer_plugins(settings.tool_output_transform_plugins),
@@ -128,9 +127,15 @@ def build_agent_middleware(context: MiddlewareContext) -> list[Any]:
     from synapse.runtime.filesystem_tool_prompt_middleware import (
         build_filesystem_tool_prompt_middleware,
     )
+    from synapse.runtime.session_header_middleware import (
+        build_session_header_middleware,
+    )
 
     settings = context.settings
     middleware: list[Any] = [
+        # Publish the active thread id so the httpx layer can stamp
+        # X-Session-ID / Session-Id on every model request (gateway affinity).
+        build_session_header_middleware(),
         build_agent_md_middleware(context.project_root),
         build_filesystem_tool_prompt_middleware(),
         build_describe_image_middleware(
@@ -145,22 +150,11 @@ def build_agent_middleware(context: MiddlewareContext) -> list[Any]:
             service=context.goal_service,
         ),
     ]
-    transform_enabled = bool(getattr(settings, "enable_tool_output_transform", True))
-    if context.turbo:
-        # Turbo routes tool outputs through the headroom proxy for compression;
-        # running the built-in transformer too would double-compress and break
-        # the proxy's content-type routing.
-        transform_enabled = False
-    middleware.append(
-        build_tool_output_transform_middleware(
-            context.output_repository,
-            threshold_bytes=getattr(settings, "tool_output_transform_threshold_bytes", 512),
-            pipeline=build_tool_output_pipeline(settings),
-            enabled=transform_enabled,
-        )
-    )
-    if transform_enabled:
-        middleware.append(build_tool_output_usage_middleware(context.output_repository))
+    # Deprecated: the reversible tool-output transform middleware
+    # (build_tool_output_transform_middleware / build_tool_output_usage_middleware)
+    # is intentionally NOT registered for now. The pipeline and read_tool_result
+    # stay in place for compatibility; re-enable by appending those builders
+    # here when the feature is revived.
     middleware.extend(
         [
             build_tool_error_recovery_middleware(),
@@ -173,7 +167,6 @@ def build_agent_middleware(context: MiddlewareContext) -> list[Any]:
         [
             build_strip_redundant_prompt_blocks(),
             build_compact_tool_descriptions(),
-            build_tool_response_truncate_middleware(settings),
             build_model_request_compression_middleware(context.output_repository),
         ]
     )

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/commands/model.py

@@ -18,7 +18,7 @@ def handle_model(
     thread_id: str | None = None,
     apply_thinking_inplace: Callable[[Any, Any, str], bool],
     rebuild_agent: Callable[..., Any],
-    persist_model_binding: Callable[[Any, str | None], None],
+    persist_model_binding: Callable[[Any, str | None], str | None],
     mcp_attach_pending: Callable[[Any], bool],
 ) -> SlashResult:
     from synapse.models.registry import (
@@ -73,6 +73,9 @@ def handle_model(
             label = apply_thinking_to_settings(settings, args[1], allowed=allowed)
         except ValueError as exc:
             return SlashResult(handled=True, lines=[str(exc)], error=True)
+        # Persist before the (slow, fallible) rebuild so the user's choice is
+        # not lost when the rebuild fails or the process exits mid-switch.
+        persist_error = persist_model_binding(settings, thread_id)
         model_name = settings.active_model or reg.default
         new_agent = None
         note = ""
@@ -88,15 +91,16 @@ def handle_model(
                     defer_mcp_reconnect=True,
                 )
             except Exception as exc:  # noqa: BLE001
-                return SlashResult(
-                    handled=True,
-                    lines=[f"thinking update failed: {exc}"],
-                    error=True,
-                )
-        persist_model_binding(settings, thread_id)
+                lines = [f"thinking update failed: {exc}"]
+                if persist_error:
+                    lines.append(persist_error)
+                return SlashResult(handled=True, lines=lines, error=True)
+        lines = [f"thinking set to {label}{note}  ({format_model_status(settings)})"]
+        if persist_error:
+            lines.append(persist_error)
         return SlashResult(
             handled=True,
-            lines=[f"thinking set to {label}{note}  ({format_model_status(settings)})"],
+            lines=lines,
             agent=new_agent,
             settings_changed=True,
             mcp_attach_pending=bool(new_agent is not None and mcp_attach_pending(settings)),
@@ -148,6 +152,11 @@ def handle_model(
         except ValueError as exc:
             return SlashResult(handled=True, lines=[str(exc)], error=True)
 
+    # Persist the selection before the (slow, fallible) agent rebuild so a
+    # rebuild failure or an immediate quit cannot silently drop the user's
+    # model choice (it would otherwise live only in process memory).
+    persist_error = persist_model_binding(settings, thread_id)
+
     try:
         new_agent = rebuild_agent(
             settings,
@@ -157,16 +166,21 @@ def handle_model(
             defer_mcp_reconnect=True,
         )
     except Exception as exc:  # noqa: BLE001
+        lines = [f"model switch failed: {exc}"]
+        if persist_error:
+            lines.append(persist_error)
         return SlashResult(
             handled=True,
-            lines=[f"model switch failed: {exc}"],
+            lines=lines,
             error=True,
         )
-    persist_model_binding(settings, thread_id)
     mcp_attach_pending = mcp_attach_pending(settings)
+    lines = [f"model switched to {profile.name}  ({format_model_status(settings)})"]
+    if persist_error:
+        lines.append(persist_error)
     return SlashResult(
         handled=True,
-        lines=[f"model switched to {profile.name}  ({format_model_status(settings)})"],
+        lines=lines,
         agent=new_agent,
         settings_changed=True,
         mcp_attach_pending=mcp_attach_pending,

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/commands/slash_cmds.py

@@ -141,12 +141,17 @@ def _store(settings: Any) -> SessionStore:
     return SessionStore(settings.resolved_sessions_path())
 
 
-def _persist_model_binding(settings: Any, thread_id: str | None) -> None:
+def _persist_model_binding(settings: Any, thread_id: str | None) -> str | None:
+    """Persist the current settings as the session's model binding.
+
+    Returns an error message when the write failed (never raises), else None.
+    """
     try:
         store = _store(settings)
         store.save_model_binding(thread_id, binding_from_settings(settings), also_last=True)
-    except Exception:  # noqa: BLE001
-        pass
+    except Exception as exc:  # noqa: BLE001 - persistence must never raise into the caller
+        return f"failed to persist model binding: {exc}"
+    return None
 
 
 def _restore_thread_model(
@@ -352,7 +357,9 @@ def handle_slash(
             and result.thread_id != thread_id
             and cmd in {"/switch", "/session", "/new"}
         ):
-            _persist_model_binding(settings, thread_id)
+            persist_error = _persist_model_binding(settings, thread_id)
+            if persist_error:
+                result.lines = [*result.lines, persist_error]
         # When switching sessions, restore that session's model binding.
         if (
             result.handled

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/integrations/http_clients.py

@@ -142,9 +142,14 @@ def build_openai_async_http_client(
         )
         if capture_store is not None:
             transport = _CapturingAsyncTransport(transport, capture_store)
+        # Inject session-affinity headers (X-Session-ID / Session-Id) from the
+        # active thread context so gateways can keep one session on one key.
+        from synapse.runtime.session_headers import attach_session_headers
+
         return httpx.AsyncClient(
             transport=transport,
             timeout=DEFAULT_TIMEOUT if timeout is None else timeout,
+            event_hooks={"request": [attach_session_headers]},
         )
 
     client = runtime.run(_build())

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/models/rust_openai.py

@@ -13,6 +13,7 @@ Requests are sent as raw JSON so non-standard fields (DeepSeek
 from __future__ import annotations
 
 import asyncio
+import contextvars
 import json
 import os
 from collections.abc import AsyncIterator, Iterator
@@ -33,6 +34,8 @@ from langchain_core.messages import (
 from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
 from pydantic import Field, PrivateAttr, SecretStr
 
+from synapse.runtime.session_headers import get_session_id, session_header_values
+
 _OPENAI_ROLE_MAP = {
     "system": "system",
     "human": "user",
@@ -234,6 +237,10 @@ class RustOpenAIChatModel(BaseChatModel):
     parallel_tool_calls: bool | None = None
 
     _client: Any = PrivateAttr(default=None)
+    # Session id the cached native client was built for; a mismatch means the
+    # client must be rebuilt so the session-affinity headers are stamped on
+    # every request of the conversation.
+    _client_session_id: str | None = PrivateAttr(default=None)
 
     @property
     def _llm_type(self) -> str:
@@ -256,15 +263,29 @@ class RustOpenAIChatModel(BaseChatModel):
         return params
 
     def _ensure_client(self) -> Any:
-        if self._client is None:
+        # Not thread-safe for concurrent calls from *different* sessions on the
+        # same instance (a rebuild race could serve the wrong session headers);
+        # registry caches one model per agent session, so this does not happen
+        # in practice. Within one session the same client is reused.
+        session_id = get_session_id()
+        if self._client is None or self._client_session_id != session_id:
             from synapse_core_tool import RustOpenAIClient
 
+            # The native client bakes headers into the async-openai config at
+            # construction, so a per-session client is (re)built whenever the
+            # active session id changes; within one session the same client
+            # (and its reqwest connection pool) is reused.
+            headers = dict(self.default_headers)
+            session_headers = session_header_values()
+            if session_headers:
+                headers.update(session_headers)
             self._client = RustOpenAIClient(
                 api_key=self.api_key.get_secret_value() if self.api_key else None,
                 base_url=self.base_url,
-                headers=self.default_headers,
+                headers=headers,
                 timeout_secs=self.timeout,
             )
+            self._client_session_id = session_id
         return self._client
 
     def _build_request(
@@ -379,7 +400,12 @@ class RustOpenAIChatModel(BaseChatModel):
             finally:
                 queue.put_nowait(None)
 
-        task = asyncio.get_running_loop().run_in_executor(None, _run)
+        # run_in_executor does not propagate contextvars, so the sync streamer
+        # would lose the session id published by the agent middleware. Copy the
+        # current context explicitly so the native client stamps session
+        # headers on streaming requests too.
+        ctx = contextvars.copy_context()
+        task = asyncio.get_running_loop().run_in_executor(None, lambda: ctx.run(_run))
         while True:
             chunk = await queue.get()
             if chunk is None:
@@ -423,7 +449,9 @@ class RustOpenAIChatModel(BaseChatModel):
                     converted.append(tool)
         bound._bound_tools = converted  # type: ignore[attr-defined]
         bound._bound_tool_choice = tool_choice  # type: ignore[attr-defined]
-        # Shallow copy shares the native client; a new client is not needed.
+        # Shallow copy takes a snapshot of the native client at bind time; a
+        # new client is only needed when the session changes, in which case
+        # ``_ensure_client`` rebuilds on the bound instance too.
         bound._client = self._client
         return bound
 
@@ -434,6 +462,7 @@ class RustOpenAIChatModel(BaseChatModel):
 
     def close(self) -> None:
         self._client = None
+        self._client_session_id = None
 
     async def aclose(self) -> None:
         self.close()

{synapse_cli_agent-0.1.38 → synapse_cli_agent-0.1.39}/src/synapse/observability/debug_server.py

@@ -25,6 +25,7 @@ import threading
 import time
 import webbrowser
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from importlib.resources import files as package_files
 from pathlib import Path
 from typing import Any
 from urllib.parse import parse_qs, urlparse
@@ -224,7 +225,13 @@ def _record_summary(record: Any, index: int) -> dict[str, Any]:
 # Inspector page
 # ---------------------------------------------------------------------------
 
-_PAGE_HTML = (Path(__file__).with_name("debug_inspector.html")).read_text(encoding="utf-8")
+# Read via importlib.resources so the page also works under PyInstaller
+# (frozen) builds, where ``Path(__file__)`` points into the unpacked bundle.
+_PAGE_HTML = (
+    package_files("synapse.observability")
+    .joinpath("debug_inspector.html")
+    .read_text(encoding="utf-8")
+)
 
 
 # ---------------------------------------------------------------------------