synapse-cli-agent 0.1.37__tar.gz → 0.1.38__tar.gz

{synapse_cli_agent-0.1.37 → synapse_cli_agent-0.1.38}/AGENTS.md

@@ -83,11 +83,21 @@ uv build                        # after packaging/entry-point changes
 
 ## Rust/PyO3 native crates
 
-One optional crate; the Python app must work without its wheel and must keep `ImportError`/`OSError` fallbacks.
+The Python app must work without the optional wheels and must keep `ImportError`/`OSError` fallbacks.
 
-- `rust/synapse-core-tool` (`synapse-core-tool` on PyPI): native filesystem tools (read/edit/patch/search) and math rendering. Release with tag `synapse-core-tool-v*`; `native-core-tool-wheels.yml` builds wheels and publishes to GitHub Release + PyPI. Keep the `pyproject.toml` version in sync with `Cargo.toml` — maturin names wheels from `pyproject.toml`.
+- `rust/synapse-core-tool` (`synapse-core-tool` on PyPI): native filesystem tools (read/edit/patch/search), math rendering, and the native OpenAI-compatible chat client (`RustOpenAIClient`, backed by `async-openai` byot). Release with tag `synapse-core-tool-v*`; `native-core-tool-wheels.yml` builds wheels and publishes to GitHub Release + PyPI. Keep the `pyproject.toml` version in sync with `Cargo.toml` — maturin names wheels from `pyproject.toml`.
 - `rust/synapse-tool-compress-core` (`synapse-tool-compress-core` on PyPI): native tool-output compression core. **Required dependency** of `synapse-cli-agent`. Release with tag `synapse-tool-compress-core-v*`; `native-compression-wheels.yml` builds wheels and publishes to GitHub Release + PyPI. Keep the `pyproject.toml` version in sync with `Cargo.toml`. Because the main package hard-depends on it, publish the compress-core wheels to PyPI before tagging the main `v*` release.
 
+### Building and installing a native crate locally
+
+`uv` builds the path dependencies declared in `[tool.uv.sources]` (`rust/synapse-core-tool` and `rust/synapse-tool-compress-core`). After changing Rust code, rebuild + reinstall the affected crate with:
+
+```powershell
+uv sync --reinstall-package synapse-core-tool
+```
+
+This runs maturin under the hood and installs the freshly built wheel into the venv. Do **not** rely on `maturin develop` — a plain `uv sync` afterwards re-copies the cached (stale) wheel over the develop install. To avoid that surprise, either always rebuild via `uv sync --reinstall-package ...`, or run every command with `uv run --no-sync` after a manual `maturin develop`.
+
 When changing either crate, at least run:
 
 ```powershell
@@ -95,7 +105,7 @@ cargo test --manifest-path rust/<crate>/Cargo.toml
 cargo fmt --manifest-path rust/<crate>/Cargo.toml --check
 ```
 
-If Python bindings/APIs change, build or install the local extension and run the related Python tests.
+If Python bindings/APIs change, rebuild via `uv sync --reinstall-package <crate>` and run the related Python tests.
 
 Keep the Apache-2.0 SPDX headers, attribution, `LICENSE`, and `NOTICE` under `rust/synapse-tool-compress-core/src/headroom_port/`; do not introduce excluded network calls or model downloads.
 

{synapse_cli_agent-0.1.37 → synapse_cli_agent-0.1.38}/CHANGELOG.md

@@ -8,6 +8,28 @@ All entries are written in English.
 
 ---
 
+## v0.1.38
+
+### New Features
+
+- Added a native Rust OpenAI chat transport (`RustOpenAIClient`) backed by `synapse-core-tool`, used as an OpenAI-compatible model backend.
+- Added an interactive HITL approval block: turn events emit `APPROVAL_REQUIRED`, the transcript mounts an `ApprovalBlock`, and decisions route through `resume_hitl` without treating `WAITING_APPROVAL` as a blocking busy state.
+- Added a `session-crash-repair` skill for recovering crashed sessions.
+
+### Bug Fixes
+
+- Fixed `usage_metadata` mapping in the OpenAI model: `prompt_tokens_details`/`completion_tokens_details` now map to `input_token_details`/`output_token_details` (cache_read and reasoning were always 0), `prompt_cache_hit_tokens` is preferred for cache reads on DeepSeek-compatible gateways, and the async run manager is passed through `_astream` so `on_llm_new_token` fires and TTFT/token rates update live.
+- Fixed the TUI thought block collapsing when clicking the expanded reasoning body; the toggle is now restricted to the header row(s) and clicks that finish a text selection are ignored.
+
+### Engineering
+
+- Reused the OpenAI HTTP client across requests in the native Rust transport (`synapse-core-tool` v0.1.4).
+- Deferred markdown parsing of long restored transcript bodies to a worker (bounded plain-text preview first) and mounted restored blocks in chunks of 12 to avoid one giant layout pass.
+- Cached syntax highlighting across renders (`lru_cache`, cap 16k chars) so repeated redraws of code fences skip re-tokenizing.
+- Cached `SelectableStatic` renders so selection drags skip markdown re-rendering.
+
+---
+
 ## v0.1.37
 
 ### Engineering

{synapse_cli_agent-0.1.37 → synapse_cli_agent-0.1.38}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: synapse-cli-agent
-Version: 0.1.37
+Version: 0.1.38
 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/approval-behavior.md

@@ -0,0 +1,46 @@
+# 审批（Approval）行为说明
+
+结论：打开审批后，agent 在执行**写操作类工具前会暂停等待用户决定**，而不是直接拒绝或跳过。
+
+依据代码：`src/synapse/runtime/safety.py`、`src/synapse/runtime/hitl.py`、`src/synapse/app/agent.py`、`src/synapse/ui/turn/controller.py`、`src/synapse/commands/slash_cmds.py`。
+
+## 如何正确打开审批
+
+| 方式 | 说明 |
+| --- | --- |
+| 环境变量 `AGENT_SAFETY_PROFILE=dev-approve` | 推荐方式（可在 `.env` 或系统环境变量中设置） |
+| TUI 内 `/safety dev-approve` | 切换后重建 agent |
+| 档位别名 | `approve` / `hitl` / `dev_approve` 均映射到 `dev-approve` |
+
+> 注意：仅设 `AGENT_REQUIRE_APPROVAL=true` 或 CLI `--require-approval` **不会生效**。
+> agent 构建时 `apply_safety_to_settings` 会按 `safety_profile`（默认 `dev-autopass`）把 `require_approval` 强制改回 `false`。
+> 必须切换到 `dev-approve` 档。
+
+## 打开后的行为
+
+| 阶段 | 表现 |
+| --- | --- |
+| 需要审批的工具 | 仅 `execute`、`write_file`、`edit_file`、`patch` 四类（`build_interrupt_on` 返回）；`read_file`、`search_files` 等只读工具照常执行，不拦截 |
+| 触发时机 | agent 一旦尝试调用上述工具，LangGraph 在**执行前**中断（interrupt），回合状态变为 `WAITING_APPROVAL` |
+| TUI 界面 | 顶部活动区显示 `waiting / approval`；转录区打印待批清单：工具名、参数预览、描述，末尾提示 `decide: /approve | /reject [reason]` |
+| 用户操作 | `/approve` 放行；`/reject [reason]` 拒绝（可附理由） |
+
+## 审批决策的两种结果
+
+| 决策 | 行为 |
+| --- | --- |
+| `/approve` | 恢复图执行，工具真正运行，agent 继续后续步骤 |
+| `/reject [reason]` | 工具不执行，拒绝信息（默认 `Rejected by user via /reject` 或自定义 reason）作为反馈回传给模型，agent 会据此调整方案或换工具重试 |
+
+## 其他要点
+
+| 项 | 说明 |
+| --- | --- |
+| 多个工具同时待批 | `build_decisions` 会按待批数量生成等量决策；一次 `/approve` 放行全部 |
+| 无待批时 | `/approve` 或 `/reject` 会提示 `no pending approval`，不做事 |
+| 黑名单 | `dev-approve` 档 `enable_command_blacklist=True`，但 `check_command` 目前仅定义、未接入运行时；`rm -rf` 等危险命令不会自动拦截，仍走人工审批 |
+| ACP 客户端 | 通过 `_request_permission_decisions` 请求权限，超过 `_max_permission_turns` 时报 `maximum permission resume turns exceeded` |
+
+## 一句话总结
+
+开启审批（`dev-approve`）后，agent 每次写文件或执行命令前都会停下等待用户 `/approve` 或 `/reject`，只读操作不受影响。

{synapse_cli_agent-0.1.37 → synapse_cli_agent-0.1.38}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "synapse-cli-agent"
-version = "0.1.37"
+version = "0.1.38"
 description = "Local coding agent built on LangChain Deep Agents (LocalShell, no sandbox)"
 readme = "README.md"
 license = { file = "LICENSE" }

synapse_cli_agent-0.1.38/release_body.md

@@ -0,0 +1,21 @@
+## v0.1.38
+
+### New Features
+
+- Added a native Rust OpenAI chat transport (`RustOpenAIClient`) backed by `synapse-core-tool`, used as an OpenAI-compatible model backend.
+- Added an interactive HITL approval block: turn events emit `APPROVAL_REQUIRED`, the transcript mounts an `ApprovalBlock`, and decisions route through `resume_hitl` without treating `WAITING_APPROVAL` as a blocking busy state.
+- Added a `session-crash-repair` skill for recovering crashed sessions.
+
+### Bug Fixes
+
+- Fixed `usage_metadata` mapping in the OpenAI model: `prompt_tokens_details`/`completion_tokens_details` now map to `input_token_details`/`output_token_details` (cache_read and reasoning were always 0), `prompt_cache_hit_tokens` is preferred for cache reads on DeepSeek-compatible gateways, and the async run manager is passed through `_astream` so `on_llm_new_token` fires and TTFT/token rates update live.
+- Fixed the TUI thought block collapsing when clicking the expanded reasoning body; the toggle is now restricted to the header row(s) and clicks that finish a text selection are ignored.
+
+### Engineering
+
+- Reused the OpenAI HTTP client across requests in the native Rust transport (`synapse-core-tool` v0.1.4).
+- Deferred markdown parsing of long restored transcript bodies to a worker (bounded plain-text preview first) and mounted restored blocks in chunks of 12 to avoid one giant layout pass.
+- Cached syntax highlighting across renders (`lru_cache`, cap 16k chars) so repeated redraws of code fences skip re-tokenizing.
+- Cached `SelectableStatic` renders so selection drags skip markdown re-rendering.
+
+---