@omnicrawl/cli-linux-x64 0.2.8 → 0.3.0

package/host/rust/assets/extensions/onejev_worker.py

@@ -0,0 +1,293 @@
+"""OneJev 本地决策 worker：在命名管道（Windows）/ Unix socket 上提供 System One 判定。
+
+为什么不用 `qev serve`：官方那个服务经 uvicorn/FastAPI 暴露 HTTP，本机调用方因此要为
+一次纯内存判定多付一层 TCP + ASGI 中转。这个 worker 直接持有 `qev` 的引擎对象，
+请求从管道进来、答案从管道出去，中间只有一次 JSON 编解码。
+
+协议（NDJSON，一行一个 JSON 对象）：
+
+    请求   {"model": "...", "state": {...}, "questions": {...}}
+    成功   {"model": "...", "answers": {...}, "usage": {...}, "qev": {...}}
+    失败   {"error": "可读原因"}
+
+响应形状故意与 HTTP 端点一致，Rust 侧读答案的代码因此不必区分两种传输。
+
+服务是**本机共用**的常驻进程，由 OmniCrawl 宿主以脱离进程的方式拉起（见
+`omnicrawl-onejev` 的 `server` 模块）：加载权重、创建端点、打出就绪标记，之后一直服务；
+停止只能显式做（宿主按 PID 回收整棵进程树）。
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+import os
+import socket
+import sys
+import threading
+import time
+from pathlib import Path
+
+# 就绪标记：宿主按它认定「服务已经能接请求」（写进服务日志）。
+READY_MARKER = "[onejev] 服务就绪"
+# 单次连接的等待上限（秒）：客户端总是连上就发，静默超时说明对端已经不在了。
+CONNECTION_TIMEOUT_SECONDS = 60.0
+# 管道缓冲区：一次判定请求的 state 上限是 64k 字符级别，留一倍余量。
+PIPE_BUFFER_BYTES = 128 * 1024
+
+log = logging.getLogger("onejev.worker")
+
+
+class Decider:
+    """引擎 + 判定；串行化——一次前向独占前缀缓存，并发只会互相踩。"""
+
+    def __init__(self, engine, model_name: str) -> None:
+        self.engine = engine
+        self.model_name = model_name
+        self._lock = threading.Lock()
+
+    def answer(self, body: dict) -> dict:
+        from qev.schema import SystemOneRequest
+
+        request = SystemOneRequest.model_validate(body)
+        with self._lock:
+            response, meta = self.engine.decide(request)
+        response.model = self.model_name
+        payload = response.model_dump()
+        payload["qev"] = meta
+        return payload
+
+
+def _handle_connection(decider: Decider, read_line, write_line, close) -> None:
+    """一次连接 = 一次判定：读一行请求，回一行响应（成功或错误），然后收线。"""
+    try:
+        line = read_line()
+        if not line:
+            return
+        payload = decider.answer(json.loads(line))
+    except Exception as error:  # 请求不合法 / 引擎报错：回一行可读原因，不猜答案
+        log.warning("判定失败：%s", error)
+        message = str(error) or error.__class__.__name__
+        try:
+            write_line(json.dumps({"error": message}, ensure_ascii=False))
+        except Exception:
+            pass
+    else:
+        write_line(json.dumps(payload, ensure_ascii=False))
+    finally:
+        close()
+
+
+# --------------------------------------------------------------------------
+# Windows：命名管道
+# --------------------------------------------------------------------------
+
+
+def windows_pipe_io(endpoint: str, decider: Decider) -> None:
+    import ctypes
+    import ctypes.wintypes as wintypes
+    import msvcrt
+
+    kernel32 = ctypes.WinDLL("kernel32", use_last_error=True)
+
+    pipe_access_duplex = 0x00000003
+    pipe_type_byte = 0x00000000
+    pipe_readmode_byte = 0x00000000
+    pipe_wait = 0x00000000
+    pipe_unlimited_instances = 255
+    invalid_handle = wintypes.HANDLE(-1).value
+    error_pipe_connected = 535
+
+    kernel32.CreateNamedPipeW.restype = wintypes.HANDLE
+    kernel32.CreateNamedPipeW.argtypes = [
+        wintypes.LPCWSTR, wintypes.DWORD, wintypes.DWORD, wintypes.DWORD,
+        wintypes.DWORD, wintypes.DWORD, wintypes.DWORD, ctypes.c_void_p,
+    ]
+    kernel32.ConnectNamedPipe.restype = wintypes.BOOL
+    kernel32.ConnectNamedPipe.argtypes = [wintypes.HANDLE, ctypes.c_void_p]
+    kernel32.PeekNamedPipe.restype = wintypes.BOOL
+    kernel32.PeekNamedPipe.argtypes = [
+        wintypes.HANDLE, ctypes.c_void_p, wintypes.DWORD,
+        ctypes.POINTER(wintypes.DWORD), ctypes.POINTER(wintypes.DWORD), ctypes.POINTER(wintypes.DWORD),
+    ]
+    kernel32.CloseHandle.restype = wintypes.BOOL
+    kernel32.CloseHandle.argtypes = [wintypes.HANDLE]
+
+    def create_instance() -> int:
+        return kernel32.CreateNamedPipeW(
+            endpoint,
+            pipe_access_duplex,
+            pipe_type_byte | pipe_readmode_byte | pipe_wait,
+            pipe_unlimited_instances,
+            PIPE_BUFFER_BYTES,
+            PIPE_BUFFER_BYTES,
+            0,
+            None,
+        )
+
+    def serve_instance(handle: int) -> None:
+        fd = msvcrt.open_osfhandle(handle, os.O_RDWR | os.O_BINARY)
+        stream = os.fdopen(fd, "r+b", buffering=0)
+
+        def read_line() -> str:
+            buffer = bytearray()
+            deadline = time.monotonic() + CONNECTION_TIMEOUT_SECONDS
+            while b"\n" not in buffer:
+                available = wintypes.DWORD(0)
+                kernel32.PeekNamedPipe(
+                    wintypes.HANDLE(handle), None, 0, None, ctypes.byref(available), None
+                )
+                if available.value:
+                    chunk = stream.read(available.value)
+                    if not chunk:
+                        break
+                    buffer += chunk
+                    continue
+                if time.monotonic() > deadline:
+                    break
+                time.sleep(0.005)
+            return bytes(buffer).split(b"\n", 1)[0].decode("utf-8", "replace")
+
+        def write_line(text: str) -> None:
+            stream.write(text.encode("utf-8") + b"\n")
+
+        # 关闭 fd 即关闭 handle（`open_osfhandle` 把它交给了 fd）。
+        _handle_connection(decider, read_line, write_line, stream.close)
+
+    # 实例先建出来再报就绪：宿主看到标记时端点必须已经能被连上。
+    first = create_instance()
+    if first == invalid_handle:
+        raise SystemExit(f"创建命名管道失败：{ctypes.get_last_error()}（{endpoint}）")
+    print(f"{READY_MARKER} endpoint={endpoint}", flush=True)
+
+    pending = first
+    while True:
+        handle = pending
+        if not kernel32.ConnectNamedPipe(wintypes.HANDLE(handle), None):
+            error = ctypes.get_last_error()
+            # ERROR_PIPE_CONNECTED 说明客户端已经连上了这个实例，照旧交给服务线程。
+            if error != error_pipe_connected:
+                kernel32.CloseHandle(wintypes.HANDLE(handle))
+                pending = create_instance()
+                continue
+        # 一个实例只服务一个连接；下一个实例提前建好，客户端不会撞到「管道不存在」。
+        pending = create_instance()
+        threading.Thread(target=serve_instance, args=(handle,), daemon=True).start()
+
+
+# --------------------------------------------------------------------------
+# Unix：Unix domain socket（没有 TCP 端口，也没有 HTTP 中转）
+# --------------------------------------------------------------------------
+
+
+def unix_socket_io(endpoint: str, decider: Decider) -> None:
+    path = Path(endpoint)
+    if path.exists():
+        path.unlink()
+    server = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+    server.bind(endpoint)
+    server.listen(64)
+    os.chmod(endpoint, 0o600)
+    print(f"{READY_MARKER} endpoint={endpoint}", flush=True)
+
+    def serve_connection(connection: socket.socket) -> None:
+        stream = connection.makefile("rwb", buffering=0)
+
+        def read_line() -> str:
+            connection.settimeout(CONNECTION_TIMEOUT_SECONDS)
+            line = stream.readline()
+            return line.decode("utf-8", "replace").rstrip("\n") if line else ""
+
+        def write_line(text: str) -> None:
+            stream.write(text.encode("utf-8") + b"\n")
+
+        def close() -> None:
+            try:
+                stream.close()
+            finally:
+                connection.close()
+
+        _handle_connection(decider, read_line, write_line, close)
+
+    while True:
+        connection, _ = server.accept()
+        threading.Thread(target=serve_connection, args=(connection,), daemon=True).start()
+
+
+# --------------------------------------------------------------------------
+# 引擎装配（与 `qev serve` 同口径，但不启 HTTP 服务）
+# --------------------------------------------------------------------------
+
+
+def build_engine(args: argparse.Namespace):
+    from qev.calibrate import Calibration
+    from qev.mm_engine import local_model_dir
+
+    model_path = local_model_dir(args.model)
+    local = Path(model_path) / "calibration.json"
+    calib_path = args.calibration or (str(local) if local.exists() else None)
+    calibration = Calibration.load(calib_path) if calib_path else None
+    if calib_path:
+        log.info("calibration: %s", calib_path)
+
+    config = json.loads((Path(model_path) / "config.json").read_text())
+    if args.multimodal or config.get("vision_config") is not None:
+        from qev.mm_engine import MMDecisionEngine
+
+        return MMDecisionEngine(
+            model_path,
+            device=args.device,
+            dtype=args.dtype,
+            calibration=calibration,
+            fork_mode=args.fork_mode,
+            head_dtype=args.head_dtype,
+            max_branch_tokens=args.max_branch_tokens,
+            max_request_tokens=args.max_request_tokens,
+            gpu_preprocess=not args.no_gpu_preprocess,
+            cuda_graphs=not args.no_cuda_graphs,
+        )
+    from qev.engine import DecisionEngine
+
+    return DecisionEngine(
+        model_path,
+        device=args.device,
+        dtype=args.dtype,
+        calibration=calibration,
+        fork_mode=args.fork_mode,
+        head_dtype=args.head_dtype,
+        max_branch_tokens=args.max_branch_tokens,
+        max_request_tokens=args.max_request_tokens,
+    )
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(prog="onejev_worker")
+    parser.add_argument("--model", required=True, help="本地权重目录（或 Hub id）")
+    parser.add_argument("--name", default=None, help="对外的模型名")
+    parser.add_argument("--endpoint", required=True, help="命名管道名或 Unix socket 路径")
+    parser.add_argument("--device", default="cuda")
+    parser.add_argument("--dtype", default="bfloat16")
+    parser.add_argument("--head-dtype", default="float32")
+    parser.add_argument("--fork-mode", default="auto")
+    parser.add_argument("--multimodal", action="store_true")
+    parser.add_argument("--no-cuda-graphs", action="store_true")
+    parser.add_argument("--no-gpu-preprocess", action="store_true")
+    parser.add_argument("--calibration", default=None)
+    parser.add_argument("--max-branch-tokens", type=int, default=32768)
+    parser.add_argument("--max-request-tokens", type=int, default=65536)
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO, format="%(asctime)s %(name)s %(levelname)s %(message)s")
+
+    engine = build_engine(args)
+    decider = Decider(engine, args.name or Path(args.model).name)
+
+    if sys.platform == "win32":
+        windows_pipe_io(args.endpoint, decider)
+    else:
+        unix_socket_io(args.endpoint, decider)
+
+
+if __name__ == "__main__":
+    main()

package/host/rust/assets/templates/system_prompt.md

@@ -24,6 +24,10 @@ OmniCrawl 是一个能够长期处理本地项目任务的 Agent。
 - 写入前先搜索；填写 frontmatter（title、created、updated、project、tags、type note|meeting|decision|log|research|reference、status draft|done|archived）。文件应存放在 `projects/<project>/`、`topics/<topic>/` 或 `daily/YYYY/MM/` 下；绝不手动编辑 `INDEX.md`。
 - 不得写入凭据或个人敏感信息；笔记必须是 UTF-8 Markdown。不要使用知识库存储会话进度、项目技术事实或稳定用户偏好；不要将知识库与 `omnicrawl://docs/` 或工作区文件混淆。
 
+Skill 使用协议：
+- Skill 目录不注入上下文：需要时先用 `skill_search` 检索索引（`query` 写任务关键词，`scope` 取 all/project/user/enterprise/path 收窄范围），命中后用 `read` 打开结果里的 `location` 读正文，再按其中的流程执行。
+- Skill 目录常在 `~/.omnicrawl/skills/` 或 `.omnicrawl/skills/`（工作区之外），不要用 `find`、`grep` 去找 Skill。用户用 `/skill:<名>` 指定时，该 Skill 正文已在上下文里。
+
 Todo 规划协议：
 - 对于任何多步骤项目任务，先调用 `update_todos`，提交简洁、有序且具体的步骤列表，然后再检查或编辑文件。每完成一个重要步骤，就重新提交列表并将对应步骤的 `completed` 标记为 true；提交空列表表示清除计划。不要只用文字计划替代该工具调用。
 
@@ -57,7 +61,7 @@ omnicrawl文档（共 14 篇，均以 `omnicrawl://docs/<文件名>` 读取，
 - 顾问策略（零参数 `advisor` 工具、`[advisor]` 配置、`/advisor` 命令、设置面板「顾问设置」）：读取 `omnicrawl://docs/advisor_design.md`；实现位于 `rust/crates/omnicrawl-controllers/src/advisor.rs` 和 `rust/crates/omnicrawl-config/src/features/advisor.rs`。
 - 工具输出压缩（外接小模型压缩工具结果、`[tool_output_compression]` 配置、设置面板「工具输出压缩」）：读取 `omnicrawl://docs/tool_output_compression_design.md`；实现位于 `rust/crates/omnicrawl-controllers/src/compression.rs`。
 - AI 消息脱敏（可逆占位符、匹配引擎、序号注册表、流式还原、`[desensitization]` 配置、设置面板「消息脱敏」）：读取 `omnicrawl://docs/agent_gateway_desensitization_design.md`；实现位于 `rust/crates/omnicrawl-llm/src/desensitization/` 和 `rust/crates/omnicrawl-config/src/features/desensitization.rs`。
-- TTS 语音合成（MOSS-TTS-Nano ONNX CPU：配置、tts_synthesize 工具、CLI、语音克隆、模型下载）：读取 `omnicrawl://docs/TTS.md`；实现位于 `rust/crates/omnicrawl-tts/` 和 `rust/crates/omnicrawl-config/src/features/tts.rs`。
+- TTS 语音合成（Kokoro-82M ONNX CPU：配置、tts_synthesize 工具、CLI、音素化、模型下载）：读取 `omnicrawl://docs/TTS.md`；实现位于 `rust/crates/omnicrawl-tts/` 和 `rust/crates/omnicrawl-config/src/features/tts.rs`。
 - 决策接口（结构化决策模型的本地 REST 服务：启动、无鉴权的回环访问、decide/choice/rank/review 端点、常驻与复用、脱敏与安全边界）：读取 `omnicrawl://docs/decision_api.md`；实现位于 `rust/crates/omnicrawl-decision/` 和 `rust/crates/omnicrawl-config/src/features/decision_model.rs`。
 - FSAPP 接入（文件系统应用侧的接入约定）：读取 `omnicrawl://docs/FSAPP.md`。
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@omnicrawl/cli-linux-x64",
-  "version": "0.2.8",
+  "version": "0.3.0",
   "description": "OmniCrawl 平台包（Linux x64：内核 + 完整宿主）",
   "repository": {
     "type": "git",

package/version.json

@@ -1,10 +1,10 @@
 {
   "name": "@omnicrawl/cli-linux-x64",
-  "version": "0.2.8",
+  "version": "0.3.0",
   "os": "linux",
   "cpu": "x64",
   "kernelVersion": "0.0.1",
-  "hostVersion": "0.2.8",
+  "hostVersion": "0.3.0",
   "hostPlatform": "linux-x64",
-  "builtAt": "2026-10-07T18:36:14.676Z"
+  "builtAt": "2026-10-18T23:17:22.646Z"
 }