fix(desktop): fetch trimmed lark_oapi bundle instead of pip install

.github/workflows/publish-feishu-vendor.yml

@@ -0,0 +1,68 @@
+name: Build Feishu SDK Bundle
+
+# The desktop build ships without lark_oapi; clients download a trimmed bundle
+# the first time a user enables Feishu. This workflow builds that bundle
+# reproducibly and smoke-tests it.
+#
+# Run it only when VENDOR_VERSION in channel/feishu/lark_install.py changes:
+# the artifact is published once per SDK version and reused from then on.
+# The bundle is pure Python, so one artifact serves every platform.
+#
+# Publishing is a manual step -- download the artifact from this run and put it
+# on the CDN, then pin the printed sha256 in lark_install.py.
+
+on:
+  workflow_dispatch:
+  pull_request:
+    paths:
+      - "desktop/build/build-feishu-vendor.py"
+      - "channel/feishu/lark_install.py"
+      - ".github/workflows/publish-feishu-vendor.yml"
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        # Matches the Python the desktop backend is frozen with.
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Build and smoke-test the bundle
+        run: |
+          python -m pip install --upgrade pip
+          # The bundle expects these from the desktop build rather than
+          # carrying its own copies; the smoke test needs them too.
+          python -m pip install requests pycryptodome
+          python desktop/build/build-feishu-vendor.py --out dist/feishu-vendor.zip
+
+      - name: Check the pinned checksum matches
+        run: |
+          expected=$(python -c "import re,pathlib; \
+            print(re.search(r'VENDOR_SHA256 = \"([0-9a-f]*)\"', \
+            pathlib.Path('channel/feishu/lark_install.py').read_text()).group(1))")
+          actual=$(sha256sum dist/feishu-vendor.zip | cut -d' ' -f1)
+          echo "pinned in lark_install.py: ${expected:-<unset>}"
+          echo "built artifact           : $actual"
+          if [ -n "$expected" ] && [ "$expected" != "$actual" ]; then
+            echo "::error::VENDOR_SHA256 is stale. Set it to: $actual"
+            exit 1
+          fi
+
+      - name: Rename artifact to its published name
+        run: |
+          version=$(python -c "import re,pathlib; \
+            print(re.search(r'VENDOR_VERSION = \"([^\"]+)\"', \
+            pathlib.Path('channel/feishu/lark_install.py').read_text()).group(1))")
+          mv dist/feishu-vendor.zip "dist/feishu-vendor-$version.zip"
+
+      - name: Upload artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: feishu-vendor
+          path: dist/feishu-vendor-*.zip

.gitignore

@@ -54,6 +54,7 @@ desktop/build/*
 !desktop/build/cowagent-backend.spec
 !desktop/build/requirements-desktop.txt
 !desktop/build/build-backend.sh
+!desktop/build/build-feishu-vendor.py
 !desktop/build/entitlements.mac.plist
 !desktop/build/notarize-dmg.sh
 

.preview_secret

@@ -0,0 +1 @@
+9f77a3afd2e0488aa6e67cc7aa9c0772d12e1aac56814615ac6d4499c9f1c051

channel/feishu/feishu_channel.py

@@ -51,8 +51,8 @@ URL_VERIFICATION = "url_verification"
 
 # Lazy import of the lark_oapi SDK. The full `import lark_oapi` pulls in 10k+
 # files and takes 4-10s, so we defer the actual import to where it is needed.
-# In desktop mode the SDK is not bundled; the first import triggers an on-demand
-# pip install into a writable per-user dir (see channel/feishu/lark_install.py).
+# In desktop mode the SDK is not bundled; the first import downloads a trimmed
+# bundle into a writable per-user dir (see channel/feishu/lark_install.py).
 from channel.feishu import lark_install
 
 lark = None  # will be populated on first use via _ensure_lark_imported()
@@ -61,8 +61,8 @@ lark = None  # will be populated on first use via _ensure_lark_imported()
 def _ensure_lark_imported():
     """Import lark_oapi on first use (takes 4-10s due to 10k+ source files).
 
-    In desktop mode, if the SDK is missing this triggers an on-demand install
-    (requires network the first time) instead of failing outright.
+    In desktop mode, if the SDK is missing this fetches it on demand (requires
+    network the first time) instead of failing outright.
     """
     global lark
     if lark is None:

channel/feishu/lark_install.py

@@ -1,24 +1,37 @@
 """
-On-demand installer for the ``lark_oapi`` SDK (Feishu channel).
+On-demand provisioning of the ``lark_oapi`` SDK (Feishu channel).
 
 Why this exists
 ---------------
 The desktop client (PyInstaller build) intentionally does **not** bundle
-``lark_oapi`` (~116MB + dependencies) so the base download stays small
-(see zhayujie/CowAgent#2987 — "客户端没有飞书通道"). Instead, the first time a
-user enables the Feishu channel in desktop mode, we pip-install ``lark_oapi``
-(and its dependencies) into a writable per-user directory and add it to
-``sys.path``. The module is then imported lazily as before.
+``lark_oapi``: the published SDK is ~122MB unpacked because it ships models for
+all 59 Feishu open-platform domains, which is why the channel was dropped from
+the desktop build (see zhayujie/CowAgent#2987 — "客户端没有飞书通道").
 
-For source / non-desktop installs ``lark_oapi`` is expected to be present
-already (it lives in ``requirements.txt``); in that case this module is a
-no-op that simply imports it.
+Instead the first time a user enables Feishu in desktop mode we fetch a trimmed,
+pure-Python bundle (~1MB) built by ``desktop/build/build-feishu-vendor.py``,
+unpack it into a writable per-user directory and put that on ``sys.path``. The
+bundle is published once per SDK version and mirrored (see ``VENDOR_URLS``).
+
+Running ``pip`` at this point is not an option: the frozen app has no real
+Python interpreter — ``sys.executable`` is the bundled backend itself — so
+``sys.executable -m pip`` does not behave like pip. Unpacking a prebuilt archive
+needs nothing but the import machinery, which works normally in a frozen build.
+
+For source / non-desktop installs ``lark_oapi`` comes from ``requirements.txt``
+and this module simply confirms it is importable.
 """
+import hashlib
+import importlib
 import importlib.util
+import io
 import logging
 import os
-import subprocess
+import shutil
 import sys
+import tempfile
+import urllib.request
+import zipfile
 
 try:
     from common.log import logger
@@ -28,90 +41,167 @@ except Exception:  # pragma: no cover - common may be unavailable in isolation
 LARK_PKG = "lark-oapi"
 LARK_MIN = "1.5.5"
 
+# Bumped whenever a new bundle is published: "<lark version>-<bundle revision>".
+VENDOR_VERSION = "1.7.1-1"
+
+# sha256 of the published archive. The builder emits a byte-identical zip for a
+# given version, so this pin is stable; it is what makes the download safe to
+# trust. Leave empty only for local experiments (verification is then skipped
+# with a warning).
+VENDOR_SHA256 = "a96de70291e43b4829a5f717035806835f116bf4dc1d0a2d2ed551908a825381"
+
+# Tried in order; the same mirrors the desktop installers use. Feishu is a
+# China-only product, so the domestic CDN comes first. Any mirror is safe
+# because the payload is checked against VENDOR_SHA256.
+VENDOR_URLS = (
+    "https://cdn.link-ai.tech/desktop/vendor/feishu-vendor-{version}.zip",
+    "https://cdn.cowagent.ai/desktop/vendor/feishu-vendor-{version}.zip",
+)
+
+DOWNLOAD_TIMEOUT = 120
+
 # Per-user, writable, persistent location. Mirrors the browser tool's ~/.cow
 # layout so everything CowAgent owns lives under one roof.
 _VENDOR_SUBDIR = os.path.join(".cow", "feishu_vendor")
 
 
+def _install_error(reason: str = "") -> ImportError:
+    detail = f" ({reason})" if reason else ""
+    return ImportError(
+        f"lark_oapi is required for the Feishu channel but is not available{detail}. "
+        f"Install it with: pip install -U '{LARK_PKG}>={LARK_MIN}'"
+    )
+
+
 def vendor_dir() -> str:
-    """Return (and create) the directory where lark_oapi is installed on demand."""
+    """Return the directory the trimmed SDK is unpacked into.
+
+    Versioned so a newer bundle never has to overwrite a directory that a
+    running process may still be importing from.
+    """
     base = os.environ.get("COW_DATA_DIR") or os.path.expanduser("~")
-    path = os.path.join(base, _VENDOR_SUBDIR)
-    os.makedirs(path, exist_ok=True)
-    return path
+    return os.path.join(base, _VENDOR_SUBDIR, VENDOR_VERSION)
+
+
+def vendor_urls() -> list:
+    """Candidate download URLs, most preferred first."""
+    override = os.environ.get("COW_FEISHU_VENDOR_URL")
+    if override:
+        return [override]
+    return [url.format(version=VENDOR_VERSION) for url in VENDOR_URLS]
 
 
 def is_available() -> bool:
-    """True if lark_oapi can already be imported."""
+    """True if lark_oapi can be imported right now."""
     return importlib.util.find_spec("lark_oapi") is not None
 
 
-def ensure(allow_install: bool = True) -> None:
-    """Ensure ``lark_oapi`` can be imported.
+def _activate(path: str) -> bool:
+    """Put an unpacked bundle on sys.path and report whether it took effect."""
+    if not os.path.isdir(path):
+        return False
+    if path not in sys.path:
+        sys.path.insert(0, path)
+        # find_spec consults cached finders, which were populated before the
+        # directory existed.
+        importlib.invalidate_caches()
+    return is_available()
 
-    - If already importable, return immediately (source installs, or a previous
-      on-demand install that persisted).
-    - In desktop mode (``COW_DESKTOP=1``) and when ``allow_install`` is True,
-      pip-install ``lark_oapi`` (and its dependencies) on demand into
-      :func:`vendor_dir` and add that directory to ``sys.path``.
-    - Otherwise (non-desktop without the package, or install disabled) raise
-      ``ImportError`` with actionable guidance.
+
+def _fetch(url: str) -> bytes:
+    logger.info("[FeiShu] downloading Feishu SDK bundle from %s", url)
+    with urllib.request.urlopen(url, timeout=DOWNLOAD_TIMEOUT) as resp:
+        return resp.read()
+
+
+def _verify(payload: bytes) -> None:
+    if not VENDOR_SHA256:
+        logger.warning("[FeiShu] VENDOR_SHA256 is unset; skipping integrity check")
+        return
+    digest = hashlib.sha256(payload).hexdigest()
+    if digest != VENDOR_SHA256:
+        raise ValueError(f"checksum mismatch: expected {VENDOR_SHA256}, got {digest}")
+
+
+def _safe_extract(payload: bytes, dest: str) -> None:
+    """Unpack the archive, refusing entries that escape ``dest``."""
+    with zipfile.ZipFile(io.BytesIO(payload)) as zf:
+        for name in zf.namelist():
+            target = os.path.realpath(os.path.join(dest, name))
+            if not target.startswith(os.path.realpath(dest) + os.sep):
+                raise ValueError(f"refusing to extract outside the vendor dir: {name}")
+        zf.extractall(dest)
+
+
+def _download() -> bytes:
+    """Fetch the bundle, falling back through the mirrors on failure."""
+    last = None
+    for url in vendor_urls():
+        try:
+            payload = _fetch(url)
+            _verify(payload)
+            return payload
+        except Exception as exc:
+            last = exc
+            logger.warning("[FeiShu] could not fetch the bundle from %s: %s", url, exc)
+    raise last
+
+
+def _provision(target: str) -> None:
+    """Download and unpack the bundle into ``target`` atomically."""
+    parent = os.path.dirname(target)
+    os.makedirs(parent, exist_ok=True)
+    payload = _download()
+
+    staging = tempfile.mkdtemp(prefix=".incomplete-", dir=parent)
+    try:
+        _safe_extract(payload, staging)
+        try:
+            os.rename(staging, target)
+        except OSError:
+            # Another process won the race, or the platform refuses to rename
+            # onto an existing directory; its copy is equivalent.
+            if not os.path.isdir(target):
+                raise
+            shutil.rmtree(staging, ignore_errors=True)
+    except BaseException:
+        shutil.rmtree(staging, ignore_errors=True)
+        raise
+
+
+def ensure(allow_install: bool = True) -> None:
+    """Make ``lark_oapi`` importable, fetching the bundle if needed.
+
+    - Already importable (source install, or a previous download): return.
+    - Desktop mode with ``allow_install``: download the trimmed bundle and add
+      it to ``sys.path``.
+    - Otherwise: raise ``ImportError`` with actionable guidance.
     """
-    # Fast path: already importable.
     if is_available():
         return
 
+    target = vendor_dir()
+    # A previous run may have already unpacked it.
+    if _activate(target):
+        return
+
     if not allow_install:
-        raise ImportError(
-            "lark_oapi is required for the Feishu channel but is not installed. "
-            "Install it with: pip install -U 'lark-oapi>=%s'" % LARK_MIN
-        )
+        raise _install_error()
 
-    # Source / non-desktop installs should already have it via requirements.txt.
+    # Source / non-desktop installs get it from requirements.txt.
     if os.environ.get("COW_DESKTOP") != "1":
-        raise ImportError(
-            "lark_oapi is required for the Feishu channel but is not installed. "
-            "Install it with: pip install -U 'lark-oapi>=%s'" % LARK_MIN
-        )
-
-    # Desktop mode: install on demand into the user-writable vendor dir.
-    vdir = vendor_dir()
-    if vdir not in sys.path:
-        sys.path.insert(0, vdir)
-    # Re-check: a previous run may have already installed it here.
-    if is_available():
-        return
+        raise _install_error()
 
     logger.warning(
-        "[FeiShu] lark_oapi not found; installing on demand into %s "
-        "(requires network, ~%s+). This only happens once.",
-        vdir, LARK_MIN,
+        "[FeiShu] Feishu SDK not present; fetching the bundle into %s. "
+        "This happens once and needs a network connection.", target,
     )
     try:
-        subprocess.run(
-            [
-                sys.executable, "-m", "pip", "install",
-                "--target", vdir,
-                "--upgrade",
-                "lark-oapi>=%s" % LARK_MIN,
-            ],
-            check=True,
-            capture_output=True,
-            text=True,
-            timeout=600,
-        )
-    except Exception as exc:  # noqa: BLE001 - surface a clear, actionable error
-        logger.error("[FeiShu] on-demand lark_oapi install failed: %s", exc)
-        raise ImportError(
-            "Failed to install lark_oapi on demand. A network connection is "
-            "required the first time you enable Feishu. Please connect to the "
-            "internet and retry, or install it manually with: "
-            "pip install -U 'lark-oapi>=%s'" % LARK_MIN
-        ) from exc
+        _provision(target)
+    except Exception as exc:
+        logger.error("[FeiShu] failed to provision the Feishu SDK bundle: %s", exc)
+        raise _install_error(str(exc)) from exc
 
-    if not is_available():
-        raise ImportError(
-            "lark_oapi could not be imported after on-demand installation. "
-            "Try installing manually: pip install -U 'lark-oapi>=%s'" % LARK_MIN
-        )
-    logger.info("[FeiShu] lark_oapi installed on demand into %s", vdir)
+    if not _activate(target):
+        raise _install_error("bundle unpacked but lark_oapi is still not importable")
+    logger.info("[FeiShu] Feishu SDK bundle ready at %s", target)

channel/web/web_channel.py

@@ -4319,8 +4319,8 @@ class FeishuRegisterHandler:
 
         def _worker():
             try:
-                # Desktop builds don't bundle lark_oapi; install it on demand
-                # the first time the user enables Feishu (requires network).
+                # Desktop builds don't bundle lark_oapi; fetch it on demand the
+                # first time the user enables Feishu (requires network).
                 from channel.feishu import lark_install
                 lark_install.ensure(allow_install=True)
                 import lark_oapi as lark
@@ -4328,7 +4328,7 @@ class FeishuRegisterHandler:
                 with cls._lock:
                     cls._state["status"] = "error"
                     cls._state["error"] = (
-                        "lark-oapi SDK 未安装或安装失败，请联网后重试，"
+                        "飞书 SDK 不可用，请联网后重试，"
                         "或手动执行 pip install -U 'lark-oapi>=1.5.5'（%s）" % e
                     )
                 return

desktop/build/build-feishu-vendor.py

@@ -0,0 +1,276 @@
+#!/usr/bin/env python3
+"""Build the trimmed ``lark_oapi`` bundle downloaded on demand by the desktop app.
+
+The published SDK ships every Feishu open-platform domain (59 of them, ~122MB
+unpacked / ~32MB zipped). The Feishu channel only needs the websocket long
+connection, the event dispatcher and the ``im`` event models -- messages are
+sent through the REST API with plain ``requests`` -- so almost all of that is
+dead weight.
+
+This script installs the SDK with a pinned dependency set, strips it down to
+what the channel actually imports, and emits a pure-Python zip (~1MB) that works
+on every platform and Python version.
+
+Usage:
+    python desktop/build/build-feishu-vendor.py --out dist/feishu-vendor.zip
+
+Print the resulting sha256 into ``channel/feishu/lark_install.py`` when
+publishing a new bundle.
+"""
+import argparse
+import hashlib
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+import zipfile
+
+# Pinned because the trimming below patches SDK internals; bumping the SDK means
+# re-running this script and re-checking the smoke test.
+LARK_VERSION = "1.7.1"
+
+# The only API domain the channel imports (lark.im.v1.P2ImMessage*).
+KEEP_API = {"im"}
+
+# Installed with --no-deps and pinned exactly, because the sha256 recorded in
+# lark_install.py has to survive rebuilds. Letting pip resolve the tree instead
+# ties the artifact to the building interpreter (anyio only wants
+# typing_extensions below Python 3.13) and to whatever was released that day.
+#
+# requests, urllib3 and pycryptodome are deliberately missing: the desktop build
+# imports them directly, and pycryptodome is a native extension this bundle
+# cannot carry anyway. The rest travel along so the bundle also works outside a
+# frozen app -- inside one, PyInstaller's importer takes precedence over
+# sys.path, so the frozen copies keep winning and versions cannot drift apart.
+REQUIREMENTS = [
+    f"lark-oapi=={LARK_VERSION}",
+    "anyio==4.14.2",
+    "h11==0.16.0",
+    "httpcore==1.0.9",
+    "httpx==0.28.1",
+    "requests-toolbelt==1.0.0",
+    "typing_extensions==4.16.0",
+    "websockets==15.0.1",
+]
+
+
+def log(msg):
+    print(f"[feishu-vendor] {msg}", flush=True)
+
+
+def pip_install(target):
+    subprocess.run(
+        [sys.executable, "-m", "pip", "install", "--target", target,
+         "--no-compile", "--no-deps", *REQUIREMENTS],
+        check=True,
+    )
+
+
+def trim_api_domains(pkg):
+    """Delete every API domain except KEEP_API and rewrite the eager __init__."""
+    api_dir = os.path.join(pkg, "api")
+    removed = 0
+    for name in sorted(os.listdir(api_dir)):
+        path = os.path.join(api_dir, name)
+        if os.path.isdir(path) and name not in KEEP_API and name != "__pycache__":
+            shutil.rmtree(path)
+            removed += 1
+    with open(os.path.join(api_dir, "__init__.py"), "w", encoding="utf-8") as f:
+        for name in sorted(KEEP_API):
+            f.write(f"from . import {name}\n")
+    log(f"dropped {removed} api domains, kept {sorted(KEEP_API)}")
+
+
+def stub_client(pkg):
+    """Replace lark.Client, which wires up 57 services the channel never calls."""
+    with open(os.path.join(pkg, "client.py"), "w", encoding="utf-8") as f:
+        f.write(
+            '"""Stub: lark.Client is not part of the trimmed bundle."""\n\n\n'
+            "_MSG = (\n"
+            '    "lark_oapi.Client is not available in the trimmed Feishu bundle; "\n'
+            '    "call the Feishu REST API directly."\n'
+            ")\n\n\n"
+            "class Client:\n"
+            "    def __init__(self, *args, **kwargs):\n"
+            "        raise NotImplementedError(_MSG)\n\n"
+            "    @staticmethod\n"
+            "    def builder(*args, **kwargs):\n"
+            "        raise NotImplementedError(_MSG)\n"
+        )
+
+
+def trim_dispatcher(pkg):
+    """Keep only the processors of the surviving domains.
+
+    The builder declares one ``register_*`` method per event type, annotated
+    with model classes from every domain. Annotations are evaluated when the
+    class body runs, so the dropped names would raise NameError; PEP 563 defers
+    them to strings. The method bodies still reference dropped processors, but
+    they only run when called, which never happens for dropped domains.
+    """
+    path = os.path.join(pkg, "event", "dispatcher_handler.py")
+    with open(path, encoding="utf-8") as f:
+        lines = f.read().splitlines()
+
+    def keep(line):
+        m = re.match(r"from lark_oapi\.api\.(\w+)\.", line)
+        return m is None or m.group(1) in KEEP_API
+
+    kept = [l for l in lines if keep(l)]
+    with open(path, "w", encoding="utf-8") as f:
+        f.write("from __future__ import annotations\n" + "\n".join(kept) + "\n")
+
+
+def strip_binaries_and_caches(root):
+    """Remove __pycache__ and native extensions.
+
+    The only compiled module in the pinned set is the optional ``websockets``
+    speedup, which falls back to pure Python. Dropping it makes the bundle
+    independent of platform and Python version, so one artifact serves everyone.
+    """
+    removed = 0
+    for dirpath, dirnames, filenames in os.walk(root):
+        for d in list(dirnames):
+            if d == "__pycache__":
+                shutil.rmtree(os.path.join(dirpath, d))
+                dirnames.remove(d)
+        for fn in filenames:
+            if fn.endswith((".so", ".pyd", ".dylib")):
+                os.remove(os.path.join(dirpath, fn))
+                removed += 1
+    log(f"stripped {removed} native extension(s) -> pure Python")
+
+
+# Exercises every lark_oapi symbol channel/feishu/feishu_channel.py touches,
+# so trimming mistakes surface here instead of at runtime.
+SMOKE_TEST = r'''
+import json
+import lark_oapi as lark
+
+assert VENDOR in lark.__file__, "loaded the wrong lark_oapi: " + lark.__file__
+
+lark.im.v1.P2ImMessageReceiveV1
+lark.im.v1.P2ImMessageRecalledV1
+lark.JSON.marshal({"a": 1})
+lark.LogLevel.WARNING
+assert callable(lark.register_app)
+
+from lark_oapi.event.callback.model.p2_card_action_trigger import (
+    P2CardActionTriggerResponse,
+)
+P2CardActionTriggerResponse({"toast": {"type": "info", "content": "ok"}})
+
+handler = (
+    lark.EventDispatcherHandler.builder("", "")
+    .register_p2_im_message_receive_v1(lambda d: None)
+    .register_p2_im_message_recalled_v1(lambda d: None)
+    .register_p2_card_action_trigger(lambda d: None)
+    .build()
+)
+lark.ws.Client("id", "secret", event_handler=handler, log_level=lark.LogLevel.WARNING)
+
+import lark_oapi.ws.client as ws_mod
+assert hasattr(ws_mod, "loop")
+
+seen = {}
+(
+    lark.EventDispatcherHandler.builder("", "")
+    .register_p2_im_message_receive_v1(
+        lambda d: seen.update(json.loads(lark.JSON.marshal(d)))
+    )
+    .build()
+)._do_without_validation(json.dumps({
+    "schema": "2.0",
+    "header": {"event_id": "e", "event_type": "im.message.receive_v1", "token": "",
+               "create_time": "0", "tenant_key": "t", "app_id": "a"},
+    "event": {"sender": {"sender_id": {"open_id": "ou_x"}, "sender_type": "user"},
+              "message": {"message_id": "om_x", "chat_id": "oc_x", "chat_type": "p2p",
+                          "message_type": "text", "content": "{}", "create_time": "0"}},
+}).encode())
+assert seen["event"]["message"]["message_id"] == "om_x"
+print("OK")
+'''
+
+
+def verify(zip_path):
+    """Import the bundle in a subprocess and exercise the channel's API surface."""
+    workdir = tempfile.mkdtemp(prefix="feishu-vendor-check-")
+    try:
+        with zipfile.ZipFile(zip_path) as zf:
+            zf.extractall(workdir)
+        env = dict(os.environ, PYTHONPATH=workdir)
+        proc = subprocess.run(
+            [sys.executable, "-c", f"VENDOR = {workdir!r}\n" + SMOKE_TEST],
+            env=env, capture_output=True, text=True,
+        )
+        if proc.returncode != 0:
+            log("smoke test FAILED:\n" + proc.stdout + proc.stderr)
+            raise SystemExit(1)
+        log("smoke test passed")
+    finally:
+        shutil.rmtree(workdir, ignore_errors=True)
+
+
+def build(out_path):
+    staging = tempfile.mkdtemp(prefix="feishu-vendor-")
+    try:
+        log(f"installing lark-oapi=={LARK_VERSION}")
+        pip_install(staging)
+
+        pkg = os.path.join(staging, "lark_oapi")
+        trim_api_domains(pkg)
+        stub_client(pkg)
+        trim_dispatcher(pkg)
+
+        # Optional high-level layer, never imported by the channel.
+        shutil.rmtree(os.path.join(pkg, "channel"), ignore_errors=True)
+        # Metadata mostly indexes files that no longer exist after trimming.
+        for entry in os.listdir(staging):
+            if entry.endswith((".dist-info", ".egg-info")):
+                shutil.rmtree(os.path.join(staging, entry), ignore_errors=True)
+
+        # Console scripts installed alongside the packages; nothing imports them.
+        shutil.rmtree(os.path.join(staging, "bin"), ignore_errors=True)
+        strip_binaries_and_caches(staging)
+
+        os.makedirs(os.path.dirname(os.path.abspath(out_path)) or ".", exist_ok=True)
+        # Sorted entries and a fixed timestamp keep the archive byte-identical
+        # across rebuilds, so the sha256 pinned in lark_install.py stays valid.
+        members = []
+        for dirpath, _, filenames in os.walk(staging):
+            for fn in filenames:
+                full = os.path.join(dirpath, fn)
+                members.append((os.path.relpath(full, staging), full))
+        with zipfile.ZipFile(out_path, "w", zipfile.ZIP_DEFLATED) as zf:
+            for arcname, full in sorted(members):
+                info = zipfile.ZipInfo(arcname, date_time=(1980, 1, 1, 0, 0, 0))
+                info.compress_type = zipfile.ZIP_DEFLATED
+                info.external_attr = 0o644 << 16
+                with open(full, "rb") as fh:
+                    zf.writestr(info, fh.read())
+    finally:
+        shutil.rmtree(staging, ignore_errors=True)
+
+    with open(out_path, "rb") as fh:
+        digest = hashlib.sha256(fh.read()).hexdigest()
+    size_mb = os.path.getsize(out_path) / 1024 / 1024
+    log(f"wrote {out_path} ({size_mb:.1f}MB)")
+    log(f"sha256 = {digest}")
+    log("update VENDOR_SHA256 in channel/feishu/lark_install.py with the value above")
+
+
+def main():
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--out", default="dist/feishu-vendor.zip")
+    ap.add_argument("--skip-verify", action="store_true",
+                    help="skip the post-build smoke test")
+    args = ap.parse_args()
+    build(args.out)
+    if not args.skip_verify:
+        verify(args.out)
+
+
+if __name__ == "__main__":
+    main()

desktop/build/cowagent-backend.spec

@@ -29,10 +29,9 @@ def rp(*parts):
 # PyInstaller's static analysis can't see them. List every channel we ship.
 # Feishu (channel.feishu.feishu_channel) is bundled so the channel code is
 # present; its heavy lark_oapi SDK is deliberately NOT bundled — see the
-# `excludes` note below and channel/feishu/lark_install.py, which installs it
-# on demand the first time the user enables Feishu (fixes
-# zhayujie/CowAgent#2987: "客户端没有飞书通道"). pip/setuptools/wheel are
-# bundled (small) so that on-demand install can run inside the frozen app.
+# `excludes` note below and channel/feishu/lark_install.py, which downloads a
+# trimmed bundle the first time the user enables Feishu (fixes
+# zhayujie/CowAgent#2987: "客户端没有飞书通道").
 hiddenimports = [
     # channels (dynamic import in channel/channel_factory.py)
     'channel.web.web_channel',
@@ -48,12 +47,6 @@ hiddenimports = [
     'channel.telegram.telegram_channel',
     'channel.slack.slack_channel',
     'channel.discord.discord_channel',
-    # pip is used by the Feishu on-demand installer (channel/feishu/lark_install.py)
-    # to fetch lark_oapi + its deps inside the frozen desktop app.
-    'pip',
-    'pip._internal',
-    'setuptools',
-    'wheel',
 ]
 
 # Agent tools and model providers are imported lazily in places; collect their
@@ -162,17 +155,17 @@ datas += collect_data_files('playwright', include_py_files=True)
 # --- Excludes -------------------------------------------------------------
 # Keep the bundle lean: drop tests/docs and dev-only packages.
 #
-# Feishu's lark_oapi (~116MB + deps) is intentionally EXCLUDED here. The desktop
-# client does not bundle it; instead, the first time a user enables the Feishu
-# channel, channel/feishu/lark_install.py pip-installs lark_oapi (and its deps)
-# into a writable per-user dir (~/.cow/feishu_vendor) and adds it to sys.path.
-# This keeps the base download ~116MB smaller while Feishu still works out of
-# the box after a one-time, on-demand download (fixes zhayujie/CowAgent#2987).
-# pip/setuptools/wheel are NOT excluded (see hiddenimports) so that on-demand
-# install can run inside the frozen app.
+# Feishu's lark_oapi (~122MB unpacked) is intentionally EXCLUDED here. The first
+# time a user enables the Feishu channel, channel/feishu/lark_install.py
+# downloads a trimmed, pure-Python bundle (~1MB, built by
+# desktop/build/build-feishu-vendor.py) into ~/.cow/feishu_vendor and adds it to
+# sys.path. This keeps the base download small while Feishu still works out of
+# the box after a one-time fetch (fixes zhayujie/CowAgent#2987).
 excludes = [
-    'lark_oapi',          # Feishu SDK — installed on demand, NOT bundled (see above)
+    'lark_oapi',          # Feishu SDK — fetched on demand, NOT bundled (see above)
     'tests',
+    'pip',
+    'wheel',
     'pytest',
     # NOTE: playwright is now BUNDLED (pure-Python package + Node driver, ~10-15MB)
     # so the browser tool works out of the box on desktop. The heavy Chromium

desktop/build/requirements-desktop.txt

@@ -4,12 +4,11 @@
 # (which Electron talks to) plus the agent core; the remaining IM channels are
 # cheap (~27MB total) so we keep them.
 #
-# Feishu's `lark-oapi` (~116MB + deps) is NOT listed here on purpose: the desktop
-# client installs it on demand the first time the user enables the Feishu channel
-# (channel/feishu/lark_install.py), keeping the base download ~116MB smaller.
-# pip/setuptools/wheel are required by that on-demand installer and must be
-# present in the build/run environment (they ship inside the frozen app too — see
-# the spec's hiddenimports), so they stay available.
+# Feishu's `lark-oapi` (~122MB unpacked) is NOT listed here on purpose: the
+# desktop client downloads a trimmed ~1MB bundle the first time the user enables
+# the Feishu channel (channel/feishu/lark_install.py), keeping the base download
+# small. That bundle relies on `requests` and `pycryptodome` below rather than
+# shipping its own copies.
 
 # ---- core ----
 numpy>=1.21
@@ -53,7 +52,7 @@ python-pptx
 # Chrome/Edge, or downloads Chromium on demand into ~/.cow at first use.
 playwright==1.52.0
 
-# ---- IM channels (kept; lightweight). Feishu/lark-oapi is installed on demand
+# ---- IM channels (kept; lightweight). Feishu/lark-oapi is fetched on demand
 #      by the desktop client when the user enables the channel (see header note). ----
 wechatpy
 pycryptodome
@@ -62,7 +61,3 @@ websocket-client>=1.4.0
 python-telegram-bot
 slack_bolt
 discord.py
-# pip/setuptools/wheel back the Feishu on-demand installer; ensure they are present.
-pip
-setuptools
-wheel

tests/test_feishu_lark_install.py

@@ -1,92 +1,165 @@
-"""Tests for channel.feishu.lark_install (on-demand lark_oapi installer)."""
+"""Tests for channel.feishu.lark_install (on-demand Feishu SDK bundle)."""
+import hashlib
 import importlib.util
-import sys
-from unittest import mock
+import io
+import os
+import zipfile
 
 import pytest
 
 from channel.feishu import lark_install
 
 
+def _make_bundle():
+    """Build an in-memory zip shaped like a vendor bundle."""
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w") as zf:
+        zf.writestr("lark_oapi/__init__.py", b"# stub\n")
+    return buf.getvalue()
+
+
+def _unpacked(path):
+    """Stand-in for _activate: the bundle counts as usable once extracted."""
+    return os.path.isdir(os.path.join(path, "lark_oapi"))
+
+
 @pytest.fixture
-def clean_vendor(monkeypatch, tmp_path):
-    """Point the vendor dir at a temp location and drop any lark_oapi from path."""
+def offline(monkeypatch, tmp_path):
+    """Isolate the vendor dir and pretend lark_oapi is not installed."""
     monkeypatch.setenv("COW_DATA_DIR", str(tmp_path))
     monkeypatch.delenv("COW_DESKTOP", raising=False)
-    # Ensure lark_oapi is not importable in this test process.
     monkeypatch.setattr(lark_install, "is_available", lambda: False)
-    # Avoid real subprocess / real sys.path pollution.
-    monkeypatch.setattr(subprocess_run := lark_install.subprocess, "run", mock.MagicMock())
-    yield tmp_path
+    return tmp_path
 
 
-def test_is_available_true_when_spec_found(monkeypatch):
+def test_is_available_reflects_import_machinery(monkeypatch):
     monkeypatch.setattr(
-        importlib.util, "find_spec", lambda name: object() if name == "lark_oapi" else None
+        importlib.util, "find_spec",
+        lambda name: object() if name == "lark_oapi" else None,
     )
     assert lark_install.is_available() is True
-    assert lark_install.is_available() is False or True  # second call path-independent
+
+
+def test_vendor_dir_is_versioned(monkeypatch, tmp_path):
+    monkeypatch.setenv("COW_DATA_DIR", str(tmp_path))
+    assert lark_install.vendor_dir().endswith(lark_install.VENDOR_VERSION)
+
+
+def test_vendor_urls_prefer_the_china_mirror(monkeypatch):
+    monkeypatch.delenv("COW_FEISHU_VENDOR_URL", raising=False)
+    urls = lark_install.vendor_urls()
+    assert len(urls) > 1, "expected a fallback mirror"
+    assert "link-ai" in urls[0], "the China CDN should be tried first"
+    assert all(lark_install.VENDOR_VERSION in u for u in urls)
+
+
+def test_vendor_url_can_be_overridden(monkeypatch):
+    monkeypatch.setenv("COW_FEISHU_VENDOR_URL", "https://example.test/b.zip")
+    assert lark_install.vendor_urls() == ["https://example.test/b.zip"]
+
+
+def test_download_falls_back_to_next_mirror(monkeypatch):
+    payload = _make_bundle()
+    monkeypatch.setattr(lark_install, "VENDOR_SHA256",
+                        hashlib.sha256(payload).hexdigest())
+    tried = []
+
+    def flaky(url):
+        tried.append(url)
+        if len(tried) == 1:
+            raise OSError("mirror unreachable")
+        return payload
+
+    monkeypatch.setattr(lark_install, "_fetch", flaky)
+    assert lark_install._download() == payload
+    assert len(tried) == 2, "should have retried against the fallback mirror"
+
+
+def test_download_raises_when_every_mirror_fails(monkeypatch):
+    def dead(url):
+        raise OSError("unreachable")
+
+    monkeypatch.setattr(lark_install, "_fetch", dead)
+    with pytest.raises(OSError):
+        lark_install._download()
 
 
 def test_ensure_returns_when_already_available(monkeypatch):
     monkeypatch.setattr(lark_install, "is_available", lambda: True)
-    # Should not attempt any install.
-    with mock.patch.object(lark_install.subprocess, "run") as run:
-        lark_install.ensure(allow_install=True)
-    run.assert_not_called()
+    called = []
+    monkeypatch.setattr(lark_install, "_provision", lambda t: called.append(t))
+    lark_install.ensure(allow_install=True)
+    assert not called, "should not fetch anything when the SDK is importable"
 
 
-def test_ensure_non_desktop_raises(monkeypatch):
-    monkeypatch.setattr(lark_install, "is_available", lambda: False)
-    monkeypatch.delenv("COW_DESKTOP", raising=False)
+def test_ensure_non_desktop_raises(offline):
     with pytest.raises(ImportError):
         lark_install.ensure(allow_install=True)
 
 
-def test_ensure_allow_install_false_raises(monkeypatch):
-    monkeypatch.setattr(lark_install, "is_available", lambda: False)
+def test_ensure_allow_install_false_raises(offline, monkeypatch):
     monkeypatch.setenv("COW_DESKTOP", "1")
     with pytest.raises(ImportError):
         lark_install.ensure(allow_install=False)
 
 
-def test_ensure_desktop_installs_on_demand(monkeypatch, tmp_path):
-    # Simulate: not available before, available after install.
-    state = {"installed": False}
+def test_ensure_reuses_previously_unpacked_bundle(offline, monkeypatch):
+    """An existing vendor dir is activated without any download."""
+    os.makedirs(os.path.join(lark_install.vendor_dir(), "lark_oapi"))
 
-    def fake_available():
-        return state["installed"]
+    def unexpected(target):
+        raise AssertionError("should not have downloaded anything")
 
-    monkeypatch.setattr(lark_install, "is_available", fake_available)
-    monkeypatch.setenv("COW_DESKTOP", "1")
-    monkeypatch.setattr(lark_install, "vendor_dir", lambda: str(tmp_path))
-    calls = []
-
-    def fake_run(cmd, **kwargs):
-        calls.append(cmd)
-        state["installed"] = True  # pretend pip succeeded
-        return mock.MagicMock()
-
-    monkeypatch.setattr(lark_install.subprocess, "run", fake_run)
-
-    # Should not raise; install command must target the vendor dir.
+    monkeypatch.setattr(lark_install, "_provision", unexpected)
+    monkeypatch.setattr(lark_install, "_activate", _unpacked)
     lark_install.ensure(allow_install=True)
-    assert calls, "expected a pip install to have been attempted"
-    cmd = calls[0]
-    assert str(tmp_path) in cmd  # --target points at the vendor dir
-    assert "install" in cmd
-    assert any("lark-oapi" in str(part) for part in cmd)  # lark-oapi>=1.5.5 spec
 
 
-def test_ensure_desktop_install_failure_raises(monkeypatch, tmp_path):
-    monkeypatch.setattr(lark_install, "is_available", lambda: False)
+def test_ensure_downloads_and_unpacks(offline, monkeypatch):
     monkeypatch.setenv("COW_DESKTOP", "1")
-    monkeypatch.setattr(lark_install, "vendor_dir", lambda: str(tmp_path))
+    payload = _make_bundle()
+    monkeypatch.setattr(lark_install, "VENDOR_SHA256",
+                        hashlib.sha256(payload).hexdigest())
+    monkeypatch.setattr(lark_install, "_fetch", lambda url: payload)
+    monkeypatch.setattr(lark_install, "_activate", _unpacked)
+    target = lark_install.vendor_dir()
 
-    def fake_run(cmd, **kwargs):
-        raise RuntimeError("no network")
+    lark_install.ensure(allow_install=True)
 
-    monkeypatch.setattr(lark_install.subprocess, "run", fake_run)
+    assert os.path.isfile(os.path.join(target, "lark_oapi", "__init__.py"))
+    # No partially extracted staging dirs left behind.
+    leftovers = [d for d in os.listdir(os.path.dirname(target))
+                 if d.startswith(".incomplete-")]
+    assert not leftovers
+
+
+def test_ensure_rejects_checksum_mismatch(offline, monkeypatch):
+    monkeypatch.setenv("COW_DESKTOP", "1")
+    monkeypatch.setattr(lark_install, "VENDOR_SHA256", "0" * 64)
+    monkeypatch.setattr(lark_install, "_fetch", lambda url: _make_bundle())
 
     with pytest.raises(ImportError):
         lark_install.ensure(allow_install=True)
+    assert not os.path.isdir(lark_install.vendor_dir())
+
+
+def test_ensure_download_failure_raises(offline, monkeypatch):
+    monkeypatch.setenv("COW_DESKTOP", "1")
+
+    def boom(url):
+        raise OSError("no network")
+
+    monkeypatch.setattr(lark_install, "_fetch", boom)
+    with pytest.raises(ImportError):
+        lark_install.ensure(allow_install=True)
+
+
+def test_extract_rejects_path_traversal(tmp_path):
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w") as zf:
+        zf.writestr("../escaped.py", b"pwned\n")
+
+    dest = tmp_path / "dest"
+    dest.mkdir()
+    with pytest.raises(ValueError):
+        lark_install._safe_extract(buf.getvalue(), str(dest))