codebee 0.1.65 → 0.1.67

package/app/core/market_remote.py

@@ -24,16 +24,17 @@ import os
 import ipaddress
 import json
 import re
+import secrets
 import shutil
 import socket
 import subprocess
 import tarfile
 import tempfile
 import time
-import urllib.error
-import urllib.request
-import urllib.parse
-import zipfile
+import urllib.error
+import urllib.request
+import urllib.parse
+import zipfile
 from pathlib import Path
 from urllib.parse import urlparse
 
@@ -94,8 +95,8 @@ _CAP_DOWNLOAD = 80 * 1024 * 1024
 _CAP_UNPACKED = 120 * 1024 * 1024
 _CAP_FILES = 500
 _CAP_FILE_TEXT = 512 * 1024
-_CAP_TOTAL_TEXT = 4 * 1024 * 1024
-_MAX_REDIRECTS = 3
+_CAP_TOTAL_TEXT = 4 * 1024 * 1024
+_MAX_REDIRECTS = 3
 
 
 def _cache_dir():
@@ -132,52 +133,52 @@ def assert_public_url(url):
     return host, port
 
 
-def _fetch(url, cap=_CAP_MANIFEST):
+def _fetch(url, cap=_CAP_MANIFEST):
     """带体量上限的 https GET；返回 bytes。
 
     两段式网络策略：先走默认通道（Windows 上 urllib 会读注册表系统代理，
     依赖代理上网的环境靠它），失败再自动直连重试一次（实测本机代理对部分
     CDN 文件返回 404/篡改，直连正常；反过来需要代理的网络第一段就成功）。
     体量超限属确定性错误，不重试。"""
-    assert_public_url(url)
-
-    class _NoRedirect(urllib.request.HTTPRedirectHandler):
-        def redirect_request(self, req, fp, code, msg, headers, newurl):
-            return None
-
-    def _close_quietly(resp):
-        try:
-            resp.close()
-        except Exception:
-            pass
-
-    def _open(opener, target):
-        """逐跳 GET；重定向目标在发起下一跳前重新过 SSRF 网关。"""
-        current = target
-        for hop in range(_MAX_REDIRECTS + 1):
-            assert_public_url(current)
-            req = urllib.request.Request(current, headers={
-                "User-Agent": "CodeBee-Market/1.0",
-                "Accept": "*/*",
-            })
-            try:
-                resp = opener.open(req, timeout=30)
-                code = getattr(resp, "status", getattr(resp, "code", 200))
-            except urllib.error.HTTPError as e:
-                code, resp = e.code, e
-            if code not in (301, 302, 303, 307, 308):
-                if isinstance(resp, urllib.error.HTTPError):
-                    raise resp
-                return resp
-            location = resp.headers.get("Location")
-            if not location:
-                _close_quietly(resp)
-                raise ValueError("重定向缺少 Location: %s" % current)
-            _close_quietly(resp)
-            current = urllib.parse.urljoin(current, location)
-            if hop >= _MAX_REDIRECTS:
-                raise ValueError("重定向次数超过上限: %s" % target)
-        raise ValueError("重定向失败: %s" % target)
+    assert_public_url(url)
+
+    class _NoRedirect(urllib.request.HTTPRedirectHandler):
+        def redirect_request(self, req, fp, code, msg, headers, newurl):
+            return None
+
+    def _close_quietly(resp):
+        try:
+            resp.close()
+        except Exception:
+            pass
+
+    def _open(opener, target):
+        """逐跳 GET；重定向目标在发起下一跳前重新过 SSRF 网关。"""
+        current = target
+        for hop in range(_MAX_REDIRECTS + 1):
+            assert_public_url(current)
+            req = urllib.request.Request(current, headers={
+                "User-Agent": "CodeBee-Market/1.0",
+                "Accept": "*/*",
+            })
+            try:
+                resp = opener.open(req, timeout=30)
+                code = getattr(resp, "status", getattr(resp, "code", 200))
+            except urllib.error.HTTPError as e:
+                code, resp = e.code, e
+            if code not in (301, 302, 303, 307, 308):
+                if isinstance(resp, urllib.error.HTTPError):
+                    raise resp
+                return resp
+            location = resp.headers.get("Location")
+            if not location:
+                _close_quietly(resp)
+                raise ValueError("重定向缺少 Location: %s" % current)
+            _close_quietly(resp)
+            current = urllib.parse.urljoin(current, location)
+            if hop >= _MAX_REDIRECTS:
+                raise ValueError("重定向次数超过上限: %s" % target)
+        raise ValueError("重定向失败: %s" % target)
 
     def _read(open_fn):
         with open_fn() as resp:
@@ -192,17 +193,17 @@ def _fetch(url, cap=_CAP_MANIFEST):
                 chunks.append(chunk)
         return b"".join(chunks)
 
-    try:
-        # 默认 opener 仍读取系统代理，但显式禁用自动重定向。
-        opener = urllib.request.build_opener(
-            _NoRedirect(), urllib.request.HTTPSHandler(context=tlsctx.context()))
-        return _read(lambda: _open(opener, url))
-    except (urllib.error.HTTPError, urllib.error.URLError):
-        # 直连重试：ProxyHandler({}) 显式清空代理；每一跳仍经过 SSRF 校验。
-        opener = urllib.request.build_opener(
-            _NoRedirect(), urllib.request.ProxyHandler({}),
-            urllib.request.HTTPSHandler(context=tlsctx.context()))
-        return _read(lambda: _open(opener, url))
+    try:
+        # 默认 opener 仍读取系统代理，但显式禁用自动重定向。
+        opener = urllib.request.build_opener(
+            _NoRedirect(), urllib.request.HTTPSHandler(context=tlsctx.context()))
+        return _read(lambda: _open(opener, url))
+    except (urllib.error.HTTPError, urllib.error.URLError):
+        # 直连重试：ProxyHandler({}) 显式清空代理；每一跳仍经过 SSRF 校验。
+        opener = urllib.request.build_opener(
+            _NoRedirect(), urllib.request.ProxyHandler({}),
+            urllib.request.HTTPSHandler(context=tlsctx.context()))
+        return _read(lambda: _open(opener, url))
 
 
 # ---------------------------------------------------------------- 清单解析
@@ -343,11 +344,13 @@ def _entries_from_cache(cache=None):
     return entries
 
 
-def view(offset=0, limit=_PAGE_LIMIT, source=None, q=None):
+def view(offset=0, limit=_PAGE_LIMIT, source=None, q=None, installed=None):
     """外部目录视图（服务端过滤 + 分页）。缓存全在本地，对全量条目过滤再切页
     零成本——搜索/来源筛选不再受「已加载页」限制，滚动翻页逛完全部目录。
+    installed=真值时只看已安装（「已安装 N」入口的过滤器）；installed_total
+    始终是**未过滤前**的全量已装数（徽章计数与过滤条件无关）。
     返回 {sources, entries(本页), total(过滤后全量), offset, has_more,
-    categories}。"""
+    categories, installed_total}。"""
     cache = _load_cache()
     entries = _entries_from_cache(cache)
     sources = []
@@ -360,6 +363,9 @@ def view(offset=0, limit=_PAGE_LIMIT, source=None, q=None):
     for e in entries:
         if e["category"] and e["category"] not in cats:
             cats.append(e["category"])
+    installed_total = sum(1 for e in entries if e["installed"])
+    if installed:
+        entries = [e for e in entries if e["installed"]]
     if source:
         entries = [e for e in entries if e["source_id"] == source]
     qq = str(q or "").strip().lower()
@@ -372,7 +378,7 @@ def view(offset=0, limit=_PAGE_LIMIT, source=None, q=None):
     page = entries[offset:offset + limit]
     return {"sources": sources, "entries": page, "total": len(entries),
             "offset": offset, "has_more": offset + limit < len(entries),
-            "categories": cats}
+            "categories": cats, "installed_total": installed_total}
 
 
 def _fetch_clawhub_catalog():
@@ -659,7 +665,9 @@ def _safe_extract(zf, dest):
 
 
 def _download_zip(entry, tmp):
-    """下载 zip（校验 sha256）→ 安全解包 → 返回插件根目录。"""
+    """下载 zip（校验 sha256）→ 安全解包 → 返回 (插件根目录, 完整性说明)。
+    完整性：清单带 sha256 且校验通过 = verified；清单没带 = unverified
+    （靠 https 传输 + 解包后的纯技能白名单兜底，如实告知不装懂）。"""
     data = _fetch(entry["install"]["url"], cap=_CAP_DOWNLOAD)
     want = (entry["install"].get("sha256") or "").lower()
     if want:
@@ -667,7 +675,7 @@ def _download_zip(entry, tmp):
         if got != want:
             raise ValueError("sha256 校验不符（期望 %s，实际 %s）" % (want[:12], got[:12]))
     dest = _safe_extract(zipfile.ZipFile(io.BytesIO(data)), tmp / "unzip")
-    return _plugin_root(dest)
+    return _plugin_root(dest), ("verified" if want else "unverified")
 
 
 def _gh_split(url):
@@ -810,22 +818,22 @@ def _gh_fetch_files(entry, tmp):
     raise ValueError("jsdelivr 拉取失败: %s/%s（%s）" % (owner, repo, last_err))
 
 
-def _download_git(entry, tmp):
+def _download_git(entry, tmp):
     """git 浅克隆取子目录（Anthropic 生态的 git-subdir 来源）。"""
     if shutil.which("git") is None:
         raise ValueError("本机没有 git，无法安装 git-subdir 来源的插件")
-    inst = entry["install"]
-    # 非 GitHub 来源没有 codeload/jsdelivr 的安全收口，必须在 clone 前
-    # 复用同一公网 HTTPS 网关，避免 git 自己跟随重定向或访问内网。
-    assert_public_url(inst.get("url") or "")
+    inst = entry["install"]
+    # 非 GitHub 来源没有 codeload/jsdelivr 的安全收口，必须在 clone 前
+    # 复用同一公网 HTTPS 网关，避免 git 自己跟随重定向或访问内网。
+    assert_public_url(inst.get("url") or "")
     dst = tmp / "git"
     sub = str(inst.get("path") or "").replace("\\", "/")
     if sub.startswith("/") or ":" in sub or ".." in _rel_parts(sub):
         raise ValueError("插件子目录路径可疑: %s" % inst.get("path"))
-    # git 会自行跟随 HTTP 重定向，而重定向目标无法再经过 assert_public_url；
-    # ��装来源宁可明确失败，也不能让 clone 跳进内网。
-    cmd = ["git", "-c", "http.followRedirects=false", "clone",
-           "--depth", "1", "--single-branch", "--quiet"]
+    # git 会自行跟随 HTTP 重定向，而重定向目标无法再经过 assert_public_url；
+    # ��装来源宁可明确失败，也不能让 clone 跳进内网。
+    cmd = ["git", "-c", "http.followRedirects=false", "clone",
+           "--depth", "1", "--single-branch", "--quiet"]
     if inst.get("ref"):
         cmd += ["--branch", inst["ref"]]
     cmd += [inst["url"], str(dst)]
@@ -870,7 +878,7 @@ def _download_clawhub(entry, tmp):
         _CLAWHUB_BASE, q(inst["slug"], safe=""), q(inst["reference"] or inst["slug"], safe=""))
     data = _fetch(url, cap=_CAP_DOWNLOAD)
     dest = _safe_extract(zipfile.ZipFile(io.BytesIO(data)), tmp / "clawhub")
-    return _plugin_root(dest)
+    return _plugin_root(dest), "unverified"
 
 
 def _plugin_root(base):
@@ -961,9 +969,20 @@ def build_files(entry, root):
     return out, None, stripped
 
 
-def install_remote(entry_id):
-    """安装外部目录插件：下载 → 纯技能检查 → 转换 → 复用 market 安装通道。
-    返回 (结果 dict, 错误)。"""
+def _fetch_root(entry, tmp):
+    """按来源下载并解包，返回 (插件根目录, 完整性说明)。
+    verified=zip 带 sha256 且校验通过；unverified=直链无哈希（https+白名单兜底）；
+    git-ref/git-head=git 类来源按 ref/HEAD 取树。"""
+    kind = entry["install"]["kind"]
+    if kind == "zip":
+        return _download_zip(entry, tmp)
+    if kind == "clawhub":
+        return _download_clawhub(entry, tmp)
+    return _download_git_any(entry, tmp), \
+        ("git-ref" if entry["install"].get("ref") else "git-head")
+
+
+def _resolve_entry(entry_id):
     entries = _entries_from_cache()
     entry = next((e for e in entries if e["id"] == entry_id), None)
     if not entry:
@@ -972,35 +991,114 @@ def install_remote(entry_id):
         return None, entry["block_reason"] or "该插件不适配 CodeBee"
     if entry["install"]["kind"] == "unsupported":
         return None, "来源类型不支持（仅支持 zip 直链、git 子目录与 ClawHub）"
+    return entry, None
+
+
+# 预览→确认安装两阶段之间的已检内容缓存：token 一次性、TTL 10 分钟、最多 8 份
+# （files 是纯文本 dict、总量有 _CAP_TOTAL_TEXT 上限，缓存占用有界）。
+_PREVIEW_TTL = 600
+_PREVIEW_MAX = 8
+_PREVIEW_CACHE = {}
+
+
+def _preview_purge():
+    now = time.time()
+    for k in [k for k, v in _PREVIEW_CACHE.items() if now - v["ts"] > _PREVIEW_TTL]:
+        _PREVIEW_CACHE.pop(k, None)
+    while len(_PREVIEW_CACHE) > _PREVIEW_MAX:
+        _PREVIEW_CACHE.pop(min(_PREVIEW_CACHE, key=lambda k: _PREVIEW_CACHE[k]["ts"]))
+
+
+def _skill_names_from_files(files, entry):
+    """从转换后的 files 键推技能名：market-<id>.md 是主技能（用条目名），
+    market-<id>__<名>.md 是多技能包里的其余技能（用目录名）。"""
+    names = []
+    for k in files:
+        if "/" in k or not k.endswith(".md") or not k.startswith("market-"):
+            continue
+        base = k[len("market-"):-3]
+        names.append("__" in base and base.split("__", 1)[1] or entry["title"])
+    return sorted(set(names))
+
+
+def preview_remote(entry_id):
+    """安装前预览（先看后装）：下载 → 剥离式检查 → 转换，但**不落盘安装**。
+    返回 (预览 dict, 错误)。dict 带 token，确认安装时把 token 传回
+    install_remote 即安装这份已检内容，不重复下载。"""
+    entry, err = _resolve_entry(entry_id)
+    if err:
+        return None, err
     tmp = Path(tempfile.mkdtemp(prefix="codebee-mkt-"))
     try:
-        kind = entry["install"]["kind"]
-        if kind == "zip":
-            root = _download_zip(entry, tmp)
-        elif kind == "clawhub":
-            root = _download_clawhub(entry, tmp)
-        else:
-            root = _download_git_any(entry, tmp)
+        root, integrity = _fetch_root(entry, tmp)
         files, blocked, stripped = build_files(entry, root)
         if blocked:
             return None, blocked
-        res, err = market.install_files(entry_id, entry["title"], files, extra={
-            "remote": {"source": entry["source_id"], "name": entry["name"],
-                       "version": entry["version"], "homepage": entry["homepage"]},
-        })
-        if err:
-            return None, err
-        res["skills"] = sum(1 for k in files if k.endswith(".md") and "/" not in k)
-        res["stripped"] = sorted(stripped)
-        return res, None
+        _preview_purge()
+        token = secrets.token_hex(12)
+        _PREVIEW_CACHE[token] = {"entry_id": entry_id, "files": files,
+                                 "stripped": stripped, "ts": time.time()}
+        file_list = sorted(files.items(), key=lambda kv: kv[0])
+        return {
+            "token": token, "id": entry_id, "title": entry["title"],
+            "source": entry["source_name"], "version": entry["version"],
+            "homepage": entry["homepage"], "integrity": integrity,
+            "skills": _skill_names_from_files(files, entry),
+            "files_total": len(files),
+            "total_chars": sum(len(v) for v in files.values()),
+            "file_list": [{"path": k, "chars": len(v)} for k, v in file_list[:60]],
+            "stripped": sorted(stripped)[:40], "stripped_total": len(stripped),
+            "expires_in": _PREVIEW_TTL,
+        }, None
     except ValueError as e:
         return None, str(e)
-    except Exception as e:   # 网络/解包等意外错误也要兜成用户可读的一句话
+    except Exception as e:
         return None, "下载或解析插件失败: %s" % e
     finally:
         shutil.rmtree(tmp, ignore_errors=True)
 
 
+def install_remote(entry_id, token=""):
+    """安装外部目录插件：下载 → 纯技能检查 → 转换 → 复用 market 安装通道。
+    token 非空且命中预览缓存时直接安装那份已检内容（不重复下载）；token
+    一次性（装成功即焚，防重放）。返回 (结果 dict, 错误)。"""
+    entry, err = _resolve_entry(entry_id)
+    if err:
+        return None, err
+    cached = None
+    if token:
+        _preview_purge()
+        c = _PREVIEW_CACHE.get(str(token))
+        if c and c["entry_id"] == entry_id:
+            cached = c
+    if cached:
+        files, stripped = cached["files"], cached["stripped"]
+    else:
+        tmp = Path(tempfile.mkdtemp(prefix="codebee-mkt-"))
+        try:
+            root, _integrity = _fetch_root(entry, tmp)
+            files, blocked, stripped = build_files(entry, root)
+            if blocked:
+                return None, blocked
+        except ValueError as e:
+            return None, str(e)
+        except Exception as e:   # 网络/解包等意外错误也要兜成用户可读的一句话
+            return None, "下载或解析插件失败: %s" % e
+        finally:
+            shutil.rmtree(tmp, ignore_errors=True)
+    res, err = market.install_files(entry_id, entry["title"], files, extra={
+        "remote": {"source": entry["source_id"], "name": entry["name"],
+                   "version": entry["version"], "homepage": entry["homepage"]},
+    })
+    if err:
+        return None, err
+    if cached:
+        _PREVIEW_CACHE.pop(str(token), None)
+    res["skills"] = sum(1 for k in files if k.endswith(".md") and "/" not in k)
+    res["stripped"] = sorted(stripped)
+    return res, None
+
+
 def remove_remote(entry_id):
     """卸载外部插件：记账与标记齐全时走 market 既有卸载；否则报错。"""
     return market.remove(entry_id)

package/app/core/mocks.py

@@ -1,64 +1,64 @@
-# -*- coding: utf-8 -*-
-"""内置演示/测试智能体（mock）：不消耗任何配额即可跑通全流程。"""
-from __future__ import annotations
-
-import hashlib
-
-
-def _seed(agent_id):
-    return int(hashlib.sha256(agent_id.encode("utf-8")).hexdigest()[:6], 16)
-
-
-def draft_manuscript(task, round_no):
-    """生成一份示例稿件（mock 起草/修订直接落盘的内容）。"""
-    title = task.get("title") or "未命名章节"
-    goal = task.get("goal") or ""
-    return (
-        "# {title}\n\n"
-        "这是第 {r} 轮 mock 草稿（演示用，不消耗模型配额）。\n\n"
-        "## 任务目标\n{goal}\n\n"
-        "## 正文\n"
-        "夜色沿着山脊线缓慢退去，主角终于看清了石门上的刻痕——那不是文字，"
-        "而是一段被反复涂抹又反复显影的名字。他伸手触碰的瞬间，耳边响起的"
-        "却是自己昨夜说过的话。{extra}\n\n"
-        "（本轮修订关注：{focus}）\n"
-    ).format(
-        title=title, r=round_no, goal=goal,
-        extra="这一段在第 2 轮修订中补充了人物动机的伏笔。" if round_no >= 2 else "",
-        focus="节奏与悬念铺排" if round_no == 1 else "人物弧光与首尾呼应",
-    )
-
-
-def implement_note(task):
-    return "（mock 实现）已在目标目录写入 mock-impl.txt 作为变更占位。"
-
-
-def critique(agent_id, round_no, dims, threshold):
-    """确定性评分：第 1 轮略低于阈值，第 2 轮达标，用于验证发布门禁逻辑。"""
-    s = _seed(agent_id)
-    scores = {}
-    for i, d in enumerate(dims):
-        jitter = ((s >> i) % 5) * 0.1
-        scores[d] = round(min(9.5, 4.8 + 1.2 * round_no + jitter), 1)
-    passed = all(v >= threshold for v in scores.values())
-    issues = []
-    if round_no == 1:
-        issues = [
-            {"dim": dims[0], "severity": "major", "note": "开篇信息密度偏低，钩子出现太晚（mock 意见）"},
-            {"dim": dims[-1], "severity": "minor", "note": "段落长短交替不够，读感偏平（mock 意见）"},
-        ]
-    return {
-        "scores": scores,
-        "issues": issues,
-        "summary": "mock 总评：第 %d 轮稿件%s发布阈值" % (round_no, "已达到" if passed else "尚未达到"),
-    }
-
-
-def review(task, verify_passed):
-    return {
-        "pass": bool(verify_passed),
-        "scores": {"正确性": 8.0 if verify_passed else 5.0, "可维护性": 8.0, "安全": 8.5},
-        "issues": [] if verify_passed else [
-            {"severity": "major", "title": "验证命令未通过", "detail": "请检查测试输出（mock 意见）"}],
-        "summary": "mock 代码评审：%s" % ("通过" if verify_passed else "未通过"),
-    }
+# -*- coding: utf-8 -*-
+"""内置演示/测试智能体（mock）：不消耗任何配额即可跑通全流程。"""
+from __future__ import annotations
+
+import hashlib
+
+
+def _seed(agent_id):
+    return int(hashlib.sha256(agent_id.encode("utf-8")).hexdigest()[:6], 16)
+
+
+def draft_manuscript(task, round_no):
+    """生成一份示例稿件（mock 起草/修订直接落盘的内容）。"""
+    title = task.get("title") or "未命名章节"
+    goal = task.get("goal") or ""
+    return (
+        "# {title}\n\n"
+        "这是第 {r} 轮 mock 草稿（演示用，不消耗模型配额）。\n\n"
+        "## 任务目标\n{goal}\n\n"
+        "## 正文\n"
+        "夜色沿着山脊线缓慢退去，主角终于看清了石门上的刻痕——那不是文字，"
+        "而是一段被反复涂抹又反复显影的名字。他伸手触碰的瞬间，耳边响起的"
+        "却是自己昨夜说过的话。{extra}\n\n"
+        "（本轮修订关注：{focus}）\n"
+    ).format(
+        title=title, r=round_no, goal=goal,
+        extra="这一段在第 2 轮修订中补充了人物动机的伏笔。" if round_no >= 2 else "",
+        focus="节奏与悬念铺排" if round_no == 1 else "人物弧光与首尾呼应",
+    )
+
+
+def implement_note(task):
+    return "（mock 实现）已在目标目录写入 mock-impl.txt 作为变更占位。"
+
+
+def critique(agent_id, round_no, dims, threshold):
+    """确定性评分：第 1 轮略低于阈值，第 2 轮达标，用于验证发布门禁逻辑。"""
+    s = _seed(agent_id)
+    scores = {}
+    for i, d in enumerate(dims):
+        jitter = ((s >> i) % 5) * 0.1
+        scores[d] = round(min(9.5, 4.8 + 1.2 * round_no + jitter), 1)
+    passed = all(v >= threshold for v in scores.values())
+    issues = []
+    if round_no == 1:
+        issues = [
+            {"dim": dims[0], "severity": "major", "note": "开篇信息密度偏低，钩子出现太晚（mock 意见）"},
+            {"dim": dims[-1], "severity": "minor", "note": "段落长短交替不够，读感偏平（mock 意见）"},
+        ]
+    return {
+        "scores": scores,
+        "issues": issues,
+        "summary": "mock 总评：第 %d 轮稿件%s发布阈值" % (round_no, "已达到" if passed else "尚未达到"),
+    }
+
+
+def review(task, verify_passed):
+    return {
+        "pass": bool(verify_passed),
+        "scores": {"正确性": 8.0 if verify_passed else 5.0, "可维护性": 8.0, "安全": 8.5},
+        "issues": [] if verify_passed else [
+            {"severity": "major", "title": "验证命令未通过", "detail": "请检查测试输出（mock 意见）"}],
+        "summary": "mock 代码评审：%s" % ("通过" if verify_passed else "未通过"),
+    }