codebee 0.1.71 → 0.1.72

package/app/core/market.py

@@ -120,6 +120,31 @@ def _is_market_file(path, pack_id):
     return "source: market" in head and ("market_id: %s" % pack_id) in head
 
 
+def _pack_digest(files):
+    """包内容指纹（sha256 前 16）：升级突变对账用——内容一致则同值。"""
+    import hashlib
+    blob = json.dumps({k: files[k] for k in sorted(files)}, ensure_ascii=False)
+    return hashlib.sha256(blob.encode("utf-8")).hexdigest()[:16]
+
+
+def _similar_names(name, installed, cutoff=0.82):
+    """typosquatting 特征近似名对账（difflib 序列相似度）：与已装技能名过近的
+    返回 [(已装名, 相似度)]。降本求准——cutoff 以上才算「过近」。"""
+    import difflib
+    low = str(name or "").strip().lower()
+    if not low:
+        return []
+    out = []
+    for pid, rec in (installed or {}).items():
+        prev = str((rec or {}).get("name") or "").strip().lower()
+        if not prev or prev == low:
+            continue
+        ratio = difflib.SequenceMatcher(None, low, prev).ratio()
+        if ratio >= cutoff:
+            out.append(((rec or {}).get("name") or pid, round(ratio, 2)))
+    return sorted(out, key=lambda x: -x[1])[:3]
+
+
 # ---------------------------------------------------------------- 安装状态（持久化）
 
 def _load_registry():
@@ -260,6 +285,14 @@ def install_files(pack_id, name, files, extra=None):
         udir = _user_pack_dir()
         reg = _load_registry()
         installed = reg.setdefault("installed", {})
+        # 跨卸载的内容记忆（AST10 #2 供应链投毒对账）：同名/同 id 包重装时
+        # 与上次见过的内容比对，突变即警示——卸载重装也骗不过对账。
+        seen = reg.setdefault("seen", {})
+        digest = _pack_digest(files)
+        prev_digest = str(seen.get(pack_id) or "")
+        changed_vs_seen = bool(prev_digest and prev_digest != digest)
+        # typosquatting 特征（AST10 #2）：与已装技能名过近的新包点名
+        near = _similar_names(name, installed)
         rec = installed.get(pack_id)
         primary = "market-%s.md" % pack_id
         target = (rec or {}).get("file") or ""
@@ -283,13 +316,24 @@ def install_files(pack_id, name, files, extra=None):
                     written.append(name_i)
             except OSError as e:
                 return None, "写入用户技能库失败: %s" % e
-            record = {"file": target, "files": written, "installed_at": _now()}
+            record = {"file": target, "files": written, "installed_at": _now(),
+                      "digest": digest, "name": str(name or "")[:60]}
             if scan_note:
                 record["scan"] = scan_note   # 危险模式扫描结果随包记账
             if extra:
                 record.update(extra)
             installed[pack_id] = record
+        else:
+            installed[pack_id]["digest"] = digest
+        seen[pack_id] = digest        # seen 永远跟最新一次见到对齐
         _save_registry(reg)
+    warnings = []
+    if changed_vs_seen:
+        warnings.append("⚠ 内容较上次安装发生变化（内容指纹 %s → %s），"
+                        "建议核对diff后再启用" % (prev_digest, digest))
+    for near_name, ratio in near:
+        warnings.append("⚠ 与已装技能「%s」名称过近（相似度 %.0f%%），"
+                        "注意甄别仿冒（typosquatting）" % (near_name, ratio * 100))
     # 装后冒烟验证（SkillForge 证据驱动借鉴）：装完重新走 skills 的解析链，
     # 证明该包能被加载、名字对得上、正文非空——写盘成功 ≠ 技能可用，
     # frontmatter 缺失/正文为空都能在这里当场暴露，而不是下次任务注入时静默丢失。
@@ -317,7 +361,49 @@ def install_files(pack_id, name, files, extra=None):
             reg["installed"][pack_id]["smoke"] = smoke
             _save_registry(reg)
     return {"ok": True, "id": pack_id, "name": name, "file": target,
-            "already": already, "scan": scan_note, "smoke": smoke}, None
+            "already": already, "scan": scan_note, "smoke": smoke,
+            "digest": digest, "changed": changed_vs_seen,
+            "warnings": warnings}, None
+
+
+def skill_card(pack_id):
+    """技能行为卡（ClawHub Skill Card 模式）：安装时把「这个技能声明做什么 /
+    从哪来 / 指纹是什么 / 安检结论」拼成一张卡。经验库与市场卡共用这一份
+    真源——行为可见性是投毒防线的最后一环（装了什么必须随时可查）。
+
+    返回 dict 或 None（非市场安装）。"""
+    with _LOCK:
+        rec = dict((_load_registry().get("installed") or {}).get(pack_id) or {})
+    if not rec:
+        return None
+    body_head = ""
+    target = rec.get("file") or ""
+    if target:
+        try:
+            raw = (_user_pack_dir() / target).read_text(encoding="utf-8",
+                                                        errors="replace")
+            meta, body = skills._parse_frontmatter(raw)
+            rec.setdefault("name", meta.get("name") or pack_id)
+            scopes = meta.get("scopes")
+            if isinstance(scopes, list):
+                rec.setdefault("scopes", scopes)
+            body_head = body.strip()[:200]
+        except Exception:
+            body_head = ""
+    remote = rec.get("remote") or {}
+    return {
+        "id": pack_id,
+        "name": rec.get("name") or pack_id,
+        "declares": body_head,                       # 技能正文开头=自述行为
+        "scopes": rec.get("scopes") or [],
+        "digest": rec.get("digest") or "",
+        "scan": rec.get("scan") or "",               # 装前安检结论（含风险分级）
+        "smoke": rec.get("smoke") or "",
+        "installed_at": rec.get("installed_at") or "",
+        "source": {"market": "内置市场" if not remote else str(remote.get("name") or ""),
+                   "version": str(remote.get("version") or ""),
+                   "homepage": str(remote.get("homepage") or "")},
+    }
 
 
 def remove(pack_id):

package/app/core/market_remote.py

@@ -272,6 +272,9 @@ def _normalize(source_meta, entry):
         "version": str(entry.get("version") or "").strip(),
         "homepage": str(entry.get("homepage") or "").strip(),
         "keywords": [str(k).lower() for k in (entry.get("keywords") or []) if isinstance(k, str)],
+        # 信任卡元数据（ClawTrust 思路）：来源自报的下载量——自报数据只作
+        # 热度参考，不作安全结论
+        "downloads": str(((entry.get("stats") or {}).get("downloads")) or ""),
         "source_id": source_meta["id"],
         "source_name": source_meta["name"],
         "install": install,
@@ -325,6 +328,8 @@ def _entries_from_cache(cache=None):
     cache = cache if cache is not None else _load_cache()
     src_meta = {s["id"]: s for s in SOURCES}
     installed = market.installed_ids()
+    # 信任卡：已装包的安检摘要（真注册表，非 id 集合）
+    registry = market._load_registry().get("installed") or {}
     entries = []
     for sid in [s["id"] for s in SOURCES]:
         c = cache.get(sid)
@@ -339,6 +344,7 @@ def _entries_from_cache(cache=None):
             e["block_reason"] = _compat_block(e) or ""
             e["installed"] = e["id"] in installed
             e["installable"] = e["compat"] == "ok"
+            e["scan_note"] = str((registry.get(e["id"]) or {}).get("scan") or "")
             entries.append(e)
     entries.sort(key=lambda x: (x["source_id"], x["name"].lower()))
     return entries
@@ -1048,6 +1054,15 @@ def preview_remote(entry_id):
             risk_label = skill_scan.risk_label(risk_findings)
         except Exception:
             risk_findings, risk_label = [], ""
+        # 供应链对账（AST10 #2）：与上次见过的内容/已装名比对（只读不写）
+        try:
+            digest = market._pack_digest(files)
+            reg = market._load_registry()
+            changed = str((reg.get("seen") or {}).get(entry_id) or "") not in ("", digest)
+            near = market._similar_names(entry.get("title") or "",
+                                         reg.get("installed") or {})
+        except Exception:
+            digest, changed, near = "", False, []
         return {
             "token": token, "id": entry_id, "title": entry["title"],
             "source": entry["source_name"], "version": entry["version"],
@@ -1058,6 +1073,7 @@ def preview_remote(entry_id):
             "file_list": [{"path": k, "chars": len(v)} for k, v in file_list[:60]],
             "stripped": sorted(stripped)[:40], "stripped_total": len(stripped),
             "risk_label": risk_label, "risk_findings": risk_findings[:8],
+            "digest": digest, "changed": changed, "near_names": near,
             "expires_in": _PREVIEW_TTL,
         }, None
     except ValueError as e:

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"),
+    }

package/app/core/modelhub.py

@@ -2414,8 +2414,9 @@ def _wire_endpoint_compatible(proto, base_url):
 def bindable_protocols(agent_kind_or_id):
     """该 CLI 可绑定的 wire 协议（与 resolve_binding 的 allowed 一致）。
     供死链告警/失败文案解释「为什么绑不上」：claude 只认 anthropic，
-    codex/dsh 只认 openai，其余开放双协议（含 wire 适配）。"""
-    if _deepseek_env_target(agent_kind_or_id) or agent_kind_or_id in ("codex-cli", "codex"):
+    codex/dsh/qwen 只认 openai，其余开放双协议（含 wire 适配）。"""
+    if (_deepseek_env_target(agent_kind_or_id)
+            or agent_kind_or_id in ("codex-cli", "codex", "qwencode", "qwen")):
         return ("openai",)
     if agent_kind_or_id in ("claude-code", "claude"):
         return ("anthropic",)