Skill flagged — suspicious patterns detected

ClawHub Security flagged this skill as suspicious. Review the scan results before using.

Huo15 Openclaw Enhance

v5.7.10

火一五·克劳德·龙虾增强插件 v5.7.8 — 全面适配 openclaw 2026.4.24:peerDep ^4.24 + build/compat 同步到 4.24 + 14 处 api.on 全部去掉 as any 改成 typed hook(hookName 联合类型 + handler 自动推断 Pl...

0· 335·1 current·1 all-time
byJob Zhao@zhaobod1

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for zhaobod1/huo15-openclaw-enhance.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "Huo15 Openclaw Enhance" (zhaobod1/huo15-openclaw-enhance) from ClawHub.
Skill page: https://clawhub.ai/zhaobod1/huo15-openclaw-enhance
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install huo15-openclaw-enhance

ClawHub CLI

Package manager switcher

npx clawhub@latest install huo15-openclaw-enhance
Security Scan
Capability signals
CryptoRequires sensitive credentials
These labels describe what authority the skill may exercise. They are separate from suspicious or malicious moderation verdicts.
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
medium confidence
Purpose & Capability
Name/description describe an OpenClaw enhancement plugin; code and SKILL.md implement hooks, local filesystem scans (~/ .openclaw, ~/.claude, agent workspaces), session/search tools, and skill-recommender behavior that are coherent with that purpose. No unrelated cloud credentials or unrelated binaries are requested.
Instruction Scope
SKILL.md and docs include developer-run steps (decompiling Claude.app, npm publish examples) and many read-only operations (scanning skill frontmatter, sessions JSONL, package.json of installed plugins, .git/config). Those read operations are consistent with features (skill-recommender, config-doctor, transcript-search), but they are privacy-sensitive (scan many user paths). The SKILL.md also contains large development instructions and examples which are not runtime agent actions but could influence maintainers; the pre-scan found unicode-control-chars inside SKILL.md (prompt-injection pattern) which may be used to try to influence automated processing — not expected for core plugin runtime and worth caution.
Install Mechanism
No install spec supplied (instruction-only metadata). There are many source files in the package (JS/TS, shell scripts) but no download-from-arbitrary-URL or remote installer in the metadata. The presence of scripts/setup.sh and publish commands in docs is normal for a developer repo but the package manifest does not declare an automatic external installer. Extra care: these scripts are present but not automatically run by platform install according to metadata.
Credentials
The skill declares no required environment variables, no credentials, and no required binaries. README and PLAN mention publishing commands that use tokens (npm/CLAWHUB) but the plugin does not require them at runtime. The plugin reads local configuration and files under user directories, which is proportional to features like skill discovery, config-doctor, and transcript search.
Persistence & Privilege
The plugin's manifest sets enabledByDefault: true, so when installed it will be enabled by default (not the same as system-wide 'always:true'). The package does not request always:true. It registers persistent hooks and writes to an enhance SQLite DB (its own storage), which is in line with its functionality but means it will keep local state (memories, logs). Consider that enabledByDefault gives immediate active presence upon install.
Scan Findings in Context
[unicode-control-chars] unexpected: The SKILL.md contains Unicode control characters detected by the scanner (a prompt-injection pattern). This is not necessary for the plugin's runtime features and could be an attempt to influence automated parsing or evaluation. Treat it as a red flag to inspect SKILL.md content before trusting automation that ingests it.
Assessment
This package appears internally consistent with its stated purpose (an OpenClaw enhancement plugin) and does not request external credentials. Key points to consider before installing: - Privacy: the plugin intentionally scans many local locations (~/.openclaw, ~/.claude, agent workspaces, sessions JSONL, installed skill package.json files, .git/config) to implement skill-recommender, config-doctor and transcript-search. That behavior is expected for the advertised features but can surface sensitive metadata (installed-skill names, local session contents). If you run it on a shared machine or on data you want to keep private, review code paths that read those locations. - Prompt-injection artifact: a unicode-control-chars pattern was detected in SKILL.md. While the rest of the package looks coherent, inspect SKILL.md for hidden control characters or unexpected instructions before allowing automated tools to ingest it. - Enabled-by-default: manifest sets enabledByDefault=true so the plugin will be active once installed. If you prefer manual enable, disable it immediately after install until you've audited it. - Scripts present but not auto-run: repository includes scripts (e.g., scripts/setup.sh) and release/publish examples referencing tokens; the package metadata doesn't show an automatic installer fetching code from external URLs. However, do not run included shell scripts unless you have reviewed them. - Network behavior: README references a local dashboard (http://localhost:18789). Review runtime code that opens network listeners or makes outbound requests (not all files were shown). If you require, run the plugin in a restricted environment or sandbox, and monitor outbound network connections on first run. - If you are not comfortable: (1) audit the SKILL.md and modules that read home directories and sessions (skill-recommender, config-doctor, transcript-search, native-memory-surfacer); (2) install in a test profile/VM; (3) disable enabledByDefault or block network access until audited. If you want, I can list the specific files and code locations that read user paths and where to look for any network or child_process invocation to help you audit further.
!
src/modules/dashboard.ts:27
File read combined with network send (possible exfiltration).
About static analysis
These patterns were detected by automated regex scanning. They may be normal for skills that integrate with external APIs. Check the VirusTotal and OpenClaw results above for context-aware analysis.

Like a lobster shell, security has layers — review code before you run it.

Runtime requirements

🦞 Clawdis
latestvk975rzwrz3qq32tc8wra2y4mfx85qp5j
335downloads
0stars
25versions
Updated 13h ago
v5.7.10
MIT-0

火一五·克劳德·龙虾增强插件 v5.7.8

简介

@huo15/openclaw-enhanceOpenClaw 2026.4.24+非侵入式增强插件,对标 Claude Code 的 Agent Harness 体验。

核心原则:凡是龙虾原生有的功能一律不复制,重叠处以龙虾为准;只补龙虾没有的 Claude-Code 体验。

v5.7.8 全面适配 openclaw 2026.4.24(2026-04-26 同日)

跑完整 SOP 第 1+2 步发现 openclaw 4.24 的 api.on完全 typed<K extends PluginHookName>(hookName: K, handler: PluginHookHandlerMap[K])),enhance 之前 14 处 api.on(...as any) 都能去掉 cast。这是真正的"全面适配"——不是简单升 peerDep,而是让 enhance 利用 SDK 的全部类型信息。

维度v5.7.7v5.7.8
peerDependencies.openclaw^2026.4.22^2026.4.24
build.openclawVersion2026.4.11(落后 13 patch)2026.4.24
compat.pluginApi>=2026.4.11>=2026.4.24
api.on(...as any) 使用次数14 处0 处
(event: any, ctx: any) 使用次数5 处0 处
(ctx as any)?.agentId 模式(hook 内部)9 处0 处(仅余 4 处 helper 函数内部,工具 ctx 用)
typecheck 错误数00
openclaw.plugin.json 顶层字段58(加 enabledByDefault / uiHints / activation

隐藏 bug 修复

去掉 as any 后 typecheck 暴露 self-check.ts 长期被屏蔽的类型不匹配:之前 return {}; 试图返回 PluginHookBeforeAgentReplyResult,但该类型 handled: boolean 是必填的——空对象不合规。修法:所有"不接管"分支改成 return;(void),仅"阻断空回复"分支返回 { handled: true, reply: ..., reason: ... }

不破坏 openclaw 原生

  • 所有 hook handler return undefined(void) — 不返回 {block, prependContext} 等控制信号时 → enhance 仅观察+附加,绝不改变 openclaw 决策
  • typed hook handler 实际行为跟 untyped 完全一致 — 只是 TS 编译期能 narrow 类型,运行时无差异
  • manifest 新加的 enabledByDefault / uiHints / activation 都是 openclaw 4.x 已有字段,不引入新依赖

v5.7.7 session-lifecycle(2026-04-26 同日,跑完整 gap 调研后落地)

调研依据:跑了一次完整 SOP 第 1+2 步(Claude Code 官方 hooks 文档 + 反编译 Claude.app + openclaw 4.22 SDK)。发现 openclaw 4.22 暴露 29 个 hook,enhance 之前只用 4 个。落地最高 ROI 的 5 个 hook 闭环 session 生命周期:

Hookenhance 行为落地表
session_startidle > 30min 时插入"🚀 会话开始/续启"章节占位chapters
session_end加"🏁 会话结束"章节 + flush in_progress todo 到 project memory(tag=session-flush, importance=4)chapters + memories
before_resetreset 前最后机会抢救最近 3 章节 + 全部未完成 todo 到 decision memory(tag=reset-rescue, importance=6)+ 推 notificationmemories
subagent_spawned派生子 agent 时加"🤖 派生子 agent: X"章节chapters
subagent_ended子 agent 结束加"✅/❌ 子 agent 结束: X"章节chapters

防 noise factory 三层防御(吸收 v5.7.1 教训):30 秒 dedup + 低 importance + 专用 tag(不进黑名单,用户下次会想恢复)。

v5.7.5 skill-recommender(2026-04-26 同日)

用户反馈:"新增自动根据用户的需求自动挑选已经安装的技能,如果没有技能就把规划方案给出来。看看 Claude 是如何做的"

调研:反编译 /Applications/Claude.app/Contents/Resources/app.asar,发现 Claude 的 skill auto-discovery 本质是把所有 skill 的 name+description 拼成 "Available skills: ${list}." 注入到 specialist agent 的 system prompt —— 没有复杂算法,让 LLM 自己挑。

enhance 改造:照搬 name+description 匹配思路,但改成按需工具避免每轮 prompt 占 schema。新增模块 skill-recommender + 工具 enhance_skill_recommend(query, limit?, includeUninstalled?, includePlanning?)

  1. 启动期扫多路径(WeCom / DingTalk 多 agent 场景关键):
    • ~/.openclaw/skills/
    • ~/.openclaw/workspace/skills/
    • ~/.openclaw/workspace-*/skills/WeCom 多 agent 动态 workspace(一开始漏扫,烟测才发现)
    • ~/.openclaw/agents/*/skills/
    • <cwd>/.claude/skills/~/.claude/skills/
    • 实测用户机器扫到 56 个 skill 跨 27 个路径
  2. 解析 SKILL.md frontmatter(轻量正则,无 yaml 依赖):name + description + aliases
  3. CJK 双字滑窗 + alias 强 boost 评分:
    • JS \w 不含中日韩,直接 split \s\W 会让"代码简化"分成空数组
    • 解决:CJK 连续段当整体 phrase + 长 ≥4 时滑动 2-grams
    • alias 的 token 严格命中(如 "规划" === alias "规划")→ 保底 0.7 分
  4. 三段式输出
    • 🎯 已装 skill(命中 ≥ threshold=0.25)+ 召唤建议
    • 📦 ClawHub 上未装的 huo15-* 候选(含 openclaw skills install 命令)
    • 🛠️ 都没合适 → 自建 skill 规划:建议 slug + frontmatter 模板 + 触发关键词 + 内容大纲 + 红线 #3 提醒(必须先 ClawHub publish 再让 enhance 引用 slug,插件不内嵌 skill 内容)

实测

查询命中分数
"帮我 review 这个 PR"huo15-openclaw-code-review0.60
"设计一个 Web UI 原型"huo15-openclaw-frontend-design0.94
"代码简化"huo15-openclaw-simplify1.00
"做安全审查"huo15-openclaw-security-review0.96
"规划这个任务"huo15-openclaw-plan-mode0.70(alias exact 命中保底)

模块 tier=2(balanced 默认启用,minimal 不暴露 — 用户多半已知道用什么 skill)。

v5.7.4 config-doctor 扩展:扫已装插件 bare pluginApi(2026-04-26 同日)

用户反馈:"提示插件要求 2026.2.24,但是我的 openclaw 已经是 2026.4.22"

根因:openclaw plugin compat 规则要求 compat.pluginApi 必须是 ranged spec(>=X.Y.Z / ^X.Y.Z / ~X.Y.Z)。bare 字符串(如 "2026.2.24" 没前缀)= 精确匹配,与当前 openclaw 不匹配时启动失败。用户实测:

  • ~/.openclaw/extensions/tips/package.json v1.0.0 → pluginApi: "2026.4.11"
  • ~/.openclaw/node_modules/@huo15/huo15-huihuoyun-odoo/package.json v1.2.0(npm peerDep 残留)→ pluginApi: "2026.2.24"

新增config-doctor 启动期扫描 ~/.openclaw/extensions/* + ~/.openclaw/node_modules/@huo15/* + 无 scope 的 node_modules/*,对每个声明 openclaw.extensions 的包检查 compat.pluginApi。bare 命中 → 推仪表盘 + log warn + 给可粘贴 fix 命令。

⚠️ [plugin-bare-pluginApi] 已装插件 @huo15/wecom-tips 的 openclaw.compat.pluginApi="2026.4.11" 是 bare 版本,会被解读为精确匹配...
   → 修复: python3 -c "..."(一行 inline)

v5.7.3 config-doctor(2026-04-26 同日)

直击用户高频反馈"装上插件还是 'Context limit exceeded'" — 根因往往不在插件,而在 ~/.openclaw/openclaw.json 的两处陷阱:

  1. 缺失 agents.defaults.compaction.reserveTokensFloor — openclaw 4.22 把这个字段嵌套到 agents.defaults 里(4.11 时是顶层 compaction),老用户配置文件没自动迁移,用 4.22 默认值(很小)→ 长 session 必爆
  2. 某个 model 的 maxTokenscontextWindow 一半以上 — 例如 MiniMax-M2.7 默认 maxTokens=131072 / contextWindow=204800,每轮预留输出就吃掉 64% budget。openclaw 把 maxTokens 当作"必须留给输出的 reserve",剩 73k 给 input/tools/memory,任意几轮就爆

新增

  • src/modules/config-doctor.ts — 启动期 sync 读 openclaw.json 检查上述两类陷阱,发现后用 notifyQueue.emit("config-doctor", ...) 推到仪表盘 + log warn + 给可粘贴的 fix 命令(python3 一行原地改 JSON,不调 child_process
  • 工具:enhance_config_doctor — 无参数,agent / 用户随时调一下,重新跑诊断(修完了配置可以再跑确认 ✅)
  • 配置项:config.configDoctorenabled / minReserveTokensFloor(默认 5000)/ maxReserveTokensFloor(默认 100000)/ maxModelMaxTokens(默认 32000)

红线遵守

  • 完全只读 ~/.openclaw/openclaw.json(红线 #1:不侵入式修改 openclaw)
  • 不调 child_process(红线 #4) — 修复命令是 python3 inline,由用户/cron-cli 执行
  • 不暴露在 minimal 之外?反过来:tier=1 minimal 也启用 — 这是关键的"防爆 context"诊断,每个用户都该有

v5.7.2 hardening(2026-04-26 同日)

继 v5.7.1 hot-fix 之后,对全代码库做了一次审计,修复 4 类潜在 bug:

  • 进程内 Map LRU 上限mode-gatemodeState / plannedActionssession-recaplastRecapAt 之前 keyed by agentId::sessionId 跨 session 永不清。WeCom 多用户场景下 100+ session 会无限累积。现在加 200/200/500 三档 LRU cap,活跃 session 重新插入刷新顺序,老 session 自动淘汰
  • safety_log / notifications 启动期 TTLgetDb() 时跑一次 DELETE WHERE created_at < datetime('now', '-90 days'),避免长期运行库无限增长。新增 purgeOldSafetyLogs(retentionDays) helper 给运维调
  • memory corpus tag 黑名单auto-compact / auto-checkpoint / audit / internal 这 4 个保留 tag 在 scoreRelevance() 入口直接 return 0,永不召回到 prompt(防御未来 hook 万一又写入 noise)
  • enhance_memory_store 拒收保留 tag — 用户/agent 显式调 store 时若 tags 含保留词,立即返回错误而非写入

bump openclaw peerDep ^2026.4.22

之前 peerDep >=2026.4.11,但 npm global 已升到 2026.4.22(差 11 个 patch)。本地 SDK 类型定义同步升级;hook 名验证全部仍存在(before_prompt_build / before_tool_call / after_tool_call / before_compaction / before_agent_reply),无破坏性变更。

v5.7.1 hot-fix(2026-04-26)

  • 删除 before_compaction 噪音 hook — 之前每次 openclaw auto-compact 都会以 decision 类、auto-compact tag 写入一条「[auto-compact] 对话上下文已压缩…」记忆。实测单 agent 24 小时积累 613 条全是噪音,关键词命中率虚高 0.4-0.5(过 corpus pruner 默认 0.5 阈值),把真正的 user/project/feedback 决策记忆挤出 prompt 上下文
  • 新增工具 enhance_memory_purge — 按 tag / category / contentLike 批量清理当前 agent 记忆,dry_run 默认 true(仅预览匹配数)。一键清理历史噪音:enhance_memory_purge tag="auto-compact" dry_run=false
  • 首次启动自动迁移:升级到 5.7.1 后即不再生成新噪音;旧噪音留待用户用 purge 工具或直接 SQL 清

v5.7 新特性(2026-04-25)

  • 历史会话搜索(transcript-search) — 照搬 Claude Desktop transcriptSearchWorker 算法(解包 /Applications/Claude.app/Contents/Resources/app.asar 抽出参考实现):
    • 流式扫 ~/.openclaw/agents/<agentId>/sessions/*.jsonl,行级 JSON.parse
    • extractText 兼容 string / [{type:"text", text}] 数组
    • indexOf 子串匹配 + ±80 字符 snippet
    • 79 个 session 中扫 30 个 → 3–5 ms 找到 5 个 hits(实测)
    • 完全只读、不建索引、不建表 — 不动 openclaw 任何东西
  • 工具:enhance_transcript_searchquery 必填;可选 agentId / limit / includeReset / caseSensitive
  • 模块 tier=2,默认 balanced/full 即可见(minimal 下不暴露)

v5.6 新特性(2026-04-24)

  • 工具分层(toolTier) — 按 minimal/balanced/full 三档暴露工具 schema,降低每轮 prompt 固定底座(解决长会话 context 提早爆满)
    • minimal(10 工具):仅核心层 — 记忆 / 状态栏 / spawn / 模式 / 章节 / installer / integrator
    • balanced(18 工具,默认):+ todo / 章节 / 定时任务桥
    • full(26 工具):+ workflow / safety / task-planner / session-recap / skill-doctor
  • Workflow 5→2 工具合并enhance_workflow_define / _list / _delete / _tasks / enhance_task 合并为 enhance_workflow(action=define/list/delete/tasks)+ enhance_task(保留独立 action 派发器)
  • 工具描述全面压缩 — 26 个工具描述从 ~4610 字符 → ~1750 字符(-62%),每轮 prompt 节省约 1400 token,prompt cache 更稳

一键安装

openclaw plugins install @huo15/openclaw-enhance
openclaw restart

安装后访问仪表盘:http://localhost:18789/plugins/enhance/

核心能力

  • 分类记忆(corpus supplement + 两段式) — user / project / feedback / reference / decision 五类,额外支持 why(背景/约束)和 howToApply(套用时机),对齐 Claude Code feedback/project 记忆体例;通过 registerMemoryCorpusSupplement 合入龙虾原生 memory 搜索,不自建第二套向量库;搜索结果 deterministic 排序(score → importance → updated → id)保持 prompt cache 稳定
  • 工具安全观察员 — 只分类错误 + 建议退避,完全尊重龙虾 tools.allow/deny
  • 任务 / 章节 / 模式闸门 + ExitPlanMode — TodoWrite / mark_chapter / plan-explore 模式;enhance_exit_plan_mode 在 plan 模式下提交计划给用户审批,自动把计划期间被拦截的写入意图打包成 decision 记忆
  • 状态栏(含可观测性)enhance_statusline 额外展示当前模型、思考档、fast 模式、消息通道、会话 ID;HTTP 端点 /plugins/enhance/api/statusline 输出 JSON 供仪表盘嵌入
  • 技能巡检 / 子任务一键派发enhance_spawn_task 返回可直接粘贴到终端的 openclaw agent CLI 命令,支持跨 agent 派发和思考档选择
  • 定时任务桥 — 登记定时工作流时返回一条 openclaw cron add 命令;调度归龙虾 cron-cli,插件只管触发时注入 instructions
  • 多 Agent 隔离 — 完美适配 WeCom 插件的动态 Agent,记忆/任务/章节/宠物全部按 agentId 隔离
  • 增强仪表盘 — 小火苗宠物 + 记忆/任务/章节/定时全景

与龙虾原生的关系

能力龙虾原生enhance 策略
记忆向量库✅ 龙虾负责不复制,改为 corpus supplement 并入搜索
工具 allow/deny✅ 龙虾负责只观察,不拦截
Cron 调度✅ 龙虾 cron-cli不管理调度,只在触发时注入上下文
技能安装✅ ClawHub只读巡检,不擅自安装

增强技能(自动注入 workspace/skills/

工作流模式(4 个)

  • huo15-openclaw-plan-mode — 结构化规划模式
  • huo15-openclaw-explore-mode — 深度探索模式
  • huo15-openclaw-verify-mode — 验证检查模式
  • huo15-openclaw-memory-curator — 记忆整理

设计能力(v5.4 新增,对标 Anthropic frontend-design + huashu-design 生态)

  • huo15-openclaw-frontend-design — 高保真 Web UI 原型 + 5 美学流派 + 反 AI Slop 硬红线
  • huo15-openclaw-design-director — 设计方向顾问(3 方向反差对比 + 强制推荐)
  • huo15-openclaw-brand-protocol — 品牌规范抓取(Ask/Search/Download/Verify/Codify 5 步)
  • huo15-openclaw-design-critique — 5 维设计评审(美学/可用性/品牌/内容/实现)

开发辅助(v5.5.1 新增,对标 Claude Code /simplify / /security-review / /review)

  • huo15-openclaw-simplify — 代码简化三维审查(复用/质量/效率)+ 分级修复清单
  • huo15-openclaw-security-review — 六类漏洞矩阵(密钥/注入/XSS/SSRF/权限/依赖)+ CVSS 分级
  • huo15-openclaw-code-review — PR 五维综合评审(设计/实现/测试/安全/可维护)+ 可粘贴评论

详见 README.mdCHANGELOG.md

链接

Comments

Loading comments...