Skill flagged — suspicious patterns detected

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

Dev Team

v0.1.4

Multi-agent development team orchestration. Use when managing coding agents (Codex, Claude Code, Gemini, Cursor) for automated software development: (1) Spaw...

0· 545·4 current·4 all-time
byVint@vintlin
Security Scan
VirusTotalVirusTotal
Suspicious
View report →
OpenClawOpenClaw
Suspicious
medium confidence
Purpose & Capability
The name/description match the files: the package provides scripts to spawn/monitor AI coding agents, run reviews, manage git worktrees, and send notifications. That functionality explains the presence of many git/agent/monitoring scripts. One mismatch: the SKILL.md and metadata do not declare required credentials (e.g., GitHub CLI auth tokens, OpenClaw/Feishu channels) even though the scripts call 'gh pr', may auto-merge, and rely on OpenClaw/Feishu for notifications. This omission is important operationally but not necessarily malicious.
!
Instruction Scope
Runtime instructions and scripts operate on arbitrary repository paths (--repo-path), create and remove git worktrees/branches (including forced deletions), and instruct users to add cron/LaunchAgent jobs that run the maintenance scripts. The SKILL.md and scripts also reference invoking agent CLIs with flags that permit shell/file access. These instructions give broad filesystem and repo-modification privileges and instruct persistent scheduled execution; they go beyond passive orchestration and can make destructive changes automatically if misconfigured.
Install Mechanism
No external install spec (no downloads) — the package is instruction/code-only, which avoids fetching remote binaries. However the bundle contains many executable scripts and node/python code that will run locally. There are no remote-install URLs to flag, but installing/using this skill means running these local scripts on your machine.
!
Credentials
The skill declares no required env vars, but its behavior implicitly requires privileged environment capabilities: authenticated 'gh' CLI (GitHub token), tmux, git access to repositories, and optional OpenClaw/Feishu integration. Additionally, config/agents.json intentionally configures agent CLIs with flags like '--dangerously-bypass-approvals-and-sandbox', '--dangerously-skip-permissions', and Gemini '--approval-mode yolo' with allowed-tools including run_shell_command/write_file/read_file — enabling agents to execute arbitrary shell/file operations. Those capabilities are consistent with the orchestrator role but are high-privilege and not explicitly justified by declared env requirements.
!
Persistence & Privilege
The skill recommends adding cron jobs or LaunchAgents to run check/cleanup/prune scripts periodically and suggests OpenClaw cron integration. That creates persistent, autonomous system activity (scheduling automatic merges, cleanups, branch deletions). Although the skill is not force-enabled (always: false), following its initialization guidance will give it recurring execution privileges that can alter repositories and remove branches automatically — combine this with auto-merge/auto-cleanup features and the blast radius increases.
What to consider before installing
This skill is coherent with its stated purpose (orchestrating multi-agent development) but has multiple high-impact behaviors you should deliberately accept or change before using: - Review all scripts (especially spawn-agent.sh, check-agents.sh, cleanup-worktrees.sh, prune scripts, and claim/enqueue scripts). They will create/remove worktrees and branches, and can force-delete branches. - Check and, if needed, remove or change the 'dangerous' agent CLI flags in config/agents.json (examples: '--dangerously-bypass-approvals-and-sandbox', '--dangerously-skip-permissions', Gemini '--approval-mode yolo' and allowed-tools). Those permit agents to run shell commands and write files and dramatically increase risk if untrusted prompts are used. - Understand authentication needs: the package uses the GitHub CLI ('gh') and assumes authenticated access; it does not declare or manage tokens. Ensure any automation has least-privilege GitHub credentials and that auto-merge is disabled unless you fully trust the workflow. - Be cautious with persistence: the SKILL.md suggests cron/LaunchAgent entries to run scripts periodically. Only enable scheduled jobs after you test scripts manually and disable auto-merge/auto-cleanup until you confirm behavior in a safe environment. - Test in an isolated environment first: run against a disposable repo or a fork, with auto-merge off, and review logs and assets/active-tasks.json to confirm actions. Use dry-run flags where present. - If you intend to trust the skill: lock its repo paths (avoid using --repo-path that points to critical repos), tighten allowedAgents in config/user.json, and audit who can submit prompts or add queue tasks. Why 'suspicious' rather than 'malicious': the code and instructions appear designed for the declared orchestration purpose, but they grant broad destructive and autonomous capabilities (forced branch deletion, scheduled auto-operations, enabling agent tool exec). Those are proportionate for a fully trusted internal tool but risky if used without careful configuration or with untrusted agents/prompts. Additional provenance (who published it, signatures, or a changelog), a manifest of what external CLIs are required and why, and documented safeguards (defaults disabling auto-merge/auto-cleanup) would increase confidence.

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

latestvk972090f5ntb4z4dd3rpvb0nrs826nt6
545downloads
0stars
4versions
Updated 7h ago
v0.1.4
MIT-0

Dev Team

One-person dev team orchestration using Levi as the orchestrator.

Main Agent Rules (入口摘要)

主 Agent 是编排器(orchestrator),不是默认主力编码器。

关键准则:

  • PR 优先:正式开发默认使用 --completion-mode pr
  • 禁止 SubAgent 在 main/master 上直接开发(spawn-agent.sh 已拦截)
  • 主 Agent 使用最小上下文:优先看状态机与聚合结果,而不是全量日志
  • 阶段化编排:明确 --phase build|review|fixup
  • 自动化负责推进与收口(check/review/fixup/cleanup/prune),人工保留最终产品决策与 GitHub approve
  • SubAgent DONE 不等于任务完成:合并前必须做 post-task review(DoD + 真实验证 + 指标门禁)

详细说明见:

  • references/AGENTS.md(主 Agent 行动准则)
  • (已合并入 references/AGENTS.md)编排模型与状态机
  • (已合并入 references/AGENTS.md)主 Agent 最小上下文操作手册

Quick Start

首次使用

# 运行设置检查
scripts/setup-check.sh

启动内置 dev-board 控制面板

dev-team 已内置 scripts/dev-board/,可直接启动:

./scripts/run-dev-board.sh

默认地址:http://localhost:4310

Spawn an Agent

正式功能开发(推荐 PR 流程):

scripts/spawn-agent.sh \
  --agent auto \
  --phase build \
  --repo-path ./scripts/dev-board \
  --branch feat/auth-timeout-fix \
  --description "修复登录超时错误" \
  --completion-mode pr \
  --prompt-file /tmp/task-prompt.txt

本地 smoke / 临时探索(仅无 PR 场景使用 session):

scripts/spawn-agent.sh \
  --agent auto \
  --repo-path ./scripts/dev-board \
  --branch fix/auth-error \
  --description "修复登录超时错误" \
  --phase build \
  --completion-mode session \
  --prompt-file /tmp/task-prompt.txt

Check Agent Status

cat assets/active-tasks.json

Monitor Agents

scripts/check-agents.sh

Mandatory: Post-task Review Gate (合并前必做)

当 SubAgent 报告完成(例如 TASK_*_DONE)后,主 Agent 必须执行:

# 1) 看任务状态与日志摘要
scripts/check-agents.sh

# 2) 审核 subagent 提交内容(示例)
git -C <subagent-worktree> show --name-only --oneline -n 1 <commit>

# 3) 在目标仓库执行任务 DoD 规定的真实验证命令(示例)
# 例如:真实 ingestion/validate、真实 API 调用、非 mock 链路

通过标准:

  • DoD 条款全部满足
  • 至少 1 条真实链路验证通过(非 mock)
  • 指标达到任务阈值(如 successRate/evidenceBindingRate/citationNonEmptyRate/fallbackRate)

不通过处理:

  • 不合并
  • 立即派发 fixup(同分支或新分支)
  • 在 docs 中记录“未通过项 + 修复输入”

Supported Agents

AgentUse CaseCommand
Codex后端逻辑、复杂 bug、多文件重构codex exec --dangerously-bypass-approvals-and-sandbox -C "/path" "prompt"
Claude Code前端、git 操作、快速任务claude --dangerously-skip-permissions -p "prompt"
GeminiUI 设计、生成规范gemini -p "prompt"gemini -y -p "prompt" (YOLO 模式)
Cursor前端产品化、IDE 协同、界面 polish;版本前差异分析(plan)cursor agent -f -p --workspace "/path" --model composer-1.5 "prompt"

Scripts

spawn-agent.sh

创建新任务并启动代理。

参数:

  • --agent (必需): codex | claude | gemini | cursor | auto(推荐)
  • --repo (必需二选一): 仓库名(位于 skill 上级目录时使用)
  • --repo-path (必需二选一,推荐): 仓库绝对/相对路径;用于 skill 根目录外的仓库
  • --branch (必需): 分支名
  • --description (必需): 任务描述
  • --prompt (三选一): 代理提示词(短文本时可用)
  • --prompt-file (三选一,推荐): 从文件读取提示词,避免 shell 转义问题(尤其包含反引号、$()、多行内容)
  • --prompt-b64 (三选一): Base64 提示词,适合脚本自动化调用
  • --agent-model (可选): 指定模型(例如 --agent cursor --agent-model composer-1.5
  • --cursor-mode (可选): dev|plan(仅 cursor 生效,默认 dev
    • plan 时会自动使用模型 gpt-5.3-codex
  • --phase (可选): build|review|fixup(默认 build);会受 config/user.json.agentPolicy 限制
  • --completion-mode (可选): pr(默认) | session;本地测试仓库无 PR 时建议用 session
  • --auto-merge (可选,pr 模式): 当 CI 与 review 满足条件后触发 gh pr merge --auto
  • --merge-method (可选,pr 模式): squash|merge|rebase,默认 squash
  • --cleanup-after-seconds (可选): 临时 worktree 清理 TTL;session 模式默认 3600

check-agents.sh

监控所有运行中的代理。

功能:

  • 检查 tmux 会话是否存活
  • 检查 PR 是否已创建(completionMode=pr
  • 跟踪 PR 状态机(waiting_checks / checks_failed / waiting_review / changes_requested / merge_ready / merge_queued / merged
  • 检查 required CI checks(gh pr checks --required
  • 检查 reviewDecision(gh pr view --json reviewDecision
  • 可选自动触发 gh pr merge --auto(任务级 autoMergeconfig/user.json.pr.autoMerge
  • 失败自动重试(按配置延迟,最多3次;依赖任务记录中的命令元数据)
  • 写入通知到 notifications.json (由 OpenClaw 发送飞书通知)
  • 结束后可自动触发 cleanup-worktrees.shprune-history.sh(避免 worktree / active-tasks 膨胀)

注意:

  • session exit mode 仅表示会话结束,不代表质量达标。
  • 质量达标必须由主 Agent 的 post-task review 判定。

review-agent.sh

多 reviewer 代码审查(默认三审:codex + gemini + claude)。

参数:

  • --repo (必需): 仓库名
  • --branch (必需): 分支名或 PR 号
  • --reviewers: 审查者 (默认: codex,gemini,claude;默认要求至少 3 个)
  • --allow-fewer-reviewers: 调试用,允许少于 3 个 reviewer
  • --no-post: 只生成本地审查产物,不回写 PR 评论

行为(新版):

  • 每个 reviewer 生成结构化审查结论(Verdict / Findings / Validation / Next Steps)
  • 自动抓取 gh pr diff 并注入 reviewer prompt(截断版)
  • 每个 reviewer 单独在 PR 下发表评论(正文带 [dev-team][reviewer=...] 标识)
  • 会按 config/user.json.agentPolicy 过滤禁用/阶段不允许的 reviewer
  • 生成聚合结果 assets/logs/reviews/review-<pr>-<ts>.aggregate.json
  • 聚合策略会结合 reviewer 权重、severity 和 reviewer 角色偏好,输出:
    • waiting_human_approve
    • review_changes_requested
    • review_human_attention
  • 写回 assets/active-tasks.json
    • checks.reviewAggregate
    • checks.reviewSummary
    • checks.fixupSuggested
    • checks.fixupTarget(用于后续优先回原 subagent 返工)

aggregate-reviews.sh

聚合多 reviewer 的审查产物,输出单一决策(通过 / 需修改 / 人工关注)。

示例:

./scripts/aggregate-reviews.sh --file assets/logs/reviews/review-1-1234567890.json

request-fixup.sh

根据 review-agent 聚合结果,为原 PR 负责 subagent 生成返工 prompt,并优先发回原 subagent(如会话仍存活)。

参数:

  • --branch <branch>--pr <number>:定位任务
  • --dry-run:只生成返工 prompt,不执行
  • --spawn-if-dead:原 subagent 会话已退出时,按同 branch 重新派工(谨慎使用)

行为:

  • 读取 checks.reviewAggregate / checks.reviewAggregateFile
  • fixupSuggested=true,生成结构化返工 prompt(按 severity 排序)
  • 写回 checks.fixupRequestedAt / checks.fixupRequestMode
  • 默认优先回原 subagent(同 branch owner)处理

Agent Adaptation & Policy

  • 统一 agent 适配与全流程使用建议:references/agent-adapters.md
  • 可配置 agent 启用开关与阶段允许列表:config/user.json -> agentPolicy
  • 推荐器支持阶段参数(会过滤禁用 agent):
./scripts/recommend-agent.sh --description "前端看板优化" --phase build
./scripts/recommend-agent.sh --description "PR 代码审核" --phase review

cleanup-worktrees.sh

清理已合并的 worktree。

补充说明:

  • completionMode=sessioncleanupMode=session_ttl 的任务,会在 cleanupEligibleAt 到期后自动清理 worktree
  • 清理后任务状态会写回为 cleaned(保留历史记录)

prune-history.sh

归档 assets/active-tasks.json 中过旧的已完成任务(done/failed/cancelled/cleaned/...),避免注册表无限增长。

常用参数:

  • --keep-days <days>: 仅归档早于 N 天的记录(当前默认配置为 0,更偏热数据)
  • --keep-count <n>: 无论时间如何,至少保留最近 N 条非失败/取消终态记录(当前默认配置为 12
  • --keep-fail-cancel-days <days>: failed/cancelled 独立归档天数(默认可更激进)
  • --keep-fail-cancel-count <n>: 至少保留最近 N 条 failed/cancelled
  • --dry-run: 预览归档数量,不落盘

自动化建议:

  • cleanup-worktrees.sh 执行后自动触发 prune-history.sh(可在 config/user.json.archive 配置)
  • 也可单独配置每日归档 cron

说明:

  • cleanup-worktrees.sh 会清理孤儿 worktree(任务已归档但同级目录仍残留)
  • session_ttl 会受全局 TTL 上限约束,避免历史任务长时间占用磁盘

setup-check.sh

首次使用检查脚本,验证配置完整性。

recommend-agent.sh

根据任务描述/提示词关键词、当前运行负载、CLI 健康状态推荐最合适的 agent(含理由)。

示例:

./scripts/recommend-agent.sh \
  --description "dev-board 前端 UI 产品化改版,优化筛选与日志面板" \
  --prompt-file /tmp/task-prompt.txt \
  --format json

list-agent-models.sh

列出 agent 模型(当前主要用于 Cursor)。

示例:

./scripts/list-agent-models.sh cursor

enqueue-task.sh / list-queue.sh / claim-task.sh

Queue/Claim 模式(推荐演进路径,现已提供最小可用版本):

  • enqueue-task.sh:主 Agent 添加任务到队列(assets/tasks.json
  • list-queue.sh:查看队列(支持状态/阶段过滤)
  • claim-task.sh:认领一个 queued 任务并调用 spawn-agent.sh 派工
  • sync-queue-status.sh:根据 assets/active-tasks.json 回写队列任务状态(running/done/cleaned/...
  • prune-queue-history.sh:归档 assets/tasks.json 中过旧终态队列任务到 assets/logs/archives/queue-history-YYYY-MM.jsonl

示例:

./scripts/enqueue-task.sh \
  --repo-path ./scripts/dev-board \
  --branch feat/dev-board-action-history \
  --description "dev-board 操作历史列表" \
  --phase build \
  --completion-mode pr \
  --prompt-file /tmp/task.txt

./scripts/list-queue.sh --status queued
./scripts/claim-task.sh --agent auto

Task Registry

位置:assets/active-tasks.json

编码约定:

  • 所有任务与通知 JSON 文件统一使用 UTF-8 保存
  • 写入时使用 ensure_ascii=False,中文会以可读形式保存(不再显示 \\uXXXX

归档约定:

  • assets/active-tasks.json 保存热数据(运行中 + 近期终态任务)
  • 历史记录归档到 assets/logs/archives/task-history-YYYY-MM.jsonl

任务状态以脚本实现为准,详见 references/state-machine.md

补充字段(新任务会记录):

  • requestedAgent: 用户请求的 agent(可能是 auto
  • agentModel: 实际使用模型(如 composer-1.5
  • agentSelectionReason: 自动派工的简要理由(关键词/健康回退等)

References


Troubleshooting

1) Claude 子任务启动失败(常见于复杂 prompt)

常见根因:

  • 在 shell 中直接使用 --prompt "...",内容包含反引号、$() 等,导致 shell 提前展开,实际传给脚本的 prompt 被污染。

修复方式(推荐顺序):

  1. 使用 --prompt-file(首选)
  2. 自动化调用使用 --prompt-b64
  3. 仅短 prompt 使用 --prompt

2) Cursor agent 无法使用 / 很少被选中

如果 cursor agent status 显示 Not logged inspawn-agent 会在启动前直接失败并提示:

cursor agent login

你提到常用模型 composer 1.5,当前已在 config/agents.json 中设为 Cursor 默认模型(composer-1.5)。 如需覆盖:

./scripts/spawn-agent.sh --agent cursor --agent-model composer-1.5 ...

Cursor plan 模式(用于版本前差异分析):

./scripts/spawn-agent.sh \
  --agent cursor \
  --cursor-mode plan \
  --repo-path /path/to/repo \
  --branch feat/version-plan \
  --description "版本规划与差异分析" \
  --completion-mode session \
  --prompt-file /tmp/plan.txt

说明:

  • 该模式等价于 cursor agent --mode plan(只读规划,不改代码)
  • spawn-agent 会强制模型为 gpt-5.3-codex

3) PR / CI / review 流程不推进

必要前置条件:

  1. 仓库已配置 GitHub origin remote
  2. gh auth status 有效(token 未过期)
  3. subagent 的 prompt 明确要求:
    • 提交 commit
    • git push -u origin <branch>
    • gh pr create ...

如果缺一项,check-agents.sh 无法推进 PR 状态机。

tmux

tmux new-session -d -s <name> -c <dir> "command"
tmux send-keys -t <session> "text" Enter
tmux list-sessions
tmux kill-session -t <session>

Comments

Loading comments...