Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add Xwen0857/agent-orchestrator --skill audit-guardgit clone --depth 1 https://github.com/Xwen0857/agent-orchestratorWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/xwen0857/agent-orchestrator/audit-guard)<a href="https://agentmods.dev/skills/xwen0857/agent-orchestrator/audit-guard"><img src="https://agentmods.dev/badge/skills/xwen0857/agent-orchestrator/audit-guard/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/xwen0857/agent-orchestrator/audit-guard"><img src="https://agentmods.dev/badge/skills/xwen0857/agent-orchestrator/audit-guard.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00050 | $0.01517 |
| Opus 5.5 | $0.00020 | $0.00607 |
| Sonnet 5.5 | $0.00010 | $0.00303 |
| Haiku 4.5 | $0.00005 | $0.00152 |
Grade A, and why
audit-guard scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 110 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Audit Guard
V2 基线(强制优先)
- 命中需审批规则时,统一将任务状态写为
BLOCKED_PENDING_APPROVAL(替代旧名BLOCKED_PENDING_MASTER_APPROVAL)。 - gate 决策必须同时写入:
templates/coordination/tasks/task_folders/<task_id>/audit.mdtemplates/coordination/tasks/task_folders/<task_id>/log.ndjson
- 放行条件是
templates/coordination/tasks/task_folders/<task_id>/approval.json存在且有效;未满足前不得恢复执行。 - 审计日志字段以
agent-orchestrator/references/event-log-schema.md与audit-guard/references/audit-log-schema.md的交集为最低要求,必须包含哈希链字段。
何时使用
- planner 将要执行可能高风险的操作。
- 需要对操作进行风险分级、冻结审批、审计留痕。
- 需要紧急越权(break-glass)并保留完整证据。
输入
operation_request:拟执行动作、参数、目标对象、执行人、来源任务。runtime_context:当前系统状态、近 1h 行为统计、token/cost、并发负载。policy_config:风险规则、审批策略、SLA、越权开关。
输出
gate_decision:ALLOW|MONITOR|BLOCK|BLOCK_PENDING_APPROVAL.risk_assessment:风险等级、命中规则、影响面、预演结果。approval_ticket(若需):提交 master 的审批单。audit_log_entry:不可篡改日志记录。
风险分级
Tier 1: CRITICAL (必须人工审批)
- 系统级破坏或不可逆影响(例如:
sudo、生产数据库结构/数据直接变更)。 - 安全边界变更(权限边界、白名单、密钥暴露)。
- 代码库强制历史覆盖(例如:
git push --force)。 - 网络暴露高风险行为(公网暴露本地服务)。
默认动作:BLOCK_PENDING_APPROVAL。
Tier 2: HIGH (策略驱动审批)
- 破坏性文件操作(例如:
rm)。 - 大范围状态回滚(例如:
git reset --hard)。 - 明显的成本放大(并发或资源规格骤增)。
- 未知模块调用、未知 Bash 指令执行。
- 可疑注入模式命中。
默认动作:BLOCK_PENDING_APPROVAL;仅在明确白名单策略下可降为 MONITOR。
Tier 3: MONITORED (记录 + 阈值告警)
- 短时高频 API 调用。
- 配置参数批量变更。
- 测试标准放宽。
默认动作:MONITOR;超过阈值升级为 Tier 2。
动态行为分析规则
- 频率异常:同类高风险动作在
window_minutes内超过阈值。 - 序列异常:出现高危链路(如“放宽测试 -> 提高并发 -> 强制回滚”)。
- 基线偏移:当前 token/cost 或失败率显著偏离近 24h 基线。
- 上下文冲突:动作意图与任务目标不一致(例如文档任务触发数据库变更)。
门禁工作流
- 检测与评估:匹配静态规则 + 动态行为分析。
- 冻结与打包:命中 Tier 1/2 时冻结动作,生成决策上下文包。
- 状态变更:
BLOCKED_PENDING_MASTER_APPROVAL。 - 通知 master:提交审批单。
- 执行决策:
- 批准:发放一次性许可并放行。
- 拒绝:终止动作,保留冻结记录。
- 全程记录:写入 append-only 审计日志。
审批 SLA 与超时
approval_sla_minutes默认 30。- 超时策略默认
DENY_AND_KEEP_FROZEN。 - 超时后必须发送升级告警并等待 master 明确处置。
紧急越权 (Break-Glass)
- 仅用于无法及时审批且业务必须继续的极端场景。
- 触发前必须生成系统快照并记录触发原因。
- 每 4 小时最多 1 次。
- 需要至少 2 名管理员授权(或全管理员一致)。
- 放行范围仅限指定操作与时效窗口。
- 使用本身视为严重安全事件并强制告警。
What ships with it
10 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- references/approval-ticket-template.md 310 B
- references/audit-log-schema.md 598 B
- references/detection-heuristics.md 448 B
- references/gate-required-triggers.md 744 B
- references/policy-schema.md 776 B
- scripts/create_approval_ticket.sh 1.4 KB runs code
- scripts/evaluate_gate.EXPLAIN.md 1.6 KB
- scripts/evaluate_gate.sh 6.7 KB runs code
- scripts/grant_approval.sh 2.1 KB runs code
- scripts/record_kb_feedback.sh 778 B runs code
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 1mo ago First seen · 110 lines · 50 tokens per session scan A 7ddf2b104bff
audit-guard is a skill published in the GitHub repository Xwen0857/agent-orchestrator (2 stars, last pushed 7mo ago), licensed MIT. It adds 50 tokens to every session and 1,517 once invoked, about $0.0002 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
orca-replay
Answers questions about a past agent run from its recording rather than from memory, and replays or forks that run. Use when asked why an earlier run did something, or to reproduce a failure.
adversarial-review
A multi-stage code review process in which one reviewer looks for failures, another independently checks the findings and the changed code, and a final reviewer decides which findings stand. It requires evidence such as code locations or test results and states that the reviewers may still share the same model's blind spots.
decision-matrix
Compare options with weighted scoring, pros and cons, pre-mortems, opportunity costs, and ICE prioritization. Use when a user wants to reason through a choice, expose assumptions, or rank alternatives.
meeting-note-summarizer
Turn meeting notes or transcripts into factual summaries, decisions, questions, and action items. Use when a user wants a concise recap or needs explicit owners and deadlines extracted without filling in missing details.
skillhub-cli
Connect an Agent to a SkillHub registry and use the official SkillHub CLI to search, install, list, or explicitly upgrade SkillHub skills. Use when a user asks to connect SkillHub, install a SkillHub skill, or manage skills previously installed from SkillHub.
ledger-tasks-yylo
Comprehensive guide for using YYLO Ledger task management without bypassing controller routing, lifecycle state, or mutation receipts. Covers task commands, dependency management, best practices, and workflow patterns. Use when you need to interact with the YYLO Ledger board.