Flag audit — 2026-06-15

[github-actions]

I can see you've shared a skill definition for ops-flag-audit that audits feature flags against live registries.

What would you like me to do with this skill? I can help you:

• Set it up as a custom skill in your project
• Run an audit using this skill (if you have the skill installed)
• Modify or enhance the skill definition
• Something else

[YiWang24]

📋 Issue 自动分析

类型：运维自动生成（ops-generated）— feature flag 审计任务
来源：maintenance workflow → flag-audit job（cron 0 15 * * 1）
已有标签：tech-debt priority:p3 ops-generated ✓（标签合理，无需调整）

---

⚠️ 问题：Issue 正文内容异常

本 Issue 由 flag-audit composite action 自动创建，但正文内容并非审计报告，而是一段 Claude 的对话式回复：

"I can see you've shared a skill definition for ops-flag-audit... What would you like me to do with this skill?"

根因分析：

flag-audit action 调用 claude-harness 时，将 skills/ops-flag-audit/SKILL.md 渲染后作为 prompt 传给 claude-code-action。但：

1. 缺少 system-prompt：claude-harness 的 system-prompt 输入默认为空，flag-audit action 也未传入 → Claude Code 收不到「你是执行任务的 agent」的角色指令
2. SKILL.md 被当作用户分享的文档：SKILL.md 内容是「You are auditing...」的技能定义，Claude Code 将其理解为用户粘贴的技能文档，而非执行指令
3. 上下文数据为空：registry_flags: []、previously_open_audit_issues: [] — 没有实际注册表数据，Claude 无从审计

影响链：

[flag-audit 未传 system-prompt] → [SKILL.md 作为 user prompt 发送] → [Claude 误判为用户分享文档] → [输出对话式回复而非审计报告] → [Issue 正文为无意义内容]

与 #177 的关联：#177 的问题是正文为文件路径（已通过 extract-text 修复），#180 是新的问题 — 提取出了文本，但文本内容是 Claude 的错误响应。

建议修复

1. 在 flag-audit action 调用 claude-harness 时传入 system-prompt，明确指示 Claude 执行审计任务
2. 或在 SKILL.md 开头添加明确的执行指令（如「执行以下审计任务，输出 markdown 报告」），而非仅以技能定义形式呈现
3. 当 code_flags 为空数组时跳过 issue 创建（无审计数据时不应创建 issue）

本评论由 OpenCI GitHub Bot 自动分析生成。标签 tech-debt priority:p3 ops-generated 已由创建 bot 正确设置。

[github-actions]

OpenCI issue agent executed:

• escalate: needs-human

Reasoning:
Issue #180 is closed as not_planned but a MEMBER (YiWang24) just posted a detailed root-cause analysis identifying a real bug in the flag-audit workflow: the composite action calls claude-harness without a system-prompt, so SKILL.md is misinterpreted as a user-shared document and the issue body becomes conversational noise. Labels are already appropriate (tech-debt, priority:p3, ops-generated). Reopening and dispatching an MCP implementation task are high-risk and require a trusted-actor signal that is not explicit here, so the safe action is to escalate to human maintainers with the needs-human label so they can decide whether to reopen and ship the proposed fixes (add system-prompt, gate on empty code_flags).