Human-AI Closed Loop

人机闭环认知 Skill —— AI 整理清单→人类实践证伪→想象力注入→AI 结构化输出的持续进化循环

Security Scan

The skill's documentation describes a Python runtime API, persistent logging, and operational behaviors but no runtime code or install spec is included — the declared capabilities don't match the provided artifacts, so proceed with caution.

Purpose & Capability

The declared purpose (a four-phase human-AI closed loop) is coherent with the text, but the SKILL.md demonstrates a Python class API (skills.human_ai_closed_loop.HumanAIClosedLoop), persistent log writes (loop_log/session_...), and runtime behaviors (heartbeats, reminders). However there are no code files implementing that API, and no install spec to fetch such code. That mismatch (claims of executable interface without provided implementation) is a material incoherence.

Instruction Scope

Instructions reference reading/writing persistent JSON logs, heartbeats, sending reminders, and preventing user-data exfiltration, but they don't specify mechanisms or endpoints. The spec asks the agent to write to loop_log/ and create session_{id}_round_{n}.json files — reasonable for a logging skill — but because no implementation is provided, it's unclear how reminders are sent or how/where data might be transmitted. The SKILL.md does not instruct reading unrelated system files or asking for extra env vars, but the lack of implementation details is a risk (unknown network calls could be added later).

Install Mechanism

No install spec is present (instruction-only), which is low risk in itself. However the README/SKILL.md show install commands (clawhub install or cp -r to ~/.openclaw/skills/) and example Python imports, which imply that an install would place executable code under skills/. The registry metadata exists, but the package as provided contains only docs — this inconsistency warrants caution because a future or remote install step could fetch code not shown here.

Credentials

The skill declares no required environment variables, no credentials, and no privileged config paths. That is proportionate to the described functionality (local checklist generation and logging). Still, the VERIFICATION_PROTOCOL calls out preventing outbound leakage but provides no enforcement mechanism — a gap to verify before trusting injected user data.

Persistence & Privilege

The skill expects to persist session logs in loop_log/ and maintain heartbeat state; it does not request 'always: true' and does not declare system-wide config changes. Writing project-local logs is reasonable, but you should verify (once code is available) that writes are confined to intended paths and that the skill does not modify other skills or global agent config.

Scan Findings in Context

[no_code_files_detected] unexpected: The registry package contains SKILL.md and docs but no implementation for the documented Python API. The regex scanner found nothing to analyze because there are no code files. For a runtime skill that shows import examples, an absent implementation is unexpected.

What to consider before installing

Do not install or enable this skill in a production environment yet. Key things to verify before trusting it: 1) Obtain and review the actual runtime code that implements skills.human_ai_closed_loop — the package currently contains only documentation; 2) If you plan to use clawhub install, inspect exactly what that command would download and from which repository/URL; 3) Confirm where reminders/notifications would be sent (no channels/endpoints are specified in the docs); 4) Verify that log writes are confined to the intended workspace path (loop_log/) and that the code does not access other system files or credentials; 5) Ensure injected human data is not transmitted to external services (the protocol promises interception but gives no enforcement details); 6) Prefer running the skill in a sandboxed environment and disable autonomous invocation for the first trials, or require a signed release/source repository and presence of the implementing code before enabling. If the author can supply the missing implementation or a link to a trusted repository containing the code, re-evaluate after reviewing that code and any install scripts.

Human-AI Closed Loop Cognition Skill

元数据

字段	值
名称	human-ai-closed-loop
版本	1.0.0
作者	KingOfZhao
发布日期	2026-03-31
置信度	96%

核心能力

实现四阶段人机闭环持续进化：

Phase 1: AI 整理清单
  └─ 将任务分解为「已验证事实」+「待证伪假设」+「未知盲区」

Phase 2: 人类实践证伪
  └─ 人类在真实世界运行清单，收集反例和边界条件

Phase 3: 人类想象力注入
  └─ 人类输入直觉、经验、创意（AI 无法独立产生的洞见）

Phase 4: AI 结构化输出
  └─ 吸收人类输入，更新世界模型，输出升级版清单
  └─ 置信度重新标注，循环开始下一轮

安装命令

clawhub install human-ai-closed-loop
# 或手动安装
cp -r skills/human-ai-closed-loop ~/.openclaw/skills/

调用方式

from skills.human_ai_closed_loop import HumanAIClosedLoop

loop = HumanAIClosedLoop(workspace=".", session_id="project_xyz")

# Phase 1: AI 整理清单
checklist = loop.generate_checklist(task="优化包装生产线良率")

# Phase 4: 人类反馈注入后重新输出
loop.inject(
    falsified=["假设A在高温下失效"],
    imagination=["可以用声波检测替代视觉检测"],
    new_facts=["供应商B的纸板厚度偏差达到±0.3mm"]
)
updated = loop.synthesize()
print(updated.confidence, updated.checklist_v2)