Self-Governor

LLM 通用内部自裁决技能。在关键节点判断"当前这一层最优的下一步动作是什么"，再让主链继续执行。触发条件：(1) 路径分叉时——多个可行方案且无明显优先级；(2) 高代价动作前——搜索/生成/发布等消耗军费或不可逆操作；(3) 连续两步无明显增益时——进展停滞、输出质量未提升。禁止：改写主任务、输出并列动作、长...

Security Scan

The skill's instructions, inputs/outputs, and requirements are consistent with its stated purpose (a lightweight decision filter) and it doesn't request extra binaries, secrets, or installs, though provenance is unclear so caution is warranted.

Purpose & Capability

Name and description describe an internal decision layer; the SKILL.md, action set, triggers, and I/O all implement exactly that behavior. There are no unrelated env vars, binaries, or config paths requested — the requested capabilities align with the stated purpose.

Instruction Scope

Runtime instructions only describe receiving a small structured input, returning a single action+reason, and expecting the host (main chain) to perform an Update. The skill does not instruct reading arbitrary files, accessing environment variables, or sending data to external endpoints. It does require the main chain to supply Update callbacks to prevent loop-triggering, which is a legitimate integration requirement.

Install Mechanism

This is instruction-only with no install spec and no code files executed at runtime. That is the lowest-risk install profile and matches the declared metadata.

Credentials

No environment variables, credentials, or config paths are required. The skill does not request any secrets or broad access, and the declared inputs are limited and appropriate for the decision role.

Persistence & Privilege

always is false and the skill is user-invocable (normal). Autonomous invocation (disable-model-invocation=false) is the platform default; by itself this is not a problem given the skill's narrow scope. There is no request to modify other skills or system-wide settings.

Assessment

This skill appears coherent and low-risk: it only implements a small decision filter and asks nothing extra. Before installing, verify provenance (no homepage or identifiable author), test it in a safe/staging agent to ensure the host correctly enforces the required Update callback (to avoid loop triggers), and confirm you trust the agent's orchestration logic to enforce the skill's 'must not' rules. If you allow autonomous invocation, monitor logs initially to ensure the skill doesn't get inserted in inappropriate places. If provenance or publisher identity matters to you, ask the publisher for more info or prefer a skill with an auditable source.

在关键决策节点插入一个轻量自裁决层，输出一个动作，让主链继续。

核心闭环

Goal → Gap → Action → Update → (下一轮)

不是复杂协议，是一个关键节点微循环。

触发时机

以下三种信号出现时，主链立即插入一次裁决：

1. 路径分叉 — 2+ 可行方向，无明显优先级
2. 高代价动作前 — 不可逆操作或稀缺资源消耗
3. 连续两步无明显增益 — stage_goal 完成度未提升

详见 references/triggers.md

输入输出

详见 references/input-output.md

最小输入（3字段）

• stage_goal — 当前阶段目标
• current_gap — 最大缺口
• available_actions — 候选动作子集

最小输出

• next_action — 唯一动作（必填）
• reason — 一句话（必填，≤30字）
• fallback_if_fail — 失败兜底（可选）

Update 环

主链强制义务: 每次 self-governor 输出动作并执行后，主链必须在下一步开始前完成 Update 回写。不回写则触发器3（连续无增益）将失效。

回写字段: goal_achievement / gap_change / actions_taken / total_steps_in_stage

详见 references/input-output.md

固定动作集合

6 个动作，详见 references/actions.md：

anchor · time_bind · search · clean · synthesize · degrade_continue

禁止项

• ❌ 改写主任务（你是裁决器，不是任务规划器）
• ❌ 输出多个并列动作（只返回一个）
• ❌ 生成长篇规划（reason 一句话）
• ❌ 变成审批器（不要"建议暂停等待确认"）
• ❌ 连续两轮默认要求补资料（最多一轮，第二轮必须 degrade_continue）

反官僚原则

• ✅ 默认选择动作，不默认停住
• ✅ 每次只返回一个动作
• ✅ reason 保持一句话
• ✅ 条件不理想时，优先 degrade_continue 而非空转

不做什么

• 不写业务内容
• 不给最终结论
• 不替代主链分析
• 不充当审批器
• 不做多动作排序/评分/长链规划/状态机

它只负责：先停一下，再选一步。

示例

详见 references/examples.md