autonomous-decisions

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Autonomous Decisions

自主决策

Activate only on the user's explicit instruction, for the specified decision set. Expansion to all decisions in the task requires an explicit user instruction. An accepted decision permits the coordinator to act within existing task authorization; deliberation creates no additional execution permission.

仅在用户明确指示后，针对指定的决策集启动本流程。若要扩展至任务中的所有决策，同样需要用户明确指示。决策获得接受后，coordinator 可在现有任务授权范围内采取行动；审议过程本身不会产生额外的执行权限。

Standing rules

长期规则

Bring high-risk or irreversible choices, conflicts with previous directions or established values, and unresolved material uncertainty about these boundaries to the user. Apply this gate whenever relevant facts change.
Use this process for choices between viable alternatives that materially affect behavior, experience, cost, or maintenance. Carry out mechanically determined choices directly within the authorized task.
Before every subagent dispatch, apply the available
```
model-selection
```
skill independently to that assignment and follow its delegation rules. Resolve it from the skill catalog; in this repository, use model-selection. Set the actual model and reasoning effort explicitly and disclose them at dispatch.
Give every subagent a fresh context, its current role instructions, necessary evidence, and only the authorized read-only resources. Use read-only tool restrictions when supported; otherwise enforce the boundary in the assignment and describe isolation accurately. Subagents return findings without mutations, external messages, or further delegation. Keep each decider isolated from peer deliberations; give judges and verifiers only their explicitly assigned artifacts.
Complete at most two initial decider assignments and one additional decider assignment per decision. The coordinator owns execution and verification.

Use the same decision protocol in Codex and Claude Code. Native isolated agents and fresh CLI sessions are execution options. Before using a CLI, read only its adapter: Codex or Claude.

将高风险或不可逆的选择、与先前指示或既定价值观的冲突，以及对这些边界仍未解决的重要不确定性提交给用户。相关事实发生变化时，需重新应用此门控。
对于存在于多个可行方案之间，且会实质影响行为、体验、成本或维护的选择，使用此流程。对于通过机械规则确定的选择，直接在已授权的任务范围内执行。
在每次分派 subagent 前，都需针对该任务独立应用现有的
```
model-selection
```
skill，并遵循其委派规则。从 skill 目录中确定适用项；在本仓库中，使用 model-selection。明确设置实际模型和 reasoning effort，并在分派时披露。
为每个 subagent 提供全新的上下文、当前角色说明、必要证据，以及仅限获授权的只读资源。如果系统支持，则使用只读工具限制；否则，在任务说明中强制执行该边界，并准确描述隔离方式。Subagent 仅返回调查结论，不执行任何变更、发送外部消息或继续委派。让每个决策者与同伴审议相隔离；仅向评审者和验证者提供明确分配的材料。
每项决策最多完成两项初始决策者分派和一项额外决策者分派。Coordinator 负责执行和验证。

在 Codex 和 Claude Code 中使用相同的决策协议。原生隔离 Agent 和全新的 CLI 会话可作为执行方式。使用 CLI 前，只读取其适配器文档：Codex 或 Claude。

Begin

开始

For multiple activated decisions, begin with batch.md. For one decision, begin with prepare.md. Load each subsequent reference only when the current stage's completion criterion names it. Give subagents their role and inputs rather than the coordinator's routing documents. Splitting files controls disclosure; fresh subagent contexts provide the actual context boundary.

对于多项已启动的决策，从 batch.md 开始。对于单项决策，从 prepare.md 开始。仅当当前阶段的完成标准明确指定后续参考文件时，才加载该文件。向 subagent 提供其角色和输入，而非 coordinator 的路由文档。通过拆分文件控制信息披露范围；全新的 subagent 上下文提供实际的上下文边界。