orchestrate

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Orchestrate

任务编排

Keep the root agent responsible for decomposition, user communication, approvals, synthesis, and the final claim. Delegate outcomes, not vague help.

由根代理负责任务分解、用户沟通、审批、结果合成以及最终结论输出。委派的是明确的任务成果，而非模糊的协助请求。

Decide Whether to Delegate

判断是否需要委派

Delegate only when at least one condition holds:

Two or more workstreams can proceed independently.
A slow command, transfer, test suite, monitor, or data pull would otherwise make the root unavailable.
A bounded investigation can gather evidence while the root handles the main path.
An independent review would materially reduce implementation, production, security, payment, or release risk.

Stay single-agent when the task is small, the next step depends on the immediately previous result, multiple agents would touch the same mutable files, or delegation overhead is comparable to doing the work directly.

仅当至少满足以下任一条件时才进行委派：

存在两个或多个可独立推进的工作流。
若不委派，执行缓慢的命令、传输、测试套件、监控或数据拉取操作会导致根代理无法响应。
可通过有界调查收集证据，同时根代理处理主任务路径。
独立评审能显著降低实现、生产、安全、支付或发布风险。

当任务规模小、下一步依赖上一步的即时结果、多个代理会修改相同的可变文件，或者委派的开销与直接完成任务相当，应保持单代理模式。

Select the Model

选择模型

Choose by ambiguity, coupling, and consequence rather than task size alone.

Model	Use when	Typical effort	Do not use for
Luna	The assignment can be completed as a bounded leaf. Examples: evidence gathering, CI inspection, repository or documentation research, workflow execution, scoped implementation, review, browser verification, or multi-source reconciliation.	`xhigh` by default; `max` when quality matters more than latency or a weak result would cause rework.	Final high-stakes judgment or coordination of other agents.
Terra	The assignment benefits from a stronger collaborative peer, a higher coding ceiling than Luna, or delegated coordination of its own independent workstream.	`max` by default.	Routine leaf work that Luna xhigh or max can complete.
Sol	The assignment needs independent senior judgment under ambiguity or high consequence. Examples: architecture, competing incident hypotheses, security or payment correctness, release-risk analysis, or an adversarial audit of a proposed plan.	`high` by default; `max` for the hardest quality-first work.	Routine scouting, deterministic workflows, or ordinary isolated coding.

Default to Luna xhigh for leaf work. Escalate to Luna max for difficult leaves. Use Luna high, medium, or low only as an explicit latency optimization for completely mechanical work whose result is deterministic and cheaply verified. Move to Terra max when peer coordination, its higher ceiling, or lower wall-clock time than a long Luna max run justifies the additional cost. Move to Sol high or max when the assignment must resolve the central ambiguity or make a high-consequence judgment.

Use Luna only as a leaf and explicitly tell it not to delegate. Sol and Terra can act as collaborative peers and may delegate further only when the root explicitly assigns coordination ownership, the child work is independently scoped, and slots remain available. Otherwise make them leaves too.

选择模型时需考虑任务的模糊性、耦合度和影响程度，而非仅看任务规模。

模型	使用场景	典型算力投入	不适用于
Luna	任务为有界叶子任务。例如：证据收集、CI检查、仓库或文档调研、工作流执行、限定范围的实现、评审、浏览器验证或多源数据核对。	默认 `xhigh` ；当质量优先于延迟，或结果不佳会导致返工的场景下使用 `max` 。	高风险的最终决策或协调其他代理的任务。
Terra	任务需要更强的协作能力、比Luna更高的编码上限，或需要委派协调独立工作流。	默认 `max` 。	Luna的xhigh或max算力即可完成的常规叶子任务。
Sol	任务需要在模糊或高影响场景下做出独立的资深判断。例如：架构设计、竞争事故假设、安全或支付正确性验证、发布风险分析，或对拟议方案的对抗性审计。	默认 `high` ；针对最注重质量的高难度任务使用 `max` 。	常规侦察、确定性工作流或普通独立编码任务。

叶子任务默认使用Luna xhigh。难度较高的叶子任务升级为Luna max。仅当任务完全机械、结果可确定且易于验证，需要明确优化延迟时，才使用Luna high、medium或low。当需要协作能力、更高的能力上限，或相比长时间运行的Luna max能缩短耗时，且额外成本合理时，使用Terra max。当任务必须解决核心模糊点或做出高影响判断时，使用Sol high或max。

仅将Luna用作叶子代理，并明确告知其不得委派任务。Sol和Terra可作为协作伙伴，仅当根代理明确赋予协调权限、子任务范围独立且仍有可用资源时，才可进一步委派任务；否则也需作为叶子代理。

Disclose Every Spawn

披露所有代理生成操作

Immediately before every

spawn_agent

call, send a concise commentary update that tells the user:

The agent task name.
The exact requested model slug, such as
```
gpt-5.6-luna
```
.
The exact requested reasoning effort, such as
```
high
```
or
```
max
```
.
The assigned role and whether the agent is a leaf or an authorized coordinator.

Group simultaneous spawns into one compact update when useful. For example:

text

Spawning:
- `evidence_scout` — `gpt-5.6-luna`, reasoning `xhigh` — read-only leaf
- `senior_critic` — `gpt-5.6-sol`, reasoning `max` — senior-review leaf

Describe these as requested configuration, not verified runtime identity. If the runtime later reports a different actual model or reasoning effort, disclose the correction. Prefer

fork_turns: "none"

with explicit

model

and

reasoning_effort

values so the configuration is visible. When a history fork requires inherited values, say they are inherited and name the parent's exact model and reasoning effort when known; if the runtime does not expose them, state that explicitly and never guess.

在每次调用

spawn_agent

之前，立即向用户发送简洁的更新说明，内容包括：

代理任务名称。
确切请求的模型标识，例如
```
gpt-5.6-luna
```
。
确切请求的推理算力级别，例如
```
high
```
或
```
max
```
。
分配的角色，以及该代理是叶子代理还是获授权的协调者。

当有多个代理同时生成时，可将更新合并为一条简洁信息。例如：

text

生成以下代理：
- `evidence_scout` — `gpt-5.6-luna`，推理算力`xhigh` — 只读叶子代理
- `senior_critic` — `gpt-5.6-sol`，推理算力`max` — 资深评审叶子代理

描述内容应与请求的配置一致，而非已验证的运行时身份。若运行时后续报告的实际模型或推理算力与请求不符，需披露修正信息。优先使用

fork_turns: "none"

并明确指定

model

和

reasoning_effort

值，以便配置可见。当任务分支需要继承值时，说明其为继承值，并在已知时指明父代理的确切模型和推理算力；若运行时未暴露这些信息，需明确说明，切勿猜测。

Use These Reusable Personas

使用可复用角色模板

Treat these as assignment shapes, not permanent project agents.

将这些视为任务分配模板，而非永久项目代理。

Evidence scout — Luna xhigh

证据侦察员 — Luna xhigh

Gather a defined evidence set and return paths, commands, timestamps, exact failures, and confidence. Keep it read-only. Good for service health, Sentry or log excerpts, GitHub checks, current diffs, official documentation, and locating relevant code.

收集指定的证据集，返回路径、命令、时间戳、确切错误信息和置信度。保持只读权限。适用于服务健康检查、Sentry或日志片段提取、GitHub检查、当前代码差异、官方文档以及相关代码定位。

Workflow operator — Luna xhigh

工作流操作员 — Luna xhigh

Run and monitor an already-defined, user-authorized workflow with explicit inputs and completion checks. Good for long transfers, transcription and packaging, CI watching, deterministic exports, or repeatable validation. Do not invent scope or make approval decisions.

运行并监控已定义、经用户授权的工作流，需包含明确的输入和完成检查条件。适用于长时间传输、转录与打包、CI监控、确定性导出或可重复验证任务。不得自行扩展任务范围或做出审批决策。

Implementation or review leaf — Luna max

实现或评审叶子代理 — Luna max

Own one bounded code surface or review target, complete the requested work, run targeted checks, and report changed files, evidence, and residual risk. Give distinct file or component ownership.

负责一个有界的代码范围或评审目标，完成请求的工作，运行针对性检查，并报告修改的文件、证据和剩余风险。需明确文件或组件的所有权。

Collaborative workstream owner — Terra max

协作工作流所有者 — Terra max

Own a substantial independent workstream that benefits from a stronger coding peer or further decomposition. Coordinate children only when explicitly authorized; otherwise remain a leaf. Return one integrated result for the assigned workstream.

负责一个重要的独立工作流，该工作流需要更强的编码协作能力或进一步分解。仅在明确获授权时才可协调子代理；否则需作为叶子代理。返回分配工作流的整合结果。

Senior critic — Sol high or max

资深评审员 — Sol high或max

Challenge architecture, diagnosis, security, financial correctness, or release readiness. Ask for the strongest competing explanation and the evidence that would distinguish it. Return a recommendation; keep the final decision with the root and user.

对架构、诊断结果、安全性、财务正确性或发布就绪性提出质疑。要求提供最具竞争力的解释以及区分不同解释的证据。返回建议；最终决策由根代理和用户做出。

Apply Common Delegation Patterns

应用常见委派模式

Production incident: send Luna xhigh for live health/logs and Luna max for the relevant code path; use Terra max only if the investigation becomes a substantial coordinated workstream. Have the root separate underlying cause from resilience mitigation.
PR or release: send Luna xhigh for exact-head diff/check/CI evidence and Luna max for substantive review; keep approval, merge, deploy, and final live verification with the root unless the user explicitly delegates those actions.
Analytics or payments: use separate Luna xhigh scouts for independent sources when useful, then Luna max to reconcile definitions and mismatches; use Sol high or max for money-sensitive ambiguity.
UI work: give Luna xhigh ordinary bounded implementation and Luna max difficult implementation or adaptive browser verification; use Terra max for a larger independent workstream that needs further decomposition.
Long-running creator workflow: give Luna xhigh the established preflight, transfer, transcription, packaging, and monitoring sequence so the root remains available for corrections and approvals.
Consulting or research: use Luna xhigh for source collection, Luna max for evidence normalization, and Sol or the root for the recommendation and tradeoffs.

生产环境事故： 派遣Luna xhigh获取实时健康/日志数据，Luna max分析相关代码路径；仅当调查成为重要的协调工作流时，使用Terra max。由根代理区分根本原因与弹性缓解措施。
PR或发布： 派遣Luna xhigh获取精确的代码差异/检查/CI证据，Luna max进行实质性评审；除非用户明确委派，否则审批、合并、部署和最终实时验证需由根代理负责。
分析或支付： 必要时为独立数据源分别派遣Luna xhigh侦察员，然后由Luna max核对定义和差异；针对涉及资金的模糊场景，使用Sol high或max。
UI工作： 常规有界实现任务交给Luna xhigh，难度较高的实现或自适应浏览器验证交给Luna max；若需要进一步分解的大型独立工作流，使用Terra max。
长时间运行的创建工作流： 将既定的预检查、传输、转录、打包和监控序列交给Luna xhigh，以便根代理可随时响应用户的修正和审批请求。
咨询或研究： 使用Luna xhigh收集数据源，Luna max归一化证据，由Sol或根代理输出建议和权衡分析。

Write the Assignment Contract

编写任务分配协议

Default to

fork_turns: "none"

. Pass only the context required to succeed:

One concrete objective and why it matters.
Exact scope: repository, paths, systems, date range, or source set.
Ownership boundary, including files or surfaces the worker may change.
Constraints from the user, repository, and applicable skills.
Required validation and the expected return format.
Whether the task is read-only or authorized to mutate state.
Whether the worker is a leaf or an explicitly authorized Sol/Terra coordinator. Luna is always a leaf.
The exact requested model and reasoning effort that were disclosed to the user.

Use a history fork only when the subtask genuinely depends on conversation context too dense or fragile to restate. Never send two agents overlapping implementation ownership in the same checkout.

默认使用

fork_turns: "none"

。仅传递任务成功所需的上下文：

一个具体的目标及其重要性。
确切的范围：仓库、路径、系统、日期范围或数据源集合。
所有权边界，包括工作代理可修改的文件或范围。
来自用户、仓库和相关技能的约束条件。
所需的验证步骤和预期的返回格式。
任务是只读还是获授权修改状态。
工作代理是叶子代理还是获明确授权的Sol/Terra协调者。Luna始终为叶子代理。
已向用户披露的确切请求模型和推理算力级别。

仅当子任务确实依赖过于密集或易变的对话上下文，无法重新表述时，才使用历史分支。切勿让两个代理在同一检查点拥有重叠的实现所有权。

Coordinate and Integrate

协调与整合

Respect the current slot limit and use only the agents that have independent work. One precise worker is better than forced fan-out.
Keep at least one useful path moving locally while agents work, unless the delegated operation is the whole task.
Treat every result as evidence. Reopen authoritative live surfaces before claiming merged, deployed, fixed, paid, published, or delivered status.
Follow up with an existing agent when the next task depends on its context; spawn a new one when independence is more valuable.
Stop or redirect work when ownership overlaps, assumptions diverge, or a worker expands scope.
Synthesize disagreements explicitly. State which evidence wins and why.
Preserve user approval boundaries for destructive, production, financial, publishing, messaging, merge, and deployment actions.

Return one integrated answer. Do not dump parallel reports on the user.

遵守当前的代理数量限制，仅使用有独立工作任务的代理。一个精准的代理优于强制拆分任务。
除非委派的操作是整个任务，否则在代理工作期间至少保持一条本地有用的任务路径推进。
将每个结果视为证据。在声称任务已合并、部署、修复、支付、发布或交付之前，重新确认权威实时数据源。
当下一任务依赖现有代理的上下文时，跟进该代理；当独立性更重要时，生成新代理。
当所有权重叠、假设分歧或代理扩展任务范围时，停止或重定向工作。
明确整合不同意见。说明哪项证据更具说服力及原因。
针对破坏性操作、生产环境操作、财务操作、发布操作、消息发送、合并和部署操作，严格遵守用户的审批边界。

返回一个整合后的答案。不得向用户直接输出多个并行报告。