fable-opus

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Fable Mode — Opus (v3, agent-routed)

Fable Mode — Opus（v3版本，Agent路由式）

v3 change: delegation is enforced structurally, not requested in prose. The orchestrator is a real agent definition (

agents/fable-orchestrator.md

) with no Write/Edit tool — it cannot do the work inline, so "spawn a worker" stops being a suggestion the model can skip. (Change prompted by field report: prose-level "you may spawn workers" almost always ran inline on the main thread.)

If a task has one obvious correct approach and fits in a single pass, skip this loop and do it directly.

v3版本变更：任务委派通过结构强制实现，而非通过文本请求。Orchestrator是一个真实的Agent定义（

agents/fable-orchestrator.md

），无写入/编辑工具——它无法在线完成工作，因此「生成Worker」不再是模型可以忽略的建议。（此次变更源于实际场景反馈：文本层面的「你可以生成Worker」几乎总是在主线程中在线执行。）

如果任务有明确的正确解法且可通过单轮完成，请跳过此流程直接处理。

How to run it

运行方法

Confirm the fable agents are installed (
```
fable-orchestrator
```
,
```
fable-worker-sonnet
```
,
```
fable-worker-haiku
```
,
```
fable-verifier
```
appear in the available agent types). If they are not, fall back to the inline method: spawn a general-purpose Opus agent and pass it the Core loop and operational rules verbatim from
```
agents/fable-orchestrator.md
```
.
Spawn @fable-orchestrator via the Task tool (
```
subagent_type: "fable-orchestrator"
```
). Brief it with: the user's task, the output directory, relevant session context, and any user-set limits (warning threshold, worker cap, deadline).
Do not restate the Core Loop or operational rules in the briefing — the orchestrator's agent definition carries them. Brief the task, not the method.
When it returns, relay the result, every stage it marked unverified, and its recommendations (surfaced scope it did not build).

确认已安装fable系列Agent（
```
fable-orchestrator
```
、
```
fable-worker-sonnet
```
、
```
fable-worker-haiku
```
、
```
fable-verifier
```
已出现在可用Agent类型列表中）。若未安装，则回退至在线执行方法：生成一个通用Opus Agent，并将
```
agents/fable-orchestrator.md
```
中的核心流程和操作规则原封不动地传递给它。
通过Task工具生成**@fable-orchestrator**（
```
subagent_type: "fable-orchestrator"
```
）。向其简要说明：用户任务、输出目录、相关会话上下文，以及所有用户设定的限制条件（警告阈值、Worker数量上限、截止时间）。
请勿在说明中重复核心流程或操作规则——这些已包含在Orchestrator的Agent定义中。只需说明任务内容，无需说明方法。
当它返回结果时，转达最终成果、所有标记为未校验的阶段，以及它的建议（未构建的已识别范围）。

Known limitation

已知限制

Write-removal closes the front door, not the side door: the orchestrator keeps Bash for running verification commands, and Bash can technically create files. Its definition forbids that use; if audits show it writing through Bash, the next tightening is removing Bash and routing even check-execution through workers (cost: one extra hop per check).

移除写入权限仅关闭了正门，而非侧门：Orchestrator保留了用于运行校验命令的Bash，而从技术上讲Bash可以创建文件。其定义禁止这种用法；若审计发现它通过Bash写入内容，下一步将进一步限制：移除Bash，甚至将校验执行也路由至Worker处理（代价：每次校验多一次跳转）。