fable-council

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Fable Council

Convene a council of AI coding agents before finalizing an architectural plan. Propose → hear independent assessments from each installed agent CLI → revise, repeated across bounded deliberation rounds.

在最终确定架构计划之前召集 AI 编码 Agent 委员会进行审议。提出方案 → 获取每个已安装 Agent CLI 的独立评估 → 修改方案，并按指定的审议轮次重复这一过程。

Purpose

用途

Enforce multi-agent deliberation and cross-model consensus on complex plans. Leverage complementary models (Claude, Codex, Cursor, Gemini, Grok, OpenCode) to detect blind spots, challenge flawed assumptions, and converge on robust engineering plans.

在复杂方案中落实多 Agent 审议与跨模型共识。借助不同模型的优势（Claude、Codex、Cursor、Gemini、Grok、OpenCode）发现盲点、质疑有缺陷的假设，并形成稳健的工程计划。

When to Use

使用场景

When planning high-stakes architectural changes, major refactorings, or migrations.
When the user asks to convene the council or get second opinions ("ask the council", "consult the other agents").
When deliberating between multiple competing implementation strategies.

规划高风险的架构变更、大规模重构或迁移时。
用户要求召集委员会或获取第二意见时（“ask the council”、“consult the other agents”）。
在多个相互 competing 的实施策略之间进行审议时。

When NOT to Use

不适用场景

Bounded single-file bugfixes or routine implementation tasks (use
```
fable-execute
```
or
```
fable-tdd
```
).
When no other agent CLIs are installed on the local system (fall back to solo
```
fable-plan
```
).
Post-implementation diff verification or code review (use
```
fable-review
```
or
```
fable-judge
```
).

范围有限的单文件 Bug 修复或常规实施任务（使用
```
fable-execute
```
或
```
fable-tdd
```
）。
本地系统未安装其他 Agent CLI 时（回退到单独使用
```
fable-plan
```
）。
实施后的差异验证或代码审查（使用
```
fable-review
```
或
```
fable-judge
```
）。

Inputs

输入

```
draft_plan
```
: The lead agent's initial architectural proposal.
```
member_clis
```
: List of detected or user-specified agent CLI binaries on the host machine.
```
rounds_count
```
: Number of deliberation turns requested (default: 1).

```
draft_plan
```
：主导 Agent 最初提出的架构方案。
```
member_clis
```
：在主机上检测到或由用户指定的 Agent CLI 二进制文件列表。
```
rounds_count
```
：请求的审议轮数（默认：1）。

Expected Outputs

预期输出

```
counciled_plan
```
: The final synthesized plan refined through multi-agent critiques.
```
deliberation_transcript
```
: Structured conversational record of member critiques, adopted changes, and rejected arguments.

```
counciled_plan
```
：经过多 Agent 批评与完善后综合形成的最终计划。
```
deliberation_transcript
```
：记录成员批评、已采纳变更和已拒绝论点的结构化对话记录。

Procedure

执行流程

Parse Request: Extract the core task, requested number of turns (X), and identify the host lead runtime (
```
SELF
```
).
Detect Members: Probe available CLIs on PATH (
```
claude
```
,
```
codex
```
,
```
cursor-agent
```
,
```
gemini
```
,
```
grok
```
,
```
opencode
```
) minus
```
SELF
```
.
Initialize Session: Create
```
~/.council/<timestamp>-<slug>
```
and persist the verbatim task to
```
task.md
```
and initial plan to
```
plan-v1.md
```
.
Deliberation Loop (1..X):
- Generate prompt per member with strict read-only execution flags.
- Invoke all members concurrently in background processes with a ~300s timeout.
- Collect and parse member responses.
- Revise the plan: adopt valid critiques, reject flawed suggestions with explicit technical reasoning.
- Produce
```
plan-v(R+1).md
```
  and
```
changes-v(R+1).md
```
  .
Format Final Answer: Present the deliberation conversation followed by the finalized consensus plan.

解析请求：提取核心任务、请求的轮数（X），并确定主机主导运行时（
```
SELF
```
）。
检测成员：探测 PATH 中可用的 CLI（
```
claude
```
、
```
codex
```
、
```
cursor-agent
```
、
```
gemini
```
、
```
grok
```
、
```
opencode
```
），并排除
```
SELF
```
。
初始化会话：创建
```
~/.council/<timestamp>-<slug>
```
，将原始任务保存到
```
task.md
```
，将初始计划保存到
```
plan-v1.md
```
。
审议循环（1..X）：
- 使用严格的只读执行参数为每位成员生成提示词。
- 在后台进程中并行调用所有成员，超时时间约为 300 秒。
- 收集并解析成员回复。
- 修改计划：采纳合理的批评，并以明确的技术理由拒绝有缺陷的建议。
- 生成
```
plan-v(R+1).md
```
  和
```
changes-v(R+1).md
```
  。
格式化最终答复：先展示审议对话，然后展示最终形成的共识计划。

Decision Rules

决策规则

The lead agent never implements changes during a council run; planning and deliberation only.
Council members are invoked in read-only mode (
```
--mode plan
```
,
```
-s read-only
```
,
```
--skip-trust
```
).
All member suggestions are evaluated on technical merit, not model authority.
Retaining original positions is permitted when supported by concrete technical rationale.

委员会运行期间，主导 Agent 不得实施任何变更；仅进行规划和审议。
委员会成员必须以只读模式调用（
```
--mode plan
```
、
```
-s read-only
```
、
```
--skip-trust
```
）。
所有成员建议均依据技术价值进行评估，而非依据模型的权威性。
如果有具体技术依据支持，可以保留原有立场。

Tool Policy

工具策略

Execute member CLI binaries in parallel using non-interactive flags.
Enforce strict timeouts (e.g. via Perl alarm or process managers on macOS).
Pass system prompt overrides to headless models (e.g. Grok) to prevent unauthorized tool-use stalls.

使用非交互式参数并行执行成员 CLI 二进制文件。
实施严格的超时控制（例如在 macOS 上使用 Perl alarm 或进程管理器）。
向无界面模型传递系统提示词覆盖项（例如 Grok），以防止未经授权的工具调用导致停滞。

Evidence Requirements

证据要求

Deliberation session artifacts saved under
```
~/.council/
```
.
Verbatim quotes and rationale for accepted vs rejected recommendations.

审议会话产物保存在
```
~/.council/
```
下。
保留原话引用，以及接受或拒绝各项建议的技术依据。

Failure Handling

失败处理

If a member CLI times out or errors, record the timeout and proceed without blocking the round.
If no secondary agent CLIs exist on the machine, explicitly declare a solo un-counciled plan.

如果某个成员 CLI 超时或报错，则记录超时情况，并在不阻塞该轮审议的情况下继续执行。
如果机器上不存在其他 Agent CLI，则明确声明这是未经委员会审议的独立计划。

Completion Criteria

完成标准

All X deliberation rounds completed without deadlock.
Deliberation transcript and final synthesized plan presented clearly.
Deliverable ready for human sign-off before any execution begins.

所有 X 轮审议均已完成，且未发生死锁。
清晰展示审议记录和最终综合计划。
交付物已准备好，可在任何执行开始前交由人工确认。