workbench-goal-mode-v2

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Workbench Goal Mode v2

工作台目标模式v2

Use this skill when an issue contains

/goal

GOAL_MODE: yes

GOAL_MODE_V2

, or asks the workbench to complete an objective autonomously across multiple agents, turns, and evidence gates until a real blocker appears.

Goal Mode v2 is a persistent conductor, not a single-agent persistence wrapper. It splits the workflow into two cooperating layers so the human does not need to babysit windows, monitor every agent, or manually switch contexts.

当问题包含

/goal

、

GOAL_MODE: yes

、

GOAL_MODE_V2

，或者要求工作台跨多个Agent、多轮任务和验证节点自主完成目标，直到出现真实阻塞时，使用此技能。

目标模式v2是一个持久化指挥系统，而非单Agent持久化包装器。它将工作流拆分为两个协作层，因此人类无需监控窗口、跟踪每个Agent或手动切换上下文。

Session Communication Bootstrap

会话通信引导

At conductor startup, load

docs/agent-communication-profile.md

and apply:

text

Apply communication profile docs/agent-communication-profile.md.
Tone: human, direct, bilingual, pushback-ok.

This controls communication style only. Runtime identity, model, tools, and MCP availability still require live proof.

指挥系统启动时，加载

docs/agent-communication-profile.md

并应用：

text

Apply communication profile docs/agent-communication-profile.md.
Tone: human, direct, bilingual, pushback-ok.

此配置仅控制沟通风格。运行时身份、模型、工具和MCP可用性仍需实时验证。

Architecture

架构

mermaid

flowchart TB
  H["Operator shorthand / issue"] --> GC["GOAL_CAPTURED"]
  GC --> DSN["DESIGNING"]
  DSN --> DP["DECISION_PACKET"]
  DP --> DSP["DISPATCHING"]
  DSP --> OBS["OBSERVING"]
  OBS -->|agent done| REV["REVIEWING"]
  OBS -->|new evidence| DP
  REV -->|PASS| LRN["LEARNING / ARCHIVING"]
  REV -->|FLAG| DSP
  REV -->|BLOCK| BC["BLOCKER_CLASSIFIED"]
  BC -->|operator| OP["OPERATOR_NEEDED"]
  BC -->|external| CD["COOLDOWN → retry"]
  BC -->|permanent| ARC["ARCHIVED"]
  LRN --> NG["NEXT_GOAL or DONE"]
  OP -->|human responds| DSN
  CD -->|cooldown elapsed| DSN

mermaid

flowchart TB
  H["Operator shorthand / issue"] --> GC["GOAL_CAPTURED"]
  GC --> DSN["DESIGNING"]
  DSN --> DP["DECISION_PACKET"]
  DP --> DSP["DISPATCHING"]
  DSP --> OBS["OBSERVING"]
  OBS -->|agent done| REV["REVIEWING"]
  OBS -->|new evidence| DP
  REV -->|PASS| LRN["LEARNING / ARCHIVING"]
  REV -->|FLAG| DSP
  REV -->|BLOCK| BC["BLOCKER_CLASSIFIED"]
  BC -->|operator| OP["OPERATOR_NEEDED"]
  BC -->|external| CD["COOLDOWN → retry"]
  BC -->|permanent| ARC["ARCHIVED"]
  LRN --> NG["NEXT_GOAL or DONE"]
  OP -->|human responds| DSN
  CD -->|cooldown elapsed| DSN

Layer 1: Design / Decision

第一层：设计/决策

Owned by the conductor (Admin or delegated design agent). Produces decision packets, not raw execution spam.

Responsibilities:

Refine operator intent into a scoped objective.
Pull constraints, non-goals, blockers, and evidence expectations from durable memory (SYNTHESIS.md, DECISIONS.md, prior issue evidence).
Choose the Friction Tier and applicable gates (SDD, L2 Pressure, Capy check).
Maintain taste, scope, blocker semantics, and evidence standards.
Ask the operator only when a real design trade-off needs human judgment.

由指挥器（管理员或委派的设计Agent）负责。生成决策包，而非原始执行指令。

职责：

将操作者意图细化为限定范围的目标。
从持久化内存（SYNTHESIS.md、DECISIONS.md、过往问题证据）中提取约束条件、非目标、阻塞因素和证据预期。
选择摩擦层级（Friction Tier）和适用的验证节点（SDD、L2 Pressure、Capy检查）。
维持风格、范围、阻塞语义和证据标准。
仅当需要人类判断真实的设计权衡时，才向操作者询问。

Layer 2: Dispatch / Operations

第二层：调度/操作

Owned by the dispatcher (Admin or Supervisor). Converts decision packets into bounded Multica issues, monitors them, and routes follow-ups.

Responsibilities:

Convert one decision packet into one or more bounded issues.
Assign one owner per issue. Mention specialists only for narrow review.
Monitor active / in_review / blocked / done states via
```
multica issue list
```
.
Re-route only when new evidence changes state — never re-sweep unchanged work.
Stop when only operator-call or external-platform blockers remain.

由调度器（管理员或主管）负责。将决策包转换为限定范围的Multica问题，监控这些问题并路由后续任务。

职责：

将一个决策包转换为一个或多个限定范围的问题。
为每个问题分配一位负责人。仅在需要专项评审时提及专家。
通过
```
multica issue list
```
监控活跃/评审中/阻塞/完成状态。
仅当新证据改变状态时重新路由——绝不重复处理未变更的工作。
当仅剩余操作者呼叫或外部平台阻塞时停止。

State Machine

状态机

State	Meaning	Next
`GOAL_CAPTURED`	Shorthand or issue received, not yet designed	`DESIGNING`
`DESIGNING`	Conductor is pulling context, refining intent, checking memory	`DECISION_PACKET` or `OPERATOR_NEEDED`
`DECISION_PACKET`	Scoped route produced, ready to dispatch	`DISPATCHING`
`DISPATCHING`	One or more bounded issues created/assigned	`OBSERVING`
`OBSERVING`	Waiting for agent evidence; periodic status scan	`REVIEWING` , `DECISION_PACKET` (re-route), or `BLOCKER_CLASSIFIED`
`REVIEWING`	Evidence evaluated against closeout gates	`LEARNING` (PASS), `DISPATCHING` (FLAG), or `BLOCKER_CLASSIFIED` (BLOCK)
`BLOCKER_CLASSIFIED`	Blocker typed: operator / external / permanent	`OPERATOR_NEEDED` , cooldown → `DESIGNING` , or `ARCHIVED`
`LEARNING / ARCHIVING`	Evidence harvested, noise archived, memory updated	`NEXT_GOAL` or `DONE`
`OPERATOR_NEEDED`	Waiting for human input	`DESIGNING` (on response)
`DONE`	Objective achieved, all gates passed	terminal

状态	含义	下一状态
`GOAL_CAPTURED`	收到简写指令或问题，尚未开始设计	`DESIGNING`
`DESIGNING`	指挥器正在获取上下文、细化意图、检查内存	`DECISION_PACKET` 或 `OPERATOR_NEEDED`
`DECISION_PACKET`	已生成限定范围的执行路径，准备调度	`DISPATCHING`
`DISPATCHING`	已创建/分配一个或多个限定范围的问题	`OBSERVING`
`OBSERVING`	等待Agent提交证据；定期扫描状态	`REVIEWING` 、 `DECISION_PACKET` （重新路由）或 `BLOCKER_CLASSIFIED`
`REVIEWING`	根据收尾验证节点评估证据	`LEARNING` （通过）、 `DISPATCHING` （标记）或 `BLOCKER_CLASSIFIED` （阻塞）
`BLOCKER_CLASSIFIED`	已分类阻塞类型：操作者/外部/永久	`OPERATOR_NEEDED` 、冷却后→ `DESIGNING` 或 `ARCHIVED`
`LEARNING / ARCHIVING`	已收集证据、归档无效信息、更新内存	`NEXT_GOAL` 或 `DONE`
`OPERATOR_NEEDED`	等待人类输入	`DESIGNING` （收到响应后）
`DONE`	已达成目标，所有验证节点通过	终止

Decision Packet Format

决策包格式

Every decision packet must be posted as a structured comment before any issues are created:

text

DECISION_PACKET
goal_id: <goal-shorthand-or-issue-id>
intent: <one-sentence objective>
route: <chosen execution path — which agents, skills, lanes>
owner: <primary executing agent>
reviewer: <who verifies evidence>
constraints: <non-negotiables, boundaries, blocked lanes>
evidence_expectations: <what proves success — exact artifacts or checks>
non_goals: <explicit exclusions>
blocker_conditions: <what would block this route>
tier: fast | standard | heavy
dedupe_key: <canonical key to prevent duplicate dispatch>
max_active: <max concurrent in_progress issues for this goal>
cooldown_minutes: <minimum minutes before next sweep or re-route>
verdict: READY_TO_DISPATCH | NEEDS_DESIGN | OPERATOR_NEEDED

在创建任何问题之前，必须以结构化评论的形式发布每个决策包：

text

DECISION_PACKET
goal_id: <goal-shorthand-or-issue-id>
intent: <one-sentence objective>
route: <chosen execution path — which agents, skills, lanes>
owner: <primary executing agent>
reviewer: <who verifies evidence>
constraints: <non-negotiables, boundaries, blocked lanes>
evidence_expectations: <what proves success — exact artifacts or checks>
non_goals: <explicit exclusions>
blocker_conditions: <what would block this route>
tier: fast | standard | heavy
dedupe_key: <canonical key to prevent duplicate dispatch>
max_active: <max concurrent in_progress issues for this goal>
cooldown_minutes: <minimum minutes before next sweep or re-route>
verdict: READY_TO_DISPATCH | NEEDS_DESIGN | OPERATOR_NEEDED

Dedupe, Cooldown, and Noise Controls

去重、冷却与无效信息控制

These prevent the controller from accumulating endless sweeper issues (per DAS-741 / DAS-743 findings):

这些机制可防止控制器累积无尽的扫描任务（基于DAS-741 / DAS-743的研究结果）：

Dedupe

去重

Before creating any issue, check if an equivalent issue already exists using the
```
dedupe_key
```
. Match on title prefix + active status (not done/cancelled).
If an equivalent issue exists and is
```
in_progress
```
or
```
in_review
```
, do not create a duplicate. Post a status note on the existing issue instead.
If an equivalent issue exists and is
```
blocked
```
, check if the blocker changed before re-dispatching.

在创建任何问题之前，使用
```
dedupe_key
```
检查是否已存在等效问题。匹配标题前缀+活跃状态（非完成/已取消）。
如果等效问题存在且处于
```
in_progress
```
或
```
in_review
```
状态，则不创建重复问题，而是在现有问题上发布状态说明。
如果等效问题存在且处于
```
blocked
```
状态，在重新调度前检查阻塞因素是否已变更。

Cooldown

冷却

After dispatching a decision packet, set
```
cooldown_minutes
```
. Do not scan or re-route for this goal until the cooldown elapses.
Default cooldown: 15 minutes for standard work, 30 minutes for heavy work.
Cooldown resets when a new decision packet is dispatched.

调度决策包后，设置
```
cooldown_minutes
```
。在冷却时间结束前，不扫描或重新路由此目标的任务。
默认冷却时间：标准任务15分钟，重型任务30分钟。
当新决策包被调度时，重置冷却时间。

Max Active

最大活跃任务数

No more than
```
max_active
```
(default: 5) issues may be
```
in_progress
```
for a single goal at any time.
If the limit is reached, new dispatch waits until an active issue moves to
```
in_review
```
or
```
done
```
.

单个目标同时处于
```
in_progress
```
状态的任务不得超过
```
max_active
```
（默认：5）个。
如果达到限制，新的调度需等待活跃任务转为
```
in_review
```
或
```
done
```
状态。

Archive and Cancel Semantics

归档与取消语义

Completed (
```
done
```
) issues: harvest evidence, then archive — do not re-read or re-route.
Stale
```
in_review
```
issues (no activity for 48h): post one nudge, then flag for Supervisor.
Noise issues (created in error, duplicate, superseded): set to
```
cancelled
```
with a one-line reason. Do not leave them in
```
blocked
```
.
The conductor itself must have a self-cancel condition: if the conductor issue is the only remaining active issue and all dispatched work is
```
done
```
or
```
cancelled
```
, post the final verdict and stop.

已完成（
```
done
```
）的任务：收集证据后归档——不再重新读取或路由。
停滞的
```
in_review
```
任务（48小时无活动）：发送一次提醒，然后标记给主管。
无效任务（错误创建、重复、已被取代）：设置为
```
cancelled
```
并附上一行原因说明。不要将其留在
```
blocked
```
状态。
指挥器自身必须具备自取消条件：如果指挥器问题是唯一剩余的活跃问题，且所有已调度任务均为
```
done
```
或
```
cancelled
```
，则发布最终结论并停止运行。

Autonomy Guards

自主操作防护机制

The conductor may act autonomously when:

The decision packet is
```
READY_TO_DISPATCH
```
.
No operator-call conditions are triggered.
Dedupe confirms no duplicate work exists.
The route does not touch blocked lanes (DAS-696, DAS-693, or operator-declared frozen lanes).
Workbench Max is not in the route.

The conductor MUST call the operator when:

A design trade-off needs human taste judgment (not just more context).
A permission, secret, payment, or runtime mutation is required.
A blocked lane is the only viable route.
Two prior attempts at the same sub-task failed with different approaches.
The dedupe key matches an existing active issue with a different decision packet (route conflict).

当满足以下条件时，指挥器可自主操作：

决策包状态为
```
READY_TO_DISPATCH
```
。
未触发操作者呼叫条件。
去重检查确认无重复工作。
执行路径未涉及阻塞通道（DAS-696、DAS-693或操作者声明的冻结通道）。
执行路径中不包含Workbench Max。

当出现以下情况时，指挥器必须呼叫操作者：

设计权衡需要人类的主观判断（而非仅需更多上下文）。
需要权限、密钥、支付或运行时变更。
唯一可行的执行路径涉及阻塞通道。
同一子任务的两次尝试采用不同方法均失败。
去重键匹配现有活跃问题，但决策包不同（路径冲突）。

Composition With Existing Layers

与现有层的组合

Goal Mode v2 composes with, but does not replace:

Layer	How Goal Mode v2 uses it
Friction Tier Router	Selected during DESIGNING; encoded in decision packet tier field
SDD	Decision packet may route heavy work through SDD stages; design layer handles Raw Requirement + Product Design
L2 Pressure Gate	Applied during DESIGNING when objective touches remote/HarnessMax/leaderboard work
Goal Mode v1	v2 replaces the persistence wrapper; v1 skill remains for simple single-agent /goal tasks
Auto Review Sweeper	Dispatch layer uses sweeper results during OBSERVING → REVIEWING transition
Self-Awareness	Run once at GOAL_CAPTURED for heavy-path goals before DESIGNING begins
Completion Cooling	Applied to dispatched child issues at 75/85/90/100 thresholds

目标模式v2与现有层组合使用，但不替代它们：

层	目标模式v2的使用方式
Friction Tier Router	在DESIGNING阶段选择；编码到决策包的tier字段中
SDD	决策包可将重型任务路由至SDD阶段；设计层负责原始需求+产品设计
L2 Pressure Gate	当目标涉及远程/HarnessMax/排行榜任务时，在DESIGNING阶段应用
Goal Mode v1	v2取代了持久化包装器；v1技能仍适用于简单的单Agent /goal任务
Auto Review Sweeper	调度层在OBSERVING → REVIEWING转换期间使用扫描器结果
Self-Awareness	对于重型路径目标，在DESIGNING开始前于GOAL_CAPTURED阶段运行一次
Completion Cooling	在已调度子任务的75/85/90/100进度阈值处应用

Activation

激活方式

When activated by

/goal

GOAL_MODE: yes

, the conductor must:

Apply
```
docs/agent-communication-profile.md
```
before the first user-facing conductor note.
Post
```
GOAL_LOCK
```
with objective, owner, non_goals, closeout gates, and operator-call conditions.
If Heavy Path: post
```
SELF_AWARENESS_BOOTSTRAP
```
before designing.
If L2_PRESSURE: post
```
RV_PRESSURE_CHECK
```
before routing.
Enter
```
DESIGNING
```
→ produce
```
DECISION_PACKET
```
→ enter
```
DISPATCHING
```
.
Continue the observe → review → archive → next-goal loop until DONE or OPERATOR_NEEDED.

当通过

/goal

或

GOAL_MODE: yes

激活时，指挥器必须：

在发布第一条面向用户的指挥器说明之前，应用
```
docs/agent-communication-profile.md
```
。
发布
```
GOAL_LOCK
```
，包含目标、负责人、非目标、收尾验证节点和操作者呼叫条件。
如果是重型路径：在设计前发布
```
SELF_AWARENESS_BOOTSTRAP
```
。
如果是L2_PRESSURE：在路由前发布
```
RV_PRESSURE_CHECK
```
。
进入
```
DESIGNING
```
状态→生成
```
DECISION_PACKET
```
→进入
```
DISPATCHING
```
状态。
继续观察→评审→归档→下一目标的循环，直到DONE或OPERATOR_NEEDED状态。

Closeout Contract

收尾协议

text

GOAL_MODE_V2_CLOSEOUT
goal_id:
objective:
state_machine_path: <trace of states visited, e.g. GOAL_CAPTURED→DESIGNING→...→DONE>
decision_packets_produced: <count>
issues_dispatched: <count and IDs>
evidence_harvested: <summary of completed evidence>
noise_cancelled: <count and IDs of cancelled duplicates>
operator_calls: <count and reasons, if any>
residual_risk:
archive_actions_taken:
verdict: PASS | FLAG | BLOCK

text

GOAL_MODE_V2_CLOSEOUT
goal_id:
objective:
state_machine_path: <trace of states visited, e.g. GOAL_CAPTURED→DESIGNING→...→DONE>
decision_packets_produced: <count>
issues_dispatched: <count and IDs>
evidence_harvested: <summary of completed evidence>
noise_cancelled: <count and IDs of cancelled duplicates>
operator_calls: <count and reasons, if any>
residual_risk:
archive_actions_taken:
verdict: PASS | FLAG | BLOCK

Validation

验证

Before claiming DONE, verify:

Every dispatched issue has a verdict (PASS/FLAG/BLOCK) or is properly cancelled.
No duplicate active issues share the same dedupe_key.
The conductor's self-cancel condition is satisfied.
Evidence is posted on the goal issue, not scattered across child comments.
```
OPERATOR_NEEDED
```
was raised for every case where the conductor lacked authority.

在宣告DONE之前，需验证：

每个已调度的问题均有结论（PASS/FLAG/BLOCK）或已被正确取消。
没有重复的活跃问题共享相同的dedupe_key。
指挥器的自取消条件已满足。
证据已发布在目标问题上，而非分散在子评论中。
所有指挥器缺乏权限的情况均已触发
```
OPERATOR_NEEDED
```
。