workbench-self-awareness-infra

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Workbench Self-Awareness Infra

工作台自我感知基础设施

Use this skill when the Friction Tier Router selects Heavy Path, when repo/runtime ownership is ambiguous, or when Standard Path work depends on current runtime capability.

Self-awareness is the workbench's heavy-risk boot layer. It prevents agents from confusing memory, old sessions, model assumptions, or job-start events with current evidence without forcing low-risk Fast Path work through ceremony.

当Friction Tier Router选择Heavy Path、代码库/运行时归属权不明确，或者Standard Path工作依赖当前运行时能力时，使用此技能。

自我感知是工作台的高风险启动层。它可防止Agent将内存、旧会话、模型假设或任务启动事件与当前证据混淆，同时无需让低风险的Fast Path工作遵循繁琐流程。

Activation

激活条件

Use this layer when any of these apply:

Friction Tier Router selected
```
HEAVY_PATH
```
;
repo, runtime, branch, issue, or owner is ambiguous;
```
STANDARD_PATH
```
evidence depends on live runtime capability;

the issue contains

GOAL_MODE: yes

L2_PRESSURE: yes

/goal

HarnessMax

remote

VM

Research Vault

MCP

, or autonomous language;

the task may change a live skill, agent, autopilot, runtime, deploy surface, payment/OAuth/secrets boundary, branch/merge state, or public proof surface;
the agent is starting from a new session, rerun, stale context, or external handoff and cannot cheaply verify the current anchor;
the task asks for the highest-yield route, leaderboard pressure, or full auto execution.

Do not use this layer for

FAST_PATH

work unless repo/runtime ownership is ambiguous.

满足以下任一条件时使用此层：

Friction Tier Router 选择了
```
HEAVY_PATH
```
；
代码库、运行时、分支、问题或归属权不明确；
```
STANDARD_PATH
```
的证据依赖实时运行时能力；

问题中包含

GOAL_MODE: yes

、

L2_PRESSURE: yes

、

/goal

、

HarnessMax

、

remote

、

VM

、

Research Vault

、

MCP

或自主执行相关表述；

任务可能会修改实时技能、Agent、自动驾驶仪、运行时、部署面、支付/OAuth/密钥边界、分支/合并状态或公开验证面；
Agent从新会话、重跑、过时上下文或外部交接启动，且无法低成本验证当前锚点；
任务要求最高收益路径、排行榜压力或完全自动执行。

除非代码库/运行时归属权不明确，否则不要将此层用于

FAST_PATH

工作。

Required Block

必填模块

Post or maintain this block before routing, implementation, or review when the selected tier requires it:

text

SELF_AWARENESS_BOOTSTRAP
runtime_identity:
role_boundary:
repo_anchor:
tool_envelope:
mcp_envelope:
memory_sources_checked:
current_state_proof:
risk_envelope:
routing_decision:
success_metric:
operator_call_conditions:
verdict: READY | FLAG | BLOCK

当所选层级要求时，在路由、实施或评审前发布或维护以下模块：

text

SELF_AWARENESS_BOOTSTRAP
runtime_identity:
role_boundary:
repo_anchor:
tool_envelope:
mcp_envelope:
memory_sources_checked:
current_state_proof:
risk_envelope:
routing_decision:
success_metric:
operator_call_conditions:
verdict: READY | FLAG | BLOCK

Field Rules

字段规则

```
runtime_identity
```
: name the runtime family and execution cell, without exposing private IDs, tokens, direct IPs, or raw environment output.
```
role_boundary
```
: name what this agent owns and what it must not take over.
```
repo_anchor
```
: name the project-bound repo/resource, branch, and whether any local path is authoritative or only fallback evidence.
```
tool_envelope
```
: list the relevant tools actually available or checked.
```
mcp_envelope
```
: list the relevant MCP/connectors visible for this run; mark unavailable tools as
```
missing
```
instead of pretending they exist.
```
memory_sources_checked
```
: distinguish current repo state from advisory memory.
```
current_state_proof
```
: include small proof such as
```
git status
```
, issue JSON, run status, docs read, or command exit status.
```
risk_envelope
```
: state secrets, destructive actions, public/private boundary, runtime mutation, cost, and irreversible release risks.
```
routing_decision
```
: decide inline execution, Multica parent issue, child issues, remote runtime, VM lane, or Supervisor review.
```
success_metric
```
: state the artifact that counts: merged PR, verified run, build/test pass, shipped doc, closed issue, or proven blocker.
```
operator_call_conditions
```
: list the few cases that justify stopping for the human.

```
runtime_identity
```
：指定运行时家族和执行单元，不得暴露私有ID、令牌、直接IP或原始环境输出。
```
role_boundary
```
：明确此Agent的管辖范围以及不得接管的内容。
```
repo_anchor
```
：指定项目绑定的代码库/资源、分支，以及本地路径是否为权威数据源或仅作为 fallback 证据。
```
tool_envelope
```
：列出实际可用或已检查的相关工具。
```
mcp_envelope
```
：列出本次运行可见的相关MCP/连接器；将不可用的工具标记为
```
missing
```
，而非假装其存在。
```
memory_sources_checked
```
：区分当前代码库状态与参考内存。
```
current_state_proof
```
：包含小型验证证据，如
```
git status
```
、问题JSON、运行状态、已读文档或命令退出状态。
```
risk_envelope
```
：说明密钥、破坏性操作、公/私边界、运行时变更、成本以及不可逆发布风险。
```
routing_decision
```
：决定内联执行、Multica父问题、子问题、远程运行时、VM通道或主管评审。
```
success_metric
```
：明确判定成功的产物：合并的PR、已验证的运行、构建/测试通过、已发布的文档、已关闭的问题或已确认的阻塞点。
```
operator_call_conditions
```
：列出少数需要暂停并寻求人工介入的情况。

Source Order

数据源优先级

Use the newest current source first:

Active issue description, latest relevant comments, and current run status.
Project-bound GitHub repo or explicitly attached repo resource.
Runtime-local
```
git status
```
, branch, recent commit, and changed files.
Live tool/MCP inventory for the current runtime.
Research Vault or workbench memory as advisory pressure, not as truth.
Historical docs, logs, and old summaries only when the task needs them.

If memory conflicts with current repo or issue evidence, current evidence wins and the conflict becomes residual risk.

优先使用最新的当前数据源：

活跃问题描述、最新相关评论和当前运行状态。
项目绑定的GitHub代码库或明确附加的代码库资源。
运行时本地的
```
git status
```
、分支、最近提交和已更改文件。
当前运行时的实时工具/MCP清单。
Research Vault或工作台内存仅作为参考依据，而非事实。
仅当任务需要时才使用历史文档、日志和旧摘要。

如果内存与当前代码库或问题证据冲突，以当前证据为准，冲突将成为残留风险。

Routing Rules

路由规则

Use the bootstrap to route work instead of becoming the bottleneck.

If two or more independent tasks exist, create or use Multica child issues.
If the task is high-pressure remote or HarnessMax work, run L2 Pressure after this bootstrap and before implementation.
If the task needs a disposable browser, GUI, sandbox, or screenshot-backed proof, route to the VM lane with a lease.
If the task needs product or architecture clarification, use SDD before code.
If the task is already implemented and waiting for evidence, route to Supervisor or QA review instead of re-implementing it.
If the task only needs a small local patch, execute inline and report evidence.

使用引导模块来路由工作，而非成为瓶颈。

如果存在两个或更多独立任务，创建或使用Multica子问题。
如果任务是高压力远程或HarnessMax工作，在引导之后、实施之前运行L2 Pressure。
如果任务需要一次性浏览器、GUI、沙箱或基于截图的验证，路由到带租赁的VM通道。
如果任务需要产品或架构澄清，在编码前使用SDD。
如果任务已实施并等待验证，路由到主管或QA评审，而非重新实施。
如果任务仅需要小型本地补丁，内联执行并报告证据。

Readiness Verdicts

就绪判定

READY

means the agent has enough current evidence, tools, repo anchor, risk boundary, and success metric to proceed.

FLAG

means the agent can proceed with a bounded caveat. Examples: fallback memory source, missing non-critical MCP, repo anchor is usable but not ideal, or one verification gate must be deferred with rationale.

BLOCK

means execution must not proceed until a real external blocker is fixed. Examples: missing credentials, invalid repo anchor for a repo-changing task, destructive action needs approval, or the required runtime/tool is unavailable.

READY

表示Agent拥有足够的当前证据、工具、代码库锚点、风险边界和成功指标，可以继续执行。

FLAG

表示Agent可以在有限制条件下继续执行。例如：使用fallback内存源、缺少非关键MCP、代码库锚点可用但非最优，或需延迟一个验证门并说明理由。

BLOCK

表示在解决真实外部阻塞点之前不得继续执行。例如：缺少凭据、针对修改代码库的任务使用无效锚点、破坏性操作需要批准，或所需运行时/工具不可用。

Anti-Patterns

反模式

Do not treat "job started" as success.
Do not treat a scheduled tick as the main execution path when the session can run the first slice now.
Do not paste raw environment dumps, tokens, cookies, request payloads, or full logs into durable docs.
Do not say a tool or MCP exists because it existed in a different runtime.
Do not use historical memory as proof of current branch, PR, issue, or CI state.
Do not silently widen into runtime, daemon, Desktop UI, or preserved-agent mutation.

不要将“任务已启动”视为成功。
当会话现在可以运行第一部分时，不要将定时调度作为主要执行路径。
不要将原始环境转储、令牌、Cookie、请求负载或完整日志粘贴到持久文档中。
不要因为某工具或MCP在其他运行时存在就声称其在当前运行时存在。
不要将历史内存作为当前分支、PR、问题或CI状态的证据。
不要擅自扩大到运行时、守护进程、桌面UI或保留Agent的变更操作。

Closeout

收尾工作

For tasks that used this layer, final reports should say whether the bootstrap changed routing or risk posture:

text

SELF_AWARENESS_CLOSEOUT
bootstrap_verdict:
routing_used:
capability_gap_found:
artifact_or_blocker:
residual_risk:
next_slice_started:

The strongest closeout is not a larger summary. It is a shipped artifact, real verification, merged PR, closed issue, or a precise blocker with the smallest operator action needed.

对于使用过此层的任务，最终报告应说明引导模块是否改变了路由或风险态势：

text

SELF_AWARENESS_CLOSEOUT
bootstrap_verdict:
routing_used:
capability_gap_found:
artifact_or_blocker:
residual_risk:
next_slice_started:

最有力的收尾不是冗长的总结，而是已交付的产物、真实的验证、合并的PR、已关闭的问题，或一个明确的阻塞点以及所需的最小人工操作。