codex-model-routing-team

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Codex 模型路由团队

Codex Model Routing Team

主 Agent 保持当前模型,负责任务拆分、所有权、集成和最终验收。两个以上 Worker 先编译 TeamPlan;默认路由为
native_subagent/gpt-5.6-luna/xhigh/standard
,高风险升
max
。App Thread 是有状态工作区与耐久任务的专用 Surface,不再是 Luna 的兼容入口。
The main Agent retains the current model and is responsible for task splitting, ownership, integration, and final acceptance. For tasks requiring two or more Workers, compile TeamPlan first; the default routing is
native_subagent/gpt-5.6-luna/xhigh/standard
, and upgrade to
max
for high-risk tasks. App Thread is a dedicated Surface for stateful workspaces and durable tasks, and is no longer a compatibility entry for Luna.

不使用

Do Not Use

简单问答、状态查询、单文件小改、强顺序任务和不可逆外部操作留在主任务;Worker 只能准备外部动作材料。
Simple Q&A, status queries, small single-file modifications, strongly sequential tasks, and irreversible external operations remain in the main task; Workers can only prepare materials for external actions.

执行模式

Execution Modes

  • native-v2
    (默认):用原生 V2 创建 Luna leaf Worker;fresh context 使用
    fork_turns="none"
    ,需要少量最近上下文时写正整数。三类 JSON 默认从 stdin 校验。
  • durable-app
    :任务需要独立 worktree、侧栏可见、跨任务恢复、长期监督,或原生路径缺少精确 live 能力时,使用 App Thread 与耐久账本。
  • 上游 Skill 已定义拆分、阶段和产物时,遵守 适配协议,不重做 Scale、阶段门或第二套业务账本。
  • native-v2
    (default): Use native V2 to create Luna leaf Workers; use
    fork_turns="none"
    for fresh context, and write a positive integer when a small amount of recent context is needed. Three types of JSON are validated from stdin by default.
  • durable-app
    : Use App Thread and durable ledger when the task requires independent worktree, sidebar visibility, cross-task recovery, long-term supervision, or when the native path lacks precise live capabilities.
  • When the upstream Skill has defined splitting, phases, and deliverables, comply with the Adapter Protocol and do not rework Scale, phase gates, or a second set of business ledgers.

执行流程

Execution Process

  1. 确认显式或长期授权;只有两个以上独立交付物且净收益为正时派遣,否则
    lead_only
  2. 两个以上 Worker 按 TeamPlan 协议 编译 unit、依赖、所有权、交付物、验收和集成顺序,并运行
    scripts/validate_team_plan.py
    ;已有上游计划时只编译、不重写。
  3. registryProvider路由Surface 策略固定候选链。
  4. 每个 unit 生成
    schema_version: "3.0"
    RoutePlan,写
    surface_intent: parent_integrated|durable_app
    并运行
    scripts/validate_route_plan.py
    。原生候选必须显式写
    fork_turns
    和 tuple-bound
    runtime_evidence
    ;Fast 还必须有当前 live
    service_tier=priority
    证据。
  5. 任务包 写 unit、唯一
    task_id
    、权限、验收和禁止下级派遣;简报精确路由、职责、fallback 与 reserved slots。
  6. 原生路径遵守 生命周期;App 路径遵守 Thread 生命周期监督协议
  7. TeamPlan 默认
    standard
    6/8/3;
    expanded
    12/16/6 需 live 容量门、2 个 reserved slots 和 10 分钟内的结构化证据。child slots 等于总槽位减协调者与活动 Worker,按较小值切波。
  8. 每 unit 最多 2 次 attempt、一次 follow-up;失败只沿 预声明链。结构变化才在收口后修订 TeamPlan。
  9. 主 Agent 验证集成;原生 Worker 按 live 能力 close 或 completed-idle 后写
    RELEASED
    ,App Thread 过门后归档。运行
    scripts/validate_team_ledger.py
  1. Confirm explicit or long-term authorization; dispatch only when there are two or more independent deliverables with positive net benefits, otherwise use
    lead_only
    .
  2. For two or more Workers, compile unit, dependencies, ownership, deliverables, acceptance, and integration sequence according to the TeamPlan Protocol, and run
    scripts/validate_team_plan.py
    ; only compile without rewriting when an upstream plan already exists.
  3. Fix the candidate chain according to registry, Provider, Routing and Surface Selection strategies.
  4. Generate
    schema_version: "3.0"
    RoutePlan for each unit, write
    surface_intent: parent_integrated|durable_app
    and run
    scripts/validate_route_plan.py
    . Native candidates must explicitly write
    fork_turns
    and tuple-bound
    runtime_evidence
    ; Fast must also have current live
    service_tier=priority
    evidence.
  5. The Task Packet includes unit, unique
    task_id
    , permissions, acceptance criteria, and prohibits subordinate dispatch; the briefing specifies precise routing, responsibilities, fallback, and reserved slots.
  6. Native path complies with Lifecycle; App path complies with Thread Lifecycle and Supervision Protocol.
  7. TeamPlan uses
    standard
    6/8/3 by default;
    expanded
    12/16/6 requires live capacity gate, 2 reserved slots, and structured evidence within 10 minutes. Child slots equal total slots minus coordinator and active Workers, and split waves by the smaller value.
  8. Each unit allows a maximum of 2 attempts and 1 follow-up; failure only follows the Pre-declared Chain. Revise TeamPlan only after收口(这里应该是“conclusion”?不对,原文是“收口后”,应该是“after conclusion”?不,看原文是“结构变化才在收口后修订 TeamPlan”,翻译成“Revise TeamPlan only after conclusion when structural changes occur.”)
  9. The main Agent validates integration; after native Workers close or enter completed-idle state according to live capabilities, write
    RELEASED
    ; archive App Thread after passing the gate. Run
    scripts/validate_team_ledger.py
    .

硬门

Hard Gates

  • registry 决定策略范围;live schema 只证明当前 host 接受精确组合。requested/accepted/observed 分开记录,未回显为
    unknown
  • V2 父 Agent 可创建 picker 可见且未禁用的 V1 leaf model;Luna 可走原生 V2但不获协作工具。Sol/Terra 也禁止下级派遣。
  • 不加
    model: luna
    frontmatter;编排入口留在协作父 Agent,Luna 只做 Worker。
  • Luna 最低 XHigh,高风险 Max;Sol 最低 High;Terra 仅显式首项;Grok 过门;Gemini blocked。禁止旧模型、Ultra 和低强度 fallback。
  • Fast 即
    service_tier=priority
    ;live schema 无字段时一律 Standard,不把 catalog 或请求值冒充 observed Fast。
  • app_thread
    只用于 worktree、侧栏、跨任务恢复、耐久监督或预声明 fallback。
  • Worker 不得继续派生任务,也不得执行发布、发送、付款、删除、账户或生产变更。主 Agent 不切换自身模型。
  • TeamPlan 默认不创建 Planner、不调用重型计划、不落持久文件;同波写冲突、依赖环、超预算、计划外 Worker 或下放最终验收必须拒绝。
  • 未确认返回值或
    pendingWorktreeId
    不得当正式身份;
    UNKNOWN
    禁止追问、归档、fallback、重复创建或改数据库。
  • Registry determines the scope of strategies; live schema only proves that the current host accepts the exact combination. Record requested/accepted/observed separately, and mark unechoed items as
    unknown
    .
  • V2 parent Agent can create V1 leaf models that are visible to picker and not disabled; Luna can use native V2 but does not get collaboration tools. Sol/Terra also prohibit subordinate dispatch.
  • Do not add
    model: luna
    frontmatter; the orchestration entry remains in the collaborative parent Agent, and Luna only acts as a Worker.
  • Luna requires at least XHigh, Max for high-risk tasks; Sol requires at least High; Terra only for explicit first items; Grok requires passing gates; Gemini is blocked. Prohibit old models, Ultra, and low-intensity fallback.
  • Fast means
    service_tier=priority
    ; if there is no such field in the live schema, use Standard uniformly, and do not pass off catalog or requested values as observed Fast.
  • app_thread
    is only used for worktree, sidebar, cross-task recovery, durable supervision, or pre-declared fallback.
  • Workers shall not continue to derive tasks, nor execute release, send, payment, deletion, account, or production changes. The main Agent shall not switch its own model.
  • TeamPlan does not create Planner, call heavy planning, or write persistent files by default; must reject same-wave write conflicts, dependency loops, over-budget, unplanned Workers, or delegated final acceptance.
  • Do not treat unconfirmed return values or
    pendingWorktreeId
    as official identity;
    UNKNOWN
    prohibits follow-up questions, archiving, fallback, repeated creation, or database modification.

输出契约

Output Contract

交付必须完整、自洽并经主 Agent 验证,包含可审计的 Surface、模型、推理强度、速度、上下文范围、Provider 门、尝试、fallback、采纳及收尾状态。行为边界见 验证案例
evals/
Deliverables must be complete, self-consistent, and verified by the main Agent, including auditable Surface, model, inference intensity, speed, context scope, Provider gates, attempts, fallback, adoption, and closing status. See Validation Cases and
evals/
for behavioral boundaries.