routing

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese
Pick the agent by its description, then its row under Models. Always pass
model
as
<id>#<effort>
.
先根据 Agent 的描述选择 Agent,再查看 模型 下对应的行。始终将
model
传为
<id>#<effort>
。

Handoffs

交接

In every handoff, ask: "Return proof that the work satisfies the request."
A spec passes the spec test: a careful newcomer could finish from the prompt and tree alone, making no design decisions. Send it to
general
, ending with: "Follow the spec. If its approach doesn't fit the code, stop and report why."
A brief gives the outcome, constraints, and completion criteria, leaving design decisions to the child. Send it to
specialist
, asking for objections with evidence.
A second opinion goes to a fresh
critic
whose maker differs from that of the model that wrote the plan or diff under review.
在每次交接中,都要要求:“返回能够证明工作满足请求的证据。”
若一位谨慎的新成员仅凭 prompt 和 tree 就能完成工作,且无需做出任何设计决策,则该 spec 通过 spec 测试。将其交给
general
,并在末尾附上:“遵循该 spec。如果其方案不适合代码,则停止并说明原因。”
brief 应包含预期结果、约束条件和完成标准,同时将设计决策留给子 Agent。将其交给
specialist
,并要求其提供有证据支持的反对意见。
第二意见 应交给一个全新的
critic
,其 Maker(创建者)必须不同于编写待审查计划或 diff 的模型。

Models

模型

Use these unless the user names another. Verify the ID in the model catalog. Maker means creator, not serving provider. Only listed efforts are allowed.
Benchmark gives each row's score, with its USD cost per task in parentheses: FrontierCode 2.0 (would a maintainer merge the change) from Cognition and Artificial Analysis (AA), or OSWorld 2.1 offline (multistep desktop tasks) as reported by OpenAI.
Model#effortMakerAgentsBenchmarkOpenCode model ID
deepseek#highDeepSeek
explore
FrontierCode: Cognition 18.2% ($0.27), AA 16.8% ($0.35)
acp-cmd/deepseek/deepseek-v4.1-flash
deepseek#maxDeepSeek
general
,
specialist
FrontierCode: Cognition 19.3% ($0.38), AA 28.5% ($0.52)
acp-cmd/deepseek/deepseek-v4.1-flash
terra#highOpenAI
computer-use
OSWorld: 68.7% ($0.51)
acp-codex/gpt-6.1-terra
sol#highOpenAI
critic
FrontierCode: Cognition 34.8% ($0.65), AA 36.7% ($0.63)
acp-codex/gpt-6.2-sol
sol#maxOpenAI
specialist
after escalation
FrontierCode: Cognition 39.0% ($1.03), AA 40.6% ($1.13)
acp-codex/gpt-6.2-sol
opus#highAnthropicFusion orchestrator;
specialist
after escalation when OpenAI is excluded;
critic
of Sol work
FrontierCode: Cognition 34.5% ($2.42), AA 35.9% ($2.49)
acp-claude/opus
除非用户指定其他模型,否则使用以下模型。在模型目录中核验 ID。Maker 指创建者,而不是服务提供商。仅允许使用表中列出的 effort。
Benchmark 给出了每一行的分数,括号中为每项任务的美元成本: 由 Cognition 和 Artificial Analysis (AA) 提供的 FrontierCode 2.0(维护者是否会合并该更改)分数,或由 OpenAI 报告的 OSWorld 2.1 离线(多步骤桌面任务)分数。
Model#effortMaker适用 AgentBenchmarkOpenCode model ID
deepseek#highDeepSeek
explore
FrontierCode: Cognition 18.2% ($0.27), AA 16.8% ($0.35)
acp-cmd/deepseek/deepseek-v4.1-flash
deepseek#maxDeepSeek
general
,
specialist
FrontierCode: Cognition 19.3% ($0.38), AA 28.5% ($0.52)
acp-cmd/deepseek/deepseek-v4.1-flash
terra#highOpenAI
computer-use
OSWorld: 68.7% ($0.51)
acp-codex/gpt-6.1-terra
sol#highOpenAI
critic
FrontierCode: Cognition 34.8% ($0.65), AA 36.7% ($0.63)
acp-codex/gpt-6.2-sol
sol#maxOpenAI升级后使用
specialist
FrontierCode: Cognition 39.0% ($1.03), AA 40.6% ($1.13)
acp-codex/gpt-6.2-sol
opus#highAnthropicFusion 编排器;排除 OpenAI 时,升级后使用
specialist
;用于审查 Sol 工作的
critic
FrontierCode: Cognition 34.5% ($2.42), AA 35.9% ($2.49)
acp-claude/opus

Escalation

升级处理

  • Before escalating, split the task and spec the parts that pass the spec test.
  • A brief that fails on deepseek#max moves to sol#max, or to opus#high when OpenAI is excluded. Past that, report the blocker.
  • A spec that fails on
    general
    gets a sharper spec, or goes to another agent as a brief.
  • 升级前,先拆分任务,并为其中通过 spec 测试的部分编写 spec。
  • 若 brief 在 deepseek#max 上未通过,则转至 sol#max;如果排除了 OpenAI,则转至 opus#high。若仍无法通过,报告阻塞原因。
  • 若 spec 在
    general
    上未通过,则应进一步明确 spec,或将其作为 brief 交给另一个 Agent。

Failures

失败处理

Provider error: retry once, then switch models. Usage limit: switch now and exclude that provider's models.
提供商错误:重试一次,然后切换模型。使用量上限:立即切换模型,并排除该提供商的模型。