← Starchild Wiki

Conductor ModeConductor Mode（模型智选）

A tiny classifier picks the model for every turn — and a versioned rules file, not code, decides how用一个极小的分类器为每一轮选模型——决定权在版本化的规则文件，而不在代码里

Formerly known as “Smart Routing”.前身为 “Smart Routing”，现更名 Conductor Mode（模型智选）。

core/smart_router.py · core/smart_router_provider.py · config/smart_routing/rules.md

01The bet这个赌注

Most turns in a real conversation are easy. Paying frontier prices for all of them is the single largest avoidable cost in an agent.真实对话里大多数回合都很简单。为所有回合支付顶级模型价格，是 Agent 中最大的一笔可避免成本。

Conductor Mode spends one call on a purpose-tuned lightweight classifier to decide which model answers the turn. The classifier is cheap enough that being right ~90% of the time already pays for itself several times over; the design work is all in making the mistakes survivable.Conductor Mode 花一次专门调校的轻量级分类器调用，决定这一轮由谁来答。分类器足够便宜，只要正确率约 90% 就已经数倍回本；真正的设计工作全在于「让判错也能活下来」。

Fig 1.图 1. One classifier call, guarded, resolving to a four-leg catalog.一次分类器调用，经守卫层校正，落到四条腿的模型目录。

02Rules live in Markdown, not in code规则在 Markdown 里，不在代码里

The routing policy is a prompt: workspace/smart_routing/rules.md. Editing it changes behaviour immediately — no deploy, and every user can tune their own copy.路由策略本身就是一段 prompt：workspace/smart_routing/rules.md。改它即刻生效——无需发版，每个用户都可以调整自己的副本。

What the rules actually judge, distilled from production experience:规则真正判断的东西，来自生产环境的沉淀：

• Structure, not surface. A short prompt is not a simple prompt. What predicts difficulty is dependency depth — whether the answer needs one pass or a chain of dependent steps. Topic, tone and length do not.看结构，不看表面。短请求不等于简单请求。真正预示难度的是依赖深度——答案是一步可得，还是需要一串相互依赖的步骤。话题、语气、长度都不算数。
• Fabrication risk routes strong. Any turn whose correctness depends on fetched data — prices, balances, live state — goes to a strong model by default, because a cheap model's confident guess is worse than an expensive model's verified answer.有编造风险的一律走 strong。凡正确性依赖抓取数据的回合——价格、余额、实时状态——默认交给强模型，因为便宜模型的自信猜测比强模型的核实答案更危险。
• Media has its own lane. Image and vision turns route to a multimodal model directly instead of burning a frontier slot.多媒体走独立车道。图像与视觉类回合直接路由到多模态模型，不占用旗舰配额。

03Guards around the classifier分类器周边的守卫

Guard守卫	What it prevents防什么
Tool-result cache工具结果缓存	A turn that is just a tool result reuses the cached lane — no classifier call. Re-classifying mid-tool-loop would let a turn switch models halfway through its own task.纯工具结果的回合复用缓存的车道，不再调分类器。在工具循环中途重新分类，会让一个回合在自己任务执行到一半时换模型。
Session-change reset会话切换重置	The cache is agent-scoped, not thread-scoped. Without a conversation fingerprint, thread A's strong-lane state leaks into thread B's first message. The router fingerprints the conversation and clears on change.缓存作用域是 agent 而非 thread。没有会话指纹的话，线程 A 的 strong 车道状态会泄漏到线程 B 的首条消息。路由器对会话做指纹，变化即清空。
Rules guard / tier snap规则守卫 / 层级吸附	If the classifier names a model that isn't in the rules, it is snapped to a rules-named model of the same catalog tier (budget < standard < premium, rounding up) instead of being taken literally.分类器若点名了规则中没有的模型，会被吸附到同一 catalog 层级中规则里有的模型（budget < standard < premium，向上取整），而不是照字面执行。
OAuth neutralityOAuth 中立	Same-name pairs like openai/gpt-5.5 and openai-codex/gpt-5.5 are disambiguated by exact id or an explicit marker — never by heuristic. The resolver never biases toward or against a user's own subscription.像 openai/gpt-5.5 与 openai-codex/gpt-5.5 这样的同名对，只按精确 id 或显式标记消歧，绝不用启发式。解析器不会偏向或排斥用户自己的订阅额度。
Media safety net多模态兜底	A complex turn carrying an image stays on the strong leg (gpt-5.6-luna, natively multimodal); video turns are substituted to qwen-3.8-max — never demoted to cheap.带图片的复杂回合留在 strong 腿（gpt-5.6-luna，原生多模态）；视频回合替换为 qwen-3.8-max——绝不降级到 cheap。

04Follow-ups inherit the lane追问继承车道

"go", "start", "redo it", "开始", a pasted error screenshot — these carry no complexity of their own. They inherit the complexity of whatever they act on. The classifier sees a short window of recent context (200 chars; benchmarked against 400 and 800, all three scored identically, so the smallest won) purely to resolve what the terse turn refers to.「go」「start」「重来」「开始」、粘贴的报错截图——它们自身没有复杂度，只继承所指向任务的复杂度。分类器会看一小段近期上下文（200 字符；与 400、800 做过对比测试，三者得分完全相同，于是取最小的那个），纯粹用来解析这句简短回合在指什么。

Choosing 200 over 800 when the accuracy is identical is not a micro-optimisation — the context window is prepended to every classifier call, so it's a permanent per-turn tax. Measure, then take the cheapest tie.在准确率相同时选 200 而非 800，不是微优化——这段上下文会前置到每一次分类器调用，是永久性的每轮税。先测，再在打平时取最便宜的那个。

05What it buys, measured实测收益

All numbers come from objective, exact-match benchmarks — no LLM judge, no vendor claims. Two independent lines of evidence:以下数字全部来自客观、精确匹配判分的基准——没有 LLM 裁判，也不是厂商宣称。两条独立证据线：

Composite score across four suites (AIME 2025 · GPQA Diamond · LiveCodeBench-100 · everyday conversation, equal-weighted), Conductor Mode vs. single frontier models on identical questions:四套件综合成绩（AIME 2025 · GPQA Diamond · LiveCodeBench-100 · 日常对话，等权重），Conductor Mode 与单一旗舰模型在完全相同的题目上对比：

Setup方案	Accuracy准确率	Cost / task单题成本	Relative cost相对成本
Kimi K3	93.4%	$0.062	7.0×
GPT-5.6 Sol	93.1%	$0.039	4.4×
Claude Opus 5	92.2%	$0.108	12.1×
Gemini 3.6 Flash	91.5%	$0.071	8.0×
Conductor Mode (v3 final)Conductor Mode（v3 定版）	89.5%	$0.0089	1×1×

Within 3.9 points of the best single model, at 1/4 to 1/12 of frontier cost. Suite results: AIME 2025 100% (majority@3) · GPQA Diamond 82.8% (198/198) · LiveCodeBench 87.9% · everyday 87.3% · multimodal (MMMU) 91.7%. Live figures; the report re-aggregates from raw run files.综合准确率与最强单模型相差 3.9 个百分点，成本仅为旗舰模型的 1/4 – 1/12。各套件成绩：AIME 2025 100%（majority@3）· GPQA Diamond 82.8%（198/198）· LiveCodeBench 87.9%· 日常对话 87.3% · 多模态（MMMU）91.7%。数据为实时值，报告页由原始运行文件重新聚合。

Full interactive results — every suite, every model, per-task costs — are published at the Conductor Mode benchmark report.完整交互式结果——每个套件、每个模型、单题成本——见 Conductor Mode 基准报告。

Everyday conversational turns — greetings, lookups, short factual questions — are where routing pays hardest: Conductor Mode answers them at $0.0011/turn, versus $0.01–0.03 for a frontier single model at comparable quality. That is a 10–25× saving on the most common kind of turn.日常对话回合——问候、查询、短事实问题——是路由收益最大的地方：Conductor Mode 单轮成本 $0.0011，对照旗舰单模型的 $0.01–0.03，质量相当。在最高频的回合类型上省 10–25 倍。

The public ceiling agrees. On RouterBench (36,497 prompts, 11 models, published data), a perfect per-prompt router reaches 91.2% accuracy at $0.24 per thousand prompts — versus GPT-4's 78.1% at $3.29. Routing's headroom is not our invention; it is a measured property of the model landscape. Conductor Mode is our production approximation of that oracle.公开数据给出同样的天花板。RouterBench（36,497 条 prompt、11 个模型的公开数据）显示：完美的逐题路由器可达 91.2% 准确率、每千题 $0.24——对照 GPT-4 的 78.1%、每千题 $3.29。路由的收益空间不是我们的发明，而是模型格局的可测属性；Conductor Mode 是这一理论上限的生产近似。

06Design principles设计原则

Principle原则	Why理由
Policy as prompt策略即 prompt	Routing behaviour is a Markdown file, not code — it evolves at the speed of an edit, and users can tune it.路由行为是一个 Markdown 文件而不是代码——以编辑的速度演进，用户也能自行调整。
Structure, not surface看结构不看表面	Dependency depth predicts difficulty; topic, tone, and length do not.依赖深度预示难度；话题、语气、长度都不。
Fabrication is a correctness failure编造是正确性失败	Anything whose correctness depends on fetched data is strong by default.凡正确性依赖抓取数据的，默认 strong。
Escalate a step, not a turn升级一步而非整轮	The hard tail gets one advisor call, not a whole conversation at frontier price.困难长尾用一次 advisor 调用解决，而不是让整段对话按顶级模型计费。
Take the cheapest tie打平时取最便宜	When two configurations score identically, the cheaper one wins — every recurring token is a permanent per-turn tax.当两个配置得分相同时，取更便宜的那个——每个重复出现的 token 都是永久性的每轮税。