feat(routing): add configurable per-prompt model routing

docs-site/src/content/docs/guides/routing-profile-editor.md

@@ -16,6 +16,34 @@ The **Models → Routing** tab in the OpenCodex dashboard can manage `config.rou
 
 Profile ids are immutable after creation. To use a different id, create a new profile and remove the old one after updating callers.
 
+## Route every prompt automatically
+
+Enable **Automatic prompt routing** when one picker entry should choose a different candidate for
+each request. Assign every candidate one or more task tiers:
+
+- **Fast** for greetings, short rewrites, translations, and other lightweight requests.
+- **Balanced** for ordinary questions and general work.
+- **Powerful** for implementation, debugging, architecture, security, multi-step work, and other
+  complex requests.
+
+All three tiers must be covered before the profile can be saved. A candidate may serve more than one
+tier, so the same model can be used for both balanced and powerful requests.
+
+After saving, select the profile's alias (for example `ocx/auto`) in Codex once. OpenCodex
+reclassifies the latest user prompt on **every request**, so a simple follow-up can use the fast
+candidate and the next complex coding request can use the powerful candidate without changing the
+picker again.
+
+Classification is local and deterministic: it does not make a separate model request and does not
+persist the raw prompt as routing metadata. Capability requirements, health, quota, cost limits, and
+the profile's normal scoring rules still apply after the task-tier filter. The selected provider,
+model, tier reason, and exclusions are available in the route-decision trace and
+`ocx logs explain`; OpenCodex does not modify the assistant response with a routing footer.
+
+Prompt routing is provider-agnostic. Any model already configured in OpenCodex can be assigned to a
+tier, including DeepSeek, Qwen, Kimi, GLM, and other OpenAI-compatible or native providers. The
+provider and model ids in the example below are illustrative.
+
 ## Validation and persistence
 
 The dashboard sends the same profile object used by `config.routingProfiles` to the management API. The server validates the complete candidate before writing it:
@@ -55,15 +83,23 @@ Example save payload:
 
 ```json
 {
-  "id": "fast",
+  "id": "smart",
   "mode": "create",
   "profile": {
-    "alias": "ocx/fast",
+    "alias": "ocx/auto",
+    "promptRouting": { "enabled": true },
     "candidates": [
-      { "provider": "anthropic", "model": "claude-sonnet-5" },
-      { "provider": "openai", "model": "gpt-5.6" }
+      {
+        "provider": "deepseek",
+        "model": "deepseek-chat",
+        "taskTiers": ["fast", "balanced"]
+      },
+      {
+        "provider": "openai",
+        "model": "gpt-5.6",
+        "taskTiers": ["powerful"]
+      }
     ],
-    "require": { "tools": true, "minContextWindow": 128000 },
     "optimize": { "latency": 0.55, "health": 0.25, "cost": 0.1, "quota": 0.1 },
     "limits": { "maxEstimatedCostUsd": 0.5, "onUnknownCost": "allow" },
     "unknownEvidence": {

docs-site/src/content/docs/reference/configuration/routing.md

@@ -110,8 +110,9 @@ namespace, or reserved bare native families (`gpt-*`, `o1-*`, `o3-*`, `o4-*`, `c
 
 | Key | Type | Default | Meaning |
 | --- | --- | --- | --- |
-| `candidates` | `{ provider: string; model: string }[]` | required | Explicit allowlist of `provider/model` refs. No implicit expansion. |
+| `candidates` | `{ provider: string; model: string; taskTiers?: ("fast" \| "balanced" \| "powerful")[] }[]` | required | Explicit allowlist of `provider/model` refs. `taskTiers` declares which prompt-complexity tiers a candidate may serve when automatic prompt routing is enabled. No implicit expansion. |
 | `alias?` | `string` | — | Optional public model id in place of `policy/<id>`. |
+| `promptRouting?` | `{ enabled?: boolean }` | disabled | Locally classify the latest user prompt on every request and restrict eligibility to candidates assigned to the resulting task tier. |
 | `require?` | object | `{}` | Hard capability requirements evaluated before scoring (see below). |
 | `optimize?` | object | latency 0.55, health 0.25, cost 0.10, quota 0.10 | Scoring weights, normalized deterministically. `health`, `quota`, and `cost` have score dimensions; the configured-priority share is `1 - health - quota - cost` (default 0.55), and `latency` folds into that priority share rather than scoring independently. |
 | `limits?` | object | — | Hard limits. `maxEstimatedCostUsd` excludes a candidate when its estimated cost is known and above the cap. When that cap is set, `onUnknownCost` (`"allow"` default, or `"exclude"`) controls unknown estimates: allow prevents a cap-specific exclusion and records `cost.capOutcome: "unknown-allowed"`; exclude emits `cost-limit-unknown` and `capOutcome: "unknown-excluded"`. `onUnknownCost` alone (no cap) is inert. Separate from `unknownEvidence.cost`, which can still exclude or penalize unknown prices via `unknown-price` / scoring. |
@@ -134,16 +135,44 @@ evidence surface for context-sensitive profiles.
 The CLI dry-run accepts request-evidence flags but cannot supply candidate capability evidence yet;
 candidate evidence is provided through the API (`POST /api/routing-profiles/dry-run`).
 
+### Automatic prompt routing
+
+With `promptRouting.enabled: true`, OpenCodex classifies the latest user prompt as `fast`,
+`balanced`, or `powerful` for every request. The classifier is deterministic and local: it makes no
+extra model call and adds only the resulting tier to routing evidence, not the raw prompt. Requests
+without usable text, such as image-only input, default to `balanced`. The requested reasoning effort
+can also raise or lower the tier.
+
+Every candidate must declare a non-empty `taskTiers` array, and the profile must cover all three
+tiers. Tier matching is an eligibility filter; the existing capability, health, quota, cost-limit,
+unknown-evidence, and scoring rules still decide among candidates eligible for that tier.
+
+Select the profile alias once in Codex. Each later request through that alias is classified again,
+so consecutive turns may use different providers or models. The selected trace reason is
+`prompt-tier-fast`, `prompt-tier-balanced`, or `prompt-tier-powerful`, visible through the
+route-decision API and `ocx logs explain`.
+
+The tier filter is provider-agnostic: any configured provider/model can participate. DeepSeek, Qwen,
+Kimi, GLM, and the ids in the example are illustrative rather than a hardcoded model catalog.
+
 ```json
 {
   "routingProfiles": {
-    "fast": {
-      "alias": "ocx/fast",
+    "smart": {
+      "alias": "ocx/auto",
+      "promptRouting": { "enabled": true },
       "candidates": [
-        { "provider": "anthropic", "model": "claude-sonnet-5" },
-        { "provider": "openai", "model": "gpt-5.6-sol" }
+        {
+          "provider": "deepseek",
+          "model": "deepseek-chat",
+          "taskTiers": ["fast", "balanced"]
+        },
+        {
+          "provider": "openai",
+          "model": "gpt-5.6-sol",
+          "taskTiers": ["powerful"]
+        }
       ],
-      "require": { "tools": true, "minContextWindow": 128000 },
       "optimize": { "latency": 0.55, "health": 0.25, "cost": 0.10, "quota": 0.10 },
       "limits": { "maxEstimatedCostUsd": 0.50, "onUnknownCost": "allow" },
       "unknownEvidence": {
@@ -158,7 +187,7 @@ candidate evidence is provided through the API (`POST /api/routing-profiles/dry-
 ```
 
 CLI: `ocx route policy list [--json]`, `ocx route policy show <id> [--json]`, and
-`ocx route policy dry-run <id> [--model-context <tokens>] [--tools] [--image] [--structured-output] [--json]`.
+`ocx route policy dry-run <id> [--task-tier <fast|balanced|powerful>] [--model-context <tokens>] [--tools] [--image] [--structured-output] [--json]`.
 Dry-run evaluates candidates without sending any upstream request.
 
 Quota evidence (`optimize.quota`, `require.minQuotaHeadroom`, `unknownEvidence.quota`) comes from

docs-site/src/content/docs/zh-cn/reference/configuration/routing.md

@@ -102,9 +102,13 @@ Codex Auth 页面将此 picker 行为作为选择加入项。关闭它会隐藏
 
 ## 路由策略配置文件（`config.routingProfiles`）
 
-显式请求的 `policy/<id>`（或配置的别名）会在固定的候选白名单中，根据硬性能力要求与确定性、可解释的评分进行选择。现有模型 ID 永远不会隐式经过配置文件。支持 `candidates`（显式白名单）、可选 `alias`、`require`（`minContextWindow`、`minQuotaHeadroom`、`tools`、`imageInput`、`structuredOutput`、`localOnly`、`remoteAllowed`、`encryptedCodexTasks`、`reasoningEffort`、`serviceTier`）、`optimize`（latency/health/cost/quota 权重）、`limits.maxEstimatedCostUsd`、`unknownEvidence`（allow/penalize/exclude）。未知不会被当作零或免费。
+显式请求的 `policy/<id>`（或配置的别名）会在固定的候选白名单中，根据硬性能力要求与确定性、可解释的评分进行选择。现有模型 ID 永远不会隐式经过配置文件。支持 `candidates`（显式白名单）、可选 `alias`、`promptRouting`、`require`（`minContextWindow`、`minQuotaHeadroom`、`tools`、`imageInput`、`structuredOutput`、`localOnly`、`remoteAllowed`、`encryptedCodexTasks`、`reasoningEffort`、`serviceTier`）、`optimize`（latency/health/cost/quota 权重）、`limits.maxEstimatedCostUsd`、`unknownEvidence`（allow/penalize/exclude）。未知不会被当作零或免费。
 
-CLI：`ocx route policy list`、`ocx route policy show <id>`、`ocx route policy dry-run <id> --model-context <tokens> --tools`、`ocx route policy evaluate <id>`。
+设置 `"promptRouting": { "enabled": true }` 后，只需在 Codex 中选择一次该配置文件的别名（例如 `ocx/auto`）。OpenCodex 会在每次请求时重新分析最新一条用户提示词，把任务确定性地分为 `fast`、`balanced` 或 `powerful`，再从声明了对应 `taskTiers` 的候选中选择模型。所有三个等级都必须有候选覆盖。该分类在本地完成，不会额外调用模型，也不会把原始提示词写入路由元数据；能力、健康度、额度、成本限制和正常评分仍会继续生效。最终模型与 `prompt-tier-*` 原因可通过路由决策记录和 `ocx logs explain` 查看。
+
+任务等级筛选与提供商无关：任何已经在 OpenCodex 中配置好的模型都能参与，包括 DeepSeek、通义千问、Kimi、智谱 GLM 等国内模型；它们不是写死在路由器里的模型清单。
+
+CLI：`ocx route policy list`、`ocx route policy show <id>`、`ocx route policy dry-run <id> --task-tier powerful --model-context <tokens> --tools`、`ocx route policy evaluate <id>`。
 
 组合是显式的有序/加权目标路由与故障转移；策略配置文件是基于证据在候选之间进行选择。
 

gui/src/i18n/de.ts

@@ -31,6 +31,12 @@ export const de: Record<TKey, string> = {
   "routing.unknownEvidence.exclude": "ausschließen",
   "routing.removeCandidate": "Kandidat {provider}/{model} entfernen",
   "routing.candidates": "Kandidaten",
+  "routing.promptRouting": "Automatisches Prompt-Routing",
+  "routing.promptRoutingHelp": "Klassifiziert den neuesten Benutzer-Prompt lokal und wählt für jede Anfrage eine geeignete Aufgabenstufe.",
+  "routing.taskTiers": "Aufgabenstufen",
+  "routing.taskTier.fast": "Schnell",
+  "routing.taskTier.balanced": "Ausgewogen",
+  "routing.taskTier.powerful": "Leistungsstark",
   "routing.require": "Harte Anforderungen",
   "routing.optimize": "Optimierungsgewichte",
   "routing.limits": "Grenzen",
@@ -39,6 +45,7 @@ export const de: Record<TKey, string> = {
   "routing.unavailable": "–",
   "routing.dryRun": "Trockenlauf-Bewertung",
   "routing.dryRunContext": "Kontextfenster der Anfrage (Tokens)",
+  "routing.dryRunTaskTier": "Aufgabenstufe der Anfrage",
   "routing.dryRunTools": "Anfrage benötigt Tools",
   "routing.dryRunImage": "Anfrage benötigt Bild-Eingabe",
   "routing.dryRunStructured": "Anfrage benötigt strukturierte Ausgabe",
@@ -52,6 +59,7 @@ export const de: Record<TKey, string> = {
   "routing.capOutcome.unknown-allowed": "unbekannt (erlaubt)",
   "routing.capOutcome.unknown-excluded": "unbekannt (ausgeschlossen)",
   "routing.exclusion.capability-unsatisfied": "Anforderung nicht erfüllt",
+  "routing.exclusion.task-tier-mismatch": "Nicht dieser Aufgabenstufe zugewiesen",
   "routing.exclusion.unknown-capability": "unbekannte Fähigkeit",
   "routing.exclusion.cost-limit": "über Kostenobergrenze",
   "routing.exclusion.cost-limit-unknown": "unbekannte Kosten unter Obergrenze",

gui/src/i18n/en.ts

@@ -74,6 +74,12 @@ export const en = {
   "routing.unknownEvidence.exclude": "exclude",
   "routing.removeCandidate": "Remove candidate {provider}/{model}",
   "routing.candidates": "Candidates",
+  "routing.promptRouting": "Automatic prompt routing",
+  "routing.promptRoutingHelp": "Classify each latest user prompt locally and select an eligible task tier for every request.",
+  "routing.taskTiers": "Task tiers",
+  "routing.taskTier.fast": "Fast",
+  "routing.taskTier.balanced": "Balanced",
+  "routing.taskTier.powerful": "Powerful",
   "routing.require": "Hard requirements",
   "routing.optimize": "Optimization weights",
   "routing.limits": "Limits",
@@ -82,6 +88,7 @@ export const en = {
   "routing.unavailable": "–",
   "routing.dryRun": "Dry-run evaluation",
   "routing.dryRunContext": "Request context window (tokens)",
+  "routing.dryRunTaskTier": "Request task tier",
   "routing.dryRunTools": "Request requires tools",
   "routing.dryRunImage": "Request requires image input",
   "routing.dryRunStructured": "Request requires structured output",
@@ -95,6 +102,7 @@ export const en = {
   "routing.capOutcome.unknown-allowed": "unknown (allowed)",
   "routing.capOutcome.unknown-excluded": "unknown (excluded)",
   "routing.exclusion.capability-unsatisfied": "capability not met",
+  "routing.exclusion.task-tier-mismatch": "not assigned to this task tier",
   "routing.exclusion.unknown-capability": "unknown capability",
   "routing.exclusion.cost-limit": "over cost cap",
   "routing.exclusion.cost-limit-unknown": "unknown cost under cap",

gui/src/i18n/ja.ts

@@ -31,6 +31,12 @@ export const ja: Record<TKey, string> = {
   "routing.unknownEvidence.exclude": "除外",
   "routing.removeCandidate": "候補 {provider}/{model} を削除",
   "routing.candidates": "候補",
+  "routing.promptRouting": "自動プロンプトルーティング",
+  "routing.promptRoutingHelp": "最新のユーザープロンプトをローカルで分類し、リクエストごとに適切なタスク階層を選択します。",
+  "routing.taskTiers": "タスク階層",
+  "routing.taskTier.fast": "高速",
+  "routing.taskTier.balanced": "バランス",
+  "routing.taskTier.powerful": "高性能",
   "routing.require": "必須要件",
   "routing.optimize": "最適化ウェイト",
   "routing.limits": "制限",
@@ -39,6 +45,7 @@ export const ja: Record<TKey, string> = {
   "routing.unavailable": "–",
   "routing.dryRun": "ドライラン評価",
   "routing.dryRunContext": "リクエストのコンテキストウィンドウ（トークン）",
+  "routing.dryRunTaskTier": "リクエストのタスク階層",
   "routing.dryRunTools": "リクエストにツールが必要",
   "routing.dryRunImage": "リクエストに画像入力が必要",
   "routing.dryRunStructured": "リクエストに構造化出力が必要",
@@ -52,6 +59,7 @@ export const ja: Record<TKey, string> = {
   "routing.capOutcome.unknown-allowed": "不明（許可）",
   "routing.capOutcome.unknown-excluded": "不明（除外）",
   "routing.exclusion.capability-unsatisfied": "能力要件未達",
+  "routing.exclusion.task-tier-mismatch": "このタスク階層に未割り当て",
   "routing.exclusion.unknown-capability": "能力不明",
   "routing.exclusion.cost-limit": "コスト上限超過",
   "routing.exclusion.cost-limit-unknown": "上限下でコスト不明",

gui/src/i18n/ko.ts

@@ -31,6 +31,12 @@ export const ko: Record<TKey, string> = {
   "routing.unknownEvidence.exclude": "제외",
   "routing.removeCandidate": "후보 {provider}/{model} 제거",
   "routing.candidates": "후보",
+  "routing.promptRouting": "자동 프롬프트 라우팅",
+  "routing.promptRoutingHelp": "최신 사용자 프롬프트를 로컬에서 분류하고 요청마다 적합한 작업 등급을 선택합니다.",
+  "routing.taskTiers": "작업 등급",
+  "routing.taskTier.fast": "빠름",
+  "routing.taskTier.balanced": "균형",
+  "routing.taskTier.powerful": "강력",
   "routing.require": "필수 요구사항",
   "routing.optimize": "최적화 가중치",
   "routing.limits": "제한",
@@ -39,6 +45,7 @@ export const ko: Record<TKey, string> = {
   "routing.unavailable": "–",
   "routing.dryRun": "드라이런 평가",
   "routing.dryRunContext": "요청 컨텍스트 창(토큰)",
+  "routing.dryRunTaskTier": "요청 작업 등급",
   "routing.dryRunTools": "요청에 도구 필요",
   "routing.dryRunImage": "요청에 이미지 입력 필요",
   "routing.dryRunStructured": "요청에 구조화된 출력 필요",
@@ -52,6 +59,7 @@ export const ko: Record<TKey, string> = {
   "routing.capOutcome.unknown-allowed": "알 수 없음(허용)",
   "routing.capOutcome.unknown-excluded": "알 수 없음(제외)",
   "routing.exclusion.capability-unsatisfied": "기능 미충족",
+  "routing.exclusion.task-tier-mismatch": "이 작업 등급에 할당되지 않음",
   "routing.exclusion.unknown-capability": "알 수 없는 기능",
   "routing.exclusion.cost-limit": "비용 상한 초과",
   "routing.exclusion.cost-limit-unknown": "상한 이하 비용 불명",

gui/src/i18n/ru.ts

@@ -31,6 +31,12 @@ export const ru: Record<TKey, string> = {
   "routing.unknownEvidence.exclude": "исключить",
   "routing.removeCandidate": "Удалить кандидата {provider}/{model}",
   "routing.candidates": "Кандидаты",
+  "routing.promptRouting": "Автоматическая маршрутизация промптов",
+  "routing.promptRoutingHelp": "Локально классифицирует последний промпт пользователя и выбирает подходящий уровень задачи для каждого запроса.",
+  "routing.taskTiers": "Уровни задач",
+  "routing.taskTier.fast": "Быстрый",
+  "routing.taskTier.balanced": "Сбалансированный",
+  "routing.taskTier.powerful": "Мощный",
   "routing.require": "Жёсткие требования",
   "routing.optimize": "Веса оптимизации",
   "routing.limits": "Лимиты",
@@ -39,6 +45,7 @@ export const ru: Record<TKey, string> = {
   "routing.unavailable": "–",
   "routing.dryRun": "Пробная оценка",
   "routing.dryRunContext": "Контекстное окно запроса (токены)",
+  "routing.dryRunTaskTier": "Уровень задачи запроса",
   "routing.dryRunTools": "Запрос требует инструменты",
   "routing.dryRunImage": "Запрос требует изображения",
   "routing.dryRunStructured": "Запрос требует структурированный вывод",
@@ -52,6 +59,7 @@ export const ru: Record<TKey, string> = {
   "routing.capOutcome.unknown-allowed": "неизвестно (разрешено)",
   "routing.capOutcome.unknown-excluded": "неизвестно (исключено)",
   "routing.exclusion.capability-unsatisfied": "требование не выполнено",
+  "routing.exclusion.task-tier-mismatch": "Не назначено этому уровню задачи",
   "routing.exclusion.unknown-capability": "неизвестная возможность",
   "routing.exclusion.cost-limit": "сверх лимита стоимости",
   "routing.exclusion.cost-limit-unknown": "неизвестная стоимость при лимите",