@weniatall/dsh-task-models 0.3.1 → 0.3.2

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Veildawn
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -34,7 +34,7 @@ dsh plugin --profile web add @weniatall/dsh-task-models
 或锁定版本，从 GitHub Release 安装 tarball：
 
 ```bash
-dsh plugin --profile web add https://github.com/veildawn/dsh-plugins/releases/download/dsh-task-models@v0.3.1/dsh-task-models-0.3.1.tgz
+dsh plugin --profile web add https://github.com/veildawn/dsh-plugins/releases/download/dsh-task-models@v0.3.2/dsh-task-models-0.3.2.tgz
 ```
 
 > 本包已发布到 npm scope `@weniatall/dsh-task-models`，后续版本由 CI 通过 Trusted
@@ -47,7 +47,7 @@ dsh plugin --profile web add https://github.com/veildawn/dsh-plugins/releases/do
 cd plugins/dsh-task-models
 npm test
 npm pack
-dsh plugin --profile web add ./weniatall-dsh-task-models-0.3.1.tgz
+dsh plugin --profile web add ./weniatall-dsh-task-models-0.3.2.tgz
 ```
 
 > 本仓库的 `AGENTS.md` 规定：除非用户明确要求，不在本机部署插件或重启服务。开发验证请只运行
@@ -163,6 +163,9 @@ task-models:
 ## 与 DSH 的边界
 
 - 不 patch `ctx.subagents.start`；只在公开的 `agent/pre-step` / `agent/request` 上工作。
+  子代理的模型改写发生在 `agent/pre-step`（通过 `agent.model.select()`）：DSH 0.2.2 起
+  `agent/request` 已废弃且会在瀑布之后强校验模型一致性，从那里改模型会直接抛错。
+  `agent/request` 只用于记录子代理实际使用的模型。
 - 不新增子代理工具，不修改 `subagent` / `workflow` 的工具 schema。
 - 不处理 ACP / Codex / Claude Code 等外部进程子代理的模型选择；这些由各自外部进程决定。
 - 主任务自动分类、识图委派、顾问复核不在本插件职责内；由宿主或相应插件负责。本插件只处理

package/lib/index.js

@@ -434,6 +434,18 @@ function toRoute(profile) {
   }
 }
 
+/** 把任意携带 provider/model 的对象收敛成路由配置；缺失时返回 undefined。 */
+function toRouteConfig(entry) {
+  if (!entry?.provider || !entry?.model) return undefined
+  return {
+    provider: entry.provider,
+    model: entry.model,
+    ...(entry.reasoningEffort === undefined || entry.reasoningEffort === null || entry.reasoningEffort === ''
+      ? {}
+      : { reasoningEffort: String(entry.reasoningEffort) }),
+  }
+}
+
 function applyRoute(config, route) {
   const { reasoningEffort: _previousEffort, ...rest } = config
   return {
@@ -554,7 +566,8 @@ export function apply(ctx, config = {}) {
     }))
   }
 
-  function keepCurrent(agent, current, reason) {
+  /** 记录「保持当前模型」的原因，并原样返回候选配置。 */
+  function applyKeep(agent, current, reason) {
     rememberRoutingDetail(agent, { status: 'current', source: 'current', reason, durationMs: 0 })
     return current
   }
@@ -946,52 +959,69 @@ export function apply(ctx, config = {}) {
     )
   })
 
-  // Capture the delegated prompt before agent/request: the Host commits the
-  // current step's user message only after the request waterfall.
-  ctx.on('agent/pre-step', async ({ agent, turn, step }, next) => {
+  ctx.on('session/event', (session, event) => {
+    if (event?.type === 'subagent/descriptor' && typeof event.data?.label === 'string') {
+      subagentLabels.set(session, event.data.label)
+    }
+    if (event?.type === 'request/header') {
+      rememberChildModel({ session }, event.data?.header?.config ?? event.data?.config, true)
+    }
+  })
+
+  // DSH 0.2.2 起 `agent/request` 已废弃，且宿主会在瀑布之后强校验模型一致性：
+  // 改了 provider/model/reasoningEffort 会直接抛错。因此子代理的路由改写移到
+  // 步骤准备期，通过 `agent.model.select()` 落地，agent/request 只负责记录结果。
+  ctx.on('agent/pre-step', async ({ agent, turn, step, signal, modelSelection }, next) => {
     const decision = await next()
     if (decision.kind !== 'enter') return decision
+    // 宿主只在请求瀑布之后才提交本步骤的用户消息，所以委派提示词要在这里抓取。
     if (isSubagent(agent) && turn === 1 && step === 1
       && agent?.options?.modelRole === undefined && !firstStepText.has(agent)) {
       const text = promptTextOf(decision.messages)
       if (text !== '') firstStepText.set(agent, text.slice(-6_000))
     }
+
+    // 决定本步骤的模型：宿主会在 pre-step 返回后把 agent.model.current() 固定为本步骤选择，
+    // 之后再从 agent/request 改模型会触发宿主的强校验错误，所以必须在这里写入。
+    const current = routeCandidateOf(agent, modelSelection)
+    const routed = await routeChildRequest(agent, turn, current, signal)
+    if (routed !== current) applyChildRoute(agent, routed)
+    signal?.throwIfAborted?.()
     return decision
   }, { prepend: true })
 
-  ctx.on('session/event', (session, event) => {
-    if (event?.type === 'subagent/descriptor' && typeof event.data?.label === 'string') {
-      subagentLabels.set(session, event.data.label)
-    }
-    if (event?.type === 'request/header') {
-      rememberChildModel({ session }, event.data?.header?.config ?? event.data?.config, true)
-    }
-  })
+  /** 读取本步骤为子代理准备的路由候选（未经过步骤准备期时退回请求头/options）。 */
+  function routeCandidateOf(agent, modelSelection) {
+    const fromSelection = toRouteConfig(modelSelection)
+    if (fromSelection !== undefined) return fromSelection
+    const header = requestHeaderOf(agent)
+    const fromHeader = toRouteConfig(header?.config)
+    if (fromHeader !== undefined) return fromHeader
+    return toRouteConfig(agent?.options)
+  }
 
-  // Prepending makes this router the outermost agent/request listener: it
-  // observes every other listener's result through next() and applies its
-  // profile last. It only touches delegated children.
-  ctx.on('agent/request', async ({ agent, signal }, next) => {
-    const current = await next()
+  /** 决定子代理本回合要用的模型；返回与 `current` 相同的对象表示保持现状。 */
+  async function routeChildRequest(agent, turn, current, signal) {
+    if (current === undefined) return current
     if (!isSubagent(agent)) return current
     throwIfAborted(signal)
     rememberChildModel(agent, current)
     const cfg = routingConfig
-    if (agent?.options?.modelRole !== undefined) return keepCurrent(agent, current, 'internal-role')
-    if (planModeActive(agent?.session?.events)) return keepCurrent(agent, current, 'plan-mode')
-    if (!cfg.routing.enabled) return keepCurrent(agent, current, 'routing-disabled')
+    if (agent?.options?.modelRole !== undefined) return applyKeep(agent, current, 'internal-role')
+    if (planModeActive(agent?.session?.events)) return applyKeep(agent, current, 'plan-mode')
+    if (!cfg.routing.enabled) return applyKeep(agent, current, 'routing-disabled')
 
     // An adopted snapshot survives edits, disablement and removal of its profile.
     // Only the explicit global switch and the existing bypasses suspend replay.
     let cached = activeDecision(agent)
     if (cached === undefined) {
-      if (![...cfg.profiles.values()].some((profile) => profile.enabled)) return keepCurrent(agent, current, 'no-profiles')
-      if (requestHeaderOf(agent) !== undefined) return keepCurrent(agent, current, 'existing-request')
-      if (cfg.routing.respectExplicitChildModel && childHasExplicitRoute(agent)) return keepCurrent(agent, current, 'explicit-child-model')
+      if (![...cfg.profiles.values()].some((profile) => profile.enabled)) return applyKeep(agent, current, 'no-profiles')
+      if (requestHeaderOf(agent) !== undefined) return applyKeep(agent, current, 'existing-request')
+      if (cfg.routing.respectExplicitChildModel && childHasExplicitRoute(agent)) return applyKeep(agent, current, 'explicit-child-model')
       const label = labelOf(agent)
       const captured = firstStepText.get(agent)
       const text = typeof captured === 'string' && captured !== '' ? captured : taskTextOf(agent)
-      if ((label ?? '') === '' && (text ?? '') === '') return keepCurrent(agent, current, 'no-task')
+      if ((label ?? '') === '' && (text ?? '') === '') return applyKeep(agent, current, 'no-task')
       decisionFor(agent, cfg, label, text, signal)
       cached = taskDecisions.get(agent)
     }
@@ -1009,7 +1039,7 @@ export function apply(ctx, config = {}) {
           ruleAnalysis: decision.ruleAnalysis, durationMs: decision.durationMs,
         })
       }
-      return keepCurrent(agent, current, bypass)
+      return applyKeep(agent, current, bypass)
     }
     rememberRoutingDetail(agent, decision)
     if (decision.route === undefined) return current
@@ -1017,5 +1047,29 @@ export function apply(ctx, config = {}) {
     cached.applied = true
     rememberChildModel(agent, routed)
     return routed
+  }
+
+  /**
+   * 在步骤准备期写入下一次请求的模型选择。
+   * 宿主没有 agent.model 时（旧版本）退回追加 model/selection 事件。
+   */
+  function applyChildRoute(agent, routed) {
+    const selection = {
+      provider: routed.provider,
+      model: routed.model,
+      ...(routed.reasoningEffort === undefined ? {} : { reasoningEffort: routed.reasoningEffort }),
+    }
+    if (typeof agent?.model?.select === 'function') {
+      agent.model.select(selection)
+      return
+    }
+    agent?.session?.append?.('model/selection', selection)
+  }
+
+  // agent/request 只做记录：子代理实际使用的模型已经由步骤准备期决定。
+  ctx.on('agent/request', async ({ agent }, next) => {
+    const current = await next()
+    if (isSubagent(agent) && current?.provider && current?.model) rememberChildModel(agent, current)
+    return current
   }, { prepend: true })
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@weniatall/dsh-task-models",
-  "version": "0.3.1",
+  "version": "0.3.2",
   "description": "子任务模型路由插件：按任务类型为子代理选择模型，显式标签优先、规则试运行与路由解释，可选 TypeSafe Jev 决策与轻量分类回退，并区分预选与请求头确认的 workflow 成员模型。",
   "type": "module",
   "main": "lib/index.js",
@@ -10,9 +10,6 @@
     "./client": "./lib/client.js",
     "./package.json": "./package.json"
   },
-  "scripts": {
-    "test": "node --test test/*.test.mjs && node ../../scripts/check-host-settings.mjs plugins/dsh-task-models task-models"
-  },
   "files": [
     "lib/index.js",
     "lib/core.js",
@@ -74,5 +71,8 @@
   "homepage": "https://github.com/veildawn/dsh-plugins/tree/main/plugins/dsh-task-models#readme",
   "bugs": {
     "url": "https://github.com/veildawn/dsh-plugins/issues"
+  },
+  "scripts": {
+    "test": "node --test test/*.test.mjs && node ../../scripts/check-host-settings.mjs plugins/dsh-task-models task-models"
   }
-}
+}