diff --git a/mateclaw-ui/src/views/ChatConsole.vue b/mateclaw-ui/src/views/ChatConsole.vue
index 5b6d8fd3..4331c2e9 100644
--- a/mateclaw-ui/src/views/ChatConsole.vue
+++ b/mateclaw-ui/src/views/ChatConsole.vue
@@ -404,6 +404,12 @@ const enabledModels = ref<ModelConfig[]>([])
 const activeModels = ref<ActiveModelsInfo | null>(null)
 // Global default model — seeds the selector for conversations with no pin yet.
 const globalDefaultModel = ref<{ providerId: string; model: string } | null>(null)
+// True once the user manually picks a model for the CURRENT conversation.
+// Guards the agent-capability re-seed (see the selectedAgentId watcher) from
+// clobbering that pick during the window where the async capability fetch
+// resolves after a fresh conversation was already seeded. Reset whenever we
+// move to a different conversation.
+const userPickedModel = ref(false)
 const pendingAttachments = ref<ChatAttachment[]>([])
 const uploadingAttachment = ref(false)
 
@@ -445,6 +451,9 @@ function onAgentPicked(value: string | number | null) {
 function selectModel(value: string) {
   const [providerId, model] = value.split('::')
   if (!providerId || !model) return
+  // Remember the explicit pick so a late-arriving agent-capability fetch
+  // doesn't re-seed over it on a not-yet-persisted conversation.
+  userPickedModel.value = true
   // Per-conversation model: switching here only affects THIS conversation.
   // We update the selector + the local list entry immediately so the UI is
   // responsive, then persist the pin to the server right away IF the
@@ -514,13 +523,62 @@ function selectModel(value: string) {
 }
 
 /**
- * Point the model selector at a conversation's pinned model, or the global
- * default when the conversation has no pin yet (fresh chat, IM, cron).
+ * The selected agent's model override, as a (provider, model) pair — or null
+ * when the agent has no usable override.
+ *
+ * Two sources, in order:
+ *   1. agentCapabilities — the backend-resolved pair from /agents/{id}/capabilities.
+ *      Authoritative: it runs the same resolveModel the runtime uses (honouring
+ *      disabled→default fallback and provider disambiguation). But it arrives
+ *      via an async fetch that resolves AFTER the synchronous seeding on an
+ *      agent switch / deep-link, and is lost entirely if that fetch fails.
+ *   2. currentAgent.modelName resolved against the enabled-model list — a
+ *      SYNCHRONOUS fallback so an agent switch, a capability-fetch failure, or a
+ *      not-yet-hydrated deep-link still honour the per-agent override instead of
+ *      silently dropping to the global default (which handleSendMessage would
+ *      then pin, re-introducing the very clobber this whole change fixes).
+ *
+ * enabledModels is viewer-accessible (/models/enabled) and loaded at mount, so
+ * the fallback works for viewers too.
+ */
+function agentSeedModel(): { providerId: string; model: string } | null {
+  const cap = agentCapabilities.value
+  if (cap?.providerId && cap?.modelName) {
+    return { providerId: cap.providerId, model: cap.modelName }
+  }
+  const name = currentAgent.value?.modelName
+  if (name) {
+    const hit = enabledModels.value.find(m => m.modelName === name)
+    if (hit?.provider) {
+      return { providerId: hit.provider, model: hit.modelName }
+    }
+  }
+  return null
+}
+
+/**
+ * Point the model selector at the model this conversation should run, honouring
+ * the same precedence the backend uses when it resolves the runtime model
+ * (see AgentGraphBuilder.resolveRuntimeBaseModel):
+ *
+ *   conversation pin  >  selected agent's model override  >  global default
+ *
+ * The middle tier is the fix for the "I set the agent's model but chat used
+ * another one" bug: without it a fresh conversation seeded the global default,
+ * and handleSendMessage then PINNED that default onto the conversation row —
+ * which outranks the agent override and silently clobbered it.
  */
 function applyConversationModel(conv?: Conversation | null) {
   if (conv?.modelProvider && conv?.modelName) {
     activeModels.value = { activeLlm: { providerId: conv.modelProvider, model: conv.modelName } }
-  } else if (globalDefaultModel.value) {
+    return
+  }
+  const agentModel = agentSeedModel()
+  if (agentModel) {
+    activeModels.value = { activeLlm: { providerId: agentModel.providerId, model: agentModel.model } }
+    return
+  }
+  if (globalDefaultModel.value) {
     activeModels.value = { activeLlm: { ...globalDefaultModel.value } }
   }
 }
@@ -1369,6 +1427,17 @@ watch(selectedAgentId, async (id) => {
   try {
     const res: any = await agentApi.getCapabilities(id)
     agentCapabilities.value = res.data || null
+    // The capability fetch is async and typically resolves AFTER the
+    // synchronous newConversation()/applyConversationModel() that ran on this
+    // same agent switch (which fell back to the global default because caps
+    // weren't loaded yet). Re-seed now that we know the agent's resolved model
+    // — but only for a conversation with no server-side pin and where the user
+    // hasn't manually picked a model, so we never clobber an explicit choice.
+    const conv = conversations.value.find(c => c.conversationId === currentConversationId.value)
+    const hasServerPin = !!(conv?.modelProvider && conv?.modelName)
+    if (!hasServerPin && !userPickedModel.value) {
+      applyConversationModel(conv)
+    }
   } catch {
     agentCapabilities.value = null
   }
@@ -1594,6 +1663,9 @@ async function selectConversation(conv: Conversation) {
   }
   currentConversationId.value = conv.conversationId
   selectedAgentId.value = conv.agentId || selectedAgentId.value
+  // Opening another conversation: its pin (or the agent/global fallback) is
+  // authoritative, so clear the previous conversation's manual-pick guard.
+  userPickedModel.value = false
   // Restore this conversation's pinned model into the selector.
   applyConversationModel(conv)
   // Reset cron placeholder state up front; the immediate fetch below repopulates
@@ -1728,7 +1800,9 @@ function newConversation() {
   resetForNewConversation()
   currentConversationId.value = `conv_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`
   messages.value = []
-  // A fresh conversation starts on the global default model.
+  // A fresh conversation defers to the selected agent's model (then the global
+  // default) until the user explicitly picks one.
+  userPickedModel.value = false
   applyConversationModel()
 }
 
@@ -1741,6 +1815,11 @@ function onConversationsDeleted(ids: string[]) {
     resetStreamingState()
     messages.value = []
     currentConversationId.value = ''
+    // Deleting the open conversation drops us back to a blank slate — clear the
+    // manual-pick guard so the next message (which mints a fresh conversation)
+    // seeds from the agent override again instead of inheriting the deleted
+    // conversation's pick. Mirrors newConversation()/selectConversation().
+    userPickedModel.value = false
   }
 }
 
diff --git a/mateclaw-ui/src/views/__tests__/chatConsoleModelSeed.test.ts b/mateclaw-ui/src/views/__tests__/chatConsoleModelSeed.test.ts
new file mode 100644
index 00000000..e4dd55c7
--- /dev/null
+++ b/mateclaw-ui/src/views/__tests__/chatConsoleModelSeed.test.ts
@@ -0,0 +1,170 @@
+// @vitest-environment happy-dom
+import { describe, it, expect } from 'vitest'
+
+/**
+ * 复刻 ChatConsole 中模型 seed 的核心判定逻辑做纯函数测试。
+ *
+ * 修复的 bug：员工编辑页选定的「模型」（per-Agent override）在网页对话里被
+ * 全局默认模型覆盖。根因是 applyConversationModel 在会话无 pin 时直接回落全局
+ * 默认，随后 handleSendMessage 把该默认 PIN 到会话行——而会话 pin 优先级高于
+ * 员工 override，导致员工选的模型被静默覆盖。
+ *
+ * 修复后 seed 优先级与后端 AgentGraphBuilder.resolveRuntimeBaseModel 对齐：
+ *   会话 pin  >  员工 modelName override  >  全局默认
+ *
+ * 员工 override 这一层有两个来源（复刻 agentSeedModel）：
+ *   1. agentCapabilities —— /agents/{id}/capabilities 的后端解析结果（异步、权威）
+ *   2. currentAgent.modelName 在 enabledModels 里的同步解析（异步失败/切换/深链兜底）
+ */
+
+type Model = { providerId: string; model: string }
+type EnabledModel = { provider: string; modelName: string }
+
+// 复刻 agentSeedModel：先用 capabilities，再用 modelName 同步兜底。
+// 关键点：capabilities 未配默认模型时后端返回的是空串 ""（见 AgentController
+// 的 catch 分支），不是 null —— 空串是 JS falsy，`cap?.providerId && cap?.modelName`
+// 会正确落空并走同步兜底。测试特意用 "" 锁住这个 wire 契约。
+function agentSeedModel(opts: {
+  capProvider?: string | null
+  capModel?: string | null
+  agentModelName?: string | null
+  enabledModels?: EnabledModel[]
+}): Model | null {
+  if (opts.capProvider && opts.capModel) {
+    return { providerId: opts.capProvider, model: opts.capModel }
+  }
+  const name = opts.agentModelName
+  if (name) {
+    const hit = (opts.enabledModels || []).find(m => m.modelName === name)
+    if (hit?.provider) {
+      return { providerId: hit.provider, model: hit.modelName }
+    }
+  }
+  return null
+}
+
+// 复刻 applyConversationModel 的取值优先级
+function resolveSeedModel(opts: {
+  convProvider?: string | null
+  convModel?: string | null
+  capProvider?: string | null
+  capModel?: string | null
+  agentModelName?: string | null
+  enabledModels?: EnabledModel[]
+  globalDefault?: Model | null
+}): Model | null {
+  if (opts.convProvider && opts.convModel) {
+    return { providerId: opts.convProvider, model: opts.convModel }
+  }
+  const agentModel = agentSeedModel(opts)
+  if (agentModel) return agentModel
+  if (opts.globalDefault) return { ...opts.globalDefault }
+  return null
+}
+
+// 复刻 selectedAgentId watcher 中的 re-seed 守卫：
+// 仅当会话无服务端 pin 且用户未手动选过模型时，才允许用刚加载的员工能力覆盖。
+function shouldReseedFromAgentCaps(opts: {
+  convProvider?: string | null
+  convModel?: string | null
+  userPickedModel: boolean
+}): boolean {
+  const hasServerPin = !!(opts.convProvider && opts.convModel)
+  return !hasServerPin && !opts.userPickedModel
+}
+
+const GLOBAL: Model = { providerId: 'openai', model: 'gpt-x' }
+const ENABLED: EnabledModel[] = [
+  { provider: 'anthropic', modelName: 'claude-x' },
+  { provider: 'volcano', modelName: 'doubao-pro' },
+]
+
+describe('resolveSeedModel — 模型 seed 优先级', () => {
+  it('会话已有 pin 时，优先用会话 pin（压过员工与全局）', () => {
+    const r = resolveSeedModel({
+      convProvider: 'volcano', convModel: 'doubao-pro',
+      capProvider: 'anthropic', capModel: 'claude-x',
+      globalDefault: GLOBAL,
+    })
+    expect(r).toEqual({ providerId: 'volcano', model: 'doubao-pro' })
+  })
+
+  it('会话无 pin 时，用 capabilities 的员工 override（不再回落全局默认）', () => {
+    const r = resolveSeedModel({
+      capProvider: 'anthropic', capModel: 'claude-x',
+      globalDefault: GLOBAL,
+    })
+    expect(r).toEqual({ providerId: 'anthropic', model: 'claude-x' })
+  })
+
+  it('capabilities 未就绪（切换/失败/深链）时，用 modelName 在 enabledModels 里同步兜底', () => {
+    const r = resolveSeedModel({
+      capProvider: null, capModel: null,
+      agentModelName: 'claude-x', enabledModels: ENABLED,
+      globalDefault: GLOBAL,
+    })
+    expect(r).toEqual({ providerId: 'anthropic', model: 'claude-x' })
+  })
+
+  it('capabilities 与同步兜底都缺失时，回落全局默认', () => {
+    const r = resolveSeedModel({
+      agentModelName: '', enabledModels: ENABLED,
+      globalDefault: GLOBAL,
+    })
+    expect(r).toEqual(GLOBAL)
+  })
+
+  it('员工的 modelName 不在 enabledModels（禁用/删除）时，同步兜底落空 → 全局默认', () => {
+    const r = resolveSeedModel({
+      agentModelName: 'ghost-model', enabledModels: ENABLED,
+      globalDefault: GLOBAL,
+    })
+    expect(r).toEqual(GLOBAL)
+  })
+
+  it('全部缺失时返回 null（无可用模型）', () => {
+    expect(resolveSeedModel({ globalDefault: null })).toBeNull()
+  })
+})
+
+describe('agentSeedModel — 空串 wire 契约（后端无默认模型时返回 "" 而非 null）', () => {
+  it('capabilities 为空串对（providerId:"" modelName:""）视为无 override，走同步兜底', () => {
+    const r = agentSeedModel({
+      capProvider: '', capModel: '',
+      agentModelName: 'claude-x', enabledModels: ENABLED,
+    })
+    expect(r).toEqual({ providerId: 'anthropic', model: 'claude-x' })
+  })
+
+  it('半残的 capabilities 对（provider 空串、model 非空）视为无 override', () => {
+    const r = agentSeedModel({
+      capProvider: '', capModel: 'claude-x',
+      agentModelName: null, enabledModels: ENABLED,
+    })
+    expect(r).toBeNull()
+  })
+
+  it('capabilities 完整时优先于同步兜底', () => {
+    const r = agentSeedModel({
+      capProvider: 'openai', capModel: 'gpt-x',
+      agentModelName: 'claude-x', enabledModels: ENABLED,
+    })
+    expect(r).toEqual({ providerId: 'openai', model: 'gpt-x' })
+  })
+})
+
+describe('shouldReseedFromAgentCaps — 能力异步到达后的 re-seed 守卫', () => {
+  it('新会话、无 pin、用户未手动选 → 允许 re-seed（修复 agent 切换竞态）', () => {
+    expect(shouldReseedFromAgentCaps({ userPickedModel: false })).toBe(true)
+  })
+
+  it('会话已有服务端 pin → 不 re-seed（不覆盖已固定的会话）', () => {
+    expect(shouldReseedFromAgentCaps({
+      convProvider: 'volcano', convModel: 'doubao-pro', userPickedModel: false,
+    })).toBe(false)
+  })
+
+  it('用户已手动选过模型 → 不 re-seed（不覆盖显式选择）', () => {
+    expect(shouldReseedFromAgentCaps({ userPickedModel: true })).toBe(false)
+  })
+})