diff --git a/mateclaw-server/src/main/java/vip/mate/llm/failover/ProviderRequirements.java b/mateclaw-server/src/main/java/vip/mate/llm/failover/ProviderRequirements.java
new file mode 100644
index 00000000..58790110
--- /dev/null
+++ b/mateclaw-server/src/main/java/vip/mate/llm/failover/ProviderRequirements.java
@@ -0,0 +1,90 @@
+package vip.mate.llm.failover;
+
+import vip.mate.llm.model.ModelProviderEntity;
+
+import java.util.Map;
+
+/**
+ * Decide which fields a {@link ModelProviderEntity} needs to be considered
+ * "configured", based on the row's columns rather than its protocol enum.
+ *
+ * <p>Why row-based: every OpenAI-compatible provider (OpenAI / Kimi / DeepSeek
+ * cloud as well as llama.cpp / lmstudio / vllm / ollama local) shares the
+ * single {@code OPENAI_COMPATIBLE} protocol value. A protocol-keyed lookup
+ * cannot distinguish "cloud, needs api key" from "local, needs base url".
+ * The discriminating signals all live on the row: {@code requireApiKey},
+ * {@code isLocal}, {@code isCustom}, {@code authType}, {@code providerId}.
+ *
+ * <p>Hint key + args (no raw text) so the frontend renders via i18n
+ * {@code t(key, args)} without leaking Chinese into the English locale.
+ */
+public final class ProviderRequirements {
+
+    /** What this provider row needs to be considered configured. */
+    public record Required(
+            boolean needsApiKey,
+            boolean needsBaseUrl,
+            String hintKey,                 // i18n key, null when no hint applies
+            Map<String, Object> hintArgs    // template params for vue-i18n; never raw text
+    ) {}
+
+    private static final Required NONE = new Required(false, false, null, Map.of());
+
+    private ProviderRequirements() {}
+
+    /**
+     * Compute the required-fields verdict for a provider row.
+     *
+     * Decision tree:
+     *   - authType == "oauth"        -> no api key, no base url (OAuth handled elsewhere)
+     *   - requireApiKey == true       -> needs api key
+     *   - isLocal || isCustom         -> needs base url (no sane SDK default)
+     *   - cloud built-ins             -> SDK ships hard-coded base url; no base url needed
+     *
+     * Hint key picked from a small providerId-substring map for the most common
+     * local providers; everything else falls back to a generic OpenAI-compatible
+     * hint so the user always sees an actionable example URL.
+     */
+    public static Required of(ModelProviderEntity provider) {
+        if (provider == null) return NONE;
+
+        // OAuth providers store credentials elsewhere (DB column or disk). Neither
+        // api key nor base url applies to the configured check.
+        if ("oauth".equals(provider.getAuthType())) {
+            return NONE;
+        }
+
+        boolean needsApiKey  = Boolean.TRUE.equals(provider.getRequireApiKey());
+        boolean isLocal      = Boolean.TRUE.equals(provider.getIsLocal());
+        boolean isCustom     = Boolean.TRUE.equals(provider.getIsCustom());
+        boolean needsBaseUrl = isLocal || isCustom;
+
+        if (!needsBaseUrl) {
+            return new Required(needsApiKey, false, null, Map.of());
+        }
+
+        // Pick a hint by providerId substring. Order matters: more specific names
+        // first so "lm-studio" doesn't accidentally match a generic prefix later.
+        String pid = provider.getProviderId() == null ? "" : provider.getProviderId().toLowerCase();
+        String hintKey;
+        Map<String, Object> hintArgs;
+        if (pid.contains("ollama")) {
+            hintKey = "provider.hint.ollamaBaseUrlExample";
+            hintArgs = Map.of("example", "http://127.0.0.1:11434");
+        } else if (pid.contains("lmstudio") || pid.contains("lm-studio") || pid.contains("lm_studio")) {
+            hintKey = "provider.hint.lmstudioBaseUrlExample";
+            hintArgs = Map.of("example", "http://127.0.0.1:1234/v1");
+        } else if (pid.contains("llamacpp") || pid.contains("llama-cpp") || pid.contains("llama_cpp")
+                || pid.contains("llama.cpp")) {
+            hintKey = "provider.hint.llamacppBaseUrlExample";
+            hintArgs = Map.of("example", "http://127.0.0.1:8080/v1");
+        } else if (pid.contains("vllm")) {
+            hintKey = "provider.hint.vllmBaseUrlExample";
+            hintArgs = Map.of("example", "http://127.0.0.1:8000/v1");
+        } else {
+            hintKey = "provider.hint.openaiCompatBaseUrlExample";
+            hintArgs = Map.of("example", "http://127.0.0.1:8080/v1");
+        }
+        return new Required(needsApiKey, true, hintKey, hintArgs);
+    }
+}
diff --git a/mateclaw-server/src/main/java/vip/mate/llm/model/ProviderInfoDTO.java b/mateclaw-server/src/main/java/vip/mate/llm/model/ProviderInfoDTO.java
index c5cfd317..1fac6e52 100644
--- a/mateclaw-server/src/main/java/vip/mate/llm/model/ProviderInfoDTO.java
+++ b/mateclaw-server/src/main/java/vip/mate/llm/model/ProviderInfoDTO.java
@@ -3,6 +3,7 @@ package vip.mate.llm.model;
 import lombok.Data;
 
 import java.util.ArrayList;
+import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Map;
 
@@ -41,4 +42,35 @@ public class ProviderInfoDTO {
     private Long cooldownRemainingMs;
     /** RFC-074: whether the user has explicitly enabled this provider. False = lives in the catalog drawer only. */
     private Boolean enabled;
+
+    // Issue #81: derived fields powering the chat-console liveness-aware popup.
+    // All six are computed from existing columns; none are persisted.
+
+    /** Credential status: CONFIGURED / MISSING / NOT_REQUIRED / OAUTH_PENDING. */
+    private String authStatus;
+
+    /** Base URL completeness: null when not applicable; true/false when applicable. */
+    private Boolean baseUrlComplete;
+
+    /** Comma-joined missing field names ("apiKey", "baseUrl"); empty when nothing missing. */
+    private String missingFields;
+
+    /**
+     * Machine-readable next-step key. Switches the chat popup's primary button text + handler.
+     * Values: fill_base_url / fill_api_key / start_oauth / configure_required_fields /
+     *         test_connection / pull_model / wait_cooldown / reprobe / none.
+     */
+    private String suggestedAction;
+
+    /**
+     * i18n key for an actionable hint (e.g. "provider.hint.llamacppBaseUrlExample").
+     * Frontend renders via t(key, args). Null when no hint applies.
+     */
+    private String suggestedActionHintKey;
+
+    /**
+     * Template parameters for {@link #suggestedActionHintKey}. Frontend passes
+     * this directly to vue-i18n. Empty map means no parameters.
+     */
+    private Map<String, Object> suggestedActionHintArgs = new LinkedHashMap<>();
 }
diff --git a/mateclaw-server/src/main/java/vip/mate/llm/service/ModelProviderService.java b/mateclaw-server/src/main/java/vip/mate/llm/service/ModelProviderService.java
index f7d2cade..b3f2b999 100644
--- a/mateclaw-server/src/main/java/vip/mate/llm/service/ModelProviderService.java
+++ b/mateclaw-server/src/main/java/vip/mate/llm/service/ModelProviderService.java
@@ -14,6 +14,7 @@ import vip.mate.llm.event.ModelConfigChangedEvent;
 import vip.mate.llm.failover.AvailableProviderPool;
 import vip.mate.llm.failover.ProviderHealthTracker;
 import vip.mate.llm.failover.ProviderInitProbe;
+import vip.mate.llm.failover.ProviderRequirements;
 import vip.mate.llm.model.*;
 import vip.mate.llm.repository.ModelProviderMapper;
 
@@ -240,12 +241,22 @@ public class ModelProviderService {
     public String getProviderUnavailableReason(String providerId) {
         ModelProviderEntity provider = getProvider(providerId);
         if (!isProviderConfigured(provider)) {
-            if (Boolean.TRUE.equals(provider.getRequireApiKey())) {
-                return "Provider 未配置有效的 API Key";
+            // Issue #81: emit a precise reason based on which row-level fields are
+            // missing, rather than the previous protocol-blind heuristic. The new
+            // frontend reads suggestedActionHintKey/Args; this string remains for
+            // logs and legacy callers.
+            ProviderRequirements.Required req = ProviderRequirements.of(provider);
+            boolean hasBaseUrl = StringUtils.hasText(provider.getBaseUrl());
+            boolean hasApiKey  = hasUsableApiKey(provider.getApiKey());
+            if (req.needsBaseUrl() && !hasBaseUrl && req.needsApiKey() && !hasApiKey) {
+                return "Provider 未配置 Base URL 和 API Key";
             }
-            if (Boolean.TRUE.equals(provider.getIsCustom()) || !Boolean.TRUE.equals(provider.getIsLocal())) {
+            if (req.needsBaseUrl() && !hasBaseUrl) {
                 return "Provider 未配置 Base URL";
             }
+            if (req.needsApiKey() && !hasApiKey) {
+                return "Provider 未配置 API Key";
+            }
             return "Provider 未完成配置";
         }
         if (!hasModels(providerId)) {
@@ -416,9 +427,85 @@ public class ModelProviderService {
         }
         dto.setModels(builtinModels);
         dto.setExtraModels(extraModels);
+        applySuggestedAction(dto, provider, providerLiveness);
         return dto;
     }
 
+    /**
+     * Issue #81: derive the chat-popup recovery hint from row + liveness, so the
+     * frontend can render a precise "next step" instead of a generic
+     * "model unavailable" toast. Six fields populated:
+     *   - authStatus: CONFIGURED / MISSING / NOT_REQUIRED / OAUTH_PENDING
+     *   - baseUrlComplete: null when not applicable, true/false otherwise
+     *   - missingFields: comma-joined ("apiKey", "baseUrl") for required-field UX
+     *   - suggestedAction: machine-readable next-step key (frontend switches on this)
+     *   - suggestedActionHintKey + suggestedActionHintArgs: i18n key/args, no raw text
+     */
+    private void applySuggestedAction(ProviderInfoDTO dto, ModelProviderEntity provider, Liveness liveness) {
+        ProviderRequirements.Required req = ProviderRequirements.of(provider);
+        boolean hasBaseUrl = StringUtils.hasText(provider.getBaseUrl());
+        boolean hasApiKey  = hasUsableApiKey(provider.getApiKey());
+        boolean hasModels  = (dto.getModels() != null && !dto.getModels().isEmpty())
+                          || (dto.getExtraModels() != null && !dto.getExtraModels().isEmpty());
+
+        // 1. authStatus
+        if ("oauth".equals(provider.getAuthType())) {
+            dto.setAuthStatus(Boolean.TRUE.equals(dto.getOauthConnected()) ? "CONFIGURED" : "OAUTH_PENDING");
+        } else if (req.needsApiKey()) {
+            dto.setAuthStatus(hasApiKey ? "CONFIGURED" : "MISSING");
+        } else {
+            dto.setAuthStatus("NOT_REQUIRED");
+        }
+
+        // 2. baseUrlComplete: null when this provider doesn't need a base URL.
+        dto.setBaseUrlComplete(req.needsBaseUrl() ? hasBaseUrl : null);
+
+        // 3. missingFields
+        java.util.List<String> missing = new ArrayList<>();
+        if (req.needsApiKey()  && !hasApiKey)  missing.add("apiKey");
+        if (req.needsBaseUrl() && !hasBaseUrl) missing.add("baseUrl");
+        dto.setMissingFields(String.join(",", missing));
+
+        // 4. suggestedAction
+        String action;
+        if (liveness == Liveness.UNCONFIGURED) {
+            if ("oauth".equals(provider.getAuthType())) {
+                action = "start_oauth";
+            } else if (missing.size() == 1 && missing.get(0).equals("baseUrl")) {
+                action = "fill_base_url";
+            } else if (missing.size() == 1 && missing.get(0).equals("apiKey")) {
+                action = "fill_api_key";
+            } else {
+                action = "configure_required_fields";
+            }
+        } else if (liveness == Liveness.REMOVED) {
+            action = "reprobe";
+        } else if (liveness == Liveness.COOLDOWN) {
+            action = "wait_cooldown";
+        } else if (liveness == Liveness.UNPROBED) {
+            action = "reprobe";
+        } else if (liveness == Liveness.LIVE && !hasModels) {
+            action = Boolean.TRUE.equals(provider.getSupportModelDiscovery())
+                    ? "pull_model"
+                    : "configure_required_fields";
+        } else {
+            action = "none";
+        }
+        dto.setSuggestedAction(action);
+
+        // 5. hint key + args (NOT raw text). Frontend renders via t(key, args).
+        // Only emit hint when it actually applies to the action; suppress for
+        // REMOVED / COOLDOWN / UNPROBED to keep the popup clean.
+        if ("fill_base_url".equals(action) || "configure_required_fields".equals(action)) {
+            dto.setSuggestedActionHintKey(req.hintKey());
+            dto.setSuggestedActionHintArgs(req.hintArgs() == null ? new java.util.LinkedHashMap<>()
+                                                                  : new java.util.LinkedHashMap<>(req.hintArgs()));
+        } else {
+            dto.setSuggestedActionHintKey(null);
+            dto.setSuggestedActionHintArgs(new java.util.LinkedHashMap<>());
+        }
+    }
+
     private boolean hasModels(String providerId) {
         return !modelConfigService.listModelsByProvider(providerId).isEmpty();
     }
@@ -427,13 +514,11 @@ public class ModelProviderService {
         if (provider == null) {
             return false;
         }
-        if (Boolean.TRUE.equals(provider.getIsLocal())) {
-            return true;
-        }
 
-        // OAuth 认证的 provider：检查 OAuth token 是否存在
+        // OAuth providers store credentials elsewhere (DB column or disk for
+        // Claude Code). Resolve them via the OAuth service rather than the
+        // base-URL / api-key columns.
         if ("oauth".equals(provider.getAuthType())) {
-            // Claude Code OAuth (RFC-062) — token lives on disk, not in DB.
             if (CLAUDE_CODE_PROVIDER_ID.equals(provider.getProviderId())) {
                 ClaudeCodeOAuthService svc = claudeCodeOAuthServiceProvider.getIfAvailable();
                 return svc != null && svc.isLoggedIn();
@@ -441,16 +526,21 @@ public class ModelProviderService {
             return StringUtils.hasText(provider.getOauthAccessToken());
         }
 
-        boolean hasBaseUrl = StringUtils.hasText(provider.getBaseUrl());
-        boolean hasApiKey = hasUsableApiKey(provider.getApiKey());
-
-        if (Boolean.TRUE.equals(provider.getIsCustom())) {
-            return hasBaseUrl && (!Boolean.TRUE.equals(provider.getRequireApiKey()) || hasApiKey);
+        // Issue #81: decide required fields from the provider row, not from the
+        // protocol enum. Every OpenAI-compatible provider (cloud or local) shares
+        // OPENAI_COMPATIBLE, so a protocol-keyed table cannot tell OpenAI cloud
+        // (needs api_key, no base url) apart from llama.cpp local (no api_key,
+        // needs base url). Without this, isLocal=true short-circuited to true
+        // for llama.cpp regardless of an empty Base URL, hiding the real cause
+        // behind a confusing REMOVED state.
+        ProviderRequirements.Required req = ProviderRequirements.of(provider);
+        if (req.needsApiKey() && !hasUsableApiKey(provider.getApiKey())) {
+            return false;
         }
-        if (Boolean.FALSE.equals(provider.getRequireApiKey())) {
-            return hasBaseUrl;
+        if (req.needsBaseUrl() && !StringUtils.hasText(provider.getBaseUrl())) {
+            return false;
         }
-        return hasApiKey;
+        return true;
     }
 
     public boolean hasUsableApiKey(String apiKey) {
diff --git a/mateclaw-ui/src/components/chat/ModelSelector.vue b/mateclaw-ui/src/components/chat/ModelSelector.vue
index 8bd869a1..b31bd5aa 100644
--- a/mateclaw-ui/src/components/chat/ModelSelector.vue
+++ b/mateclaw-ui/src/components/chat/ModelSelector.vue
@@ -34,25 +34,47 @@
               <div class="model-group-header">
                 <span class="model-group-header__name">{{ group.provider.name }}</span>
                 <span v-if="group.provider.isLocal" class="model-group-header__badge model-group-header__badge--local">Local</span>
-                <!-- RFC-073: liveness dot. UNPROBED = grey (still booting),
-                     COOLDOWN = amber (transient backoff). LIVE has no dot. -->
+                <!-- Liveness dot — UNPROBED = grey, COOLDOWN = amber, LIVE = none.
+                     v2 R3: when showAllStates, also surface UNCONFIGURED / REMOVED chips
+                     so the user can see what's wrong without leaving chat. -->
                 <span
                   v-if="group.provider.liveness === 'UNPROBED'"
                   class="model-group-header__dot model-group-header__dot--unprobed"
-                  :title="$t('chat.modelLivenessUnprobed')"
+                  :title="$t('provider.status.unprobed')"
                 ></span>
                 <span
                   v-else-if="group.provider.liveness === 'COOLDOWN'"
                   class="model-group-header__dot model-group-header__dot--cooldown"
-                  :title="$t('chat.modelLivenessCooldown', { seconds: cooldownSeconds(group.provider) })"
+                  :title="$t('provider.status.cooldown', { s: cooldownSeconds(group.provider) })"
                 ></span>
+                <span
+                  v-if="group.provider.liveness === 'UNCONFIGURED'"
+                  class="model-group-header__chip model-group-header__chip--warn"
+                >{{ $t('provider.status.unconfigured') }}</span>
+                <span
+                  v-else-if="group.provider.liveness === 'REMOVED'"
+                  class="model-group-header__chip model-group-header__chip--err"
+                >{{ $t('provider.status.removed') }}</span>
+                <span
+                  v-else-if="group.provider.liveness === 'COOLDOWN'"
+                  class="model-group-header__chip model-group-header__chip--info"
+                >{{ $t('provider.status.cooldown', { s: cooldownSeconds(group.provider) }) }}</span>
+                <button
+                  v-if="!isSelectable(group.provider)"
+                  class="model-group-header__fix"
+                  @click.stop="emitNavigateFix(group.provider)"
+                  type="button"
+                >{{ $t('chat.promptAction.fixThis') }}</button>
               </div>
               <div
                 v-for="item in group.models"
                 :key="item.value"
                 class="model-dropdown-item"
-                :class="{ active: item.value === activeValue, dimmed: group.provider.liveness === 'COOLDOWN' || group.provider.liveness === 'UNPROBED' }"
-                @click="handleSelect(item.value)"
+                :class="{
+                  active: item.value === activeValue,
+                  dimmed: !isSelectable(group.provider),
+                }"
+                @click="onItemClick(group.provider, item.value)"
               >
                 <span class="model-dropdown-item__name">{{ item.name }}</span>
                 <svg v-if="item.value === activeValue" class="model-dropdown-item__check" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5"><polyline points="20 6 9 17 4 12"/></svg>
@@ -66,7 +88,7 @@
                    state. Push them into Settings/Models with the drawer
                    pre-opened via ?addProvider=1. -->
               <template v-if="query.trim() === '' && groups.length === 0">
-                {{ $t('chat.noProvidersConfigured') }}
+                {{ emptyHint || $t('chat.noProvidersConfigured') }}
                 <RouterLink class="model-empty__cta" to="/settings/models?addProvider=1" @click="open = false">
                   {{ $t('chat.goConfigure') }}
                 </RouterLink>
@@ -106,10 +128,19 @@ const props = defineProps<{
   activeValue: string
   activeLabel: string
   saving?: boolean
+  /**
+   * v2 R3: when true, render UNCONFIGURED / REMOVED rows too (dimmed, with
+   * status chip + Fix button) so users see what's wrong without leaving chat.
+   * Defaults to false to keep the legacy filtered behavior for any other call site.
+   */
+  showAllStates?: boolean
+  /** Optional override for the empty-state text. */
+  emptyHint?: string
 }>()
 
 const emit = defineEmits<{
   select: [value: string]
+  'navigate-fix': [provider: ProviderInfo]
 }>()
 
 const open = ref(false)
@@ -143,13 +174,20 @@ function toggle() {
 }
 
 // 按 provider 分组，云端在前，本地在后
-// RFC-073: 仅过滤 UNCONFIGURED / REMOVED；UNPROBED + COOLDOWN 仍显示但视觉上区分。
+// v2 R3: when showAllStates is true, keep UNCONFIGURED/REMOVED visible (dimmed +
+// status chip + Fix button) so the user can act in-place. Default behavior
+// (showAllStates=false) preserves the original filter for non-popup contexts.
 function isHidden(p: ProviderInfo): boolean {
-  // 旧后端不返回 liveness 时退回 available 行为，避免渐进升级期间 UI 全空。
+  if (props.showAllStates) return false
   if (!p.liveness) return !p.available
   return p.liveness === 'UNCONFIGURED' || p.liveness === 'REMOVED'
 }
 
+function isSelectable(p: ProviderInfo): boolean {
+  if (!p.liveness) return p.available === true
+  return p.liveness === 'LIVE' || p.liveness === 'COOLDOWN'
+}
+
 const groups = computed<ModelGroup[]>(() => {
   const cloud: ModelGroup[] = []
   const local: ModelGroup[] = []
@@ -157,7 +195,19 @@ const groups = computed<ModelGroup[]>(() => {
   for (const provider of props.providers) {
     if (isHidden(provider)) continue
     const allModels = [...(provider.models || []), ...(provider.extraModels || [])]
-    if (allModels.length === 0) continue
+    // When showAllStates is on we still want UNCONFIGURED rows to appear even
+    // if they have no models — synthesize one placeholder so the header chip +
+    // Fix button render. Otherwise (legacy call site) skip empty groups.
+    if (allModels.length === 0) {
+      if (!props.showAllStates) continue
+      const group: ModelGroup = {
+        provider,
+        models: [],
+      }
+      if (provider.isLocal) local.push(group)
+      else cloud.push(group)
+      continue
+    }
 
     const group: ModelGroup = {
       provider,
@@ -213,6 +263,25 @@ function handleSelect(value: string) {
   emit('select', value)
 }
 
+/**
+ * v2 R3: a row is selectable iff its provider can serve a request now (LIVE) or
+ * is just throttled (COOLDOWN — backend lets the request go and the chain
+ * walker handles fallback). UNCONFIGURED / REMOVED / UNPROBED need user action,
+ * so clicking those routes to the Fix flow instead of trying to switch.
+ */
+function onItemClick(provider: ProviderInfo, value: string) {
+  if (isSelectable(provider)) {
+    handleSelect(value)
+  } else {
+    emitNavigateFix(provider)
+  }
+}
+
+function emitNavigateFix(provider: ProviderInfo) {
+  open.value = false
+  emit('navigate-fix', provider)
+}
+
 // 打开时聚焦搜索框 + 滚动到当前选中项
 watch(open, async (isOpen) => {
   if (isOpen) {
@@ -385,6 +454,52 @@ watch(open, async (isOpen) => {
   50% { opacity: 1; }
 }
 
+/* Liveness status chips — surfaced when ModelSelector is in show-all-states mode. */
+.model-group-header__chip {
+  display: inline-flex;
+  align-items: center;
+  height: 16px;
+  padding: 0 6px;
+  border-radius: 4px;
+  font-size: 10px;
+  font-weight: 600;
+  letter-spacing: 0.02em;
+  text-transform: none;
+}
+.model-group-header__chip--warn {
+  background: rgba(245, 158, 11, 0.15);
+  color: #b45309;
+}
+.model-group-header__chip--err {
+  background: rgba(239, 68, 68, 0.15);
+  color: #b91c1c;
+}
+.model-group-header__chip--info {
+  background: rgba(59, 130, 246, 0.15);
+  color: #1d4ed8;
+}
+.dark .model-group-header__chip--warn { color: #fbbf24; }
+.dark .model-group-header__chip--err  { color: #fca5a5; }
+.dark .model-group-header__chip--info { color: #93c5fd; }
+
+.model-group-header__fix {
+  margin-left: auto;
+  height: 18px;
+  padding: 0 8px;
+  border-radius: 4px;
+  border: 1px solid var(--mc-border);
+  background: transparent;
+  color: var(--mc-text-secondary);
+  font-size: 11px;
+  cursor: pointer;
+  transition: background 0.12s, color 0.12s, border-color 0.12s;
+}
+.model-group-header__fix:hover {
+  background: var(--mc-primary);
+  color: #fff;
+  border-color: var(--mc-primary);
+}
+
 /* ---- Items ---- */
 
 .model-dropdown-item {
diff --git a/mateclaw-ui/src/components/chat/RecoverableModelBanner.vue b/mateclaw-ui/src/components/chat/RecoverableModelBanner.vue
new file mode 100644
index 00000000..84f26c18
--- /dev/null
+++ b/mateclaw-ui/src/components/chat/RecoverableModelBanner.vue
@@ -0,0 +1,78 @@
+<template>
+  <div class="recoverable-banner" role="status">
+    <span class="recoverable-banner__icon">⚠️</span>
+    <span class="recoverable-banner__text">
+      {{ $t('chat.recoverableBanner.message', { name: providerName, fallback: fallbackName }) }}
+    </span>
+    <button class="recoverable-banner__dismiss" type="button" @click="$emit('dismiss')">
+      <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.4">
+        <line x1="18" y1="6" x2="6" y2="18"/>
+        <line x1="6" y1="6" x2="18" y2="18"/>
+      </svg>
+    </button>
+  </div>
+</template>
+
+<script setup lang="ts">
+/**
+ * Issue #81: shown when the active provider is unhealthy but the chain walker
+ * can fall back to a LIVE one. Non-blocking — input stays enabled, send still
+ * works; this is just a heads-up.
+ */
+defineProps<{
+  providerName: string
+  fallbackName: string
+}>()
+
+defineEmits<{
+  dismiss: []
+}>()
+</script>
+
+<style scoped>
+.recoverable-banner {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  margin: 8px 16px 0;
+  padding: 8px 12px;
+  border: 1px solid rgba(245, 158, 11, 0.35);
+  background: rgba(245, 158, 11, 0.10);
+  border-radius: 10px;
+  color: var(--mc-text-primary);
+  font-size: 13px;
+  line-height: 1.4;
+}
+
+.recoverable-banner__icon {
+  flex-shrink: 0;
+  font-size: 14px;
+}
+
+.recoverable-banner__text {
+  flex: 1;
+  min-width: 0;
+}
+
+.recoverable-banner__dismiss {
+  flex-shrink: 0;
+  width: 22px;
+  height: 22px;
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  border: none;
+  background: transparent;
+  color: var(--mc-text-tertiary);
+  border-radius: 6px;
+  cursor: pointer;
+  transition: background 0.12s, color 0.12s;
+}
+.recoverable-banner__dismiss:hover {
+  background: rgba(0, 0, 0, 0.06);
+  color: var(--mc-text-primary);
+}
+.dark .recoverable-banner__dismiss:hover {
+  background: rgba(255, 255, 255, 0.08);
+}
+</style>
diff --git a/mateclaw-ui/src/i18n/locales/en-US.ts b/mateclaw-ui/src/i18n/locales/en-US.ts
index d14f0d56..ba0e18c1 100644
--- a/mateclaw-ui/src/i18n/locales/en-US.ts
+++ b/mateclaw-ui/src/i18n/locales/en-US.ts
@@ -180,6 +180,49 @@ export default {
     // RFC-074 PR-2: empty-state inside the model dropdown
     noProvidersConfigured: 'No models available yet',
     goConfigure: 'Configure',
+    // Issue #81: liveness-aware popup state machine.
+    prompt: {
+      noActive: {
+        title: 'Pick a model first',
+        desc: 'No active model. Choose one in Model Management before chatting.',
+      },
+      unconfigured: {
+        title: '{name} is not fully configured',
+        desc: 'Missing: {fields}. {hint}',
+      },
+      removed: {
+        title: '{name} is unavailable',
+        descFallback: 'The last probe failed — check whether the service is running.',
+      },
+      cooldown: {
+        title: '{name} is temporarily unavailable',
+        desc: 'Auto-retrying in about {seconds}s.',
+      },
+      unprobed: {
+        title: 'Checking model availability',
+        desc: 'Hold on or refresh to retry.',
+      },
+      noModels: {
+        title: '{name} has no usable models',
+        desc: 'Discover models or add one manually.',
+      },
+    },
+    promptAction: {
+      fillBaseUrl: 'Fill Base URL',
+      fillApiKey: 'Fill API Key',
+      fillRequiredFields: 'Complete required fields',
+      startOAuth: 'Sign in',
+      testConnection: 'Test connection',
+      pullModel: 'Discover models',
+      waitCooldown: 'Retry now',
+      reprobe: 'Re-probe',
+      switchToModel: 'Switch to {name}',
+      fixThis: 'Fix',
+    },
+    recoverableBanner: {
+      message: '{name} is unavailable; sends will fall back to {fallback} automatically',
+      switched: 'Primary model unavailable; switched to {fallback}',
+    },
     uploadFailed: 'File upload failed',
     dropToUpload: 'Drop files or folders here',
     copyFailed: 'Copy failed',
@@ -429,6 +472,22 @@ export default {
     lastChecked: 'Checked {time} ago',
     diagnose: 'Diagnose',
   },
+  // Issue #81: provider-level hint / status text shared across chat popup and ModelSelector.
+  provider: {
+    hint: {
+      ollamaBaseUrlExample: 'Example: {example}',
+      lmstudioBaseUrlExample: 'Example: {example} (LM Studio default port)',
+      llamacppBaseUrlExample: 'Example: {example} (llama-server default port)',
+      vllmBaseUrlExample: 'Example: {example}',
+      openaiCompatBaseUrlExample: 'Example: OpenAI-compatible endpoint, e.g. {example}',
+    },
+    status: {
+      unconfigured: 'Not configured',
+      removed: 'Unavailable',
+      cooldown: 'Retry in {s}s',
+      unprobed: 'Probing',
+    },
+  },
   settings: {
     title: 'Settings',
     kicker: 'Configuration',
diff --git a/mateclaw-ui/src/i18n/locales/zh-CN.ts b/mateclaw-ui/src/i18n/locales/zh-CN.ts
index 319cb467..3b00afd1 100644
--- a/mateclaw-ui/src/i18n/locales/zh-CN.ts
+++ b/mateclaw-ui/src/i18n/locales/zh-CN.ts
@@ -180,6 +180,49 @@ export default {
     // RFC-074 PR-2: empty-state inside the model dropdown
     noProvidersConfigured: '还没有可用的模型',
     goConfigure: '去配置',
+    // Issue #81: liveness-aware popup state machine.
+    prompt: {
+      noActive: {
+        title: '请先选择一个模型',
+        desc: '当前没有激活模型，先到模型管理选一个再开始对话。',
+      },
+      unconfigured: {
+        title: '{name} 还没配置完',
+        desc: '缺少：{fields}。{hint}',
+      },
+      removed: {
+        title: '{name} 不可用',
+        descFallback: '上一次探测失败，请检查服务是否在运行。',
+      },
+      cooldown: {
+        title: '{name} 暂时不可用',
+        desc: '约 {seconds} 秒后自动重试。',
+      },
+      unprobed: {
+        title: '正在检查模型可用性',
+        desc: '稍候片刻或刷新重试。',
+      },
+      noModels: {
+        title: '{name} 下没有可用模型',
+        desc: '尝试发现模型或手动添加一个。',
+      },
+    },
+    promptAction: {
+      fillBaseUrl: '填写 Base URL',
+      fillApiKey: '填写 API Key',
+      fillRequiredFields: '完成必填项',
+      startOAuth: '登录授权',
+      testConnection: '测试连接',
+      pullModel: '发现模型',
+      waitCooldown: '立即重试',
+      reprobe: '重新探测',
+      switchToModel: '切换到 {name}',
+      fixThis: '修复',
+    },
+    recoverableBanner: {
+      message: '{name} 暂时不可用，发送时会自动用 {fallback} 兜底',
+      switched: '主模型不可用，已切换到 {fallback}',
+    },
     uploadFailed: '文件上传失败',
     dropToUpload: '拖放文件或文件夹到此处',
     copyFailed: '复制失败',
@@ -315,6 +358,22 @@ export default {
     roleUser: '用户',
     roleAdmin: '管理员',
   },
+  // Issue #81: provider-level hint / status text shared across chat popup and ModelSelector.
+  provider: {
+    hint: {
+      ollamaBaseUrlExample: '示例：{example}',
+      lmstudioBaseUrlExample: '示例：{example}（LM Studio 默认端口）',
+      llamacppBaseUrlExample: '示例：{example}（llama-server 默认端口）',
+      vllmBaseUrlExample: '示例：{example}',
+      openaiCompatBaseUrlExample: '示例：OpenAI 兼容端点，例如 {example}',
+    },
+    status: {
+      unconfigured: '未配置',
+      removed: '不可用',
+      cooldown: '{s}s 后重试',
+      unprobed: '检测中',
+    },
+  },
   settings: {
     title: '设置',
     kicker: '配置中心',
diff --git a/mateclaw-ui/src/types/index.ts b/mateclaw-ui/src/types/index.ts
index fb5d824d..3efd79f1 100644
--- a/mateclaw-ui/src/types/index.ts
+++ b/mateclaw-ui/src/types/index.ts
@@ -742,6 +742,24 @@ export interface ProviderInfo {
   cooldownRemainingMs?: number
   /** RFC-074: whether the user has explicitly opted this provider into the dropdown. */
   enabled?: boolean
+
+  // Issue #81: derived liveness fields powering the chat-console popup state machine.
+  /** CONFIGURED / MISSING / NOT_REQUIRED / OAUTH_PENDING. */
+  authStatus?: string
+  /** null = base url N/A; true/false = applicable and complete/incomplete. */
+  baseUrlComplete?: boolean | null
+  /** Comma-joined missing field names ("apiKey", "baseUrl"); empty when nothing missing. */
+  missingFields?: string
+  /**
+   * Machine-readable next-step key driving the popup primary button.
+   * fill_base_url / fill_api_key / start_oauth / configure_required_fields /
+   * test_connection / pull_model / wait_cooldown / reprobe / none.
+   */
+  suggestedAction?: string
+  /** i18n key for the actionable hint, e.g. "provider.hint.llamacppBaseUrlExample". */
+  suggestedActionHintKey?: string | null
+  /** Template params for vue-i18n t(key, args). */
+  suggestedActionHintArgs?: Record<string, unknown>
 }
 
 /**
diff --git a/mateclaw-ui/src/views/ChatConsole.vue b/mateclaw-ui/src/views/ChatConsole.vue
index cb0cf9ee..51cf4b34 100644
--- a/mateclaw-ui/src/views/ChatConsole.vue
+++ b/mateclaw-ui/src/views/ChatConsole.vue
@@ -174,18 +174,18 @@
           <div v-else class="no-agent-hint">{{ $t('chat.selectAgent') }}</div>
         </div>
         <div class="chat-header-right">
-          <!-- Model selector -->
+          <!-- Model selector — Issue #81 v2 R3: always pass full providers + show-all-states
+               so unhealthy rows render as dimmed entries with status chips and a Fix
+               button instead of disappearing entirely. -->
           <ModelSelector
-            v-if="eligibleModels.length > 0"
-            :providers="availableProviders"
+            :providers="providers"
             :active-value="activeModelValue"
             :active-label="activeModelLabel"
             :saving="modelSaving"
+            :show-all-states="true"
             @select="selectModel"
+            @navigate-fix="onModelSelectorFix"
           />
-          <button v-else class="header-btn" @click="goToModelSettings" :title="$t('chat.configModel')">
-            <el-icon><Setting /></el-icon>
-          </button>
           <!-- Overflow menu -->
           <div class="header-overflow-wrap">
             <button class="header-btn" @click="headerMenuOpen = !headerMenuOpen" :title="$t('common.more') || 'More'">
@@ -218,25 +218,46 @@
         :loading="isGenerating"
         :assistant-icon="currentAgent?.icon || '🤖'"
         :user-icon="userInitial"
-        :title="showModelPrompt ? modelPromptTitle : $t('app.title')"
-        :subtitle="showModelPrompt ? modelPromptDesc : $t('chat.subtitle')"
-        :suggestions="showModelPrompt ? [] : suggestions"
+        :title="blockingPrompt ? modelPromptText.title : $t('app.title')"
+        :subtitle="blockingPrompt ? modelPromptText.desc : $t('chat.subtitle')"
+        :suggestions="blockingPrompt ? [] : suggestions"
         @regenerate="handleRegenerate"
         @suggestion-click="sendSuggestion"
         @toggle-thinking="handleToggleThinking"
         @approve="handleApprove"
         @deny="handleDeny"
       >
-        <!-- 自定义模型提示空状态 -->
-        <template v-if="showModelPrompt" #empty>
+        <!-- Issue #81 v2 R2: blocking-only popup. Recoverable cases use the
+             non-blocking <RecoverableModelBanner> below instead. -->
+        <template v-if="blockingPrompt" #empty>
           <div class="model-prompt">
-            <div class="model-prompt-title">{{ modelPromptTitle }}</div>
-            <div class="model-prompt-desc">{{ modelPromptDesc }}</div>
-            <button class="btn-primary" @click="goToModelSettings">{{ $t('chat.goToModelSettings') }}</button>
+            <div class="model-prompt-title">{{ modelPromptText.title }}</div>
+            <div class="model-prompt-desc">{{ modelPromptText.desc }}</div>
+            <div class="model-prompt-actions">
+              <button class="btn-primary" @click="handlePrimaryAction">
+                {{ primaryActionLabel }}
+              </button>
+              <button
+                v-if="bestSwitchTarget"
+                class="btn-secondary"
+                @click="switchToBestTarget"
+              >
+                {{ $t('chat.promptAction.switchToModel', { name: bestSwitchTarget.label }) }}
+              </button>
+            </div>
           </div>
         </template>
       </MessageList>
 
+      <!-- Issue #81: non-blocking banner — active provider is unhealthy but the
+           backend fallback chain has a LIVE provider to take over. -->
+      <RecoverableModelBanner
+        v-if="recoverablePrompt && activeProvider && bestFallbackName"
+        :provider-name="activeProvider.name"
+        :fallback-name="bestFallbackName"
+        @dismiss="recoverableDismissed = true"
+      />
+
       <!-- Cron job in-flight placeholder — visible while T2 hasn't committed
            the assistant message yet. Populated by pollActivity → /cron-jobs/active-runs. -->
       <div v-if="activeCronRuns.length > 0" class="cron-running-bar">
@@ -254,7 +275,7 @@
 
       <!-- 流式处理 Loading 栏（消息和输入框之间） -->
       <StreamLoadingBar
-        :is-loading="isGenerating && !showModelPrompt"
+        :is-loading="isGenerating && !blockingPrompt"
         :tool-count="toolCallCount"
         :completion-tokens="currentGeneratingTokens"
         :prompt-tokens="currentPromptTokens"
@@ -270,7 +291,7 @@
         ref="chatInputRef"
         v-model="inputText"
         :loading="isGenerating && !hasPendingApproval"
-        :disabled="showModelPrompt || !currentAgent"
+        :disabled="blockingPrompt || !currentAgent"
         :placeholder="$t('chat.messagePlaceholder')"
         :hint="currentRuntimeModel"
         :attachments="pendingAttachments"
@@ -325,6 +346,7 @@ import type { Conversation, Agent, ModelConfig, ProviderInfo, ActiveModelsInfo,
 
 // 导入组件化组件
 import MessageList from '@/components/chat/MessageList.vue'
+import RecoverableModelBanner from '@/components/chat/RecoverableModelBanner.vue'
 import SkillIcon from '@/components/common/SkillIcon.vue'
 import { parsePrompt, deriveTagline } from '@/utils/agentPromptProfile'
 import { agentIconColor } from '@/utils/agentIconColor'
@@ -403,7 +425,13 @@ const selectedAgentId = ref<string | number>('')
 const currentConversationId = ref<string>('')
 const inputText = ref('')
 const modelSaving = ref(false)
-const showModelPrompt = ref(false)
+// Issue #81 v2 R2: split the single showModelPrompt boolean into two flags so
+// the chat surface can either hard-block (blockingPrompt) or warn but let the
+// backend fallback chain take over (recoverablePrompt). Driven by
+// recomputePromptFlags() — see the watcher below.
+const blockingPrompt = ref(false)
+const recoverablePrompt = ref(false)
+const recoverableDismissed = ref(false)
 const defaultModel = ref<ModelConfig | null>(null)
 const providers = ref<ProviderInfo[]>([])
 const activeModels = ref<ActiveModelsInfo | null>(null)
@@ -812,26 +840,107 @@ const activeProvider = computed(() => {
   return providerId ? providers.value.find((provider) => provider.id === providerId) || null : null
 })
 
-const modelPromptTitle = computed(() => {
-  if (!activeModels.value?.activeLlm?.providerId || !activeModels.value?.activeLlm?.model) {
-    return t('chat.configModelFirst')
+// Issue #81: liveness-aware popup state machine. modelPromptKind picks one of
+// six branches; modelPromptText derives title + desc; primaryActionLabel +
+// handlePrimaryAction map to the suggestedAction the backend computed.
+type ModelPromptKind = 'no-active' | 'unconfigured' | 'removed' | 'cooldown' | 'unprobed' | 'no-models'
+
+const modelPromptKind = computed<ModelPromptKind>(() => {
+  if (!activeModels.value?.activeLlm?.providerId) return 'no-active'
+  const p = activeProvider.value
+  if (!p) return 'no-active'
+  switch (p.liveness) {
+    case 'UNCONFIGURED': return 'unconfigured'
+    case 'REMOVED':      return 'removed'
+    case 'COOLDOWN':     return 'cooldown'
+    case 'UNPROBED':     return 'unprobed'
+    case 'LIVE':         return 'no-models'
+    default:             return 'no-active'
   }
-  if (activeProvider.value && !activeProvider.value.available) {
-    return t('chat.modelUnavailable')
-  }
-  return t('chat.configModelFirst')
 })
 
-const modelPromptDesc = computed(() => {
-  if (!activeModels.value?.activeLlm?.providerId || !activeModels.value?.activeLlm?.model) {
-    return t('chat.noActiveModel')
-  }
-  if (activeProvider.value && !activeProvider.value.available) {
-    return t('chat.providerNotReady', { name: activeProvider.value.name })
-  }
-  return t('chat.noAvailableModel')
+const hintText = computed(() => {
+  const p = activeProvider.value
+  if (!p?.suggestedActionHintKey) return ''
+  return t(p.suggestedActionHintKey, (p.suggestedActionHintArgs || {}) as Record<string, unknown>)
 })
 
+const modelPromptText = computed<{ title: string; desc: string }>(() => {
+  const p = activeProvider.value
+  switch (modelPromptKind.value) {
+    case 'no-active':
+      return { title: t('chat.prompt.noActive.title'), desc: t('chat.prompt.noActive.desc') }
+    case 'unconfigured':
+      return {
+        title: t('chat.prompt.unconfigured.title', { name: p?.name || '' }),
+        desc:  t('chat.prompt.unconfigured.desc', { fields: p?.missingFields || '', hint: hintText.value }),
+      }
+    case 'removed':
+      return {
+        title: t('chat.prompt.removed.title', { name: p?.name || '' }),
+        desc:  p?.unavailableReason || t('chat.prompt.removed.descFallback'),
+      }
+    case 'cooldown':
+      return {
+        title: t('chat.prompt.cooldown.title', { name: p?.name || '' }),
+        desc:  t('chat.prompt.cooldown.desc', {
+          seconds: Math.max(1, Math.ceil((p?.cooldownRemainingMs || 0) / 1000)),
+        }),
+      }
+    case 'unprobed':
+      return { title: t('chat.prompt.unprobed.title'), desc: t('chat.prompt.unprobed.desc') }
+    case 'no-models':
+      return {
+        title: t('chat.prompt.noModels.title', { name: p?.name || '' }),
+        desc:  t('chat.prompt.noModels.desc'),
+      }
+  }
+})
+
+const primaryActionLabel = computed(() => {
+  const action = activeProvider.value?.suggestedAction || 'configure_required_fields'
+  switch (action) {
+    case 'fill_base_url':              return t('chat.promptAction.fillBaseUrl')
+    case 'fill_api_key':               return t('chat.promptAction.fillApiKey')
+    case 'start_oauth':                return t('chat.promptAction.startOAuth')
+    case 'test_connection':            return t('chat.promptAction.testConnection')
+    case 'pull_model':                 return t('chat.promptAction.pullModel')
+    case 'wait_cooldown':              return t('chat.promptAction.waitCooldown')
+    case 'reprobe':                    return t('chat.promptAction.reprobe')
+    case 'configure_required_fields':
+    default:                           return t('chat.goToModelSettings')
+  }
+})
+
+function handlePrimaryAction() {
+  goToModelSettings(activeProvider.value?.id)
+}
+
+/** First eligible model that is NOT the active one — what the secondary button switches to. */
+const bestSwitchTarget = computed<{ value: string; label: string } | null>(() => {
+  for (const m of eligibleModels.value) {
+    if (m.value !== activeModelValue.value) return m
+  }
+  return null
+})
+
+/** First LIVE provider name — used by RecoverableModelBanner. */
+const bestFallbackName = computed<string>(() => {
+  const target = bestSwitchTarget.value
+  if (!target) return ''
+  const [providerId] = target.value.split('::')
+  return providers.value.find(p => p.id === providerId)?.name || ''
+})
+
+function switchToBestTarget() {
+  const t = bestSwitchTarget.value
+  if (t) selectModel(t.value)
+}
+
+function onModelSelectorFix(provider: { id: string }) {
+  goToModelSettings(provider.id)
+}
+
 const availableProviders = computed(() =>
   providers.value.filter((p) => p.available && [...(p.models || []), ...(p.extraModels || [])].length > 0)
 )
@@ -1049,19 +1158,53 @@ async function loadModelState() {
     defaultModel.value = defaultRes.data || null
     providers.value = providersRes.data || []
     activeModels.value = activeRes.data || null
-    const providerId = activeModels.value?.activeLlm?.providerId
-    const activeProviderInfo = providerId
-      ? providers.value.find((provider) => provider.id === providerId)
-      : null
-    showModelPrompt.value = !activeModels.value?.activeLlm?.providerId
-      || !activeModels.value?.activeLlm?.model
-      || (Boolean(providerId) && !activeProviderInfo?.available)
+    recomputePromptFlags()
   } catch (e) {
     ElMessage.error(t('chat.loadModelFailed'))
-    showModelPrompt.value = true
+    blockingPrompt.value = true
+    recoverablePrompt.value = false
   }
 }
 
+/**
+ * Issue #81 v2 R2: derive blocking / recoverable prompt flags from the current
+ * providers + active model snapshot. Called from loadModelState after every
+ * /providers refresh, and from a watcher when the user switches model. The
+ * runtime fallback chain in NodeStreamingChatHelper picks the first LIVE
+ * provider regardless of which one is "active", so as long as ANY provider is
+ * LIVE we should NOT block — we just hint with a banner.
+ */
+function recomputePromptFlags() {
+  const active = activeModels.value?.activeLlm
+  if (!active?.providerId || !active?.model) {
+    blockingPrompt.value = true
+    recoverablePrompt.value = false
+    recoverableDismissed.value = false
+    return
+  }
+  const ap = providers.value.find(p => p.id === active.providerId) || null
+  const apHasModels = ap
+    ? ((ap.models?.length || 0) + (ap.extraModels?.length || 0)) > 0
+    : false
+  const activeUsable = ap?.liveness === 'LIVE' && apHasModels
+  if (activeUsable) {
+    blockingPrompt.value = false
+    recoverablePrompt.value = false
+    recoverableDismissed.value = false
+    return
+  }
+  const anyUsable = providers.value.some(p =>
+    p.liveness === 'LIVE'
+    && ((p.models?.length || 0) + (p.extraModels?.length || 0)) > 0)
+  blockingPrompt.value = !anyUsable
+  recoverablePrompt.value = anyUsable && !recoverableDismissed.value
+}
+
+// Issue #81 v2 R2: keep blocking/recoverable in sync with the providers list
+// and the active model selection without forcing every mutation site to call
+// recomputePromptFlags() manually.
+watch([providers, activeModels], recomputePromptFlags, { deep: true })
+
 async function loadConversations() {
   try {
     const res: any = await conversationApi.list()
@@ -1315,8 +1458,12 @@ async function clearMessages() {
 
 // onModelChange removed — replaced by selectModel()
 
-function goToModelSettings() {
-  router.push('/settings/models')
+function goToModelSettings(providerId?: string) {
+  // Issue #81: when called with a providerId (e.g. from the unhealthy popup or
+  // ModelSelector's Fix button), pass it as a query param so a follow-up PR can
+  // scroll/focus the right card on the settings page. Today the consumer just
+  // ignores it; harmless meanwhile.
+  router.push({ path: '/settings/models', query: providerId ? { focus: providerId } : {} })
 }
 
 // ============ 计算属性：是否有待审批 ============
@@ -1370,7 +1517,7 @@ async function handleSendMessage(content: string) {
   // 允许在等待审批时发送审批命令
   const isApprovalCommand = /^\/(approve|deny)$/i.test(content.trim())
 
-  if ((!content && pendingAttachments.value.length === 0) || !selectedAgentId.value || showModelPrompt.value) return
+  if ((!content && pendingAttachments.value.length === 0) || !selectedAgentId.value || blockingPrompt.value) return
   // 不再阻止运行中发送 — useChat 会自动走 interrupt/queue 路径
 
   // 拦截 /approve 和 /deny 命令 —— 通过 SSE 流发送（和普通消息相同通道）
@@ -1421,12 +1568,15 @@ async function handleSendMessage(content: string) {
     currentConversationId.value = `conv_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`
   }
 
+  // Issue #81 v2 R2: only abort when there is genuinely no usable provider.
+  // If the active provider is unhealthy but another is LIVE, let the request
+  // through — NodeStreamingChatHelper's fallback walker will pick it up and
+  // emit a "warning" SSE delta which the input handler surfaces as a toast.
   if (!activeModels.value?.activeLlm?.providerId || !activeModels.value?.activeLlm?.model) {
-    showModelPrompt.value = true
+    blockingPrompt.value = true
     return
   }
-  if (!activeProvider.value?.available) {
-    showModelPrompt.value = true
+  if (blockingPrompt.value) {
     return
   }
 
@@ -2523,6 +2673,30 @@ function handleCodeCopy(e: MouseEvent) {
   background: var(--mc-primary-hover);
 }
 
+/* Issue #81: side-by-side primary + secondary actions in the model prompt. */
+.model-prompt-actions {
+  display: inline-flex;
+  flex-wrap: wrap;
+  gap: 8px;
+  justify-content: center;
+}
+
+.btn-secondary {
+  padding: 8px 14px;
+  background: transparent;
+  color: var(--mc-text-primary);
+  border: 1px solid var(--mc-border);
+  border-radius: 12px;
+  font-size: 14px;
+  cursor: pointer;
+  transition: background 0.15s, color 0.15s, border-color 0.15s;
+}
+
+.btn-secondary:hover {
+  background: var(--mc-panel-raised);
+  border-color: var(--mc-primary);
+}
+
 /* ===== 移动端元素（桌面端隐藏） ===== */
 .conv-backdrop {
   display: none;