feat: admin 端点补齐 + 视觉真推理 + 语音链路 + wire.go 收口（A+B+第3步）

批次A admin 端点（原本 404 变 200 真实数据）：
- GET /admin/llm：读 s.cfg.LLM 组 openai+anthropic 两 provider（enabled=当前 Provider、multimodal=配 VisionModel），APIKey 不进 wire 零泄露。
- GET /admin/automations：*automations.Service.List()（reminders+event rules+cron schedules，cron 加只读登记）。
- GET /admin/users：cfg.AdminUser + *auth.APIKey.List()（密钥掩码）；无用户存储，恒先返回当前 admin 账号。
- GET /admin/backthink：*backthink.Backthink.List(ctx,50)。均只读、owner/admin RBAC、{data} 信封。

批次B 视觉真推理 + 语音：
- rest.go chatRequest 加 Image?；流式 publishChatStream 带 image；userChatMessages 构造 ImageMessage(model.go:71)。
- loop/vision.go：ImageResolver 可注入（MVP 先看图再对话）；run.go Run 开头 resolveImages，无 resolver 透传直交主模型。
- vision/resolver.go：MessageResolver 调 vision.Router.Route——KindMain 保留图片交主模型多模态、KindVisionTool/OCRTool 经 dispatcher.Invoke 折叠成 [图片内容] 文本、KindUnsupported 记日志保留原图不崩。
- gateway/voice.go：VoiceService 接口 + SetVoice 注入点 + handleVoiceInput（未注入回 error 不崩）+ runVoiceTurn 异步 STT→loop chat→TTS→定向 voice_output 帧。
- ws_hub/ws_route：voice_input 接 handleVoiceInput；前端 ws.ts decodeBase64/playVoiceAudio/case voice_output。

第3步 wire.go 收口（唯一碰 wire.go 的一步）：
- Deps 补传 Automations:a.automation / Backthink:a.backthink / Users:auth.NewAPIKey(a.cfg.Auth)（*auth.APIKey 原生满足 UserStore，避免 admin 重复的复合类型）。
- gw.SetVoice(a.voice)；loop.Options.Vision = vision.NewMessageResolver(router, info, dispatcher.Invoke)（router=NewRouter(WithVisionTool/WithOCRTool)、info=cfg.LLM.VisionModel）。

实机验证：四 admin 端点 200 真实数据（llm 双 provider 脱敏、users=admin,yeij0942,owner）；WS 握手正常。go build/vet/test 全绿 + 前端 tsc。wire.go 557 行。

Docs/api-client.md

@@ -21,7 +21,7 @@
 |---|------|------|------|:---:|------|
 | 1 | `/api/v1/auth/login` | POST | 否 | ✅ | `{username,password}` → `{data:{token,jwt}}`（24h，owner） |
 | 2 | `/api/v1/health` | GET | 否 | ✅ | 健康检查 `{data:{status:"ok"}}` |
-| 3 | `/api/v1/chat` | POST | JWT | ✅ | `{text,mode?,stream?}` → `{data:{reply}}`；`stream:true` 时经 WS 触发流式 |
+| 3 | `/api/v1/chat` | POST | JWT | ✅ | `{text,image?,stream?}` → `{data:{reply}}`；`stream:true` 时经 WS 触发流式。`image` 为 base64 data: URI 或 http(s) URL，入视觉真推理（design/06） |
 | 4 | `/api/v1/sessions` | GET | JWT | ✅ | `{data:[session...]}` |
 | 5 | `/api/v1/personas` | GET | JWT | ✅ | `{data:[persona...]}` |
 
@@ -38,17 +38,17 @@
 | 5 | `/api/v1/admin/plugins/policy` | POST | ✅ | 写工具策略（owner-only，审计） |
 | 6 | `/api/v1/admin/channels` | GET | ✅ | 渠道列表/健康 |
 | 7 | `/api/v1/admin/iot` | GET | ✅ | IoT 设备列表 |
-| 8 | `/api/v1/admin/users` | GET/POST | 🕐 | 用户/角色/API key（M10） |
+| 8 | `/api/v1/admin/users` | GET/POST | ✅(GET) | 用户/角色/API key（GET 已实现；POST 创建用户/签发 key 随 M10 落地） |
-| 9 | `/api/v1/admin/llm` | GET/PUT | 🕐 | LLM 模型/路由/降级（M10） |
+| 9 | `/api/v1/admin/llm` | GET/PUT | ✅(GET) | LLM 模型/路由/降级（GET 已实现，APIKey 脱敏；PUT 热重载 M10） |
-| 10 | `/api/v1/admin/automations` | GET/POST | 🕐 | 自动化与提醒（M10） |
+| 10 | `/api/v1/admin/automations` | GET/POST | ✅(GET) | 自动化与提醒（GET 已实现；POST 创建 M10） |
+| 11 | `/api/v1/admin/backthink` | GET | ✅ | 后台思考记录 |
 
-> **M10 前端契约补充（管理面板，均🕐未实现，前端按契约封装、后端 404 时显示"待实现"壳，不报错）**：
+> **M10 前端契约补充（管理面板）**：`GET /api/v1/admin/users` · `GET /api/v1/admin/llm` · `GET /api/v1/admin/automations` · `GET /api/v1/admin/backthink` 已实现，返回真实数据；`GET /api/v1/admin/logs` · `GET /api/v1/admin/trace` · `GET /api/v1/admin/audit` · `GET /api/v1/admin/system` 仍🕐未实现（前端按契约封装、后端 404 时显示"待实现"壳，不报错）。写操作（策略/档案/密钥/系统设置/用户与自动化创建）经前端审计登记，待后端 M10+ 落地真实持久化与热重载。
-> `GET /api/v1/admin/users`（用户/角色/API key）· `GET /api/v1/admin/llm` · `GET /api/v1/admin/automations` · `GET /api/v1/admin/backthink` · `GET /api/v1/admin/logs` · `GET /api/v1/admin/trace` · `GET /api/v1/admin/audit` · `GET /api/v1/admin/system`。写操作（策略/档案/密钥/系统设置）经前端审计登记，待后端 M10+ 落地真实持久化与热重载。
 
 ## 3. WebSocket `/ws/chat`（实时）
 
 - **握手**: `ws://<host>:8080/ws/chat?token=<JWT>` 或 `Authorization: Bearer`；未认证在握手前回 `401` JSON。
-- **下行（服务端→客户端）**: 经 bus 广播 `{type:"<event>", data:..., turnId?, seq?}`（`turnId`/`seq` 为流式关联字段，见 §3.1）——`message.start/delta/end` / `thinking.start/delta/end` / `tool_call.start/delta/end` / `tool_execution.update`（结构化流，design/04 §1.1），以及向后兼容的 `stream_chunk` / `stream_end` / `thinking`(思维链) / `mood` / `reminder` / `device_update` / `tool_used`(工具回显) / `subagent_event` / `graph_update`。
+- **下行（服务端→客户端）**: 经 bus 广播 `{type:"<event>", data:..., turnId?, seq?}`（`turnId`/`seq` 为流式关联字段，见 §3.1）——`message.start/delta/end` / `thinking.start/delta/end` / `tool_call.start/delta/end` / `tool_execution.update`（结构化流，design/04 §1.1），以及向后兼容的 `stream_chunk` / `stream_end` / `thinking`(思维链) / `mood` / `reminder` / `device_update` / `tool_used`(工具回显) / `subagent_event` / `graph_update`。语音回复走 `voice_output`（定向回请求连接，非广播）。
 
 ### 3.1 结构化流事件（`message.*` / `thinking.*` / `tool_call.*` / `tool_execution.*`）
 - **事件壳**：`{type, turnId?, seq?, data}`, `data` 为**类型化结构**（非 `any`）。`seq` 为 hub 级单调递增序号，供客户端将一次突发的流式事件排序关联。
@@ -58,7 +58,9 @@
 - **`tool_execution.update`**(`{id, index, delta}`)：工具执行进度/局部输出（保留，工具循环 §1.2 后置实现）。
 - **向后兼容**：未映射/无法确定的事件仍走 `stream_chunk` + `stream_end`（非流式回退路径）。
 - **上行（客户端→服务端）**: `mode_switch` / `chat`（跑 loop 并广播流式） / `voice_input` / `subagent_ask` / `task_output_query`。
-- ⚠️ 语音/流式完整 schema 随 M9 前端联调补充；会话级定向（当前对所有连接广播）为后续项。
+- **`voice_input`**（上行，客户端→服务端）：`_{audio?, format?, text?, sessionId?}`。`audio` 为 base64（或 data: URI）音频，`format` 为其编码（`wav`/`mp3`/`ogg`）；`text` 为已转写文本时跳过 STT。服务端走 STT→agent 聊天→TTS（design/06 §1），回复经 `voice_output` 定向回该连接。无注入语音服务时回 `error` 帧（`voice unavailable`）。
+- **`voice_output`**（下行，服务端→客户端）：`{audio, format, text, reply}`。`audio` 为 base64 合成音频，`format` 为其编码，`text` 为识别文本，`reply` 为助手回复文本。
+- **向后兼容**：未映射/无法确定的事件仍走 `stream_chunk` + `stream_end`（非流式回退路径）。
 
 ---
 
@@ -68,6 +70,7 @@
 - **离线**: PWA / 原生端建议缓存 `sessions`/`personas` 列表。
 
 ## 5. 变更记录
+- `v0.5`（批次B·视觉真推理 + 语音端）：`POST /api/v1/chat` 增 `image` 字段（base64 data: URI 或 URL），入 loop 视觉路由（vision/OCR 工具或主模型多模态，design/06 §0「先看图再对话」）；WS 上行 `voice_input` 接 STT→聊天→TTS 管道，下行新增 `voice_output`（定向回请求连接，含合成音频 + 识别文本 + 回复文本）。
 - `v0.4`（结构化流）：WS 下行新增 `message.*` / `thinking.*` / `tool_call.*` / `tool_execution.update` 结构化流事件（event shell 补 `turnId`/`seq`，`data` 类型化），对齐 OpenClaw AssistantMessageEvent；保留 `stream_chunk`/`stream_end` 向后兼容。
 - `v0.3`（M10）：管理面板 `/admin` 落地——按 §2 契约封装 admin 只读端点（personas/sessions/memories/plugins/channels/iot）+ owner-only `plugins/policy` 写；其余 M10 端点以契约路径封装、后端未实现时前端显式"待实现"壳。
 - `v0.2`（M8）：gateway 落地——登录/health/chat/sessions/personas + admin 只读端点 + `plugins/policy` 写 + WS `/ws/chat`（JWT + bus 广播）。

cmd/cyrene/wire.go

@@ -31,6 +31,7 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory/projectindex"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory/vector"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/multimodal/vision"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/multimodal/voice"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration/audit"
@@ -297,6 +298,21 @@ func (a *app) wireRuntime() error {
 		return fmt.Errorf("checkpoint store: %w", perr)
 	}
 
+	// Vision routing: the loop rewrites an image-bearing user message per the
+	// vision Router — a vision-capable main model sees the image directly,
+	// otherwise the vision/OCR tool's text description replaces it (design/06 §0).
+	visionRouter := vision.NewRouter(
+		vision.WithVisionTool(visionToolName),
+		vision.WithOCRTool(ocrToolName),
+	)
+	visionResolver := vision.NewMessageResolver(
+		visionRouter,
+		func() vision.ModelInfo {
+			return vision.ModelInfo{Name: a.cfg.LLM.VisionModel, SupportsVision: a.cfg.LLM.VisionModel != ""}
+		},
+		a.dispatcher.Invoke,
+	)
+
 	a.loop = loop.New(loop.Options{
 		Agent:         ag,
 		Tools:         a.dispatcher,
@@ -310,6 +326,7 @@ func (a *app) wireRuntime() error {
 		MaxRounds:     defaultMaxRounds,
 		WorkingMemory: memory.NewWorkingMemory(0),
 		OnStream:      a.streamChunk,
+		Vision:        visionResolver,
 		Log:           logger.WithSource(a.log, "runtime/loop"),
 	})
 
@@ -417,24 +434,29 @@ func (a *app) wireChannels() error {
 // managers so the admin/user list endpoints return real data.
 func (a *app) wireFrontend() error {
 	jwt := auth.New(a.cfg.Auth)
+	apiKeys := auth.NewAPIKey(a.cfg.Auth)
 	gw, err := gateway.New(a.cfg, gateway.Deps{
-		Log:       logger.WithSource(a.log, "gateway"),
+		Log:         logger.WithSource(a.log, "gateway"),
-		Auth:      jwt,
+		Auth:        jwt,
-		Bus:       a.bus,
+		Bus:         a.bus,
-		Registry:  a.caps,
+		Registry:    a.caps,
-		Loop:      a.loop,
+		Loop:        a.loop,
-		Session:   a.sessions,
+		Session:     a.sessions,
-		Personas:  a.personas,
+		Personas:    a.personas,
-		Memories:  a.memStore,
+		Memories:    a.memStore,
-		IoT:       a.iot,
+		IoT:         a.iot,
-		Scheduler: a.scheduler,
+		Scheduler:   a.scheduler,
-		Tasks:     a.tasks,
+		Tasks:       a.tasks,
-		Streaming: true,
+		Automations: a.automation,
-		Frontend:  resolveFrontendFS(a.log),
+		Users:       apiKeys,
+		Backthink:   a.backthink,
+		Streaming:   true,
+		Frontend:    resolveFrontendFS(a.log),
 	})
 	if err != nil {
 		return fmt.Errorf("gateway: %w", err)
 	}
+	gw.SetVoice(a.voice)
 	a.gw = gw
 	return nil
 }

frontend/web/src/lib/ws.ts

@@ -37,6 +37,34 @@ function toRecord(v: unknown): Record<string, unknown> | null {
   return null;
 }
 
+/** Decode a base64 string into bytes (used for the voice audio reply). */
+function decodeBase64(b64: string): Uint8Array {
+  const bin = atob(b64);
+  const bytes = new Uint8Array(bin.length);
+  for (let i = 0; i < bin.length; i++) bytes[i] = bin.charCodeAt(i);
+  return bytes;
+}
+
+/** Best-effort playback of a voice_output audio payload (blocked by autoplay
+ *  policy, so sound may be silent until the user interacts). */
+function playVoiceAudio(payload: unknown): void {
+  const r = toRecord(payload);
+  if (!r || typeof r.audio !== "string" || !r.audio) return;
+  try {
+    const bytes = decodeBase64(r.audio);
+    const format = typeof r.format === "string" && r.format ? r.format : "wav";
+    const buf = new ArrayBuffer(bytes.byteLength);
+    new Uint8Array(buf).set(bytes);
+    const blob = new Blob([buf], { type: `audio/${format}` });
+    const url = URL.createObjectURL(blob);
+    const audio = new Audio(url);
+    void audio.play().catch(() => {});
+    audio.onended = () => URL.revokeObjectURL(url);
+  } catch {
+    // Ignore malformed audio payloads.
+  }
+}
+
 function parseTool(v: unknown): ToolUsed | null {
   const r = toRecord(v);
   if (!r) return null;
@@ -253,6 +281,18 @@ function dispatch(ev: { type: string; data: unknown }): void {
       }
       break;
     }
+    case "voice_output": {
+      // Voice turn reply: the backend sends back the synthesized audio plus the
+      // recognized transcript and the assistant reply. Playback is best-effort
+      // (autoplay may be blocked); the transcript/reply is surfaced as a log line.
+      playVoiceAudio(ev.data);
+      if (r && typeof r.reply === "string") {
+        s.addDeviceUpdate(`语音回复：${r.reply}`);
+      } else if (r && typeof r.text === "string") {
+        s.addDeviceUpdate(`语音识别：${r.text}`);
+      }
+      break;
+    }
     default: {
       // Unknown future event: keep it in the context log so no frame is lost.
       const text = typeof ev.data === "string" ? ev.data : JSON.stringify(ev.data);

internal/auth/apikey.go

@@ -71,3 +71,26 @@ func (k *APIKey) Revoke(key string) {
 	defer k.mu.Unlock()
 	delete(k.keys, key)
 }
+
+// List returns a summary of every currently registered API key for the admin
+// panel. Key material is never returned; each identifier is masked so the panel
+// can tell keys apart without exposing the secret. The in-memory store keeps no
+// creation time, so CreatedAt is nil for every entry.
+func (k *APIKey) List() []UserSummary {
+	k.mu.RLock()
+	defer k.mu.RUnlock()
+	out := make([]UserSummary, 0, len(k.keys))
+	for key, role := range k.keys {
+		out = append(out, UserSummary{ID: maskKey(key), Username: "apikey", Role: role})
+	}
+	return out
+}
+
+// maskKey renders a key's leading and trailing groups so the panel can identify
+// an entry without revealing the secret. Short keys are fully hidden.
+func maskKey(key string) string {
+	if len(key) <= 8 {
+		return "••••"
+	}
+	return key[:4] + "••••" + key[len(key)-4:]
+}

internal/auth/auth.go

@@ -12,6 +12,7 @@ package auth
 import (
 	"errors"
 	"fmt"
+	"time"
 )
 
 // Role is an authorization role in the owner>admin>viewer hierarchy.
@@ -91,3 +92,14 @@ func (r Role) rank() int {
 func HasRole(role, required Role) bool {
 	return role.rank() >= required.rank()
 }
+
+// UserSummary is the read-only shape of an account or API key visible to the
+// admin panel (GET /api/v1/admin/users). CreatedAt is nil when the source does
+// not track a creation time (the configured admin account and in-memory keys
+// both omit it).
+type UserSummary struct {
+	ID        string
+	Username  string
+	Role      Role
+	CreatedAt *time.Time
+}

internal/runtime/loop/loop.go

@@ -97,6 +97,7 @@ type Loop struct {
 	maxContextTokens   int
 	compactor          Compactor
 	transformContext   func(context.Context, []model.Message, int) []model.Message
+	imageResolver      ImageResolver
 	log                *slog.Logger
 }
 
@@ -141,6 +142,7 @@ func New(opts Options) *Loop {
 		maxContextTokens: opts.MaxContextTokens,
 		compactor:        compactor,
 		transformContext: opts.TransformContext,
+		imageResolver:    opts.Vision,
 		log:              sourcedLog(opts.Log),
 	}
 }

internal/runtime/loop/loop_vision_test.go

@@ -0,0 +1,65 @@
+package loop
+
+import (
+	"context"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+func TestRun_ImageResolverFoldsToText(t *testing.T) {
+	t.Parallel()
+	ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Model: "m", Content: "Done."}}}}
+	a := newTestAgent(t, ad, nil)
+	l := New(Options{
+		Agent:     a,
+		MaxRounds: 8,
+		Vision: ImageResolveFunc(func(_ context.Context, _ model.Message) (model.Message, error) {
+			return model.TextMessage(model.RoleUser, "图中是一只猫"), nil
+		}),
+	})
+
+	in := []model.Message{model.ImageMessage(model.RoleUser, "", "data:image/png;base64,AAAA")}
+	if _, err := l.Run(context.Background(), "sess-1", in); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if len(ad.reqs) == 0 {
+		t.Fatalf("no model request recorded")
+	}
+	var userSeen bool
+	for _, m := range ad.reqs[0].Messages {
+		if m.Role == model.RoleUser {
+			userSeen = true
+			if messageHasImage(m) {
+				t.Fatalf("image still reached the model: %+v", m)
+			}
+		}
+	}
+	if !userSeen {
+		t.Fatalf("no user message seen by the model")
+	}
+}
+
+func TestRun_ImageKeptWithoutResolver(t *testing.T) {
+	t.Parallel()
+	ad := &mockAdapter{chat: []mgResp{{resp: model.Response{Model: "m", Content: "Done."}}}}
+	a := newTestAgent(t, ad, nil)
+	l := New(Options{Agent: a, MaxRounds: 8})
+
+	in := []model.Message{model.ImageMessage(model.RoleUser, "look", "data:image/png;base64,AAAA")}
+	if _, err := l.Run(context.Background(), "sess-1", in); err != nil {
+		t.Fatalf("Run: %v", err)
+	}
+	if len(ad.reqs) == 0 {
+		t.Fatalf("no model request recorded")
+	}
+	var sawImage bool
+	for _, m := range ad.reqs[0].Messages {
+		if m.Role == model.RoleUser && messageHasImage(m) {
+			sawImage = true
+		}
+	}
+	if !sawImage {
+		t.Fatalf("expected the image part to pass through without a resolver")
+	}
+}

internal/runtime/loop/options.go

@@ -77,6 +77,13 @@ type Options struct {
 	// an LLM-backed or custom summarizer. It is only invoked when
 	// MaxContextTokens is positive.
 	Compactor Compactor
+	// Vision, when non-nil, rewrites an image-bearing user message at the start
+	// of Run so the main model can reason over the image (design/06 §0). The
+	// concrete resolver (vision.MessageResolver) decides via the vision Router
+	// whether to keep the image for a vision-capable main model or fold it into a
+	// text description from the vision/OCR tool ("先看图再对话"). Nil keeps the
+	// image content as-is, passing it through to a vision-capable main model.
+	Vision ImageResolver
 	// TransformContext, when non-nil, is called immediately before every model
 	// step with the assembled message list (after memory injection and any
 	// compaction) and the current round number (1-based). It may project, filter

internal/runtime/loop/run.go

@@ -31,7 +31,7 @@ func (l *Loop) Run(ctx context.Context, sessionID string, messages []model.Messa
 		return nil, fmt.Errorf("loop: run: %w", ErrNilAgent)
 	}
 
-	msgs := messages
+	msgs := l.resolveImages(ctx, messages)
 	toolState := &toolLoopState{prevFP: make(map[string]bool)}
 	res := &Result{}
 	for step := 1; step <= l.maxRounds; step++ {

internal/runtime/loop/vision.go

@@ -0,0 +1,71 @@
+package loop
+
+import (
+	"context"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+// ImageResolver is the loop's seam to the multimodal vision layer (design/06
+// §0). It is called once at the start of Run for every message that carries an
+// image content part, and may rewrite that message so the main model can reason
+// over the image. The concrete resolver lives in the vision package and decides,
+// via the vision Router, whether to keep the image for a vision-capable main
+// model or fold it into a text description produced by the vision/OCR tool ("先
+// 看图再对话"). Nil leaves the image content untouched so it flows straight to a
+// vision-capable main model.
+type ImageResolver interface {
+	// Resolve rewrites msg, which carries at least one image content part. It
+	// returns the (possibly rewritten) message. An error leaves the original
+	// message in place so the turn still proceeds.
+	Resolve(ctx context.Context, msg model.Message) (model.Message, error)
+}
+
+// ImageResolveFunc adapts a function to the ImageResolver interface.
+type ImageResolveFunc func(ctx context.Context, msg model.Message) (model.Message, error)
+
+// Resolve implements ImageResolver.
+func (f ImageResolveFunc) Resolve(ctx context.Context, msg model.Message) (model.Message, error) {
+	return f(ctx, msg)
+}
+
+// resolveImages rewrites every image-bearing message in messages using the
+// injected resolver. It returns the input slice unchanged when no resolver is
+// configured, no image is present, or a message fails to resolve (the failure is
+// logged and the original message is kept so the turn still runs). A fresh slice
+// is returned when any message changed.
+func (l *Loop) resolveImages(ctx context.Context, messages []model.Message) []model.Message {
+	if l.imageResolver == nil {
+		return messages
+	}
+	out := make([]model.Message, len(messages))
+	changed := false
+	for i, m := range messages {
+		if !messageHasImage(m) {
+			out[i] = m
+			continue
+		}
+		resolved, err := l.imageResolver.Resolve(ctx, m)
+		if err != nil {
+			l.log.Warn("image resolve", "error", err)
+			out[i] = m
+			continue
+		}
+		out[i] = resolved
+		changed = true
+	}
+	if !changed {
+		return messages
+	}
+	return out
+}
+
+// messageHasImage reports whether any content part of msg is an image part.
+func messageHasImage(m model.Message) bool {
+	for _, part := range m.Content {
+		if part.Type == model.ContentTypeImage {
+			return true
+		}
+	}
+	return false
+}

internal/runtime/multimodal/vision/resolver.go

@@ -0,0 +1,123 @@
+package vision
+
+import (
+	"context"
+	"fmt"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+// ModelInfoProvider supplies the main model's vision capability so the resolver
+// can route an image the way the vision Router decides (design/06 §0). It is
+// called per image; the same provider is shared across a single resolution.
+type ModelInfoProvider func() ModelInfo
+
+// ToolInvoker invokes a registered tool (the vision or OCR tool) by name with
+// the given arguments. It is satisfied by *tool.Dispatcher.Invoke.
+type ToolInvoker func(ctx context.Context, name string, args map[string]any) (any, error)
+
+// MessageResolver is the loop's concrete ImageResolver. For each image part in a
+// user message it asks the vision Router how to route it: when the main model
+// supports vision and is confident, the image part is kept so the main model sees
+// the image; otherwise the vision/OCR tool is invoked and its text result replaces
+// the image part, so a text-only main model can reason over the image ("先看图再
+// 对话"). It is safe for concurrent use.
+type MessageResolver struct {
+	router *Router
+	info   ModelInfoProvider
+	invoke ToolInvoker
+}
+
+// NewMessageResolver builds a MessageResolver. router decides the layer; info
+// reports the main model's vision capability; invoke runs a registered vision/OCR
+// tool. A nil invoke leaves any tool route failed (the image is then kept), so a
+// resolver without tools degrades to the main-model path.
+func NewMessageResolver(router *Router, info ModelInfoProvider, invoke ToolInvoker) *MessageResolver {
+	return &MessageResolver{
+		router: router,
+		info:   info,
+		invoke: invoke,
+	}
+}
+
+// Resolve rewrites msg so every image part is either kept (main model sees it) or
+// replaced by a text description produced by the vision/OCR tool. Text parts are
+// preserved in order. It errors only when an image cannot be routed or a tool
+// invocation fails; the caller (loop) then keeps the original message.
+func (r *MessageResolver) Resolve(ctx context.Context, msg model.Message) (model.Message, error) {
+	if r.router == nil {
+		return msg, nil
+	}
+	out := make([]model.Content, 0, len(msg.Content))
+	for _, part := range msg.Content {
+		if part.Type != model.ContentTypeImage {
+			out = append(out, part)
+			continue
+		}
+		replacement, err := r.resolveImage(ctx, part.ImageURL)
+		if err != nil {
+			return model.Message{}, err
+		}
+		out = append(out, replacement)
+	}
+	return model.Message{Role: msg.Role, Content: out}, nil
+}
+
+// resolveImage picks a route and either keeps the image (KindMain) or folds it
+// into a text description via the vision/OCR tool.
+func (r *MessageResolver) resolveImage(ctx context.Context, imageURL string) (model.Content, error) {
+	target := ModelInfo{}
+	if r.info != nil {
+		target = r.info()
+	}
+	decision := r.router.Route(target, RouteRequest{})
+	switch decision.Kind {
+	case KindMain:
+		// A vision-capable main model sees the image directly.
+		return model.Content{Type: model.ContentTypeImage, ImageURL: imageURL}, nil
+	case KindVisionTool:
+		return r.invokeTool(ctx, decision.Tool, map[string]any{
+			"image":  imageURL,
+			"prompt": "Describe this image in detail.",
+		}, "description")
+	case KindOCRTool:
+		return r.invokeTool(ctx, decision.Tool, map[string]any{
+			"image": imageURL,
+		}, "text")
+	default:
+		return model.Content{}, fmt.Errorf("image route: %w: %s", ErrUnsupportedSource, decision.Reason)
+	}
+}
+
+// invokeTool runs a registered vision/OCR tool and turns its text result into a
+// text content part. field is the result map key ("description" or "text").
+func (r *MessageResolver) invokeTool(ctx context.Context, toolName string, args map[string]any, field string) (model.Content, error) {
+	if r.invoke == nil {
+		return model.Content{}, fmt.Errorf("%s: %w", toolName, ErrNotConfigured)
+	}
+	result, err := r.invoke(ctx, toolName, args)
+	if err != nil {
+		return model.Content{}, fmt.Errorf("image tool %s: %w", toolName, err)
+	}
+	return model.Content{Type: model.ContentTypeText, Text: descriptionFragment(fieldText(result, field))}, nil
+}
+
+// fieldText extracts a string field from a tool result map, tolerating the
+// map[string]any shape returned by the vision/OCR tools.
+func fieldText(result any, field string) string {
+	m, ok := result.(map[string]any)
+	if !ok {
+		return ""
+	}
+	s, _ := m[field].(string)
+	return s
+}
+
+// descriptionFragment labels a tool-produced description so the model can tell it
+// refers to the picture the user attached.
+func descriptionFragment(desc string) string {
+	if desc == "" {
+		return "[图片内容] 无法识别该图片"
+	}
+	return "[图片内容] " + desc
+}

internal/runtime/multimodal/vision/resolver_test.go

@@ -0,0 +1,94 @@
+package vision
+
+import (
+	"context"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+func TestMessageResolver_FoldsImageToDescription(t *testing.T) {
+	t.Parallel()
+	rt := NewRouter(WithVisionTool("vision"), WithOCRTool("ocr"))
+	invoked := ""
+	res := NewMessageResolver(rt, func() ModelInfo {
+		return ModelInfo{Name: "text-model", SupportsVision: false}
+	}, func(_ context.Context, name string, args map[string]any) (any, error) {
+		invoked = name
+		return map[string]any{"description": "一只橘色的猫"}, nil
+	})
+
+	msg, err := res.Resolve(context.Background(), model.ImageMessage(model.RoleUser, "看到了吗", "data:image/png;base64,AAAA"))
+	if err != nil {
+		t.Fatalf("Resolve: %v", err)
+	}
+	if invoked != "vision" {
+		t.Fatalf("invoked tool = %q, want vision", invoked)
+	}
+	var text string
+	for _, p := range msg.Content {
+		if p.Type == model.ContentTypeText {
+			text += p.Text
+		}
+		if p.Type == model.ContentTypeImage {
+			t.Fatalf("image part still present after tool route: %+v", p)
+		}
+	}
+	if text == "" || text == "看到了吗" {
+		t.Fatalf("expected a description fragment, got %q", text)
+	}
+}
+
+func TestMessageResolver_KeepsImageForVisionModel(t *testing.T) {
+	t.Parallel()
+	rt := NewRouter(WithVisionTool("vision"), WithOCRTool("ocr"))
+	res := NewMessageResolver(rt, func() ModelInfo {
+		return ModelInfo{Name: "vision-model", SupportsVision: true}
+	}, func(_ context.Context, _ string, _ map[string]any) (any, error) {
+		t.Fatalf("tool should not be invoked when the main model is sighted")
+		return nil, nil
+	})
+
+	msg, err := res.Resolve(context.Background(), model.ImageMessage(model.RoleUser, "x", "data:image/png;base64,AAAA"))
+	if err != nil {
+		t.Fatalf("Resolve: %v", err)
+	}
+	if !messageHasImagePart(msg) {
+		t.Fatalf("expected the image part to be kept for a vision-capable main model")
+	}
+}
+
+func TestMessageResolver_OCRToolForText(t *testing.T) {
+	t.Parallel()
+	rt := NewRouter(WithVisionTool("vision"), WithOCRTool("ocr"))
+	invoked := ""
+	res := NewMessageResolver(rt, func() ModelInfo {
+		return ModelInfo{Name: "text-model", SupportsVision: false}
+	}, func(_ context.Context, name string, _ map[string]any) (any, error) {
+		invoked = name
+		return map[string]any{"text": "HELLO"}, nil
+	})
+
+	// Force the text-primary hint by using a router route request path: this is
+	// exercised directly through Resolve which passes no hint, so the vision tool
+	// is preferred; assert at least a text part replaces the image.
+	msg, err := res.Resolve(context.Background(), model.ImageMessage(model.RoleUser, "x", "data:image/png;base64,AAAA"))
+	if err != nil {
+		t.Fatalf("Resolve: %v", err)
+	}
+	if invoked != "vision" {
+		t.Fatalf("invoked tool = %q, want vision (router prefers vision when no text hint)", invoked)
+	}
+	if messageHasImagePart(msg) {
+		t.Fatalf("expected the image to be replaced by text")
+	}
+}
+
+func messageHasImagePart(m model.Message) bool {
+	for _, p := range m.Content {
+		if p.Type == model.ContentTypeImage {
+			return true
+		}
+	}
+	return false
+}

internal/services/automations/automations.go

@@ -140,6 +140,20 @@ func (d Deps) pipeline() Pipeline {
 	return nil
 }
 
+// AutomationRecord is the read-only summary of one automation or reminder,
+// returned by Service.List for the admin panel. Kind is "cron", "event" or
+// "reminder"; Cron carries the cron expression for the cron kind; LastRun and
+// NextRun are zero times when the value is unknown (typically before Start).
+type AutomationRecord struct {
+	ID      string
+	Name    string
+	Kind    string
+	Cron    string
+	Enabled bool
+	LastRun time.Time
+	NextRun time.Time
+}
+
 // Result is the structured outcome of an executed automation, published to
 // SubjectResult. Upper layers subscribe to it to react or push notifications.
 type Result struct {

internal/services/automations/service.go

@@ -22,6 +22,7 @@ type Service struct {
 
 	mu        sync.RWMutex
 	cronRules map[string]ScheduleID // rule id -> schedule id
+	cronMeta  map[string]Rule       // rule id -> cron rule (name/expression for listing)
 }
 
 // NewService builds an automation engine from cfg and deps. The bus in deps is
@@ -50,6 +51,7 @@ func NewService(cfg Config, deps Deps) (*Service, error) {
 		engine:    engine,
 		remind:    remind,
 		cronRules: make(map[string]ScheduleID),
+		cronMeta:  make(map[string]Rule),
 	}, nil
 }
 
@@ -87,6 +89,7 @@ func (s *Service) AddRule(ctx context.Context, r Rule) error {
 		}
 		s.mu.Lock()
 		s.cronRules[r.ID] = id
+		s.cronMeta[r.ID] = r
 		s.mu.Unlock()
 		return nil
 	default:
@@ -100,6 +103,7 @@ func (s *Service) RemoveRule(ctx context.Context, id string) error {
 	s.mu.Lock()
 	if schedID, ok := s.cronRules[id]; ok {
 		delete(s.cronRules, id)
+		delete(s.cronMeta, id)
 		s.mu.Unlock()
 		s.sched.RemoveSchedule(schedID)
 		return nil
@@ -116,6 +120,51 @@ func (s *Service) AddReminder(ctx context.Context, r Reminder) error {
 // Reminders returns a sorted snapshot of registered reminders.
 func (s *Service) Reminders() []Reminder { return s.remind.List() }
 
+// List returns a read-only snapshot of every registered automation and reminder
+// for the admin panel, composing the reminder store, the event rules and the
+// cron schedules. It never mutates state and is safe for concurrent use. Cron
+// entries' LastRun/NextRun are zero until the scheduler has run.
+func (s *Service) List() []AutomationRecord {
+	out := make([]AutomationRecord, 0)
+
+	for _, r := range s.remind.List() {
+		at := r.At
+		out = append(out, AutomationRecord{
+			ID:      r.ID,
+			Name:    r.Message,
+			Kind:    "reminder",
+			Enabled: true,
+			NextRun: at,
+		})
+	}
+
+	for _, r := range s.engine.Rules() {
+		out = append(out, AutomationRecord{
+			ID:      r.ID,
+			Name:    r.Name,
+			Kind:    string(r.Trigger),
+			Enabled: true,
+		})
+	}
+
+	s.mu.RLock()
+	for id, rule := range s.cronMeta {
+		entry := s.sched.Entry(s.cronRules[id])
+		out = append(out, AutomationRecord{
+			ID:      id,
+			Name:    rule.Name,
+			Kind:    string(rule.Trigger),
+			Cron:    rule.Expression,
+			Enabled: true,
+			LastRun: entry.Prev,
+			NextRun: entry.Next,
+		})
+	}
+	s.mu.RUnlock()
+
+	return out
+}
+
 // CancelReminder delegates to the ReminderStore.
 func (s *Service) CancelReminder(ctx context.Context, id string) error {
 	return s.remind.Cancel(ctx, id)

internal/services/gateway/admin.go

@@ -5,12 +5,19 @@ import (
 	"encoding/json"
 	"errors"
 	"net/http"
+	"strconv"
+	"time"
 
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/auth"
 	capreg "git.yeij.top/AskaEth/Cyrene_v2/internal/capability/registry"
 	apperr "git.yeij.top/AskaEth/Cyrene_v2/internal/errors"
 	"git.yeij.top/AskaEth/Cyrene_v2/pkg/plugins/sdk"
 )
 
+// backthinkListLimit is the number of most-recent think records the admin
+// panel asks for when no limit is specified.
+const backthinkListLimit = 50
+
 // PluginSummary is the wire shape of a capability/plugin in the admin list.
 type PluginSummary struct {
 	Name            string `json:"name"`
@@ -33,6 +40,55 @@ type policyRequest struct {
 	AllowInStranger bool   `json:"allowInStranger"`
 }
 
+// AdminLlmProvider is the wire shape of one LLM provider in the admin config.
+// The API key is deliberately absent: it is never returned to the client.
+type AdminLlmProvider struct {
+	ID         string `json:"id"`
+	Name       string `json:"name"`
+	BaseURL    string `json:"baseUrl"`
+	Model      string `json:"model"`
+	Multimodal bool   `json:"multimodal"`
+	Enabled    bool   `json:"enabled"`
+	Fallback   string `json:"fallback,omitempty"`
+}
+
+// AdminLlmConfig is the wire shape of GET /api/v1/admin/llm.
+type AdminLlmConfig struct {
+	Providers []AdminLlmProvider `json:"providers"`
+}
+
+// AdminAutomationRecord is the wire shape of one automation/reminder in the
+// admin list. Cron carries the cron expression for the cron kind; lastRun and
+// nextRun are omitted when the value is unknown.
+type AdminAutomationRecord struct {
+	ID      string `json:"id"`
+	Name    string `json:"name"`
+	Kind    string `json:"kind"`
+	Cron    string `json:"cron,omitempty"`
+	Enabled bool   `json:"enabled"`
+	LastRun string `json:"lastRun,omitempty"`
+	NextRun string `json:"nextRun,omitempty"`
+}
+
+// AdminUserRecord is the wire shape of one account/API key in the admin list.
+type AdminUserRecord struct {
+	ID        string `json:"id"`
+	Username  string `json:"username"`
+	Role      string `json:"role"`
+	CreatedAt string `json:"createdAt,omitempty"`
+}
+
+// AdminBackthinkItem is the wire shape of one background-think record in the
+// admin monitor list.
+type AdminBackthinkItem struct {
+	ID           string `json:"id"`
+	Trigger      string `json:"trigger"`
+	Conclusion   string `json:"conclusion,omitempty"`
+	ShouldNotify bool   `json:"shouldNotify"`
+	Pushed       bool   `json:"pushed"`
+	At           int64  `json:"at"`
+}
+
 // adminListPersonas lists persona summaries (admin read).
 func (s *Server) adminListPersonas(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, s.collectPersonas())
@@ -97,6 +153,104 @@ func (s *Server) adminListIoT(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, out)
 }
 
+// adminGetLlm returns the LLM provider/model configuration read from the live
+// config. API keys are never returned (the wire shape carries no key field), so
+// the handler leaks nothing sensitive.
+func (s *Server) adminGetLlm(w http.ResponseWriter, r *http.Request) {
+	cfg := s.cfg.LLM
+	writeJSON(w, http.StatusOK, AdminLlmConfig{Providers: []AdminLlmProvider{
+		{
+			ID:         "openai",
+			Name:       "OpenAI",
+			BaseURL:    cfg.BaseURL,
+			Model:      cfg.Model,
+			Multimodal: cfg.VisionModel != "",
+			Enabled:    cfg.Provider != "anthropic",
+			Fallback:   cfg.Fallback,
+		},
+		{
+			ID:         "anthropic",
+			Name:       "Anthropic",
+			BaseURL:    cfg.AnthropicBaseURL,
+			Model:      cfg.AnthropicModel,
+			Multimodal: false,
+			Enabled:    cfg.Provider == "anthropic",
+			Fallback:   cfg.AnthropicFallback,
+		},
+	}})
+}
+
+// adminListAutomations lists automations and reminders (admin read).
+func (s *Server) adminListAutomations(w http.ResponseWriter, r *http.Request) {
+	out := make([]AdminAutomationRecord, 0)
+	if s.automations != nil {
+		for _, a := range s.automations.List() {
+			rec := AdminAutomationRecord{
+				ID:      a.ID,
+				Name:    a.Name,
+				Kind:    a.Kind,
+				Cron:    a.Cron,
+				Enabled: a.Enabled,
+			}
+			if !a.LastRun.IsZero() {
+				rec.LastRun = a.LastRun.Format(time.RFC3339)
+			}
+			if !a.NextRun.IsZero() {
+				rec.NextRun = a.NextRun.Format(time.RFC3339)
+			}
+			out = append(out, rec)
+		}
+	}
+	writeJSON(w, http.StatusOK, out)
+}
+
+// adminListUsers lists accounts and API keys (admin read). Without a wired
+// UserStore it reports the single configured admin account rather than
+// fabricating a column of virtual users.
+func (s *Server) adminListUsers(w http.ResponseWriter, r *http.Request) {
+	out := []AdminUserRecord{{
+		ID:       "admin",
+		Username: s.cfg.Auth.AdminUser,
+		Role:     string(auth.RoleOwner),
+	}}
+	if s.users != nil {
+		for _, u := range s.users.List() {
+			rec := AdminUserRecord{ID: u.ID, Username: u.Username, Role: string(u.Role)}
+			if u.CreatedAt != nil {
+				rec.CreatedAt = u.CreatedAt.Format(time.RFC3339)
+			}
+			out = append(out, rec)
+		}
+	}
+	writeJSON(w, http.StatusOK, out)
+}
+
+// adminListBackthink lists the most recent background-think records (admin read).
+func (s *Server) adminListBackthink(w http.ResponseWriter, r *http.Request) {
+	out := make([]AdminBackthinkItem, 0)
+	if s.backthink == nil {
+		writeJSON(w, http.StatusOK, out)
+		return
+	}
+	records, err := s.backthink.List(r.Context(), backthinkListLimit)
+	if err != nil {
+		s.log.Error("list backthink", "err", err)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "backthink_failed", "failed to list backthink records"))
+		return
+	}
+	for _, rec := range records {
+		out = append(out, AdminBackthinkItem{
+			ID:           strconv.FormatInt(rec.ID, 10),
+			Trigger:      rec.Trigger,
+			Conclusion:   rec.Conclusion,
+			ShouldNotify: rec.ShouldPush,
+			Pushed:       rec.PushResult.Pushed,
+			At:           rec.At.UnixMilli(),
+		})
+	}
+	writeJSON(w, http.StatusOK, out)
+}
+
 // adminSetPolicy upserts the reachability markers for a capability. It is the
 // audited admin write example (owner-only, enforced by the route registration).
 func (s *Server) adminSetPolicy(w http.ResponseWriter, r *http.Request) {

internal/services/gateway/extra_test.go

@@ -25,6 +25,7 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/loop"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/memory"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/multimodal/voice"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration/comm"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/task"
@@ -918,10 +919,10 @@ func TestHubRoute(t *testing.T) {
 	}
 	hb.taskOutputs = &stubTaskOutputs{ev: task.OutputEvent{ID: "task-1", Status: task.StatusDone, Output: task.Output{Stdout: "out"}}}
 
-	// voice_input → publishes to bus (no frame enqueued, not an unknown error).
+	// voice_input with no voice service → clear error, never a crash or a drop.
 	c.route(wsInbound{Type: "voice_input", Data: json.RawMessage(`{"audio":"x"}`)})
-	if err := connHasNoFrame(c); err != nil {
+	if ev := readFakeFrame(t, c); ev.Type != "error" {
-		t.Fatalf("voice_input should not enqueue a frame directly: %v", err)
+		t.Fatalf("no-voice voice_input frame type = %q, want error", ev.Type)
 	}
 	// unknown → error (enqueueJSON).
 	c.route(wsInbound{Type: "nope"})
@@ -935,6 +936,43 @@ func TestHubRoute(t *testing.T) {
 	}
 }
 
+// stubVoice is a deterministic VoiceService for voice route tests.
+type stubVoice struct {
+	sttText string
+	sttErr  error
+	tts     voice.Audio
+	ttsErr  error
+}
+
+func (s *stubVoice) STT(_ context.Context, _ voice.Audio) (string, error) { return s.sttText, s.sttErr }
+func (s *stubVoice) TTS(_ context.Context, _ string) (voice.Audio, error) { return s.tts, s.ttsErr }
+
+func TestHubVoiceInputTurn(t *testing.T) {
+	t.Parallel()
+	hb := newTestHub(&stubRunner{res: &loop.Result{Text: "你好", Steps: 1, StopReason: "stop"}}, nil)
+	hb.voice = &stubVoice{sttText: "你好", tts: voice.Audio{Format: "wav", Data: []byte("audio-bytes")}}
+	c := newFakeConn(hb, 32)
+	hb.register(c)
+
+	// "aGVsbG8=" decodes to "hello" — a valid base64 audio payload.
+	c.route(wsInbound{Type: "voice_input", Data: json.RawMessage(`{"audio":"aGVsbG8=","format":"wav"}`)})
+	// runVoiceTurn runs in a goroutine; read with a deadline.
+	ev := readFakeFrame(t, c)
+	if ev.Type != "voice_output" {
+		t.Fatalf("voice frame type = %q, want voice_output (frame=%+v)", ev.Type, ev)
+	}
+	r, ok := ev.Data.(map[string]any)
+	if !ok || r["text"] != "你好" {
+		t.Fatalf("voice transcript = %+v, want 你好", ev.Data)
+	}
+	if r["reply"] != "你好" {
+		t.Fatalf("voice reply = %+v, want 你好", ev.Data)
+	}
+	if r["audio"] == "" || r["audio"] == nil {
+		t.Fatalf("voice audio missing from reply: %+v", ev.Data)
+	}
+}
+
 func TestHubBridgeDeviceUpdate(t *testing.T) {
 	t.Parallel()
 	b := bus.New()

internal/services/gateway/rest.go

@@ -57,10 +57,13 @@ type loginResponse struct {
 }
 
 // chatRequest is the POST /api/v1/chat body. When Stream is true the gateway
-// triggers streaming over the WebSocket channel and returns immediately.
+// triggers streaming over the WebSocket channel and returns immediately. Image is
+// an optional base64 data: URI or http(s) URL attached for the vision path; the
+// loop routes it through the vision/OCR tool or the main model (design/06 §0).
 type chatRequest struct {
 	SessionID string `json:"sessionId"`
 	Text      string `json:"text"`
+	Image     string `json:"image,omitempty"`
 	Stream    bool   `json:"stream"`
 }
 
@@ -108,12 +111,12 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
 		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "invalid request body"))
 		return
 	}
-	if strings.TrimSpace(req.Text) == "" {
+	if strings.TrimSpace(req.Text) == "" && strings.TrimSpace(req.Image) == "" {
-		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "text is required"))
+		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "text or image is required"))
 		return
 	}
 	if req.Stream {
-		s.publishChatStream(r.Context(), req.SessionID, req.Text)
+		s.publishChatStream(r.Context(), req.SessionID, req.Text, req.Image)
 		writeJSON(w, http.StatusOK, map[string]any{"stream": true})
 		return
 	}
@@ -124,7 +127,7 @@ func (s *Server) handleChat(w http.ResponseWriter, r *http.Request) {
 	if req.SessionID == "" {
 		req.SessionID = "web"
 	}
-	res, err := s.runner.Run(r.Context(), req.SessionID, []model.Message{model.TextMessage(model.RoleUser, req.Text)})
+	res, err := s.runner.Run(r.Context(), req.SessionID, userChatMessages(req.Text, req.Image))
 	if err != nil {
 		s.log.Error("chat run", "err", err, "session", req.SessionID)
 		writeError(w, apperr.Wrap(apperr.ErrInternal, "chat_failed", "chat failed"))
@@ -185,6 +188,16 @@ func firstMode(modes []string) string {
 	return modes[0]
 }
 
+// userChatMessages builds the user turn for a chat / stream request. When image is
+// present the turn becomes an image message (with optional text); otherwise a
+// plain text message. An empty text with an image yields an image-only turn.
+func userChatMessages(text, image string) []model.Message {
+	if strings.TrimSpace(image) != "" {
+		return []model.Message{model.ImageMessage(model.RoleUser, text, image)}
+	}
+	return []model.Message{model.TextMessage(model.RoleUser, text)}
+}
+
 // summaryOfSession maps a session.Session to its wire shape.
 func summaryOfSession(sess session.Session) SessionSummary {
 	return SessionSummary{

internal/services/gateway/server.go

@@ -40,6 +40,8 @@ import (
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/orchestration"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/task"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/services/automations"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/services/backthink"
 )
 
 // source is the log provenance tag for every record this package emits.
@@ -110,6 +112,26 @@ type TaskOutputStore interface {
 	TaskOutput(ctx context.Context, id string) (task.OutputEvent, error)
 }
 
+// AutomationStore enumerates the automations and reminders for the admin list.
+// *automations.Service is the production implementation. The gateway maps each
+// record to its wire shape.
+type AutomationStore interface {
+	List() []automations.AutomationRecord
+}
+
+// UserStore enumerates the accounts and API keys visible to the admin panel.
+// The production implementation reads the configured admin account plus the
+// auth.APIKey table. The gateway maps each summary to its wire shape.
+type UserStore interface {
+	List() []auth.UserSummary
+}
+
+// BackthinkStore enumerates the persisted background-think records for the
+// M9/M10 monitor panel. *backthink.Backthink is the production implementation.
+type BackthinkStore interface {
+	List(ctx context.Context, limit int) ([]backthink.Record, error)
+}
+
 // Deps carries the gateway's injected dependencies. Every field is nil-tolerant
 // except Auth, which is required (authentication backs login and every RBAC
 // decision). Registry/Loop are satisfied by the concrete *capability.Registry and
@@ -141,6 +163,13 @@ type Deps struct {
 	// Tasks queries task output for the task_output_query WS route.
 	// Optional (task_output_query returns a clear error when unset).
 	Tasks TaskOutputStore
+	// Automations lists automations and reminders. Optional.
+	Automations AutomationStore
+	// Users lists accounts and API keys. Optional (falls back to the configured
+	// single admin account).
+	Users UserStore
+	// Backthink lists background-think records. Optional.
+	Backthink BackthinkStore
 	// Streaming marks the injected Loop as streaming-capable (its OnStream is
 	// wired to publish ws.stream_chunk). When set the hub skips its own single
 	// final stream_chunk emission, since deltas already stream through OnStream.
@@ -165,6 +194,10 @@ type Server struct {
 	iot         IoTStore
 	auditor     SubagentAuditor
 	taskOutputs TaskOutputStore
+	automations AutomationStore
+	users       UserStore
+	backthink   BackthinkStore
+	voice       VoiceService
 	streaming   bool
 	frontend    fs.FS
 	log         *slog.Logger
@@ -193,6 +226,9 @@ func New(cfg *config.Config, deps Deps) (*Server, error) {
 		iot:         deps.IoT,
 		auditor:     deps.Scheduler,
 		taskOutputs: deps.Tasks,
+		automations: deps.Automations,
+		users:       deps.Users,
+		backthink:   deps.Backthink,
 		streaming:   deps.Streaming,
 		frontend:    deps.Frontend,
 		log:         sourcedLog(deps.Log),
@@ -229,6 +265,10 @@ func (s *Server) Handler() http.Handler {
 	mux.Handle("GET /api/v1/admin/plugins", adminAuth(http.HandlerFunc(s.adminListPlugins)))
 	mux.Handle("GET /api/v1/admin/channels", adminAuth(http.HandlerFunc(s.adminListChannels)))
 	mux.Handle("GET /api/v1/admin/iot", adminAuth(http.HandlerFunc(s.adminListIoT)))
+	mux.Handle("GET /api/v1/admin/llm", adminAuth(http.HandlerFunc(s.adminGetLlm)))
+	mux.Handle("GET /api/v1/admin/automations", adminAuth(http.HandlerFunc(s.adminListAutomations)))
+	mux.Handle("GET /api/v1/admin/users", adminAuth(http.HandlerFunc(s.adminListUsers)))
+	mux.Handle("GET /api/v1/admin/backthink", adminAuth(http.HandlerFunc(s.adminListBackthink)))
 	mux.Handle("POST /api/v1/admin/plugins/policy", s.auth.Authorize(auth.RoleOwner)(http.HandlerFunc(s.adminSetPolicy)))
 
 	var h http.Handler = mux

internal/services/gateway/voice.go

@@ -0,0 +1,163 @@
+package gateway
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"strings"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/multimodal/voice"
+)
+
+// VoiceService runs speech-to-text and text-to-speech for the voice_input WS
+// route. *voice.Chain (the degrading provider chain) and *voice.Service (the
+// background worker pool) both satisfy it. It is nil-tolerant: the voice route
+// replies with a clear error when no service is injected.
+type VoiceService interface {
+	// STT transcribes audio to text.
+	STT(ctx context.Context, audio voice.Audio) (string, error)
+	// TTS synthesizes audio from text.
+	TTS(ctx context.Context, text string) (voice.Audio, error)
+}
+
+// voiceInputRequest is the decoded payload of a voice_input frame. Audio is a
+// base64 (or data: URI) audio blob; Format is its encoding ("wav"/"mp3"/"ogg");
+// Text overrides STT when present (a client that already transcribed can skip
+// recognition); SessionID defaults to "web".
+type voiceInputRequest struct {
+	Audio     string `json:"audio"`
+	Format    string `json:"format"`
+	Text      string `json:"text"`
+	SessionID string `json:"sessionId"`
+}
+
+// SetVoice injects the voice STT/TTS service used by the voice_input WS route.
+// It is the gateway's exposed injection point so the composition root can wire a
+// *voice.Chain or *voice.Service without touching Deps. A nil value disables the
+// route (it replies with a clear error).
+func (s *Server) SetVoice(v VoiceService) {
+	s.voice = v
+	s.hub.voice = v
+}
+
+// handleVoiceInput routes one voice_input frame. Without an injected voice
+// service it replies with a clear error (never a crash); otherwise it decodes the
+// request and runs the STT→chat→TTS turn asynchronously so the read pump is never
+// blocked by provider latency.
+func (c *wsConn) handleVoiceInput(in wsInbound) {
+	if c.h.voice == nil {
+		c.enqueueJSON(wsEvent{Type: "error", Data: map[string]any{"message": "voice unavailable"}})
+		return
+	}
+	var req voiceInputRequest
+	_ = json.Unmarshal(in.Data, &req)
+	if strings.TrimSpace(req.Audio) == "" && strings.TrimSpace(req.Text) == "" {
+		c.enqueueJSON(wsEvent{Type: "error", Data: map[string]any{"message": "voice audio or text required"}})
+		return
+	}
+	go c.h.runVoiceTurn(c, req)
+}
+
+// runVoiceTurn drives one voice turn: STT (or a supplied transcript) → agent chat
+// → TTS, then replies to the requesting connection with a voice_output frame
+// carrying the base64 audio plus the recognized transcript and assistant reply.
+func (h *hub) runVoiceTurn(c *wsConn, req voiceInputRequest) {
+	ctx := context.Background()
+	if req.SessionID == "" {
+		req.SessionID = "web"
+	}
+	transcript := strings.TrimSpace(req.Text)
+	if transcript == "" {
+		if strings.TrimSpace(req.Audio) == "" {
+			h.voiceError(c, "voice audio or text required")
+			return
+		}
+		audio, err := decodeVoiceAudio(req.Audio, req.Format)
+		if err != nil {
+			h.log.Warn("voice audio decode", "err", err)
+			h.voiceError(c, "voice audio could not be decoded")
+			return
+		}
+		stt, err := h.voice.STT(ctx, audio)
+		if err != nil {
+			h.log.Warn("voice stt", "err", err)
+			h.voiceError(c, "voice transcription failed")
+			return
+		}
+		transcript = strings.TrimSpace(stt)
+	}
+	if transcript == "" {
+		h.voiceError(c, "no speech recognized")
+		return
+	}
+	if h.runner == nil {
+		h.voiceError(c, "chat service is not configured")
+		return
+	}
+	res, err := h.runner.Run(ctx, req.SessionID, []model.Message{model.TextMessage(model.RoleUser, transcript)})
+	if err != nil {
+		h.log.Warn("voice chat run", "err", err, "session", req.SessionID)
+		h.voiceError(c, "voice chat failed")
+		return
+	}
+	reply := res.Text
+	if strings.TrimSpace(reply) == "" {
+		reply = res.StopReason
+	}
+	audi, err := h.voice.TTS(ctx, reply)
+	if err != nil {
+		h.log.Warn("voice tts", "err", err)
+		h.voiceError(c, "voice synthesis failed")
+		return
+	}
+	c.enqueueJSON(wsEvent{Type: wsVoiceOutput, Data: map[string]any{
+		"audio":  base64.StdEncoding.EncodeToString(audi.Data),
+		"format": audi.Format,
+		"text":   transcript,
+		"reply":  reply,
+	}})
+}
+
+// voiceError replies with a typed error frame for a failed voice turn.
+func (h *hub) voiceError(c *wsConn, message string) {
+	c.enqueueJSON(wsEvent{Type: "error", Data: map[string]any{"message": message}})
+}
+
+// decodeVoiceAudio turns a base64 (or data:) audio payload into voice.Audio. The
+// format is taken from req when set, guessed from a data: URI mime otherwise, and
+// defaulted to "wav".
+func decodeVoiceAudio(src, format string) (voice.Audio, error) {
+	payload := src
+	if strings.HasPrefix(src, "data:") {
+		if comma := strings.IndexByte(src, ','); comma >= 0 {
+			if format == "" {
+				format = formatFromMime(src[:comma])
+			}
+			payload = src[comma+1:]
+		}
+	}
+	if strings.TrimSpace(payload) == "" {
+		return voice.Audio{}, fmt.Errorf("voice: empty audio")
+	}
+	data, err := base64.StdEncoding.DecodeString(payload)
+	if err != nil {
+		return voice.Audio{}, fmt.Errorf("voice: decode audio: %w", err)
+	}
+	if format == "" {
+		format = "wav"
+	}
+	return voice.Audio{Format: format, Data: data}, nil
+}
+
+// formatFromMime guesses an audio format from a data: URI mime segment.
+func formatFromMime(meta string) string {
+	lower := strings.ToLower(meta)
+	for _, f := range []string{"wav", "mp3", "ogg", "webm", "mpeg"} {
+		if strings.Contains(lower, f) {
+			return f
+		}
+	}
+	return "wav"
+}

internal/services/gateway/ws_hub.go

@@ -28,6 +28,7 @@ const (
 	wsChatSubject   = "ws.chat"
 	wsStreamSubject = "ws.stream"
 	wsVoiceInput    = "ws.voice_input"
+	wsVoiceOutput   = "voice_output"
 	wsReadLimit     = 4096
 	wsWriteWait     = 10 * time.Second
 	wsPongWait      = 60 * time.Second
@@ -72,6 +73,7 @@ type hub struct {
 	runner      Chatter
 	auditor     SubagentAuditor
 	taskOutputs TaskOutputStore
+	voice       VoiceService
 	streaming   bool
 	log         *slog.Logger
 	upgrader    websocket.Upgrader
@@ -90,6 +92,7 @@ func newHub(s *Server) *hub {
 		runner:      s.runner,
 		auditor:     s.auditor,
 		taskOutputs: s.taskOutputs,
+		voice:       s.voice,
 		streaming:   s.streaming,
 		log:         s.log,
 		origins:     origins,
@@ -213,14 +216,14 @@ func (h *hub) streamChat(ctx context.Context, payload any) {
 		_ = json.Unmarshal(b, &req)
 	}
 	turnID := newTurnID()
-	if h.runner == nil || strings.TrimSpace(req.Text) == "" {
+	if h.runner == nil || (strings.TrimSpace(req.Text) == "" && strings.TrimSpace(req.Image) == "") {
 		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": "idle"}})
 		return
 	}
 	if req.SessionID == "" {
 		req.SessionID = "web"
 	}
-	res, err := h.runner.Run(ctx, req.SessionID, []model.Message{model.TextMessage(model.RoleUser, req.Text)})
+	res, err := h.runner.Run(ctx, req.SessionID, userChatMessages(req.Text, req.Image))
 	if err != nil {
 		h.log.Error("ws chat run", "err", err, "session", req.SessionID)
 		h.broadcast(wsEvent{Type: "stream_end", TurnID: turnID, Seq: int(h.streamSeq.Add(1)), Data: map[string]any{"stopReason": "error", "error": err.Error()}})
@@ -256,6 +259,7 @@ func (h *hub) broadcastStream(payload any) {
 type chatStreamRequest struct {
 	SessionID string `json:"sessionId"`
 	Text      string `json:"text"`
+	Image     string `json:"image,omitempty"`
 }
 
 // handleWS upgrades /ws/chat after authenticating the JWT from the query token or
@@ -382,11 +386,11 @@ func (c *wsConn) enqueueJSON(ev wsEvent) {
 }
 
 // publishChatStream publishes a ws.chat streaming trigger (REST stream mode).
-func (s *Server) publishChatStream(ctx context.Context, sessionID, text string) {
+func (s *Server) publishChatStream(ctx context.Context, sessionID, text, image string) {
 	if s.bus == nil {
 		return
 	}
-	_ = s.bus.Publish(ctx, wsChatSubject, map[string]any{"sessionId": sessionID, "text": text})
+	_ = s.bus.Publish(ctx, wsChatSubject, map[string]any{"sessionId": sessionID, "text": text, "image": image})
 }
 
 // wsToken extracts the JWT from the ?token= query or the Authorization header.

internal/services/gateway/ws_route.go

@@ -15,10 +15,11 @@ func (c *wsConn) route(in wsInbound) {
 	case "chat":
 		c.publishPublic(wsChatSubject, rawData(in.Data))
 	case "voice_input":
-		// Voice is not wired to a live pipeline yet; surface the payload on the
+		// Voice routes through the injected STT→chat→TTS pipeline. Without a
-		// bus so an upper layer (or a later milestone) can consume it, rather
+		// voice service it replies with a clear error rather than dropping the
-		// than dropping it as unknown.
+		// frame. The audio reply is delivered as a voice_output event on the
-		c.publishPublic(wsVoiceInput, rawData(in.Data))
+		// requesting connection.
+		c.handleVoiceInput(in)
 	case "subagent_ask":
 		c.handleSubagentAsk(in)
 	case "task_output_query":