From 4c212aeb040bfb4680959dbec0dcf216fd6ef362 Mon Sep 17 00:00:00 2001
From: chengqian <wuchengqian5@163.com>
Date: Tue, 15 Sep 2026 19:20:15 +0800
Subject: [PATCH] =?UTF-8?q?feat(ui):=E4=BC=98=E5=8C=96=E5=9C=A8=E7=BA=BF?=
 =?UTF-8?q?=E4=BD=93=E9=AA=8C=E5=8A=9F=E8=83=BD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 CLAUDE.md                                     |   1 +
 internal/logic/common/audit.go                |  15 +-
 internal/logic/common/inline_media.go         |  97 +++
 internal/logic/common/inline_media_test.go    | 119 ++++
 internal/middleware/content_filter.go         |  15 +-
 relay/handler/relay_handler_test.go           |  57 ++
 web/tenant/src/components/common/Icon.vue     |   4 +
 .../src/views/service/PlaygroundPage.vue      | 106 ++-
 .../service/playground/AttachmentBar.vue      | 187 ++++++
 .../src/views/service/playground/ChatTab.vue  | 622 ++++++++++++++----
 .../views/service/playground/PromptInput.vue  | 226 +++++++
 .../src/views/service/playground/VideoTab.vue |  68 +-
 .../service/playground/attachmentParts.ts     | 120 ++++
 .../service/playground/useAttachments.ts      | 205 ++++++
 14 files changed, 1703 insertions(+), 139 deletions(-)
 create mode 100644 internal/logic/common/inline_media.go
 create mode 100644 internal/logic/common/inline_media_test.go
 create mode 100644 web/tenant/src/views/service/playground/AttachmentBar.vue
 create mode 100644 web/tenant/src/views/service/playground/PromptInput.vue
 create mode 100644 web/tenant/src/views/service/playground/attachmentParts.ts
 create mode 100644 web/tenant/src/views/service/playground/useAttachments.ts

diff --git a/CLAUDE.md b/CLAUDE.md
index 8573818..141cb34 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -473,6 +473,7 @@ cmd（路由注册）
 | 对比分析 | `docs/new-api-vs-sub2api对比分析报告.md` | 两者差异 |
 | 大模型实施方案 | `docs/大模型完整实施方案-v2.1.md` | 大模型代理层设计 |
 | API 格式参考 | `docs/reference/api-format-reference.md` | JSON 示例、SSE 格式、错误映射、中间件差异 |
+| 协议转换语料采集 | `docs/协议转换语料采集手册.md` | `corpus-extract` 命令用法：从渠道调试日志提取真实流量语料喂给协议转换测试 |
 | GoFrame 使用规范 | `docs/reference/goframe-conventions.md` | 框架使用规范 + 已修复的框架 bug 记录 |
 
 
diff --git a/internal/logic/common/audit.go b/internal/logic/common/audit.go
index 721f72e..a6124c6 100644
--- a/internal/logic/common/audit.go
+++ b/internal/logic/common/audit.go
@@ -100,13 +100,18 @@ func GetTenantAuditLevel(ctx context.Context, tenantID int64) string {
 
 // ApplyAuditLevel 根据审计级别处理请求体和响应体，返回处理后的内容。
 func ApplyAuditLevel(level string, requestBody, responseBody string, isStream bool, path string) (req, resp string) {
-	req = requestBody
-	resp = responseBody
+	// 多模态请求（对话内联图片/音频、视频参考图）的 base64 载荷先剥离成长度占位：
+	// multipart 体已在 relay handler 侧被挡成 "[binary body omitted]"，但内联在 JSON
+	// 里的 data URL 会原样落库，单条就能把 aud_request_logs 撑到几 MB。
+	// 此处是对话与异步任务两条审计链路的唯一收口点（DataProviderImpl.RecordAudit），
+	// 在这里剥离即可同时覆盖。none 级别直接丢弃，无需白做一次剥离。
+	if level == AuditLevelNone {
+		return "", ""
+	}
+	req = StripInlineMedia(requestBody)
+	resp = StripInlineMedia(responseBody)
 
 	switch level {
-	case AuditLevelNone:
-		req = ""
-		resp = ""
 	case AuditLevelQuestionOnly:
 		resp = ""
 	case AuditLevelMasked:
diff --git a/internal/logic/common/inline_media.go b/internal/logic/common/inline_media.go
new file mode 100644
index 0000000..c0823f0
--- /dev/null
+++ b/internal/logic/common/inline_media.go
@@ -0,0 +1,97 @@
+package common
+
+import (
+	"bytes"
+	"strconv"
+	"strings"
+)
+
+// 内联媒体（data URL）处理工具。
+//
+// 多模态请求体里的图片/音频以 `data:<mime>;base64,<载荷>` 的形式内联在 JSON 中，
+// 单个载荷动辄数 MB。这类载荷对三条链路都是纯负担：
+//   - 预扣估算按字节折算会把一张图估成几十万 token，直接冻穿租户余额
+//   - 审计落库会把整段 base64 灌进 aud_request_logs
+//   - 敏感词匹配白扫几 MB，还可能在 base64 串里误命中
+//
+// 这里提供统一的扫描与剥离能力，三处共用，避免各写一份。
+
+// inlineMediaMarker data URL 中 base64 载荷的起始标记。
+const inlineMediaMarker = ";base64,"
+
+// inlineMediaMinPayload 认定为内联媒体的最小载荷长度。
+// 真实的图片/音频 base64 远大于此，设下限是为了不误伤正文里恰好出现该标记的短文本。
+const inlineMediaMinPayload = 64
+
+// ScanInlineMediaBytes 扫描内容中的 data URL base64 载荷，返回载荷总字节数与媒体个数。
+//
+// 纯字节扫描不做 JSON 反序列化——调用点在请求热路径上，且请求体可能有数 MB。
+// base64 字母表不含 `"` 与 `\`，故以二者作为 JSON 字符串内载荷的终止符。
+func ScanInlineMediaBytes(b []byte) (payloadBytes, count int) {
+	marker := []byte(inlineMediaMarker)
+	rest := b
+	for {
+		idx := bytes.Index(rest, marker)
+		if idx < 0 {
+			return
+		}
+		rest = rest[idx+len(marker):]
+		end := bytes.IndexAny(rest, "\"\\")
+		if end < 0 {
+			end = len(rest)
+		}
+		if end >= inlineMediaMinPayload {
+			payloadBytes += end
+			count++
+		}
+		rest = rest[end:]
+	}
+}
+
+// StripInlineMediaBytes 是 StripInlineMedia 的字节切片版本。
+// 未命中内联媒体时原样返回入参切片，不产生任何拷贝——请求热路径上的绝大多数
+// 请求体是纯文本，这条快路径让它们零开销通过。
+func StripInlineMediaBytes(b []byte) []byte {
+	if !bytes.Contains(b, []byte(inlineMediaMarker)) {
+		return b
+	}
+	return []byte(StripInlineMedia(string(b)))
+}
+
+// StripInlineMedia 把 base64 载荷替换成长度占位，保留 mime 前缀便于排查：
+//
+//	data:image/png;base64,iVBORw0KGgo...  →  data:image/png;base64,<省略 2097152 字节>
+//
+// 未命中任何内联媒体时原样返回，不产生额外分配。
+func StripInlineMedia(s string) string {
+	if !strings.Contains(s, inlineMediaMarker) {
+		return s
+	}
+
+	var sb strings.Builder
+	rest := s
+	for {
+		idx := strings.Index(rest, inlineMediaMarker)
+		if idx < 0 {
+			sb.WriteString(rest)
+			break
+		}
+		head := idx + len(inlineMediaMarker)
+		sb.WriteString(rest[:head])
+		rest = rest[head:]
+
+		end := strings.IndexAny(rest, "\"\\")
+		if end < 0 {
+			end = len(rest)
+		}
+		if end >= inlineMediaMinPayload {
+			sb.WriteString("<省略 ")
+			sb.WriteString(strconv.Itoa(end))
+			sb.WriteString(" 字节>")
+		} else {
+			sb.WriteString(rest[:end])
+		}
+		rest = rest[end:]
+	}
+	return sb.String()
+}
diff --git a/internal/logic/common/inline_media_test.go b/internal/logic/common/inline_media_test.go
new file mode 100644
index 0000000..b5b6a8f
--- /dev/null
+++ b/internal/logic/common/inline_media_test.go
@@ -0,0 +1,119 @@
+package common
+
+import (
+	"strings"
+	"testing"
+)
+
+// bigPayload 构造一段长度足够被认定为内联媒体的 base64 载荷。
+func bigPayload(n int) string {
+	return strings.Repeat("A", n)
+}
+
+func TestScanInlineMediaBytes(t *testing.T) {
+	payload := bigPayload(400)
+	cases := []struct {
+		name      string
+		body      string
+		wantBytes int
+		wantCount int
+	}{
+		{"纯文本", `{"model":"gpt-4o","messages":[{"role":"user","content":"你好"}]}`, 0, 0},
+		{"空体", "", 0, 0},
+		{
+			"单图",
+			`{"content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,` + payload + `"}}]}`,
+			400, 1,
+		},
+		{
+			"双图",
+			`{"a":"data:image/png;base64,` + payload + `","b":"data:image/jpeg;base64,` + payload + `"}`,
+			800, 2,
+		},
+		{
+			"音频",
+			`{"input_audio":{"data":"data:audio/mp3;base64,` + payload + `","format":"mp3"}}`,
+			400, 1,
+		},
+		// 载荷过短：正文里恰好出现该标记，不应被当作媒体扣减
+		{"短载荷不计入", `{"text":"写法是 ;base64,abc 这样"}`, 0, 0},
+		// 无结束引号（截断的请求体）：载荷取到末尾
+		{"未闭合", `{"url":"data:image/png;base64,` + payload, 400, 1},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			gotBytes, gotCount := ScanInlineMediaBytes([]byte(c.body))
+			if gotBytes != c.wantBytes || gotCount != c.wantCount {
+				t.Errorf("ScanInlineMediaBytes() = (%d, %d), want (%d, %d)",
+					gotBytes, gotCount, c.wantBytes, c.wantCount)
+			}
+		})
+	}
+}
+
+func TestStripInlineMedia(t *testing.T) {
+	payload := bigPayload(400)
+
+	t.Run("剥离载荷保留 mime 前缀", func(t *testing.T) {
+		in := `{"url":"data:image/png;base64,` + payload + `","prompt":"描述这张图"}`
+		got := StripInlineMedia(in)
+		if strings.Contains(got, payload) {
+			t.Error("base64 载荷未被剥离")
+		}
+		if !strings.Contains(got, "data:image/png;base64,<省略 400 字节>") {
+			t.Errorf("占位文案不符: %s", got)
+		}
+		if !strings.Contains(got, `"prompt":"描述这张图"`) {
+			t.Error("载荷之外的内容被破坏")
+		}
+	})
+
+	t.Run("纯文本原样返回", func(t *testing.T) {
+		in := `{"model":"gpt-4o","messages":[{"role":"user","content":"你好"}]}`
+		if got := StripInlineMedia(in); got != in {
+			t.Errorf("StripInlineMedia() = %q, want 原样返回", got)
+		}
+	})
+
+	t.Run("多个载荷全部剥离", func(t *testing.T) {
+		in := `{"a":"data:image/png;base64,` + payload + `","b":"data:image/jpeg;base64,` + payload + `"}`
+		got := StripInlineMedia(in)
+		if strings.Contains(got, payload) {
+			t.Error("仍残留 base64 载荷")
+		}
+		if n := strings.Count(got, "<省略 400 字节>"); n != 2 {
+			t.Errorf("占位数量 = %d, want 2", n)
+		}
+	})
+
+	t.Run("短载荷保持原样", func(t *testing.T) {
+		in := `{"text":"写法是 ;base64,abc 这样"}`
+		if got := StripInlineMedia(in); got != in {
+			t.Errorf("StripInlineMedia() = %q, want 原样返回", got)
+		}
+	})
+
+	t.Run("剥离后可被字节版复用且无拷贝", func(t *testing.T) {
+		plain := []byte(`{"content":"纯文本"}`)
+		got := StripInlineMediaBytes(plain)
+		if &got[0] != &plain[0] {
+			t.Error("纯文本路径应原样返回入参切片，不做拷贝")
+		}
+	})
+}
+
+func TestApplyAuditLevelStripsInlineMedia(t *testing.T) {
+	payload := bigPayload(400)
+	body := `{"url":"data:image/png;base64,` + payload + `"}`
+
+	req, _ := ApplyAuditLevel(AuditLevelFull, body, "", false, "/v1/chat/completions")
+	if strings.Contains(req, payload) {
+		t.Error("full 级别仍把 base64 载荷落库")
+	}
+
+	req, resp := ApplyAuditLevel(AuditLevelNone, body, body, false, "/v1/chat/completions")
+	if req != "" || resp != "" {
+		t.Error("none 级别应返回空")
+	}
+}
diff --git a/internal/middleware/content_filter.go b/internal/middleware/content_filter.go
index 458f63a..b86f645 100644
--- a/internal/middleware/content_filter.go
+++ b/internal/middleware/content_filter.go
@@ -94,7 +94,12 @@ func ContentFilter(r *ghttp.Request) {
 		return
 	}
 
-	result := common.ContentFilter().Check(body)
+	// 多模态请求体里的 base64 内联媒体先剥离再匹配：几 MB 的 base64 白扫一遍纯属浪费，
+	// 且 base64 字母表可能恰好拼出敏感词，导致正常的带图请求被误 block。
+	// 纯文本请求体原样返回，不额外拷贝。
+	checkBody := common.StripInlineMediaBytes(body)
+
+	result := common.ContentFilter().Check(checkBody)
 	if !result.Matched {
 		r.Middleware.Next()
 		return
@@ -105,7 +110,7 @@ func ContentFilter(r *ghttp.Request) {
 	r.SetCtxVar(CtxKeyContentFilterWords, result.MatchedWords)
 
 	// Queue async log write
-	queueContentFilterLog(r, mode, result.MatchedWords, string(body))
+	queueContentFilterLog(r, mode, result.MatchedWords, string(checkBody))
 
 	switch mode {
 	case "log":
@@ -115,8 +120,10 @@ func ContentFilter(r *ghttp.Request) {
 	case "replace":
 		g.Log().Infof(r.Context(), "[ContentFilter] replaced words: %v", result.MatchedWords)
 		r.SetCtxVar(CtxKeyContentFilterReplaced, true)
-		r.SetCtxVar(CtxKeyContentFilterOriginal, string(body))
-		r.SetCtxVar(CtxKeyContentFilterFiltered, result.FilteredText)
+		r.SetCtxVar(CtxKeyContentFilterOriginal, string(checkBody))
+		// 替换后的请求体必须作用在原始 body 上——剥离版本丢了内联媒体，
+		// 一旦被下游拿去改写上游请求就会把图片/音频弄没。命中才多跑这一次，代价可接受。
+		r.SetCtxVar(CtxKeyContentFilterFiltered, common.ContentFilter().Check(body).FilteredText)
 		r.Middleware.Next()
 
 	case "block":
diff --git a/relay/handler/relay_handler_test.go b/relay/handler/relay_handler_test.go
index 4b5b17c..aa58234 100644
--- a/relay/handler/relay_handler_test.go
+++ b/relay/handler/relay_handler_test.go
@@ -2,6 +2,7 @@ package handler
 
 import (
 	"context"
+	"strings"
 	"testing"
 
 	"github.com/qianfree/team-api/relay/common"
@@ -60,3 +61,59 @@ func TestResolveRelayModel(t *testing.T) {
 		}
 	})
 }
+
+// TestEstimateInputTokens 多模态请求体的预扣估算：base64 内联媒体必须按固定成本计入，
+// 不能按字节折算——一张 2MB 的图按 len/4 会被估成约 70 万 token，预扣直接冻穿余额。
+func TestEstimateInputTokens(t *testing.T) {
+	// 模拟 2MB 的 base64 图片载荷
+	payload := strings.Repeat("A", 2<<20)
+
+	cases := []struct {
+		name string
+		body string
+		want int
+	}{
+		{
+			"纯文本按字节折算",
+			strings.Repeat("x", 400),
+			100,
+		},
+		{
+			"空体",
+			"",
+			0,
+		},
+		{
+			// 文本骨架 40 字节 + 1 个媒体
+			"单图不按字节折算",
+			`{"url":"data:image/png;base64,` + payload + `"}`,
+			(len(`{"url":"data:image/png;base64,`+`"}`))/4 + inlineMediaTokenCost,
+		},
+		{
+			"双图按个数累加",
+			`{"a":"data:image/png;base64,` + payload + `","b":"data:image/png;base64,` + payload + `"}`,
+			(len(`{"a":"data:image/png;base64,`+`","b":"data:image/png;base64,`+`"}`))/4 + 2*inlineMediaTokenCost,
+		},
+		{
+			// 畸形 data URL（无 base64 标记）退化为纯文本路径
+			"畸形 data URL",
+			`{"url":"data:image/png,` + strings.Repeat("A", 400) + `"}`,
+			len(`{"url":"data:image/png,`+`"}`)/4 + 100,
+		},
+	}
+
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			if got := estimateInputTokens([]byte(c.body)); got != c.want {
+				t.Errorf("estimateInputTokens() = %d, want %d", got, c.want)
+			}
+		})
+	}
+
+	// 回归保护：2MB 图片的估算必须远低于按字节折算的量级
+	naive := len(`{"url":"data:image/png;base64,`+payload+`"}`) / 4
+	got := estimateInputTokens([]byte(`{"url":"data:image/png;base64,` + payload + `"}`))
+	if got >= naive/10 {
+		t.Errorf("2MB 图片估算 %d tokens，与按字节折算的 %d 处于同一量级，预扣仍会冻穿余额", got, naive)
+	}
+}
diff --git a/web/tenant/src/components/common/Icon.vue b/web/tenant/src/components/common/Icon.vue
index 2d347ca..815a98d 100644
--- a/web/tenant/src/components/common/Icon.vue
+++ b/web/tenant/src/components/common/Icon.vue
@@ -80,6 +80,10 @@ const icons: Record<string, string> = {
 	users: 'M15 19.128a9.38 9.38 0 002.625.372 9.337 9.337 0 004.121-.952 4.125 4.125 0 00-7.533-2.493M15 19.128v-.003c0-1.113-.285-2.16-.786-3.07M15 19.128v.106A12.318 12.318 0 018.624 21c-2.331 0-4.512-.645-6.374-1.766l-.001-.109a6.375 6.375 0 0111.964-3.07M12 6.375a3.375 3.375 0 11-6.75 0 3.375 3.375 0 016.75 0zm8.25 2.25a2.625 2.625 0 11-5.25 0 2.625 2.625 0 015.25 0z',
 
 	// Files & Documents
+	photo: 'M2.25 15.75l5.159-5.159a2.25 2.25 0 013.182 0l5.159 5.159m-1.5-1.5l1.409-1.409a2.25 2.25 0 013.182 0l2.909 2.909m-18 3.75h16.5a1.5 1.5 0 001.5-1.5V6a1.5 1.5 0 00-1.5-1.5H3.75A1.5 1.5 0 002.25 6v12a1.5 1.5 0 001.5 1.5zm10.5-11.25h.008v.008h-.008V8.25zm.375 0a.375.375 0 11-.75 0 .375.375 0 01.75 0z',
+	paperclip: 'M18.375 12.739l-7.693 7.693a4.5 4.5 0 01-6.364-6.364l10.94-10.94A3 3 0 1119.5 7.372L8.552 18.32m.009-.01l-.01.01m5.699-9.941l-7.81 7.81a1.5 1.5 0 002.112 2.13',
+	musicalNote: 'M9 9l10.5-3m0 6.553v3.75a2.25 2.25 0 01-1.632 2.163l-1.32.377a1.803 1.803 0 11-.99-3.467l2.31-.66a2.25 2.25 0 001.632-2.163zm0 0V2.25L9 5.25v10.303m0 0v3.75a2.25 2.25 0 01-1.632 2.163l-1.32.377a1.803 1.803 0 01-.99-3.467l2.31-.66A2.25 2.25 0 009 15.553z',
+	film: 'M3.375 19.5h17.25m-17.25 0a1.125 1.125 0 01-1.125-1.125M3.375 19.5h7.5c.621 0 1.125-.504 1.125-1.125m-9.75 0V5.625m0 12.75v-1.5c0-.621.504-1.125 1.125-1.125m18.375 2.625V5.625m0 12.75c0 .621-.504 1.125-1.125 1.125m1.125-1.125v-1.5c0-.621-.504-1.125-1.125-1.125m0 3.75h-7.5A1.125 1.125 0 0112 18.375m9.75-12.75c0-.621-.504-1.125-1.125-1.125H3.375c-.621 0-1.125.504-1.125 1.125m19.5 0v1.5c0 .621-.504 1.125-1.125 1.125M2.25 5.625v1.5c0 .621.504 1.125 1.125 1.125m0 0h17.25m-17.25 0h7.5c.621 0 1.125.504 1.125 1.125M3.375 8.25c-.621 0-1.125.504-1.125 1.125v1.5c0 .621.504 1.125 1.125 1.125m17.25-3.75h-7.5c-.621 0-1.125.504-1.125 1.125m8.625-1.125c.621 0 1.125.504 1.125 1.125v1.5c0 .621-.504 1.125-1.125 1.125m-17.25 0h7.5m-7.5 0c-.621 0-1.125.504-1.125 1.125v1.5c0 .621.504 1.125 1.125 1.125M12 10.875v-1.5m0 1.5c0 .621-.504 1.125-1.125 1.125M12 10.875c0 .621.504 1.125 1.125 1.125m-2.25 0c.621 0 1.125.504 1.125 1.125M13.125 12h7.5m-7.5 0c-.621 0-1.125.504-1.125 1.125M20.625 12c.621 0 1.125.504 1.125 1.125v1.5c0 .621-.504 1.125-1.125 1.125m-17.25 0h7.5M12 14.625v-1.5m0 1.5c0 .621-.504 1.125-1.125 1.125M12 14.625c0 .621.504 1.125 1.125 1.125m-2.25 0c.621 0 1.125.504 1.125 1.125m0 1.5v-1.5m0 0c0-.621.504-1.125 1.125-1.125m0 0h7.5',
 	document: 'M9 12h6m-6 4h6m2 5H7a2 2 0 01-2-2V5a2 2 0 012-2h5.586a1 1 0 01.707.293l5.414 5.414a1 1 0 01.293.707V19a2 2 0 01-2 2z',
 	bookOpen: 'M12 6.042A8.967 8.967 0 006 3.75c-1.052 0-2.062.18-3 .512v14.25A8.987 8.987 0 016 18c2.305 0 4.408.867 6 2.292m0-14.25a8.966 8.966 0 016-2.292c1.052 0 2.062.18 3 .512v14.25A8.987 8.987 0 0018 18a8.967 8.967 0 00-6 2.292m0-14.25v14.25',
 	clipboard: 'M9 12h3.75M9 15h3.75M9 18h3.75m3 .75H18a2.25 2.25 0 002.25-2.25V6.108c0-1.135-.845-2.098-1.976-2.192a48.424 48.424 0 00-1.123-.08m-5.801 0c-.065.21-.1.433-.1.664 0 .414.336.75.75.75h4.5a.75.75 0 00.75-.75 2.25 2.25 0 00-.1-.664m-5.8 0A2.251 2.251 0 0113.5 2.25H15c1.012 0 1.867.668 2.15 1.586m-5.8 0c-.376.023-.75.05-1.124.08C9.095 4.01 8.25 4.973 8.25 6.108V8.25m0 0H4.875c-.621 0-1.125.504-1.125 1.125v11.25c0 .621.504 1.125 1.125 1.125h9.75c.621 0 1.125-.504 1.125-1.125V9.375c0-.621-.504-1.125-1.125-1.125H8.25z',
diff --git a/web/tenant/src/views/service/PlaygroundPage.vue b/web/tenant/src/views/service/PlaygroundPage.vue
index 6b9c67f..4ff0494 100644
--- a/web/tenant/src/views/service/PlaygroundPage.vue
+++ b/web/tenant/src/views/service/PlaygroundPage.vue
@@ -23,15 +23,18 @@ interface ModelItem {
 	async_image?: boolean
 	// 图片模型同步端点是否可用（「仅异步」厂商为 false）
 	image_sync_supported?: boolean
+	// 模型能力特性 JSON（如 {"vision":true,"audio_input":true}），用于门控附件上传
+	capabilities?: string
 }
 
+// 每个 Tab 的图标，用于分段控件左侧的视觉锚点
 const tabs = [
-	{ key: 'chat', label: '对话' },
-	{ key: 'image', label: '图像' },
-	{ key: 'video', label: '视频' },
-	{ key: 'audio', label: '语音' },
-	{ key: 'embedding', label: '嵌入' },
-	{ key: 'rerank', label: '重排' },
+	{ key: 'chat', label: '对话', icon: 'chat' },
+	{ key: 'image', label: '图像', icon: 'photo' },
+	{ key: 'video', label: '视频', icon: 'film' },
+	{ key: 'audio', label: '语音', icon: 'musicalNote' },
+	{ key: 'embedding', label: '嵌入', icon: 'cube' },
+	{ key: 'rerank', label: '重排', icon: 'chart' },
 ]
 
 const activeTab = ref('chat')
@@ -59,6 +62,13 @@ const modelsLoading = ref(false)
 let loadingTimer: number | null = null
 let loadSeq = 0
 
+// 当前 Key 的显示名与掩码前缀，供工具栏的 Key 胶囊展示
+const selectedKeyLabel = computed(() => {
+	const key = apiKeys.value.find(k => k.id === selectedKeyId.value)
+	if (!key) return ''
+	return `${key.name} · ${key.key_prefix}···`
+})
+
 // 按右上角选中的 API Key 拉取其可用模型，使左侧模型列表跟随 Key 切换。
 // 切换期间保留旧列表避免空白闪烁；加载提示延迟 200ms 才出现，请求更快时
 // 完全不显示，杜绝 spinner 瞬闪；loadSeq 丢弃过期响应，防止快速切换 Key 时旧结果覆盖。
@@ -93,15 +103,19 @@ watch(selectedKeyId, loadModels, { immediate: true })
 
 <template>
 	<div>
-		<!-- 紧凑工具栏：左侧模型类型标签 + 右侧 API Key 单行收纳 -->
+		<!-- 统一工具栏：左侧模型类型分段控件，右侧调用身份（API Key）与计费提示。
+		     计费提示从原对话区黄色横幅迁移至此，不再占用对话区高度。 -->
 		<div class="card mb-3">
-			<div class="flex flex-wrap items-center gap-x-3 gap-y-2 px-4 py-3 sm:px-5">
-				<!-- 左侧：模型类型标签 -->
-				<div class="tabs flex-wrap">
+			<div class="flex flex-wrap items-center gap-x-3 gap-y-2 px-3 py-2.5 sm:px-4">
+				<!-- 左侧：模型类型分段控件 -->
+				<div class="tabs flex-wrap" role="tablist">
 					<button v-for="tab in tabs" :key="tab.key"
-						class="tab"
+						role="tab"
+						class="tab flex items-center gap-1.5"
 						:class="{ 'tab-active': activeTab === tab.key }"
+						:aria-selected="activeTab === tab.key"
 						@click="activeTab = tab.key">
+						<Icon :name="tab.icon" size="xs" />
 						{{ tab.label }}
 					</button>
 				</div>
@@ -116,6 +130,15 @@ watch(selectedKeyId, loadModels, { immediate: true })
 
 				<!-- 右侧：API Key 区域 -->
 				<div class="ml-auto flex shrink-0 items-center gap-2">
+					<!-- 真实计费提示：悬停展示完整说明 -->
+					<span
+						class="hidden items-center gap-1 rounded-full border border-amber-200 bg-amber-50 px-2.5 py-1 text-[11px] font-medium text-amber-700 xl:inline-flex"
+						title="Playground 模式 — 使用真实 API Key 调用，产生实际费用"
+					>
+						<Icon name="bolt" size="xs" />
+						真实计费
+					</span>
+
 					<!-- 加载中 -->
 					<template v-if="keyLoading">
 						<div class="spinner h-4 w-4 text-primary-600"></div>
@@ -129,12 +152,15 @@ watch(selectedKeyId, loadModels, { immediate: true })
 					</div>
 					<!-- Key 选择器 -->
 					<template v-else>
-						<span class="text-xs font-medium text-gray-400">API Key</span>
-						<div class="w-56">
+						<div class="key-picker">
+							<Icon name="key" size="xs" class="key-picker-icon" />
 							<n-select
 								:value="selectedKeyId"
 								:options="apiKeys.map(k => ({ label: `${k.name} (${k.key_prefix}...)`, value: k.id }))"
 								placeholder="选择 API Key"
+								size="small"
+								class="key-picker-select"
+								:title="selectedKeyLabel"
 								@update:value="selectKey"
 							/>
 						</div>
@@ -156,3 +182,57 @@ watch(selectedKeyId, loadModels, { immediate: true })
 		</template>
 	</div>
 </template>
+
+<style scoped>
+/* 分段控件：激活项用主色文字与更清晰的投影，提升可辨识度 */
+.tab {
+	display: inline-flex;
+	align-items: center;
+}
+.tab-active {
+	color: #0d9488;
+	box-shadow: 0 1px 3px rgba(15, 23, 42, 0.08), 0 0 0 1px rgba(20, 184, 166, 0.18);
+}
+
+/* API Key 胶囊：key 图标 + 紧凑下拉，与左侧分段控件在同一视觉高度上 */
+.key-picker {
+	display: flex;
+	align-items: center;
+	gap: 0.25rem;
+	height: 2.25rem;
+	border: 1px solid #e5e7eb;
+	border-radius: 0.75rem;
+	background: #fff;
+	padding-left: 0.55rem;
+	padding-right: 0.25rem;
+	box-shadow: 0 1px 2px rgba(15, 23, 42, 0.04);
+	transition: border-color 180ms ease, box-shadow 180ms ease;
+}
+.key-picker:hover,
+.key-picker:focus-within {
+	border-color: rgba(20, 184, 166, 0.45);
+	box-shadow: 0 0 0 4px rgba(20, 184, 166, 0.08);
+}
+.key-picker-icon {
+	color: #f59e0b;
+	flex-shrink: 0;
+}
+.key-picker-select {
+	width: 12.5rem;
+}
+
+.fade-enter-active,
+.fade-leave-active {
+	transition: opacity 0.2s ease;
+}
+.fade-enter-from,
+.fade-leave-to {
+	opacity: 0;
+}
+
+@media (max-width: 639px) {
+	.key-picker-select {
+		width: 9rem;
+	}
+}
+</style>
diff --git a/web/tenant/src/views/service/playground/AttachmentBar.vue b/web/tenant/src/views/service/playground/AttachmentBar.vue
new file mode 100644
index 0000000..d5765f0
--- /dev/null
+++ b/web/tenant/src/views/service/playground/AttachmentBar.vue
@@ -0,0 +1,187 @@
+<script setup lang="ts">
+import { ref, computed } from 'vue'
+import Icon from '@/components/common/Icon.vue'
+import { formatBytes, type Attachment } from './useAttachments'
+
+/**
+ * 附件预览带：位于输入框上方，承载上传入口与已选附件的缩略展示。
+ * 点击某个附件 = 在输入框光标处插入它的 [@名称] 标记（比打 @ 更快的路径）。
+ */
+
+const props = withDefaults(defineProps<{
+	attachments: Attachment[]
+	/** input 的 accept 属性，如 "image/*,audio/*" */
+	accept: string
+	max: number
+	disabled?: boolean
+	/** 禁用原因，作为上传按钮的 title 与旁注展示 */
+	disabledReason?: string
+	/**
+	 * 视觉变体：
+	 * - boxed（默认）：带虚线框的独立附件带，含上传按钮与提示文案
+	 * - bare：只渲染附件缩略片，无边框无上传按钮，供父组件把输入区整合成一体时使用
+	 *   （此变体下由父组件通过 ref 调用 pick() 触发选择文件）
+	 */
+	variant?: 'boxed' | 'bare'
+}>(), {
+	disabled: false,
+	disabledReason: '',
+	variant: 'boxed',
+})
+
+const emit = defineEmits<{
+	add: [files: File[]]
+	remove: [id: number]
+	insert: [name: string]
+}>()
+
+const fileInput = ref<HTMLInputElement | null>(null)
+const dragging = ref(false)
+
+const full = computed(() => props.attachments.length >= props.max)
+const uploadDisabled = computed(() => props.disabled || full.value)
+
+function pick() {
+	if (uploadDisabled.value) return
+	fileInput.value?.click()
+}
+
+function onPicked(e: Event) {
+	const input = e.target as HTMLInputElement
+	if (input.files?.length) emit('add', Array.from(input.files))
+	// 置空 value，否则连续选同一个文件不会再触发 change
+	input.value = ''
+}
+
+function onDrop(e: DragEvent) {
+	dragging.value = false
+	if (uploadDisabled.value) return
+	const files = Array.from(e.dataTransfer?.files || [])
+	if (files.length) emit('add', files)
+}
+
+function onDragOver() {
+	if (!uploadDisabled.value) dragging.value = true
+}
+
+// bare 变体下不渲染上传按钮，选择文件由父组件调用此方法触发
+defineExpose({ pick })
+</script>
+
+<template>
+	<!-- bare：只渲染附件缩略片，无边框、无上传入口，融入父组件的输入容器 -->
+	<div v-if="variant === 'bare'" class="flex flex-wrap items-center gap-1.5">
+		<div
+			v-for="att in attachments"
+			:key="att.id"
+			class="group relative flex items-center gap-1.5 rounded-lg border border-gray-200 bg-gray-50 py-0.5 pl-0.5 pr-6 transition-colors duration-150 hover:border-primary-300 hover:bg-primary-50/40"
+		>
+			<button
+				class="flex items-center gap-1.5 cursor-pointer"
+				:title="`${att.fileName} · 点击插入 [@${att.name}]`"
+				@click="emit('insert', att.name)"
+			>
+				<img
+					v-if="att.kind === 'image'"
+					:src="att.previewUrl"
+					class="h-6 w-6 rounded object-cover"
+					alt=""
+				/>
+				<span v-else class="flex h-6 w-6 items-center justify-center rounded bg-white text-gray-500">
+					<Icon :name="att.kind === 'audio' ? 'musicalNote' : 'film'" size="xs" />
+				</span>
+				<span class="flex items-baseline gap-1 leading-none">
+					<span class="text-xs font-medium text-gray-700">{{ att.name }}</span>
+					<span class="text-[10px] text-gray-400">{{ formatBytes(att.bytes) }}</span>
+				</span>
+			</button>
+			<button
+				class="absolute right-0.5 top-1/2 -translate-y-1/2 rounded p-0.5 text-gray-300 transition-colors duration-150 hover:bg-white hover:text-red-500"
+				title="移除"
+				@click="emit('remove', att.id)"
+			>
+				<Icon name="x" size="xs" />
+			</button>
+		</div>
+
+		<!-- 文件选择器仍需挂载，供父组件调 pick() 打开 -->
+		<input
+			ref="fileInput"
+			type="file"
+			multiple
+			class="hidden"
+			:accept="accept"
+			@change="onPicked"
+		/>
+	</div>
+
+	<div
+		v-else
+		class="rounded-xl border border-dashed px-3 py-2 transition-colors duration-150"
+		:class="dragging ? 'border-primary-400 bg-primary-50/60' : 'border-gray-200 bg-gray-50/60'"
+		@dragover.prevent="onDragOver"
+		@dragleave="dragging = false"
+		@drop.prevent="onDrop"
+	>
+		<div class="flex flex-wrap items-center gap-2">
+			<!-- 已选附件 -->
+			<div
+				v-for="att in attachments"
+				:key="att.id"
+				class="group relative flex items-center gap-2 rounded-lg border border-gray-200 bg-white py-1 pl-1 pr-7 transition-colors duration-150 hover:border-primary-300"
+			>
+				<button
+					class="flex items-center gap-2 cursor-pointer"
+					:title="`${att.fileName} · 点击插入 [@${att.name}]`"
+					@click="emit('insert', att.name)"
+				>
+					<img
+						v-if="att.kind === 'image'"
+						:src="att.previewUrl"
+						class="h-8 w-8 rounded object-cover"
+						alt=""
+					/>
+					<span v-else class="flex h-8 w-8 items-center justify-center rounded bg-gray-100 text-gray-500">
+						<Icon :name="att.kind === 'audio' ? 'musicalNote' : 'film'" size="sm" />
+					</span>
+					<span class="flex flex-col items-start leading-tight">
+						<span class="text-xs font-medium text-gray-700">{{ att.name }}</span>
+						<span class="text-[11px] text-gray-400">{{ formatBytes(att.bytes) }}</span>
+					</span>
+				</button>
+				<button
+					class="absolute right-1 top-1 rounded p-0.5 text-gray-300 transition-colors duration-150 hover:bg-gray-100 hover:text-red-500"
+					title="移除"
+					@click="emit('remove', att.id)"
+				>
+					<Icon name="x" size="xs" />
+				</button>
+			</div>
+
+			<!-- 上传入口 -->
+			<button
+				class="btn btn-secondary btn-sm"
+				:disabled="uploadDisabled"
+				:title="disabled ? disabledReason : (full ? `最多 ${max} 个附件` : '可拖拽文件或在输入框粘贴图片')"
+				@click="pick"
+			>
+				<Icon name="paperclip" size="sm" />
+				上传附件
+			</button>
+
+			<span v-if="disabled && disabledReason" class="text-xs text-amber-600">{{ disabledReason }}</span>
+			<span v-else-if="!attachments.length" class="text-xs text-gray-400">
+				可拖拽文件到此处，或在输入框直接粘贴图片
+			</span>
+		</div>
+
+		<input
+			ref="fileInput"
+			type="file"
+			multiple
+			class="hidden"
+			:accept="accept"
+			@change="onPicked"
+		/>
+	</div>
+</template>
diff --git a/web/tenant/src/views/service/playground/ChatTab.vue b/web/tenant/src/views/service/playground/ChatTab.vue
index 8fe88d0..a1aa8d1 100644
--- a/web/tenant/src/views/service/playground/ChatTab.vue
+++ b/web/tenant/src/views/service/playground/ChatTab.vue
@@ -5,6 +5,11 @@ import { calculateCost } from './calculateCost'
 import Icon from '@/components/common/Icon.vue'
 import MarkdownRenderer from '@/components/common/MarkdownRenderer.vue'
 import BaseSelect from '../../../components/common/BaseSelect.vue'
+import AttachmentBar from './AttachmentBar.vue'
+import PromptInput from './PromptInput.vue'
+import { useAttachments, formatBytes, MAX_TOTAL_BYTES, type AttachmentKind } from './useAttachments'
+import { compileChatParts, type ContentPart, type AttachmentPreview } from './attachmentParts'
+import { toast } from '@/utils/toast'
 
 interface ModelItem {
 	model_id: string
@@ -14,6 +19,8 @@ interface ModelItem {
 	per_request_price?: number | null
 	input_price?: number | null
 	output_price?: number | null
+	// 能力特性 JSON（如 {"vision":true}），决定附件上传是否开放
+	capabilities?: string
 }
 const props = defineProps<{
 	models: ModelItem[]
@@ -63,13 +70,135 @@ const imageSizeSelectOptions = computed(() => imageSizeOptions.map(s => ({ value
 
 interface ChatMessage {
 	role: string
+	/** 展示文本：[@图片1] 已换成纯文字标签 */
 	content: string
 	reasoningContent?: string
+	/** 多模态消息发给 API 的完整形态；纯文本消息为空，回传时退化为 content 字符串 */
+	parts?: ContentPart[]
+	/** 气泡里的附件缩略图 */
+	previews?: AttachmentPreview[]
 }
 const messages = ref<ChatMessage[]>([])
 const inputMessage = ref('')
 const tokenUsage = reactive({ prompt: 0, completion: 0, reasoning: 0, total: 0, cost: '' })
 
+// ── 附件 ──
+
+// capabilities 是 admin 侧存的 JSON 对象字符串，解析失败按「无能力」处理
+const modelCapabilities = computed<Record<string, boolean>>(() => {
+	const raw = selectedModelItem.value?.capabilities
+	if (!raw) return {}
+	try {
+		const parsed = JSON.parse(raw)
+		return parsed && typeof parsed === 'object' ? parsed : {}
+	} catch {
+		return {}
+	}
+})
+
+// 对话只收图片与音频：ContentPart.File 虽在 DTO 里，但各协议转换器普遍不处理，
+// 传文档会被静默丢弃还照样计费。类型随模型能力动态收放。
+const attachmentKinds = computed<AttachmentKind[]>(() => {
+	const kinds: AttachmentKind[] = []
+	if (modelCapabilities.value.vision) kinds.push('image')
+	if (modelCapabilities.value.audio_input) kinds.push('audio')
+	return kinds
+})
+
+const {
+	list: attachmentList,
+	totalBytes: attachmentBytes,
+	accept: acceptTypes,
+	max: attachmentMax,
+	addFiles: addAttachments,
+	remove: removeAttachment,
+	clear: clearAttachments,
+} = useAttachments({ kinds: () => attachmentKinds.value, max: 9 })
+
+const promptInput = ref<InstanceType<typeof PromptInput> | null>(null)
+const attachmentBar = ref<InstanceType<typeof AttachmentBar> | null>(null)
+// 输入容器的拖拽高亮状态（附件栏在 bare 变体下不再自带拖放区，统一由输入容器承接）
+const dragging = ref(false)
+
+// 空态快捷提示词：点一下直接填进输入框，减少首次使用的空白焦虑
+const quickPrompts = [
+	'用三句话介绍你能做什么',
+	'写一个 Python 快速排序函数',
+	'把下面这段话翻译成英文：',
+]
+
+// 非 vision / 非音频输入的模型禁用上传：传了也是白花 token 再拿上游报错
+const uploadDisabled = computed(() => attachmentKinds.value.length === 0)
+const uploadDisabledReason = computed(() =>
+	uploadDisabled.value ? '该模型不支持图片/音频输入' : '',
+)
+
+// 历史消息里的 base64 每轮都会重新发送，多轮下很容易撞后端 8MB 的 body 上限，
+// 这里把「历史 + 待发」的合计体积摆到台面上，并在超限时拦住发送。
+const historyAttachmentBytes = computed(() =>
+	messages.value.reduce((sum, m) => {
+		if (!m.parts) return sum
+		return sum + m.parts.reduce((s, p) => {
+			if (p.type === 'image_url') return s + (p.image_url?.url.length || 0)
+			if (p.type === 'input_audio') return s + (p.input_audio?.data.length || 0)
+			return s
+		}, 0)
+	}, 0),
+)
+const pendingBytes = computed(() => historyAttachmentBytes.value + attachmentBytes.value)
+const overBudget = computed(() => pendingBytes.value > MAX_TOTAL_BYTES)
+
+function insertMention(name: string) {
+	promptInput.value?.insertMention(name)
+}
+
+// 粘贴图片走同一条上传链路；模型不支持时给出明确原因，而非落到「不支持的文件类型」
+function onPasteFiles(files: File[]) {
+	if (uploadDisabled.value) {
+		toast.warning(uploadDisabledReason.value)
+		return
+	}
+	addAttachments(files)
+}
+
+// 输入容器承接拖放：与粘贴共用同一校验与上传链路
+function onDropFiles(e: DragEvent) {
+	dragging.value = false
+	if (uploadDisabled.value) {
+		toast.warning(uploadDisabledReason.value)
+		return
+	}
+	if (e.dataTransfer?.files?.length) addAttachments(Array.from(e.dataTransfer.files))
+}
+
+function openFilePicker() {
+	if (uploadDisabled.value) {
+		toast.warning(uploadDisabledReason.value)
+		return
+	}
+	attachmentBar.value?.pick()
+}
+
+// 一键清空附件（模型切换导致能力不匹配时尤其有用）
+function clearAllAttachments() {
+	clearAttachments()
+}
+
+// Temperature 滑块的已填充比例，用于把轨道左段染成主色
+const rangePercent = computed(() => `${(params.temperature / 2) * 100}%`)
+
+// 恢复参数默认值：只重置生成参数，不动已选模型与对话内容
+function resetParams() {
+	params.temperature = 0.7
+	params.maxTokens = 2048
+	params.systemPrompt = ''
+	params.thinking = false
+	params.stream = true
+	imageConfig.aspectRatio = '1:1'
+	imageConfig.imageSize = '1K'
+	toast.success('参数已恢复默认值')
+}
+
 // 记录每条 assistant 消息的思考内容折叠状态（用 message index 作为 key）
 const expandedThinking = ref<Record<number, boolean>>({})
 
@@ -94,7 +223,7 @@ function scrollToBottom(force = false) {
 	})
 }
 
-function buildRequestBody(apiMessages: { role: string; content: string }[]) {
+function buildRequestBody(apiMessages: { role: string; content: string | ContentPart[] }[]) {
 	const body: Record<string, any> = {
 		model: selectedModel.value,
 		messages: apiMessages,
@@ -244,22 +373,38 @@ async function sendStream(requestBody: Record<string, any>, assistantIdx: number
 }
 
 async function sendMessage() {
-	const content = inputMessage.value.trim()
-	if (!content || !selectedModel.value || sending.value) return
+	const raw = inputMessage.value.trim()
+	if ((!raw && !attachmentList.value.length) || !selectedModel.value || sending.value) return
+
+	if (overBudget.value) {
+		toast.error(`附件合计 ${formatBytes(pendingBytes.value)} 已超过 ${formatBytes(MAX_TOTAL_BYTES)}，请清空对话或删除部分附件`)
+		return
+	}
 
 	if (params.systemPrompt && messages.value.length === 0) {
 		messages.value.push({ role: 'system', content: params.systemPrompt })
 	}
 
-	messages.value.push({ role: 'user', content })
+	// 正文里的 [@图片1] 换成文字标签，并在同一位置插入对应的 content part；
+	// 未被引用的附件追加到末尾。纯文本消息不带 parts，回传时退化为字符串。
+	const compiled = compileChatParts(raw, attachmentList.value)
+	messages.value.push({
+		role: 'user',
+		content: compiled.displayText,
+		parts: attachmentList.value.length ? compiled.parts : undefined,
+		previews: attachmentList.value.length ? compiled.previews : undefined,
+	})
 	inputMessage.value = ''
+	clearAttachments()
 	sending.value = true
 
 	const assistantIdx = messages.value.length
 	messages.value.push({ role: 'assistant', content: '', reasoningContent: '' })
 	scrollToBottom()
 
-	const apiMessages = messages.value.slice(0, assistantIdx).map(m => ({ role: m.role, content: m.content }))
+	const apiMessages = messages.value
+		.slice(0, assistantIdx)
+		.map(m => ({ role: m.role, content: (m.parts ?? m.content) as string | ContentPart[] }))
 	const requestBody = buildRequestBody(apiMessages)
 
 	try {
@@ -279,6 +424,7 @@ async function sendMessage() {
 
 function clearChat() {
 	messages.value = []
+	clearAttachments()
 	tokenUsage.prompt = 0
 	tokenUsage.completion = 0
 	tokenUsage.reasoning = 0
@@ -289,155 +435,313 @@ function clearChat() {
 </script>
 
 <template>
-	<!-- 高度偏移 ≈ 顶栏与页边距(约128px) + 上方工具栏卡片(约116px)，保证对话卡片完整落在视口内 -->
-	<div class="flex flex-col lg:h-[calc(100vh-15.5rem)] lg:overflow-hidden">
-		<!-- grid-rows minmax(0,1fr) 强制行高等于容器高，配合左列内部滚动，避免行高被参数卡撑破导致底部输入框被裁切 -->
-		<div class="grid grid-cols-1 lg:grid-cols-4 lg:grid-rows-[minmax(0,1fr)] gap-6 lg:min-h-0 lg:flex-1">
+	<!-- 高度偏移 ≈ 顶栏与页边距 + 上方工具栏卡片，保证对话卡片完整落在视口内 -->
+	<div class="flex flex-col lg:h-[calc(100vh-15rem)] lg:overflow-hidden">
+		<!-- grid-rows minmax(0,1fr) 强制行高等于容器高，左列卡片内部滚动，
+		     使参数面板与对话区等高对齐，不再出现左栏半截卡片下的大片空白。
+		     左列 22rem：参数区控件较多（模型下拉、Max Tokens、System Prompt、两个开关），
+		     19rem 下 System Prompt 与下拉框偏窄，加宽后可读性更好 -->
+		<div class="grid grid-cols-1 gap-4 lg:grid-cols-[22rem_minmax(0,1fr)] lg:grid-rows-[minmax(0,1fr)] lg:min-h-0 lg:flex-1">
 			<!-- Left: Parameters -->
-			<div class="lg:col-span-1 lg:min-h-0 lg:overflow-y-auto lg:pr-1">
-				<div class="card">
-					<div class="card-header">
-						<h3 class="text-sm font-semibold text-gray-900">参数配置</h3>
+			<div class="flex lg:col-span-1 lg:min-h-0">
+				<div class="card flex w-full flex-col overflow-hidden lg:h-full">
+					<div class="flex shrink-0 items-center justify-between border-b border-gray-100 px-4 py-3">
+						<div class="flex items-center gap-2">
+							<Icon name="cog" size="sm" class="text-primary-500" />
+							<h3 class="text-sm font-semibold text-gray-900">参数配置</h3>
+						</div>
+						<button
+							class="flex items-center gap-1 rounded-lg px-2 py-1 text-xs text-gray-400 transition-colors duration-150 hover:bg-gray-50 hover:text-primary-600"
+							title="恢复默认参数"
+							@click="resetParams"
+						>
+							<Icon name="refresh" size="xs" />
+							重置
+						</button>
 					</div>
-					<div class="card-body space-y-4">
+
+					<div class="flex-1 space-y-4 overflow-y-auto px-4 py-4">
+						<!-- 模型 -->
 						<div>
-							<label class="input-label">模型</label>
+							<div class="mb-2 flex items-center justify-between">
+								<label class="text-xs font-semibold text-gray-500">模型</label>
+								<span class="text-[11px] text-gray-400">共 {{ models.length }} 个可用</span>
+							</div>
 							<BaseSelect v-model="selectedModel" :options="modelOptions" />
-							<div v-if="isImageModel" class="mt-1.5 flex items-center gap-1.5">
+							<div v-if="isImageModel" class="mt-2 flex items-center gap-1.5">
 								<span class="badge badge-primary">图片模型</span>
 								<span class="text-xs text-gray-500">支持生成图片</span>
 							</div>
 						</div>
-						<div>
-							<label class="input-label">Temperature: {{ params.temperature.toFixed(1) }}</label>
-							<input v-model.number="params.temperature" type="range" min="0" max="2" step="0.1" class="w-full" />
+
+						<!-- 生成参数 -->
+						<div class="space-y-4 border-t border-gray-100 pt-4">
+							<div>
+								<div class="mb-2 flex items-center justify-between">
+									<label class="text-xs font-semibold text-gray-500">Temperature</label>
+									<span class="rounded-md bg-primary-50 px-1.5 py-0.5 font-mono text-xs font-semibold text-primary-600">
+										{{ params.temperature.toFixed(1) }}
+									</span>
+								</div>
+								<input
+									v-model.number="params.temperature"
+									type="range"
+									min="0"
+									max="2"
+									step="0.1"
+									class="range-slider"
+									:style="{ '--pct': rangePercent }"
+								/>
+								<div class="mt-1.5 flex justify-between text-[11px] text-gray-400">
+									<span>严谨</span><span>均衡</span><span>发散</span>
+								</div>
+							</div>
+
+							<div>
+								<label class="mb-2 block text-xs font-semibold text-gray-500">Max Tokens</label>
+								<n-input-number v-model:value="params.maxTokens" :min="1" :max="128000" class="w-full" />
+							</div>
 						</div>
-						<div>
-							<label class="input-label">Max Tokens</label>
-							<n-input-number v-model:value="params.maxTokens" :min="1" :max="128000" class="w-full" />
-						</div>
-						<div>
-							<label class="input-label">System Prompt</label>
+
+						<!-- 系统提示词 -->
+						<div class="border-t border-gray-100 pt-4">
+							<label class="mb-2 block text-xs font-semibold text-gray-500">System Prompt</label>
 							<n-input v-model:value="params.systemPrompt" type="textarea" :rows="3" placeholder="设置系统提示词..." />
 						</div>
-						<div v-if="!isImageModel">
-							<div class="flex items-center justify-between">
-								<label class="input-label mb-0">深度思考</label>
-								<n-switch v-model:value="params.thinking" />
+
+						<!-- 开关 -->
+						<div class="space-y-3 border-t border-gray-100 pt-4">
+							<div v-if="!isImageModel" class="flex items-start justify-between gap-3">
+								<div class="min-w-0">
+									<p class="text-sm font-medium text-gray-700">深度思考</p>
+									<p class="mt-0.5 text-[11px] leading-relaxed text-gray-400">thinking 参数控制，仅部分模型支持</p>
+								</div>
+								<n-switch v-model:value="params.thinking" size="small" class="shrink-0" />
 							</div>
-							<p class="mt-1 text-xs text-gray-400">通过 thinking 参数控制开关，仅对支持的模型生效</p>
-						</div>
-						<div class="flex items-center justify-between">
-							<label class="input-label mb-0">流式响应</label>
-							<n-switch v-model:value="params.stream" />
-						</div>
-						<template v-if="isImageModel">
-							<div class="border-t border-gray-100 pt-4">
-								<h4 class="text-xs font-semibold text-gray-500 uppercase tracking-wider mb-3">图片参数</h4>
+							<div class="flex items-start justify-between gap-3">
+								<div class="min-w-0">
+									<p class="text-sm font-medium text-gray-700">流式响应</p>
+									<p class="mt-0.5 text-[11px] leading-relaxed text-gray-400">逐字返回，实时看到输出过程</p>
+								</div>
+								<n-switch v-model:value="params.stream" size="small" class="shrink-0" />
 							</div>
+						</div>
+
+						<!-- 图片模型专属参数 -->
+						<div v-if="isImageModel" class="space-y-4 border-t border-gray-100 pt-4">
+							<p class="text-xs font-semibold text-gray-500">图片参数</p>
 							<div>
-								<label class="input-label">宽高比</label>
+								<label class="mb-2 block text-xs font-medium text-gray-500">宽高比</label>
 								<BaseSelect v-model="imageConfig.aspectRatio" :options="aspectRatioSelectOptions" />
 							</div>
 							<div>
-								<label class="input-label">分辨率</label>
+								<label class="mb-2 block text-xs font-medium text-gray-500">分辨率</label>
 								<BaseSelect v-model="imageConfig.imageSize" :options="imageSizeSelectOptions" />
 							</div>
-						</template>
+						</div>
 					</div>
 				</div>
 			</div>
 
 			<!-- Right: Chat -->
-			<div class="h-[65vh] lg:col-span-3 lg:h-auto lg:min-h-0">
-				<div class="card flex flex-col h-full overflow-hidden">
-					<div class="px-6 py-2 bg-amber-50 border-b border-amber-200 text-center">
-						<span class="text-xs text-amber-700 font-medium">Playground 模式 — 使用真实 API Key 调用，产生实际费用</span>
-					</div>
-					<div ref="messagesRef" class="flex-1 overflow-y-auto min-h-0 p-6 space-y-4">
-						<div v-if="messages.length === 0" class="empty-state">
-							<div class="empty-state-icon"><Icon name="bookOpen" size="xl" /></div>
-							<h3 class="empty-state-title">开始对话</h3>
-							<p class="empty-state-description">选择模型，输入消息，开始测试</p>
+			<div class="h-[62vh] lg:h-auto lg:min-h-0">
+				<div class="card flex h-full flex-col overflow-hidden">
+					<!-- 消息区 -->
+					<div ref="messagesRef" class="min-h-0 flex-1 overflow-y-auto px-5 py-5 lg:px-6">
+						<!-- 空态：图标磁贴 + 快捷提示词，给首屏一个视觉锚点 -->
+						<div v-if="messages.length === 0" class="flex min-h-full flex-col items-center justify-center px-4 text-center">
+							<div class="mb-5 flex h-16 w-16 items-center justify-center rounded-3xl bg-gradient-to-br from-cyan-400 via-primary-500 to-emerald-500 text-white shadow-glow">
+								<Icon name="chat" size="lg" />
+							</div>
+							<h3 class="text-lg font-semibold text-gray-900">开始对话</h3>
+							<p class="mt-1.5 max-w-sm text-sm text-gray-500">
+								<template v-if="selectedModelItem">
+									当前模型 <span class="font-medium text-gray-700">{{ selectedModelItem.model_name || selectedModelItem.model_id }}</span>，输入消息即可测试
+								</template>
+								<template v-else>请先在左侧选择一个模型</template>
+							</p>
+
+							<!-- 快捷提示词：点一下即填入输入框 -->
+							<div v-if="selectedModel" class="mt-6 flex max-w-lg flex-wrap justify-center gap-2">
+								<button
+									v-for="p in quickPrompts"
+									:key="p"
+									class="rounded-full border border-gray-200 bg-white px-3 py-1.5 text-xs text-gray-600 transition-all duration-150 hover:-translate-y-0.5 hover:border-primary-300 hover:text-primary-600 hover:shadow-sm"
+									@click="inputMessage = p; promptInput?.focus()"
+								>
+									{{ p }}
+								</button>
+							</div>
+
+							<p class="mt-7 inline-flex items-center gap-1.5 rounded-full border border-amber-200 bg-amber-50 px-3 py-1 text-[11px] text-amber-700">
+								<Icon name="exclamationTriangle" size="xs" />
+								发送将真实调用模型并产生费用
+							</p>
 						</div>
-						<div v-for="(msg, idx) in messages" :key="idx" class="flex gap-3 animate-slide-up" :class="{ 'justify-end': msg.role === 'user' }">
-							<!-- Assistant message -->
-							<div v-if="msg.role === 'assistant'" class="flex gap-3 max-w-[85%]">
-								<div class="h-8 w-8 rounded-xl bg-primary-100 flex items-center justify-center flex-shrink-0">
-									<Icon name="cog" size="sm" class="text-primary-600" />
-								</div>
-								<div class="min-w-0 flex-1 space-y-2">
-									<!-- Thinking / Reasoning section -->
-									<div v-if="msg.reasoningContent" class="rounded-xl border border-amber-200 bg-amber-50/60 overflow-hidden">
-										<button
-											class="w-full flex items-center gap-2 px-3 py-2 text-left hover:bg-amber-100/60 transition-colors duration-150"
-											@click="toggleThinking(idx)"
-										>
-											<svg
-												class="h-3.5 w-3.5 text-amber-500 transition-transform duration-200 flex-shrink-0"
-												:class="{ 'rotate-90': expandedThinking[idx] }"
-												fill="none" viewBox="0 0 24 24" stroke="currentColor" stroke-width="2"
+
+						<!-- 消息列表 -->
+						<div v-else class="space-y-5">
+							<div v-for="(msg, idx) in messages" :key="idx" class="flex gap-3 animate-slide-up" :class="{ 'justify-end': msg.role === 'user' }">
+								<!-- Assistant message -->
+								<div v-if="msg.role === 'assistant'" class="flex max-w-[85%] gap-3">
+									<div class="mt-0.5 flex h-8 w-8 shrink-0 items-center justify-center rounded-xl bg-gradient-to-br from-primary-400 to-primary-600 text-white shadow-sm shadow-primary-500/25">
+										<Icon name="cog" size="sm" />
+									</div>
+									<div class="min-w-0 flex-1 space-y-2">
+										<!-- Thinking / Reasoning section -->
+										<div v-if="msg.reasoningContent" class="overflow-hidden rounded-xl border border-amber-200 bg-amber-50/60">
+											<button
+												class="flex w-full items-center gap-2 px-3 py-2 text-left transition-colors duration-150 hover:bg-amber-100/60"
+												@click="toggleThinking(idx)"
 											>
-												<path stroke-linecap="round" stroke-linejoin="round" d="M9 5l7 7-7 7" />
-											</svg>
-											<span class="text-xs font-medium text-amber-700">思考过程</span>
-											<span class="text-xs text-amber-500">· 消耗 Reasoning Tokens</span>
-										</button>
-										<div v-if="expandedThinking[idx]" class="px-3 pb-3">
-											<div class="rounded-lg bg-white/70 border border-amber-100 px-3 py-2 text-sm text-gray-600 whitespace-pre-wrap leading-relaxed max-h-80 overflow-y-auto">
-												{{ msg.reasoningContent }}
+												<svg
+													class="h-3.5 w-3.5 shrink-0 text-amber-500 transition-transform duration-200"
+													:class="{ 'rotate-90': expandedThinking[idx] }"
+													fill="none" viewBox="0 0 24 24" stroke="currentColor" stroke-width="2"
+												>
+													<path stroke-linecap="round" stroke-linejoin="round" d="M9 5l7 7-7 7" />
+												</svg>
+												<span class="text-xs font-medium text-amber-700">思考过程</span>
+												<span class="text-xs text-amber-500">· 消耗 Reasoning Tokens</span>
+											</button>
+											<div v-if="expandedThinking[idx]" class="px-3 pb-3">
+												<div class="max-h-80 overflow-y-auto rounded-lg border border-amber-100 bg-white/70 px-3 py-2 text-sm leading-relaxed text-gray-600 whitespace-pre-wrap">
+													{{ msg.reasoningContent }}
+												</div>
+											</div>
+										</div>
+										<!-- Main content -->
+										<div class="rounded-2xl rounded-tl-md border border-gray-200 bg-white px-4 py-3 shadow-sm">
+											<MarkdownRenderer :content="msg.content" />
+										</div>
+									</div>
+								</div>
+								<!-- User message -->
+								<div v-else-if="msg.role === 'user'" class="max-w-[80%] space-y-1.5">
+									<!-- 随消息发出的附件：用 data URL 展示（输入区的 objectURL 发送后即被释放） -->
+									<div v-if="msg.previews?.length" class="flex flex-wrap justify-end gap-1.5">
+										<div
+											v-for="p in msg.previews"
+											:key="p.name"
+											class="overflow-hidden rounded-xl border border-primary-200 bg-white"
+											:title="p.name"
+										>
+											<img v-if="p.kind === 'image'" :src="p.url" class="h-20 w-20 object-cover" alt="" />
+											<div v-else class="flex h-20 w-20 flex-col items-center justify-center gap-1 text-gray-500">
+												<Icon :name="p.kind === 'audio' ? 'musicalNote' : 'film'" size="md" />
+												<span class="text-[11px]">{{ p.name }}</span>
 											</div>
 										</div>
 									</div>
-									<!-- Main content -->
-									<div class="rounded-2xl rounded-tl-sm bg-gray-100 px-4 py-3">
-										<MarkdownRenderer :content="msg.content" />
+									<div v-if="msg.content" class="rounded-2xl rounded-tr-md bg-gradient-to-br from-primary-500 to-primary-600 px-4 py-3 text-white shadow-sm shadow-primary-500/20">
+										<p class="text-sm whitespace-pre-wrap">{{ msg.content }}</p>
 									</div>
 								</div>
-							</div>
-							<!-- User message -->
-							<div v-else-if="msg.role === 'user'" class="max-w-[80%]">
-								<div class="rounded-2xl rounded-tr-sm bg-primary-500 text-white px-4 py-3">
-									<p class="text-sm whitespace-pre-wrap">{{ msg.content }}</p>
+								<!-- System message -->
+								<div v-else-if="msg.role === 'system' && idx === 0" class="text-xs text-gray-400 italic">
+									System: {{ msg.content.substring(0, 50) }}{{ msg.content.length > 50 ? '...' : '' }}
 								</div>
 							</div>
-							<!-- System message -->
-							<div v-else-if="msg.role === 'system' && idx === 0" class="text-xs text-gray-400 italic">
-								System: {{ msg.content.substring(0, 50) }}{{ msg.content.length > 50 ? '...' : '' }}
-							</div>
-						</div>
-						<!-- Sending indicator -->
-						<div v-if="sending && (params.stream ? (messages.length > 0 && messages[messages.length - 1].role === 'user') : true)" class="flex gap-3">
-							<div class="h-8 w-8 rounded-xl bg-primary-100 flex items-center justify-center">
-								<div class="spinner h-4 w-4 text-primary-600"></div>
-							</div>
-							<div class="rounded-2xl rounded-tl-sm bg-gray-100 px-4 py-3">
-								<p class="text-sm text-gray-500">请求中...</p>
+							<!-- Sending indicator -->
+							<div v-if="sending && (params.stream ? (messages.length > 0 && messages[messages.length - 1].role === 'user') : true)" class="flex gap-3">
+								<div class="flex h-8 w-8 items-center justify-center rounded-xl bg-gradient-to-br from-primary-400 to-primary-600">
+									<div class="spinner h-4 w-4 text-white"></div>
+								</div>
+								<div class="rounded-2xl rounded-tl-md border border-gray-200 bg-white px-4 py-3 shadow-sm">
+									<p class="text-sm text-gray-500">请求中...</p>
+								</div>
 							</div>
 						</div>
 					</div>
+
 					<!-- Token usage bar -->
-					<div v-if="tokenUsage.total > 0 || tokenUsage.cost" class="px-6 py-2 border-t border-gray-100 flex items-center justify-between text-xs text-gray-500">
+					<div v-if="tokenUsage.total > 0 || tokenUsage.cost || pendingBytes > 0" class="flex items-center justify-between gap-3 border-t border-gray-100 bg-gray-50/60 px-5 py-2 text-xs text-gray-500 lg:px-6">
 						<div class="flex items-center gap-4">
-							<span>Prompt: {{ tokenUsage.prompt }}</span>
-							<span>Completion: {{ tokenUsage.completion }}</span>
-							<span v-if="tokenUsage.reasoning > 0" class="text-amber-600">Reasoning: {{ tokenUsage.reasoning }}</span>
-							<span>Total: {{ tokenUsage.total }}</span>
+							<span class="flex items-center gap-1.5">Prompt <span class="font-medium text-gray-700">{{ tokenUsage.prompt }}</span></span>
+							<span class="flex items-center gap-1.5">Completion <span class="font-medium text-gray-700">{{ tokenUsage.completion }}</span></span>
+							<span v-if="tokenUsage.reasoning > 0" class="flex items-center gap-1.5 text-amber-600">Reasoning <span class="font-medium">{{ tokenUsage.reasoning }}</span></span>
+							<span class="flex items-center gap-1.5">Total <span class="font-medium text-gray-700">{{ tokenUsage.total }}</span></span>
+						</div>
+						<div class="flex items-center gap-4">
+							<!-- 历史附件每轮都会重发，体积直接决定还能不能继续对话 -->
+							<span v-if="pendingBytes > 0" :class="overBudget ? 'font-medium text-red-600' : ''">
+								附件 {{ formatBytes(pendingBytes) }} / {{ formatBytes(MAX_TOTAL_BYTES) }}
+							</span>
+							<span v-if="tokenUsage.cost" class="font-medium text-amber-600">{{ tokenUsage.cost }}</span>
 						</div>
-						<span v-if="tokenUsage.cost" class="font-medium text-amber-600">{{ tokenUsage.cost }}</span>
 					</div>
-					<!-- Input area -->
-					<div class="p-4 border-t border-gray-200">
-						<div class="flex gap-3">
-							<n-input v-model:value="inputMessage" autosize placeholder="输入消息..." class="flex-1" @keydown.enter.exact.prevent="sendMessage" />
-							<button class="btn btn-primary self-end" :disabled="sending || !inputMessage.trim()" @click="sendMessage">
-								<Icon name="edit" size="sm" />
-								发送
-							</button>
-							<button class="btn btn-secondary self-end" :disabled="sending" @click="clearChat">
-								<Icon name="refresh" size="sm" />
-								清空
-							</button>
+
+					<!-- 输入区：附件片 + 输入框 + 操作按钮整合进同一个圆角盒子，聚焦时整块高亮 -->
+					<div class="shrink-0 border-t border-gray-100 p-3 lg:p-4">
+						<div
+							class="input-shell"
+							:class="{ 'input-shell-dragging': dragging }"
+							@dragover.prevent="dragging = true"
+							@dragleave="dragging = false"
+							@drop.prevent="onDropFiles"
+						>
+							<!-- 附件缩略片（bare：无边框无上传按钮，融入本容器）；
+							     文件选择器仍挂载于此，由下方「附件」按钮通过 ref 触发 -->
+							<AttachmentBar
+								ref="attachmentBar"
+								variant="bare"
+								:class="attachmentList.length ? 'mb-2' : ''"
+								:attachments="attachmentList"
+								:accept="acceptTypes"
+								:max="attachmentMax"
+								:disabled="uploadDisabled"
+								:disabled-reason="uploadDisabledReason"
+								@add="addAttachments"
+								@remove="removeAttachment"
+								@insert="insertMention"
+							/>
+
+							<PromptInput
+								ref="promptInput"
+								v-model="inputMessage"
+								bare
+								:attachments="attachmentList"
+								placeholder="输入消息，输入 @ 引用已上传的附件..."
+								@submit="sendMessage"
+								@paste="onPasteFiles"
+							/>
+
+							<!-- 操作行 -->
+							<div class="mt-1.5 flex items-center justify-between gap-3">
+								<div class="flex min-w-0 items-center gap-1">
+									<button
+										class="input-tool"
+										:disabled="uploadDisabled"
+										:title="uploadDisabled ? uploadDisabledReason : '上传附件（也可拖拽到此处或直接粘贴）'"
+										@click="openFilePicker"
+									>
+										<Icon name="paperclip" size="xs" />
+										附件
+									</button>
+									<button
+										v-if="attachmentList.length"
+										class="input-tool"
+										title="清空已选附件"
+										@click="clearAllAttachments"
+									>
+										<Icon name="x" size="xs" />
+										清空附件
+									</button>
+									<span class="ml-1 hidden truncate text-[11px] text-gray-400 sm:inline">
+										Enter 发送 · Shift + Enter 换行
+									</span>
+								</div>
+								<div class="flex shrink-0 items-center gap-2">
+									<button class="btn btn-ghost btn-sm text-gray-500" :disabled="sending" @click="clearChat">
+										<Icon name="refresh" size="xs" />
+										清空对话
+									</button>
+									<button class="btn btn-primary btn-sm px-4" :disabled="sending || (!inputMessage.trim() && !attachmentList.length)" @click="sendMessage">
+										<Icon name="edit" size="xs" />
+										发送
+									</button>
+								</div>
+							</div>
 						</div>
 					</div>
 				</div>
@@ -445,3 +749,93 @@ function clearChat() {
 		</div>
 	</div>
 </template>
+
+<style scoped>
+/* ── 输入容器：把附件片、输入框、操作按钮收进同一个圆角盒子 ── */
+.input-shell {
+	border: 1px solid #e5e7eb;
+	border-radius: 1rem;
+	background: #fff;
+	padding: 0.625rem 0.75rem 0.5rem;
+	transition: border-color 180ms ease, box-shadow 180ms ease, background-color 180ms ease;
+}
+/* 聚焦与拖拽时整块高亮，而不是只高亮里面的 textarea */
+.input-shell:focus-within {
+	border-color: rgba(20, 184, 166, 0.55);
+	box-shadow: 0 0 0 4px rgba(20, 184, 166, 0.09);
+}
+.input-shell-dragging {
+	border-color: #2dd4bf;
+	border-style: dashed;
+	background: rgba(240, 253, 250, 0.6);
+}
+
+/* ── 输入区内的小工具按钮 ── */
+.input-tool {
+	display: inline-flex;
+	align-items: center;
+	gap: 0.25rem;
+	border-radius: 0.5rem;
+	padding: 0.25rem 0.5rem;
+	color: #6b7280;
+	font-size: 0.75rem;
+	line-height: 1.25rem;
+	transition: background-color 150ms ease, color 150ms ease;
+}
+.input-tool:hover:not(:disabled) {
+	background: #f3f4f6;
+	color: #0d9488;
+}
+.input-tool:disabled {
+	cursor: not-allowed;
+	color: #d1d5db;
+}
+
+/* ── Temperature 滑块：替换浏览器默认的紫色原生样式 ── */
+.range-slider {
+	-webkit-appearance: none;
+	appearance: none;
+	width: 100%;
+	height: 6px;
+	border-radius: 9999px;
+	outline: none;
+	cursor: pointer;
+	background: linear-gradient(
+		to right,
+		#14b8a6 0%,
+		#14b8a6 var(--pct, 35%),
+		#e2e8f0 var(--pct, 35%),
+		#e2e8f0 100%
+	);
+}
+.range-slider::-webkit-slider-thumb {
+	-webkit-appearance: none;
+	appearance: none;
+	height: 16px;
+	width: 16px;
+	border: 2px solid #14b8a6;
+	border-radius: 9999px;
+	background: #fff;
+	box-shadow: 0 2px 6px rgba(13, 148, 136, 0.35);
+	transition: transform 150ms ease;
+}
+.range-slider::-webkit-slider-thumb:hover {
+	transform: scale(1.15);
+}
+.range-slider::-moz-range-thumb {
+	height: 14px;
+	width: 14px;
+	border: 2px solid #14b8a6;
+	border-radius: 9999px;
+	background: #fff;
+	box-shadow: 0 2px 6px rgba(13, 148, 136, 0.35);
+}
+.range-slider::-moz-range-track {
+	height: 6px;
+	border-radius: 9999px;
+	background: #e2e8f0;
+}
+.range-slider:focus-visible {
+	box-shadow: 0 0 0 4px rgba(20, 184, 166, 0.15);
+}
+</style>
diff --git a/web/tenant/src/views/service/playground/PromptInput.vue b/web/tenant/src/views/service/playground/PromptInput.vue
new file mode 100644
index 0000000..c121f9c
--- /dev/null
+++ b/web/tenant/src/views/service/playground/PromptInput.vue
@@ -0,0 +1,226 @@
+<script setup lang="ts">
+import { ref, computed, nextTick, watch } from 'vue'
+import Icon from '@/components/common/Icon.vue'
+import type { Attachment } from './useAttachments'
+
+/**
+ * 带 @ 提及的提示词输入框。
+ *
+ * 用原生 textarea 而非 n-input：插入标记要精确读写 selectionStart/selectionEnd，
+ * 隔着组件库拿内部 DOM 既脆又绕。autosize 与样式在此自行实现，视觉对齐 .input。
+ */
+
+const props = withDefaults(defineProps<{
+	modelValue: string
+	attachments: Attachment[]
+	/** 关闭后不响应 @，输入框退化为普通 textarea */
+	mentions?: boolean
+	placeholder?: string
+	minRows?: number
+	maxRows?: number
+	disabled?: boolean
+	/** 是否接受粘贴图片（对话/视频都开，纯文本场景可关） */
+	pasteFiles?: boolean
+	/**
+	 * 去掉 textarea 自身的边框与内边距，交由父容器的边框统一承载
+	 * （用于把附件栏 + 输入框 + 操作按钮整合成一个视觉整体的输入容器）
+	 */
+	bare?: boolean
+}>(), {
+	mentions: true,
+	placeholder: '',
+	minRows: 2,
+	maxRows: 8,
+	disabled: false,
+	pasteFiles: true,
+	bare: false,
+})
+
+const emit = defineEmits<{
+	'update:modelValue': [value: string]
+	submit: []
+	paste: [files: File[]]
+}>()
+
+const textarea = ref<HTMLTextAreaElement | null>(null)
+const panelOpen = ref(false)
+const query = ref('')
+const activeIndex = ref(0)
+
+/** 触发条件：光标紧邻的 @ 后面跟着不含空白与方括号的片段 */
+const TRIGGER_RE = /@([^\s@[\]]*)$/
+
+const candidates = computed(() => {
+	if (!query.value) return props.attachments
+	const q = query.value.toLowerCase()
+	return props.attachments.filter(
+		a => a.name.toLowerCase().includes(q) || a.fileName.toLowerCase().includes(q),
+	)
+})
+
+function resize() {
+	const el = textarea.value
+	if (!el) return
+	el.style.height = 'auto'
+	const lineHeight = 22
+	const max = props.maxRows * lineHeight + 20
+	el.style.height = `${Math.min(el.scrollHeight, max)}px`
+}
+
+watch(() => props.modelValue, () => nextTick(resize))
+
+// 附件被删光时关掉可能还开着的候选面板
+watch(() => props.attachments.length, len => {
+	if (!len) closePanel()
+})
+
+function closePanel() {
+	panelOpen.value = false
+	query.value = ''
+	activeIndex.value = 0
+}
+
+function syncPanel() {
+	if (!props.mentions) return
+	const el = textarea.value
+	if (!el) return
+	const before = props.modelValue.slice(0, el.selectionStart)
+	const m = TRIGGER_RE.exec(before)
+	if (!m) {
+		closePanel()
+		return
+	}
+	panelOpen.value = true
+	query.value = m[1]
+	activeIndex.value = 0
+}
+
+function onInput(e: Event) {
+	emit('update:modelValue', (e.target as HTMLTextAreaElement).value)
+	nextTick(() => {
+		resize()
+		syncPanel()
+	})
+}
+
+/** insertMention 把光标前的 @片段替换成 [@名称]；无触发片段时在光标处插入 */
+function insertMention(name: string) {
+	const el = textarea.value
+	if (!el) return
+	const token = `[@${name}]`
+	// 从未聚焦过的输入框 selectionStart 恒为 0，插到开头很反直觉，改为追加到末尾
+	const touched = el.selectionStart > 0 || el.selectionEnd > 0 || document.activeElement === el
+	const pos = touched ? el.selectionStart : props.modelValue.length
+	const end = touched ? el.selectionEnd : props.modelValue.length
+	const before = props.modelValue.slice(0, pos)
+	const after = props.modelValue.slice(end)
+	const m = TRIGGER_RE.exec(before)
+	const head = m ? before.slice(0, m.index) : before
+
+	emit('update:modelValue', head + token + after)
+	closePanel()
+	nextTick(() => {
+		const caret = head.length + token.length
+		el.focus()
+		el.setSelectionRange(caret, caret)
+		resize()
+	})
+}
+
+function onKeydown(e: KeyboardEvent) {
+	if (panelOpen.value && candidates.value.length && !e.isComposing) {
+		switch (e.key) {
+			case 'ArrowDown':
+				e.preventDefault()
+				activeIndex.value = (activeIndex.value + 1) % candidates.value.length
+				return
+			case 'ArrowUp':
+				e.preventDefault()
+				activeIndex.value = (activeIndex.value - 1 + candidates.value.length) % candidates.value.length
+				return
+			case 'Enter':
+			case 'Tab':
+				e.preventDefault()
+				insertMention(candidates.value[activeIndex.value].name)
+				return
+			case 'Escape':
+				e.preventDefault()
+				closePanel()
+				return
+		}
+	}
+	// 面板未接管时，Enter（不带修饰键）提交，Shift+Enter 换行
+	if (e.key === 'Enter' && !e.shiftKey && !e.ctrlKey && !e.metaKey && !e.altKey && !e.isComposing) {
+		e.preventDefault()
+		emit('submit')
+	}
+}
+
+function onPaste(e: ClipboardEvent) {
+	if (!props.pasteFiles) return
+	const files = Array.from(e.clipboardData?.files || [])
+	if (!files.length) return
+	e.preventDefault()
+	emit('paste', files)
+}
+
+defineExpose({ insertMention, focus: () => textarea.value?.focus() })
+</script>
+
+<template>
+	<div class="relative">
+		<!-- @ 候选面板：锚在输入框上方（输入框贴底，向上展开不会被视口裁掉） -->
+		<div
+			v-if="panelOpen"
+			class="absolute bottom-full left-0 z-30 mb-2 w-64 overflow-hidden rounded-xl border border-gray-200 bg-white shadow-lg animate-scale-in"
+		>
+			<div class="border-b border-gray-100 px-3 py-1.5 text-xs font-medium text-gray-400">
+				引用附件
+			</div>
+			<template v-if="candidates.length">
+				<button
+					v-for="(att, i) in candidates"
+					:key="att.id"
+					class="flex w-full items-center gap-2 px-3 py-2 text-left transition-colors duration-150"
+					:class="i === activeIndex ? 'bg-primary-50' : 'hover:bg-gray-50'"
+					@mouseenter="activeIndex = i"
+					@mousedown.prevent
+					@click="insertMention(att.name)"
+				>
+					<img
+						v-if="att.kind === 'image'"
+						:src="att.previewUrl"
+						class="h-7 w-7 shrink-0 rounded object-cover"
+						alt=""
+					/>
+					<span v-else class="flex h-7 w-7 shrink-0 items-center justify-center rounded bg-gray-100 text-gray-500">
+						<Icon :name="att.kind === 'audio' ? 'musicalNote' : 'film'" size="xs" />
+					</span>
+					<span class="min-w-0 flex-1">
+						<span class="block text-sm text-gray-700">{{ att.name }}</span>
+						<span class="block truncate text-xs text-gray-400">{{ att.fileName }}</span>
+					</span>
+				</button>
+			</template>
+			<div v-else class="px-3 py-3 text-xs text-gray-400">
+				{{ attachments.length ? '没有匹配的附件' : '还没有附件，先在上方上传' }}
+			</div>
+		</div>
+
+		<textarea
+			ref="textarea"
+			:value="modelValue"
+			:placeholder="placeholder"
+			:rows="minRows"
+			:disabled="disabled"
+			:class="bare
+				? 'w-full resize-none border-0 bg-transparent px-1 py-0.5 text-sm leading-[22px] text-gray-900 placeholder:text-gray-400 focus:outline-none disabled:cursor-not-allowed disabled:text-gray-400'
+				: 'input resize-none leading-[22px]'"
+			@input="onInput"
+			@keydown="onKeydown"
+			@paste="onPaste"
+			@click="syncPanel"
+			@blur="closePanel"
+		/>
+	</div>
+</template>
diff --git a/web/tenant/src/views/service/playground/VideoTab.vue b/web/tenant/src/views/service/playground/VideoTab.vue
index a049c1f..e19cae5 100644
--- a/web/tenant/src/views/service/playground/VideoTab.vue
+++ b/web/tenant/src/views/service/playground/VideoTab.vue
@@ -6,6 +6,10 @@ import { createPoller } from '@/composables/usePolling'
 import Icon from '@/components/common/Icon.vue'
 import GenerationProgressCard from './GenerationProgressCard.vue'
 import BaseSelect from '../../../components/common/BaseSelect.vue'
+import AttachmentBar from './AttachmentBar.vue'
+import PromptInput from './PromptInput.vue'
+import { useAttachments } from './useAttachments'
+import { compileVideoPrompt } from './attachmentParts'
 
 interface ModelItem {
 	model_id: string
@@ -30,6 +34,25 @@ const prompt = ref('')
 const resolution = ref('1280x720')
 const duration = ref(5)
 
+// ── 参考素材 ──
+// 前端按多素材形态实现（图/视/音皆可上传、@ 列表按类型命名），但本期只有首帧参考图
+// 能随请求发出：/v1/videos 的 input_reference 仅承载一张图，其余素材待后端扩协议。
+const {
+	list: attachmentList,
+	accept: acceptTypes,
+	max: attachmentMax,
+	addFiles: addAttachments,
+	remove: removeAttachment,
+} = useAttachments({ kinds: ['image', 'video', 'audio'], max: 6 })
+
+const promptInput = ref<InstanceType<typeof PromptInput> | null>(null)
+// 提交时未能随请求发出的素材名，用于结果区提示
+const deferredNames = ref<string[]>([])
+
+function insertMention(name: string) {
+	promptInput.value?.insertMention(name)
+}
+
 const modelOptions = computed(() => props.models.map(m => ({ value: m.model_id, label: m.model_name || m.model_id })))
 
 // MiniMax-H3 系列使用官方分辨率档位（768P/2K，Max 追加 480P；与后端计费矩阵键一致），
@@ -98,15 +121,20 @@ async function submitTask() {
 	if (!prompt.value.trim() || !selectedModel.value) return
 	submitting.value = true
 	currentTask.value = null
+	deferredNames.value = []
 	stopPolling()
 	releaseVideo()
 
 	try {
 		const api = createPlaygroundApi(props.apiKey)
+		// 正文里的 [@图片1] 换成纯文字标签；首帧参考图取第一张图片附件
+		const compiled = compileVideoPrompt(prompt.value, attachmentList.value)
+		deferredNames.value = compiled.deferred.map(a => a.name)
+
 		// OpenAI Videos 协议：size 承载分辨率（像素串或命名档位如 768P/2K），seconds 承载时长
 		const body: Record<string, any> = {
 			model: selectedModel.value,
-			prompt: prompt.value,
+			prompt: compiled.prompt,
 		}
 		if (resolution.value) {
 			body.size = resolution.value
@@ -114,8 +142,13 @@ async function submitTask() {
 		if (duration.value) {
 			body.seconds = String(duration.value)
 		}
+		if (compiled.inputReference) {
+			// 走 JSON 引用对象形态：后端 parseVideosInputReferenceJSON 接受 data URL，
+			// 无需切 multipart，也不用跟 createPlaygroundApi 固定的 Content-Type 打架
+			body.input_reference = { image_url: compiled.inputReference }
+		}
 
-		const res = await api.post('/v1/videos', body, { timeout: 60_000 })
+		const res = await api.post('/v1/videos', body, { timeout: 120_000 })
 		const data = res.data
 
 		currentTask.value = {
@@ -251,9 +284,27 @@ function downloadVideo() {
 						<label class="input-label">模型</label>
 						<BaseSelect v-model="selectedModel" :options="modelOptions" />
 					</div>
+					<div>
+						<label class="input-label">参考素材</label>
+						<AttachmentBar
+							:attachments="attachmentList"
+							:accept="acceptTypes"
+							:max="attachmentMax"
+							@add="addAttachments"
+							@remove="removeAttachment"
+							@insert="insertMention"
+						/>
+					</div>
 					<div>
 						<label class="input-label">提示词</label>
-						<n-input v-model:value="prompt" type="textarea" :rows="4" placeholder="描述你想生成的视频..." />
+						<PromptInput
+							ref="promptInput"
+							v-model="prompt"
+							:attachments="attachmentList"
+							:min-rows="4"
+							placeholder="描述你想生成的视频，输入 @ 引用已上传的素材..."
+							@paste="addAttachments"
+						/>
 					</div>
 					<div>
 						<label class="input-label">分辨率</label>
@@ -299,6 +350,17 @@ function downloadVideo() {
 						<p class="empty-state-description">输入提示词并点击生成</p>
 					</div>
 
+					<!-- 本期协议带不动的素材：提交后明确告知，避免用户以为已生效 -->
+					<div v-if="deferredNames.length" class="mb-4 rounded-xl border border-amber-200 bg-amber-50 p-3">
+						<div class="flex items-center gap-2 text-amber-700">
+							<Icon name="exclamationTriangle" size="sm" />
+							<span class="text-sm font-medium">部分素材未随本次请求发送</span>
+						</div>
+						<p class="mt-1 text-xs text-amber-600">
+							{{ deferredNames.join('、') }} — 当前仅支持一张首帧参考图，多素材引用待后端支持
+						</p>
+					</div>
+
 					<!-- 提交中：卡片式占位 -->
 					<GenerationProgressCard v-if="submitting" kind="video" label="正在提交生成任务..." />
 
diff --git a/web/tenant/src/views/service/playground/attachmentParts.ts b/web/tenant/src/views/service/playground/attachmentParts.ts
new file mode 100644
index 0000000..4f5dacf
--- /dev/null
+++ b/web/tenant/src/views/service/playground/attachmentParts.ts
@@ -0,0 +1,120 @@
+import type { Attachment, AttachmentKind } from './useAttachments'
+
+/**
+ * 把「正文 + 附件」编译成各端点要的请求形态。
+ *
+ * 正文里的 [@图片1] 是位置锚点：编译时换成纯文字标签「图片1」留在句子里
+ * （多图场景下模型能用名字指代），同时在该位置插入对应的 content part。
+ * 没被引用的附件追加到末尾——上传了不打 @ 也要能发出去。
+ * 引用了已删除的附件时退化成纯文字，不报错打断输入。
+ */
+
+export interface ContentPart {
+	type: 'text' | 'image_url' | 'input_audio'
+	text?: string
+	image_url?: { url: string }
+	input_audio?: { data: string; format: string }
+}
+
+export interface AttachmentPreview {
+	name: string
+	kind: AttachmentKind
+	/** 直接用 data URL 展示：objectURL 在附件被清空时就 revoke 了，消息气泡不能依赖它 */
+	url: string
+}
+
+export interface CompiledChat {
+	/** 气泡里展示的文本（标记已换成文字标签） */
+	displayText: string
+	/** 发给 /v1/chat/completions 的 content parts */
+	parts: ContentPart[]
+	previews: AttachmentPreview[]
+}
+
+export interface CompiledVideo {
+	/** 发给 /v1/videos 的提示词（标记已换成文字标签） */
+	prompt: string
+	/** 首帧参考图的 data URL，无图时为空串 */
+	inputReference: string
+	/** 本期协议携带不了、未随请求发出的素材 */
+	deferred: Attachment[]
+}
+
+const MENTION_RE = /\[@([^\]\n]+)\]/g
+
+/** stripMentions 把 [@图片1] 还原成纯文字标签「图片1」 */
+export function stripMentions(text: string): string {
+	return text.replace(MENTION_RE, (_, name: string) => name)
+}
+
+/** hasMention 判断正文是否引用了某个附件 */
+export function hasMention(text: string, name: string): boolean {
+	return text.includes(`[@${name}]`)
+}
+
+/** toMediaPart 把附件转成对应协议的 content part；不支持的类型返回 null */
+function toMediaPart(att: Attachment): ContentPart | null {
+	if (att.kind === 'image') {
+		return { type: 'image_url', image_url: { url: att.dataUrl } }
+	}
+	if (att.kind === 'audio') {
+		// OpenAI input_audio 要的是裸 base64 + 格式名，不是完整 data URL
+		const base64 = att.dataUrl.slice(att.dataUrl.indexOf('base64,') + 'base64,'.length)
+		const format = att.mimeType.split('/')[1]?.split(';')[0] || 'mp3'
+		return { type: 'input_audio', input_audio: { data: base64, format } }
+	}
+	return null
+}
+
+export function compileChatParts(text: string, attachments: Attachment[]): CompiledChat {
+	const byName = new Map(attachments.map(a => [a.name, a]))
+	const referenced = new Set<number>()
+
+	const parts: ContentPart[] = []
+	let buf = ''
+
+	const flush = () => {
+		if (buf.trim()) parts.push({ type: 'text', text: buf })
+		buf = ''
+	}
+
+	let cursor = 0
+	MENTION_RE.lastIndex = 0
+	for (let m = MENTION_RE.exec(text); m !== null; m = MENTION_RE.exec(text)) {
+		buf += text.slice(cursor, m.index) + m[1]
+		cursor = m.index + m[0].length
+
+		const att = byName.get(m[1])
+		if (!att) continue // 失效标记：只留文字，继续累积
+
+		const part = toMediaPart(att)
+		if (!part) continue
+		flush()
+		parts.push(part)
+		referenced.add(att.id)
+	}
+	buf += text.slice(cursor)
+	flush()
+
+	for (const att of attachments) {
+		if (referenced.has(att.id)) continue
+		const part = toMediaPart(att)
+		if (part) parts.push(part)
+	}
+
+	return {
+		displayText: stripMentions(text),
+		parts,
+		previews: attachments.map(a => ({ name: a.name, kind: a.kind, url: a.dataUrl })),
+	}
+}
+
+export function compileVideoPrompt(text: string, attachments: Attachment[]): CompiledVideo {
+	// /v1/videos 的 input_reference 只能带一张图，其余素材本期发不出去
+	const firstImage = attachments.find(a => a.kind === 'image')
+	return {
+		prompt: stripMentions(text),
+		inputReference: firstImage?.dataUrl || '',
+		deferred: attachments.filter(a => a.id !== firstImage?.id),
+	}
+}
diff --git a/web/tenant/src/views/service/playground/useAttachments.ts b/web/tenant/src/views/service/playground/useAttachments.ts
new file mode 100644
index 0000000..655e996
--- /dev/null
+++ b/web/tenant/src/views/service/playground/useAttachments.ts
@@ -0,0 +1,205 @@
+import { ref, computed, onUnmounted } from 'vue'
+import { toast } from '@/utils/toast'
+
+/**
+ * Playground 附件管理
+ *
+ * 附件全部以 base64 data URL 内联进请求体（对话走 content parts，视频走 input_reference），
+ * 不落对象存储。因此体积是硬约束：后端 GoFrame server 默认 ClientMaxBodySize 为 8MB，
+ * base64 还要膨胀 4/3，这里按 6MB 的安全水位做单个与合计两道闸门。
+ */
+
+export type AttachmentKind = 'image' | 'audio' | 'video'
+
+export interface Attachment {
+	id: number
+	kind: AttachmentKind
+	/** 引用名，如「图片1」「音频1」，即 [@图片1] 标记里的那部分 */
+	name: string
+	/** 原始文件名，用于 tooltip 展示 */
+	fileName: string
+	/** 完整 data URL（data:<mime>;base64,<载荷>），直接进请求体 */
+	dataUrl: string
+	/** 本地预览地址（objectURL），仅供 <img>/<video> 展示，用后必须 revoke */
+	previewUrl: string
+	mimeType: string
+	/** data URL 的字节长度，用于体积闸门 */
+	bytes: number
+}
+
+/** 单个附件上限：图片 5MB（超限先压缩），音视频 4MB（不压缩，超限直接拒绝） */
+const MAX_IMAGE_BYTES = 5 * 1024 * 1024
+const MAX_MEDIA_BYTES = 4 * 1024 * 1024
+/** 所有附件 data URL 合计上限，与后端 8MB body 上限留出余量 */
+export const MAX_TOTAL_BYTES = 6 * 1024 * 1024
+
+/** 图片压缩目标：最长边与 JPEG 质量 */
+const COMPRESS_MAX_EDGE = 2048
+const COMPRESS_QUALITY = 0.85
+
+const KIND_LABEL: Record<AttachmentKind, string> = {
+	image: '图片',
+	audio: '音频',
+	video: '视频',
+}
+
+export function formatBytes(bytes: number): string {
+	if (bytes < 1024) return `${bytes} B`
+	if (bytes < 1024 * 1024) return `${(bytes / 1024).toFixed(0)} KB`
+	return `${(bytes / 1024 / 1024).toFixed(1)} MB`
+}
+
+function detectKind(file: File): AttachmentKind | null {
+	if (file.type.startsWith('image/')) return 'image'
+	if (file.type.startsWith('audio/')) return 'audio'
+	if (file.type.startsWith('video/')) return 'video'
+	return null
+}
+
+function readAsDataUrl(blob: Blob): Promise<string> {
+	return new Promise((resolve, reject) => {
+		const reader = new FileReader()
+		reader.onload = () => resolve(String(reader.result || ''))
+		reader.onerror = () => reject(new Error('读取文件失败'))
+		reader.readAsDataURL(blob)
+	})
+}
+
+/**
+ * compressImage 把超限图片降采样后重编码为 JPEG。
+ * GIF 不压缩（canvas 重编码会丢掉动画），超限直接判失败。
+ */
+async function compressImage(file: File): Promise<Blob | null> {
+	if (file.type === 'image/gif') return null
+
+	const bitmapUrl = URL.createObjectURL(file)
+	try {
+		const img = await new Promise<HTMLImageElement>((resolve, reject) => {
+			const el = new Image()
+			el.onload = () => resolve(el)
+			el.onerror = () => reject(new Error('图片解码失败'))
+			el.src = bitmapUrl
+		})
+
+		const scale = Math.min(1, COMPRESS_MAX_EDGE / Math.max(img.width, img.height))
+		const canvas = document.createElement('canvas')
+		canvas.width = Math.round(img.width * scale)
+		canvas.height = Math.round(img.height * scale)
+		const ctx = canvas.getContext('2d')
+		if (!ctx) return null
+		ctx.drawImage(img, 0, 0, canvas.width, canvas.height)
+
+		return await new Promise<Blob | null>(resolve => {
+			canvas.toBlob(blob => resolve(blob), 'image/jpeg', COMPRESS_QUALITY)
+		})
+	} catch {
+		return null
+	} finally {
+		URL.revokeObjectURL(bitmapUrl)
+	}
+}
+
+export interface UseAttachmentsOptions {
+	/**
+	 * 允许的附件类型，未列出的类型会被拒绝。
+	 * 传 getter 可跟随模型能力动态收放（切到非 vision 模型后就不该再收图片）。
+	 */
+	kinds: AttachmentKind[] | (() => AttachmentKind[])
+	/** 附件数量上限 */
+	max?: number
+}
+
+export function useAttachments(options: UseAttachmentsOptions) {
+	const list = ref<Attachment[]>([])
+	const max = options.max ?? 9
+
+	let nextId = 1
+	// 各类型独立递增。删除后不重排号——正文里可能已经写了 [@图片2]，
+	// 重排会让这些标记悄悄指向另一张图，比留个号码空档危险得多。
+	const counters: Record<AttachmentKind, number> = { image: 0, audio: 0, video: 0 }
+
+	const kinds = computed(() =>
+		typeof options.kinds === 'function' ? options.kinds() : options.kinds,
+	)
+	const totalBytes = computed(() => list.value.reduce((sum, a) => sum + a.bytes, 0))
+	const accept = computed(() => kinds.value.map(k => `${k}/*`).join(','))
+
+	async function addFiles(files: File[] | FileList) {
+		const incoming = Array.from(files)
+		for (const file of incoming) {
+			if (list.value.length >= max) {
+				toast.warning(`最多上传 ${max} 个附件`)
+				return
+			}
+			await addOne(file)
+		}
+	}
+
+	async function addOne(file: File) {
+		const kind = detectKind(file)
+		if (!kind || !kinds.value.includes(kind)) {
+			toast.error(`不支持的文件类型：${file.name}`)
+			return
+		}
+
+		let source: Blob = file
+		let mimeType = file.type
+
+		if (kind === 'image' && file.size > MAX_IMAGE_BYTES) {
+			const compressed = await compressImage(file)
+			if (!compressed || compressed.size > MAX_IMAGE_BYTES) {
+				toast.error(`${file.name} 超过 ${formatBytes(MAX_IMAGE_BYTES)}，压缩后仍过大`)
+				return
+			}
+			source = compressed
+			mimeType = compressed.type
+		} else if (kind !== 'image' && file.size > MAX_MEDIA_BYTES) {
+			toast.error(`${file.name} 超过 ${formatBytes(MAX_MEDIA_BYTES)}`)
+			return
+		}
+
+		let dataUrl: string
+		try {
+			dataUrl = await readAsDataUrl(source)
+		} catch {
+			toast.error(`${file.name} 读取失败`)
+			return
+		}
+
+		const bytes = dataUrl.length
+		if (totalBytes.value + bytes > MAX_TOTAL_BYTES) {
+			toast.error(`附件合计超过 ${formatBytes(MAX_TOTAL_BYTES)}，请先删除部分附件`)
+			return
+		}
+
+		counters[kind] += 1
+		list.value.push({
+			id: nextId++,
+			kind,
+			name: `${KIND_LABEL[kind]}${counters[kind]}`,
+			fileName: file.name,
+			dataUrl,
+			previewUrl: URL.createObjectURL(source),
+			mimeType,
+			bytes,
+		})
+	}
+
+	function remove(id: number) {
+		const idx = list.value.findIndex(a => a.id === id)
+		if (idx === -1) return
+		URL.revokeObjectURL(list.value[idx].previewUrl)
+		list.value.splice(idx, 1)
+	}
+
+	function clear() {
+		list.value.forEach(a => URL.revokeObjectURL(a.previewUrl))
+		list.value = []
+	}
+
+	// Playground 的 tab 由 KeepAlive 缓存，组件不会随切换销毁；
+	// 但路由离开时仍会卸载，此处兜底释放 objectURL，避免泄漏。
+	onUnmounted(clear)
+
+	return { list, totalBytes, accept, max, addFiles, remove, clear }
+}