JOJO/agent-Specialization
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: support kimi-k2.5 with multimodal thinking

Browse Source
This commit is contained in:
JOJO 2026-01-28 10:34:27 +08:00
parent d8cffa30cc
commit 8a7cc5d9c6
14 changed files with 63 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
config/model_profiles.py
View File

@ -9,6 +9,7 @@ KIMI_BASE = _env("API_BASE_KIMI", _env("AGENT_API_BASE_URL", "https://api.moonsh
KIMI_KEY = _env("API_KEY_KIMI", _env("AGENT_API_KEY", ""))
KIMI_FAST_MODEL = _env("MODEL_KIMI_FAST", _env("AGENT_MODEL_ID", "kimi-k2-0905-preview"))
KIMI_THINK_MODEL = _env("MODEL_KIMI_THINK", _env("AGENT_THINKING_MODEL_ID", "kimi-k2-thinking"))
KIMI_25_MODEL = _env("MODEL_KIMI_25", "kimi-k2.5")
# DeepSeek
DEEPSEEK_BASE = _env("API_BASE_DEEPSEEK", "https://api.deepseek.com")
@ -31,6 +32,25 @@ MODEL_PROFILES = {
"fast_only": False,
"name": "Kimi-k2"
},
"kimi-k2.5": {
"fast": {
"base_url": KIMI_BASE,
"api_key": KIMI_KEY,
"model_id": KIMI_25_MODEL,
"max_tokens": None,
"extra_params": {"thinking": {"type": "disabled"}}
},
"thinking": {
"base_url": KIMI_BASE,
"api_key": KIMI_KEY,
"model_id": KIMI_25_MODEL,
"max_tokens": None,
"extra_params": {"thinking": {"type": "enabled"}}
},
"supports_thinking": True,
"fast_only": False,
"name": "Kimi-k2.5"
},
"deepseek": {
"fast": {"base_url": DEEPSEEK_BASE, "api_key": DEEPSEEK_KEY, "model_id": DEEPSEEK_FAST_MODEL, "max_tokens": 8192},
"thinking": {
@ -77,6 +97,11 @@ MODEL_PROMPT_OVERRIDES = {
"thinking_model_line": "思考模式时，第一次请求的模型不是 Kimi-k2，而是 Kimi-k2-Thinking，一个更善于分析复杂问题、规划复杂流程的模型，在后续请求时模型会换回 Kimi-k2。",
"deep_thinking_line": "在深度思考模式中，请求的模型是 Kimi-k2-Thinking，一个更善于分析复杂问题、规划复杂流程的模型。"
},
"kimi-k2.5": {
"model_description": "你的基础模型是 Kimi-k2.5，支持图文多模态，并通过 thinking 参数开启/关闭思考能力。",
"thinking_model_line": "思考模式时使用同一个 Kimi-k2.5 模型，但会在请求中注入 thinking={\"type\": \"enabled\"} 来开启思考；快速模式则传递 thinking={\"type\": \"disabled\"}。",
"deep_thinking_line": "深度思考模式下，所有请求都会携带 thinking={\"type\": \"enabled\"}，以获得持续的推理能力。"
},
"deepseek": {
"model_description": "你的基础模型是 DeepSeek-V3.2（deepseek-chat），由 DeepSeek 提供，数学与推理能力较强，当前通过官方 API 调用。",
"thinking_model_line": "思考模式时，第一次请求使用 DeepSeek-Reasoner，一个强化推理的模型，后续请求会切回 DeepSeek-V3.2。",

12
core/main_terminal.py
View File

@ -98,7 +98,7 @@ class MainTerminal:
# 初始化组件
self.api_client = DeepSeekClient(thinking_mode=self.thinking_mode)
self.api_client.set_deep_thinking_mode(self.deep_thinking_mode)
self.model_key = "kimi"
self.model_key = "kimi-k2.5"
self.model_profile = get_model_profile(self.model_key)
self.apply_model_profile(self.model_profile)
self.context_manager = ContextManager(project_path, data_dir=str(self.data_dir))
@ -1826,8 +1826,8 @@ class MainTerminal:
}
}
]
# Qwen-VL 自带多模态能力，不向其暴露额外的 vlm_analyze 工具，避免重复与误导
if getattr(self, "model_key", None) == "qwen3-vl-plus":
# 视觉模型（Qwen-VL / Kimi-k2.5）自带多模态能力，不再暴露 vlm_analyze，改为 view_image
if getattr(self, "model_key", None) in {"qwen3-vl-plus", "kimi-k2.5"}:
tools = [
tool for tool in tools
if (tool.get("function") or {}).get("name") != "vlm_analyze"
@ -2491,7 +2491,7 @@ class MainTerminal:
def build_messages(self, context: Dict, user_input: str) -> List[Dict]:
"""构建消息列表（添加终端内容注入）"""
# 加载系统提示（Qwen-VL 使用专用提示）
prompt_name = "main_system_qwenvl" if getattr(self, "model_key", "kimi") == "qwen3-vl-plus" else "main_system"
prompt_name = "main_system_qwenvl" if getattr(self, "model_key", "kimi") in {"qwen3-vl-plus", "kimi-k2.5"} else "main_system"
system_prompt = self.load_prompt(prompt_name)
# 格式化系统提示
@ -2803,8 +2803,8 @@ class MainTerminal:
def set_model(self, model_key: str) -> str:
profile = get_model_profile(model_key)
if getattr(self.context_manager, "has_images", False) and model_key != "qwen3-vl-plus":
raise ValueError("当前对话包含图片，仅支持 Qwen-VL")
if getattr(self.context_manager, "has_images", False) and model_key not in {"qwen3-vl-plus", "kimi-k2.5"}:
raise ValueError("当前对话包含图片，仅支持 Qwen-VL 或 Kimi-k2.5")
self.model_key = model_key
self.model_profile = profile
# 将模型标识传递给底层 API 客户端，便于按模型做兼容处理

6
modules/personalization_manager.py
View File

@ -36,7 +36,7 @@ DEFAULT_PERSONALIZATION_CONFIG: Dict[str, Any] = {
"default_run_mode": None,
"auto_generate_title": True,
"tool_intent_enabled": True,
"default_model": "kimi",
"default_model": "kimi-k2.5",
}
__all__ = [
@ -109,7 +109,7 @@ def sanitize_personalization_payload(
base.update(fallback)
data = payload or {}
allowed_tool_categories = set(TOOL_CATEGORIES.keys())
allowed_models = {"kimi", "deepseek", "qwen3-max", "qwen3-vl-plus"}
allowed_models = {"kimi", "kimi-k2.5", "deepseek", "qwen3-max", "qwen3-vl-plus"}
def _resolve_short_field(key: str) -> str:
if key in data:
@ -153,7 +153,7 @@ def sanitize_personalization_payload(
if isinstance(chosen_model, str) and chosen_model in allowed_models:
base["default_model"] = chosen_model
elif base.get("default_model") not in allowed_models:
base["default_model"] = "kimi"
base["default_model"] = "kimi-k2.5"
return base

2
prompts/main_system_qwenvl.txt
View File

@ -10,7 +10,7 @@
- **自动化任务**：批量处理文件、执行重复性工作
- **视觉理解**：你自带多模态能力，用户可以直接发送图片；如需主动查看本地图片，可调用 `view_image` 指定路径，系统会代发一条包含图片的用户消息供你查看。
## 图片分析（Qwen-VL 重点）
## 图片分析（Kimi-k2.5/Qwen-VL 重点）
当用户提出“这是什么”“识别文字/表格/票据”“找瑕疵/细节”“读屏/按钮含义”等图片分析任务时，优先采用下面的方法，保证细节充分、结论可验证：
### 基本流程（先粗后细）

4
server/_socket_segment.py
View File

@ -195,8 +195,8 @@ def handle_message(data):
if not message and not images:
emit('error', {'message': '消息不能为空'})
return
if images and getattr(terminal, "model_key", None) != "qwen3-vl-plus":
emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL'})
if images and getattr(terminal, "model_key", None) not in {"qwen3-vl-plus", "kimi-k2.5"}:
emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL 或 Kimi-k2.5'})
return
print(f"[WebSocket] 收到消息: {message}")

2
server/app_legacy.py
View File

@ -918,7 +918,7 @@ def get_user_resources(username: Optional[str] = None) -> Tuple[Optional[WebTerm
terminal.admin_policy_version = policy.get("updated_at")
# 若当前模型被禁用，则回退到第一个可用模型
if terminal.model_key in disabled_models:
for candidate in ["kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
if candidate not in disabled_models:
try:
terminal.set_model(candidate)

2
server/context.py
View File

@ -216,7 +216,7 @@ def get_user_resources(username: Optional[str] = None, workspace_id: Optional[st
terminal.admin_policy_ui_blocks = policy.get("ui_blocks") or {}
terminal.admin_policy_version = policy.get("updated_at")
if terminal.model_key in disabled_models:
for candidate in ["kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
for candidate in ["kimi-k2.5", "kimi", "deepseek", "qwen3-vl-plus", "qwen3-max"]:
if candidate not in disabled_models:
try:
terminal.set_model(candidate)

4
server/socket_handlers.py
View File

@ -216,8 +216,8 @@ def handle_message(data):
if not message and not images:
emit('error', {'message': '消息不能为空'})
return
if images and getattr(terminal, "model_key", None) != "qwen3-vl-plus":
emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL'})
if images and getattr(terminal, "model_key", None) not in {"qwen3-vl-plus", "kimi-k2.5"}:
emit('error', {'message': '当前模型不支持图片，请切换到 Qwen-VL 或 Kimi-k2.5'})
return
print(f"[WebSocket] 收到消息: {message}")

12
static/src/app.ts
View File

@ -2398,10 +2398,10 @@ const appOptions = {
return;
}
if (hasImages && this.currentModelKey !== 'qwen3-vl-plus') {
if (hasImages && !['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
this.uiPushToast({
title: '当前模型不支持图片',
message: '请切换到 Qwen-VL 再发送图片',
message: '请切换到 Qwen-VL 或 Kimi-k2.5 再发送图片',
type: 'error'
});
return;
@ -2627,10 +2627,10 @@ const appOptions = {
},
async openImagePicker() {
if (this.currentModelKey !== 'qwen3-vl-plus') {
if (!['qwen3-vl-plus', 'kimi-k2.5'].includes(this.currentModelKey)) {
this.uiPushToast({
title: '当前模型不支持图片',
message: '请选择 Qwen-VL 后再发送图片',
message: '请选择 Qwen-VL 或 Kimi-k2.5 后再发送图片',
type: 'error'
});
return;
@ -2793,10 +2793,10 @@ const appOptions = {
});
return;
}
if (this.conversationHasImages && key !== 'qwen3-vl-plus') {
if (this.conversationHasImages && !['qwen3-vl-plus', 'kimi-k2.5'].includes(key)) {
this.uiPushToast({
title: '切换失败',
message: '当前对话包含图片，仅支持 Qwen-VL',
message: '当前对话包含图片，仅支持 Qwen-VL 或 Kimi-k2.5',
type: 'error'
});
return;

2
static/src/components/input/QuickMenu.vue
View File

@ -18,7 +18,7 @@
对话回顾
</button>
<button
v-if="currentModelKey === 'qwen3-vl-plus'"
v-if="['qwen3-vl-plus', 'kimi-k2.5'].includes(currentModelKey)"
type="button"
class="menu-entry"
@click.stop="$emit('pick-images')"

3
static/src/components/personalization/PersonalizationDrawer.vue
View File

@ -609,7 +609,8 @@ const policyStore = usePolicyStore();
const modelOptions = [
{ id: 'deepseek', label: 'DeepSeek', desc: '通用 + 思考强化', value: 'deepseek' },
{ id: 'kimi', label: 'Kimi', desc: '默认模型，兼顾通用对话', value: 'kimi' },
{ id: 'kimi-k2.5', label: 'Kimi-k2.5', desc: '新版 Kimi，思考开关 + 图文多模态', value: 'kimi-k2.5', badge: '图文' },
{ id: 'kimi', label: 'Kimi-k2', desc: '旧版 Kimi-k2，兼顾通用对话', value: 'kimi' },
{ id: 'qwen3-max', label: 'Qwen-Max', desc: '仅快速模式，不支持思考', value: 'qwen3-max', badge: '仅快速' },
{ id: 'qwen3-vl-plus', label: 'Qwen-VL', desc: '图文多模态，思考/快速均可', value: 'qwen3-vl-plus', badge: '图文' }
] as const;

11
static/src/stores/model.ts
View File

@ -1,6 +1,6 @@
import { defineStore } from 'pinia';
export type ModelKey = 'kimi' | 'deepseek' | 'qwen3-max' | 'qwen3-vl-plus';
export type ModelKey = 'kimi-k2.5' | 'kimi' | 'deepseek' | 'qwen3-max' | 'qwen3-vl-plus';
export interface ModelOption {
key: ModelKey;
@ -17,8 +17,15 @@ interface ModelState {
export const useModelStore = defineStore('model', {
state: (): ModelState => ({
currentModelKey: 'kimi',
currentModelKey: 'kimi-k2.5',
models: [
{
key: 'kimi-k2.5',
label: 'Kimi-k2.5',
description: '新版 Kimi，支持图文 & 思考开关',
fastOnly: false,
supportsThinking: true
},
{
key: 'kimi',
label: 'Kimi-k2',

6
static/src/stores/personalization.ts
View File

@ -65,7 +65,7 @@ const defaultForm = (): PersonalForm => ({
thinking_interval: null,
disabled_tool_categories: [],
default_run_mode: null,
default_model: 'kimi'
default_model: 'kimi-k2.5'
});
const defaultExperimentState = (): ExperimentState => ({
@ -179,7 +179,7 @@ export const usePersonalizationStore = defineStore('personalization', {
applyPersonalizationData(data: any) {
// 若后端未返回默认模型（旧版本接口），保持当前已选模型而不是回退为 Kimi
const fallbackModel =
(this.form && typeof this.form.default_model === 'string' ? this.form.default_model : null) || 'kimi';
(this.form && typeof this.form.default_model === 'string' ? this.form.default_model : null) || 'kimi-k2.5';
this.form = {
enabled: !!data.enabled,
auto_generate_title: data.auto_generate_title !== false,
@ -357,7 +357,7 @@ export const usePersonalizationStore = defineStore('personalization', {
this.clearFeedback();
},
setDefaultModel(model: string | null) {
const allowed = ['deepseek', 'kimi', 'qwen3-max', 'qwen3-vl-plus'];
const allowed = ['deepseek', 'kimi-k2.5', 'kimi', 'qwen3-max', 'qwen3-vl-plus'];
const target = typeof model === 'string' && allowed.includes(model) ? model : null;
this.form = {
...this.form,

2
utils/context_manager.py
View File

@ -1223,7 +1223,7 @@ class ContextManager:
"""构建消息列表（添加终端内容注入）"""
# 加载系统提示（Qwen-VL 使用专用提示）
model_key = getattr(self.main_terminal, "model_key", "kimi") if hasattr(self, "main_terminal") else "kimi"
prompt_name = "main_system_qwenvl" if model_key == "qwen3-vl-plus" else "main_system"
prompt_name = "main_system_qwenvl" if model_key in {"qwen3-vl-plus", "kimi-k2.5"} else "main_system"
system_prompt = self.load_prompt(prompt_name)
# 格式化系统提示