feat: Add pipeline_uuid handling across multiple adapters and update related tests

* fix(mcp): harden remote SSE fallback

* fix(mcp): preserve transport errors during fallback

---------

Co-authored-by: RockChinQ <45992437+RockChinQ@users.noreply.github.com>

README.md

@@ -51,7 +51,7 @@ LangBot is an **open-source, production-grade platform** for building AI-powered
 
 [→ Learn more about all features](https://link.langbot.app/en/docs/features)
 
-📍 Practical guides: [deploy a multi-platform AI bot in 5 minutes](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [connect DeepSeek to WeChat, Discord, and Telegram](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [run a Dify Agent in Discord, Telegram, and Slack](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/), and [build an n8n-powered chatbot](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 Practical guides: [deploy a multi-platform AI bot in 5 minutes](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [connect DeepSeek to WeChat, Discord, and Telegram](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [run a Dify Agent in Discord, Telegram, and Slack](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/), and [build an n8n-powered chatbot](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

README_CN.md

@@ -51,7 +51,7 @@ LangBot 是一个**开源的生产级平台**，用于构建 AI 驱动的即时
 
 [→ 了解更多功能特性](https://link.langbot.app/zh/docs/features)
 
-📍 实践指南：[5 分钟部署多平台 AI 机器人](https://blog.langbot.app/zh/blog/deploy-ai-bot-in-5-minutes/)、[将 DeepSeek 接入微信、企业微信与 Discord](https://blog.langbot.app/zh/blog/connect-deepseek-to-wechat/)、[让 Dify Agent 跑在 Discord、Telegram 和 Slack 上](https://blog.langbot.app/zh/blog/dify-agent-discord-telegram-slack/)，以及[用 n8n 构建多平台 AI 聊天机器人](https://blog.langbot.app/zh/blog/n8n-multi-platform-ai-chatbot/)。
+📍 实践指南：[5 分钟部署多平台 AI 机器人](https://langbot.app/zh/blog/deploy-ai-bot-in-5-minutes/)、[将 DeepSeek 接入微信、企业微信与 Discord](https://langbot.app/zh/blog/connect-deepseek-to-wechat/)、[让 Dify Agent 跑在 Discord、Telegram 和 Slack 上](https://langbot.app/zh/blog/dify-agent-discord-telegram-slack/)，以及[用 n8n 构建多平台 AI 聊天机器人](https://langbot.app/zh/blog/n8n-multi-platform-ai-chatbot/)。
 
 ---
 

README_ES.md

@@ -50,7 +50,7 @@ LangBot es una **plataforma de código abierto y grado de producción** para con
 
 [→ Conocer más sobre todas las funcionalidades](https://link.langbot.app/en/docs/features)
 
-📍 Guías prácticas: [desplegar un bot de IA multiplataforma en 5 minutos](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [conectar DeepSeek a WeChat, Discord y Telegram](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [ejecutar un Dify Agent en Discord, Telegram y Slack](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/) y [crear un chatbot con n8n](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 Guías prácticas: [desplegar un bot de IA multiplataforma en 5 minutos](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [conectar DeepSeek a WeChat, Discord y Telegram](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [ejecutar un Dify Agent en Discord, Telegram y Slack](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/) y [crear un chatbot con n8n](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

README_FR.md

@@ -50,7 +50,7 @@ LangBot est une **plateforme open-source de niveau production** pour créer des
 
 [→ En savoir plus sur toutes les fonctionnalités](https://link.langbot.app/en/docs/features)
 
-📍 Guides pratiques : [déployer un bot IA multiplateforme en 5 minutes](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [connecter DeepSeek à WeChat, Discord et Telegram](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [exécuter un Dify Agent dans Discord, Telegram et Slack](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/) et [créer un chatbot avec n8n](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 Guides pratiques : [déployer un bot IA multiplateforme en 5 minutes](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [connecter DeepSeek à WeChat, Discord et Telegram](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [exécuter un Dify Agent dans Discord, Telegram et Slack](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/) et [créer un chatbot avec n8n](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

README_JP.md

@@ -50,7 +50,7 @@ LangBot は、AI搭載のインスタントメッセージングボットを構
 
 [→ すべての機能について詳しく見る](https://link.langbot.app/ja/docs/features)
 
-📍 実践ガイド: [5分でマルチプラットフォームAIボットをデプロイ](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/)、[DeepSeekをWeChat・Discord・Telegramに接続](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/)、[Dify AgentをDiscord・Telegram・Slackで動かす](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/)、[n8n連携チャットボットを構築](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/)。
+📍 実践ガイド: [5分でマルチプラットフォームAIボットをデプロイ](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/)、[DeepSeekをWeChat・Discord・Telegramに接続](https://langbot.app/en/blog/connect-deepseek-to-wechat/)、[Dify AgentをDiscord・Telegram・Slackで動かす](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/)、[n8n連携チャットボットを構築](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/)。
 
 ---
 

README_KO.md

@@ -50,7 +50,7 @@ LangBot은 AI 기반 인스턴트 메시징 봇을 구축하기 위한 **오픈
 
 [→ 모든 기능 자세히 보기](https://link.langbot.app/en/docs/features)
 
-📍 실전 가이드: [5분 만에 멀티 플랫폼 AI 봇 배포하기](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [DeepSeek를 WeChat, Discord, Telegram에 연결하기](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [Dify Agent를 Discord, Telegram, Slack에서 실행하기](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/), [n8n 기반 챗봇 만들기](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 실전 가이드: [5분 만에 멀티 플랫폼 AI 봇 배포하기](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [DeepSeek를 WeChat, Discord, Telegram에 연결하기](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [Dify Agent를 Discord, Telegram, Slack에서 실행하기](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/), [n8n 기반 챗봇 만들기](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

README_RU.md

@@ -50,7 +50,7 @@ LangBot — это **платформа с открытым исходным к
 
 [→ Подробнее обо всех возможностях](https://link.langbot.app/en/docs/features)
 
-📍 Практические руководства: [развернуть мультиплатформенного ИИ-бота за 5 минут](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [подключить DeepSeek к WeChat, Discord и Telegram](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [запустить Dify Agent в Discord, Telegram и Slack](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/) и [создать чат-бота на n8n](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 Практические руководства: [развернуть мультиплатформенного ИИ-бота за 5 минут](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [подключить DeepSeek к WeChat, Discord и Telegram](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [запустить Dify Agent в Discord, Telegram и Slack](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/) и [создать чат-бота на n8n](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

README_TW.md

@@ -52,7 +52,7 @@ LangBot 是一個**開源的生產級平台**，用於建構 AI 驅動的即時
 
 [→ 了解更多功能特性](https://link.langbot.app/zh/docs/features)
 
-📍 實踐指南：[5 分鐘部署多平台 AI 機器人](https://blog.langbot.app/zh/blog/deploy-ai-bot-in-5-minutes/)、[將 DeepSeek 接入微信、企業微信與 Discord](https://blog.langbot.app/zh/blog/connect-deepseek-to-wechat/)、[讓 Dify Agent 跑在 Discord、Telegram 和 Slack 上](https://blog.langbot.app/zh/blog/dify-agent-discord-telegram-slack/)，以及[用 n8n 建構多平台 AI 聊天機器人](https://blog.langbot.app/zh/blog/n8n-multi-platform-ai-chatbot/)。
+📍 實踐指南：[5 分鐘部署多平台 AI 機器人](https://langbot.app/zh/blog/deploy-ai-bot-in-5-minutes/)、[將 DeepSeek 接入微信、企業微信與 Discord](https://langbot.app/zh/blog/connect-deepseek-to-wechat/)、[讓 Dify Agent 跑在 Discord、Telegram 和 Slack 上](https://langbot.app/zh/blog/dify-agent-discord-telegram-slack/)，以及[用 n8n 建構多平台 AI 聊天機器人](https://langbot.app/zh/blog/n8n-multi-platform-ai-chatbot/)。
 
 ---
 

README_VI.md

@@ -50,7 +50,7 @@ LangBot là một **nền tảng mã nguồn mở, cấp sản xuất** để x
 
 [→ Tìm hiểu thêm về tất cả tính năng](https://link.langbot.app/en/docs/features)
 
-📍 Hướng dẫn thực hành: [triển khai bot AI đa nền tảng trong 5 phút](https://blog.langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [kết nối DeepSeek với WeChat, Discord và Telegram](https://blog.langbot.app/en/blog/connect-deepseek-to-wechat/), [chạy Dify Agent trên Discord, Telegram và Slack](https://blog.langbot.app/en/blog/dify-agent-discord-telegram-slack/) và [xây dựng chatbot với n8n](https://blog.langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
+📍 Hướng dẫn thực hành: [triển khai bot AI đa nền tảng trong 5 phút](https://langbot.app/en/blog/deploy-ai-bot-in-5-minutes/), [kết nối DeepSeek với WeChat, Discord và Telegram](https://langbot.app/en/blog/connect-deepseek-to-wechat/), [chạy Dify Agent trên Discord, Telegram và Slack](https://langbot.app/en/blog/dify-agent-discord-telegram-slack/) và [xây dựng chatbot với n8n](https://langbot.app/en/blog/n8n-multi-platform-ai-chatbot/).
 
 ---
 

docs/VALKEY_SEARCH_INTEGRATION.md

@@ -0,0 +1,169 @@
+# Valkey Search Vector Database Integration
+
+This document describes how to use **Valkey Search** (the search/vector module bundled in
+`valkey/valkey-bundle`) as the vector database backend for LangBot's knowledge base (RAG)
+feature.
+
+## What is Valkey Search?
+
+**Valkey Search** is a module that adds vector similarity search and full-text search to
+[Valkey](https://valkey.io/), the open-source, BSD-licensed in-memory data store forked from
+Redis OSS. It is distributed in the `valkey/valkey-bundle` image alongside other modules
+(JSON, Bloom, LDAP).
+
+LangBot talks to Valkey through the official [`valkey-glide`](https://pypi.org/project/valkey-glide/)
+client (Rust core + async Python wrapper), using its native `ft` (search) command namespace.
+
+### Key Features
+
+- **Vector search**: ANN via HNSW or exact via FLAT, with COSINE / L2 / IP distance metrics
+- **Full-text search**: term, prefix and phrase matching over indexed text fields
+- **Hybrid search**: a metadata/text filter pre-selects candidates, then KNN ranks them
+- **In-memory speed**: vectors and documents are stored as Valkey HASH keys
+- **Auth + TLS**: optional username/password and TLS for production (toB / SaaS) deployments
+
+### Licensing
+
+- Valkey core and the Search module are **BSD-3-Clause**.
+- The `valkey-glide` client is **Apache-2.0**.
+
+Both are compatible with LangBot.
+
+## Installation
+
+Valkey Search support is included when you install LangBot — the `valkey-glide` dependency is
+declared in `pyproject.toml`. To install manually:
+
+```bash
+pip install 'valkey-glide>=2.4.1,<3.0.0'
+```
+
+You also need a running Valkey server with the Search module loaded. The simplest way is the
+bundled image:
+
+```bash
+# Run valkey-bundle (includes the Search module) on host port 6380
+podman run -d --name valkey-test-langbot -p 6380:6379 valkey/valkey-bundle:9.1.0
+# (docker run ... works identically)
+```
+
+`valkey-bundle` ships multi-arch images (linux/amd64 + linux/arm64), so it runs on both CI
+(x86_64) and Apple-silicon dev machines.
+
+## Configuration
+
+Valkey Search is **opt-in and disabled by default** — the default `vdb.use` stays `chroma`,
+so existing single-process deployments are unaffected. To enable it, edit your `config.yaml`:
+
+```yaml
+vdb:
+  use: valkey_search
+  valkey_search:
+    host: 'localhost'
+    port: 6379            # use 6380 if you started the container as shown above
+    db: 0
+    password: ''          # optional (ACL / requirepass) — never logged
+    username: ''          # optional (ACL user)
+    tls: false            # optional (toB / SaaS)
+    index_algorithm: 'HNSW'   # HNSW | FLAT
+    distance_metric: 'COSINE' # COSINE | L2 | IP
+    request_timeout: 5000     # per-request timeout in ms
+```
+
+| Option | Default | Description |
+|--------|---------|-------------|
+| `host` | `localhost` | Valkey host |
+| `port` | `6379` | Valkey port |
+| `db` | `0` | Logical database id |
+| `password` | `''` | Optional auth password (empty = no auth). Never logged. |
+| `username` | `''` | Optional ACL username. Configuring a username without a password fails closed (raises) rather than connecting unauthenticated. |
+| `tls` | `false` | Enable TLS for the connection |
+| `index_algorithm` | `HNSW` | `HNSW` (approximate) or `FLAT` (exact) |
+| `distance_metric` | `COSINE` | `COSINE`, `L2`, or `IP` |
+| `request_timeout` | `5000` | Per-request timeout in milliseconds. The valkey-glide default (250ms) is too low for vector KNN under load; raise it further for remote/cross-AZ Valkey. |
+
+### Connection behavior
+
+The backend uses a **lazy** connection (`lazy_connect=True`): the client is created on first
+use and the connection is deferred to the first command. A misconfigured or unreachable Valkey
+server therefore does **not** block LangBot from booting — knowledge-base operations will error
+at call time instead, and you can recover by switching `vdb.use` back to another backend.
+
+The connection sets a fixed `client_name` of `langbot_vector_client` so it is identifiable in
+`CLIENT LIST` and monitoring dashboards.
+
+## Supported search types
+
+| Type | Behavior |
+|------|----------|
+| `vector` | Pure KNN over the embedding field |
+| `full_text` | Term/phrase match over the indexed `document` text field |
+| `hybrid` | Metadata/text filter **pre-selects** candidates, then KNN ranks them |
+
+### ⚠️ Important: `vector_weight` is NOT honored
+
+Valkey Search hybrid queries follow a **filter-then-KNN** model: the filter (and/or full-text
+clause) narrows the candidate set, and the KNN stage ranks the survivors by vector distance.
+There is **no native weighted score fusion** (unlike, e.g., SeekDB's RRF boost).
+
+For interface compatibility the backend still accepts a `vector_weight` argument, but it is
+**ignored** — passing different weights does not change result ordering. The first time a
+non-default weight is supplied, the backend logs a one-time warning.
+
+If weighted hybrid ranking is needed in the future, it can be added **application-side** (run
+vector KNN and full-text search separately and blend the scores). That is intentionally out of
+scope for this integration.
+
+## Metadata & filtering
+
+Documents are stored as Valkey HASH keys under the prefix `kb:{collection}:{id}` with fields:
+
+- `vector` — the embedding, packed as little-endian FLOAT32
+- `document` — the raw text (indexed as TEXT for full-text/hybrid search)
+- `file_id` — promoted to an indexed TAG field so it is filterable
+- `metadata_json` — the full metadata dict, preserved verbatim as JSON
+
+Only **indexed** fields are filterable. Currently that is `file_id`. Filters referencing
+non-indexed metadata keys are dropped with a warning (the same pragmatism used by the Milvus
+and pgvector backends). All other metadata still round-trips intact via `metadata_json`.
+
+Supported filter operators (canonical Chroma-style `where` syntax): `$eq`, `$ne`, `$gt`,
+`$gte`, `$lt`, `$lte`, `$in`, `$nin`. Multiple top-level keys are AND-ed.
+
+## Testing
+
+Unit tests (filter mapping, float32 packing, reply parsing, import guard) run in the fast lane
+with no server:
+
+```bash
+uv run pytest tests/unit_tests/vector/test_valkey_search_filter.py -q
+```
+
+Integration tests are **slow-gated** on `TEST_VALKEY_URL` and require a running server:
+
+```bash
+podman run -d --name valkey-test-langbot -p 6380:6379 valkey/valkey-bundle:9.1.0
+TEST_VALKEY_URL=valkey://localhost:6380 \
+    uv run pytest tests/integration/vector/test_valkey_search.py -m slow -q
+```
+
+The default upstream fast CI lane (`-m "not slow"`) skips these, matching the existing
+PostgreSQL migration-test precedent.
+
+## Troubleshooting
+
+| Symptom | Cause / fix |
+|---------|-------------|
+| Tests skip with "Valkey Search module not available" | The server is plain Valkey without the Search module. Use the `valkey/valkey-bundle` image. |
+| `ConnectionError` at call time | Check `host`/`port`/auth; remember `lazy_connect` defers errors to first use. |
+| Empty search results right after insert | The Search indexer is asynchronous; results become visible within a short delay. The integration tests poll/retry to account for this. |
+| Hybrid ranking ignores `vector_weight` | Expected — see the caveat above. |
+
+## Production considerations
+
+- **Cluster mode**: Valkey Search in cluster mode uses an additional coordination port. This
+  integration targets standalone mode; cluster support is a future consideration.
+- **Persistence**: configure Valkey RDB/AOF persistence if the knowledge base must survive
+  restarts; otherwise an in-memory store is ephemeral.
+- **Security**: set `password`/`username` and `tls: true` for any non-local deployment.
+  Credentials are never written to logs.

pyproject.toml

@@ -70,7 +70,7 @@ dependencies = [
     "chromadb>=1.0.0,<2.0.0",
     "qdrant-client (>=1.15.1,<2.0.0)",
     "pyseekdb==1.1.0.post3",
-    "langbot-plugin==0.4.7",
+    "langbot-plugin==0.4.13",
     "asyncpg>=0.30.0",
     "line-bot-sdk>=3.19.0",
     "matrix-nio>=0.25.2",
@@ -80,6 +80,7 @@ dependencies = [
     "pgvector>=0.4.1",
     "botocore>=1.42.39",
     "litellm>=1.0.0",
+    "valkey-glide>=2.4.1,<3.0.0",
 ]
 keywords = [
     "bot",

src/langbot/libs/dify_service_api/v1/client.py

@@ -109,6 +109,62 @@ class AsyncDifyServiceClient:
                     if chunk.startswith('data:'):
                         yield json.loads(chunk[5:])
 
+    async def workflow_submit(
+        self,
+        form_token: str,
+        workflow_run_id: str,
+        inputs: dict[str, typing.Any],
+        user: str,
+        action: str = '',
+        timeout: float = 120.0,
+    ) -> typing.AsyncGenerator[dict[str, typing.Any], None]:
+        """Submit human input to resume a paused workflow, then stream events.
+
+        1. POST /form/human_input/{form_token} to submit the form
+        2. GET /workflow/{task_id}/events to stream the resumed workflow events
+        """
+
+        headers = {
+            'Authorization': f'Bearer {self.api_key}',
+            'Content-Type': 'application/json',
+        }
+
+        async with httpx.AsyncClient(
+            base_url=self.base_url,
+            trust_env=True,
+            timeout=timeout,
+        ) as client:
+            # Step 1: Submit the form
+            payload: dict[str, typing.Any] = {
+                'inputs': inputs if isinstance(inputs, dict) else {},
+                'user': user,
+                'action': action,
+            }
+
+            submit_resp = await client.post(
+                f'/form/human_input/{form_token}',
+                headers=headers,
+                json=payload,
+            )
+            if submit_resp.status_code != 200:
+                raise DifyAPIError(f'{submit_resp.status_code} {submit_resp.text}')
+
+            # Step 2: Stream resumed workflow events
+            async with client.stream(
+                'GET',
+                f'/workflow/{workflow_run_id}/events',
+                headers={'Authorization': f'Bearer {self.api_key}'},
+                params={'user': user},
+            ) as r:
+                if r.status_code != 200:
+                    body = (await r.aread()).decode(errors='replace')
+                    raise DifyAPIError(f'{r.status_code} {body}')
+                async for chunk in r.aiter_lines():
+                    if chunk.strip() == '':
+                        continue
+                    if chunk.startswith('data:'):
+                        yield json.loads(chunk[5:])
+
     async def upload_file(
         self,
         file: httpx._types.FileTypes,

src/langbot/libs/dingtalk_api/api.py

@@ -1,17 +1,48 @@
 import asyncio
 import base64
 import json
+import logging
+import os
 import time
+import typing
+import uuid
 import urllib.parse
-from typing import Callable
+from typing import Awaitable, Callable, Optional
 import dingtalk_stream  # type: ignore
 import websockets
 from .EchoHandler import EchoTextHandler
+from .card_callback import DingTalkCardActionHandler
 from .dingtalkevent import DingTalkEvent
 import httpx
 import traceback
 
 
+_stdout_logger = logging.getLogger('langbot.dingtalk_api')
+
+
+DINGTALK_OPENAPI_BASE = 'https://api.dingtalk.com'
+
+
+def _stringify_card_param_map(card_param_map: Optional[dict]) -> dict:
+    """DingTalk cardParamMap only accepts string values.
+
+    Keep callers free to pass structured values for template variables such
+    as button groups or select options, then encode them once at the API
+    boundary.
+    """
+    if not card_param_map:
+        return {}
+    result = {}
+    for key, value in card_param_map.items():
+        if value is None:
+            result[key] = ''
+        elif isinstance(value, str):
+            result[key] = value
+        else:
+            result[key] = json.dumps(value, ensure_ascii=False)
+    return result
+
+
 class DingTalkClient:
     def __init__(
         self,
@@ -21,6 +52,7 @@ class DingTalkClient:
         robot_code: str,
         markdown_card: bool,
         logger: None,
+        card_action_callback: Optional[Callable[[dict], Awaitable[None]]] = None,
     ):
         """初始化 WebSocket 连接并自动启动"""
         self.credential = dingtalk_stream.Credential(client_id, client_secret)
@@ -30,6 +62,14 @@ class DingTalkClient:
         # 在 DingTalkClient 中传入自己作为参数，避免循环导入
         self.EchoTextHandler = EchoTextHandler(self)
         self.client.register_callback_handler(dingtalk_stream.chatbot.ChatbotMessage.TOPIC, self.EchoTextHandler)
+        # STREAM-mode card action button click handler. Forwards parsed payload
+        # to the adapter so it can resume paused Dify workflows.
+        self.card_action_callback = card_action_callback
+        self.card_action_handler = DingTalkCardActionHandler(self.client, self._on_card_action)
+        self.client.register_callback_handler(
+            dingtalk_stream.handlers.CallbackHandler.TOPIC_CARD_CALLBACK,
+            self.card_action_handler,
+        )
         self._message_handlers = {
             'example': [],
         }
@@ -39,8 +79,24 @@ class DingTalkClient:
         self.access_token_expiry_time = ''
         self.markdown_card = markdown_card
         self.logger = logger
+        # Legacy access_token used by the OLD oapi.dingtalk.com endpoints
+        # (e.g. /media/upload, which is the only documented way to get an
+        # `@xxx` media_id usable in card Avatar.imageUrl). The new v1.0
+        # token doesn't work there — different auth domain.
+        self.legacy_access_token = ''
+        self.legacy_access_token_expiry_time: typing.Optional[float] = None
         self._stopped = False  # Flag to control the event loop
 
+    async def _on_card_action(self, payload: dict) -> None:
+        """Dispatch a parsed card-action payload to the adapter callback."""
+        if self.card_action_callback is None:
+            return
+        try:
+            await self.card_action_callback(payload)
+        except Exception:
+            if self.logger:
+                await self.logger.error(f'DingTalk card action callback error: {traceback.format_exc()}')
+
     async def get_access_token(self):
         url = 'https://api.dingtalk.com/v1.0/oauth2/accessToken'
         headers = {'Content-Type': 'application/json'}
@@ -429,18 +485,35 @@ class DingTalkClient:
             'Content-Type': 'application/json',
         }
 
+        # For enterprise-internal robots, robotCode == AppKey (client_id).
+        # The dedicated robot_code field is only required for scenario-group
+        # robots or third-party robots; fall back to client_id when empty so
+        # the common single-bot setup keeps working without manual config.
+        robot_code = self.robot_code or self.key
         data = {
-            'robotCode': self.robot_code,
+            'robotCode': robot_code,
             'userIds': [target_id],
             'msgKey': 'sampleText',
             'msgParam': json.dumps({'content': content}),
         }
+        _stdout_logger.info(
+            'DingTalk send_proactive_message_to_one request: robotCode=%s target_id=%s content_len=%d',
+            robot_code,
+            target_id,
+            len(content),
+        )
         try:
             async with httpx.AsyncClient() as client:
                 response = await client.post(url, headers=headers, json=data)
+                _stdout_logger.info(
+                    'DingTalk send_proactive_message_to_one response: status=%d body=%s',
+                    response.status_code,
+                    response.text[:500],
+                )
                 if response.status_code == 200:
                     return
         except Exception:
+            _stdout_logger.exception('DingTalk send_proactive_message_to_one error')
             await self.logger.error(f'failed to send proactive massage to person: {traceback.format_exc()}')
             raise Exception(f'failed to send proactive massage to person: {traceback.format_exc()}')
 
@@ -456,7 +529,7 @@ class DingTalkClient:
         }
 
         data = {
-            'robotCode': self.robot_code,
+            'robotCode': self.robot_code or self.key,
             'openConversationId': target_id,
             'msgKey': 'sampleText',
             'msgParam': json.dumps({'content': content}),
@@ -477,47 +550,334 @@ class DingTalkClient:
         quote_origin: bool = False,
         card_auto_layout: bool = False,
     ):
-        card_data = {}
+        """Create + deliver the streaming chat card for a chatbot reply.
-        card_data['config'] = json.dumps({'autoLayout': card_auto_layout})
-        card_data['content'] = ''
 
-        # 将用户的消息内容作为卡片的查询参数，方便后续处理
+        Replaces the old `dingtalk_stream.AICardReplier`-based path. Returns
-        if incoming_message.message_type == 'text':
+        `(None, out_track_id)` to keep call sites compatible with the
-            card_data['query'] = incoming_message.get_text_list()[0]
+        previous `(card_instance, card_instance_id)` shape — the first slot
+        is unused now that everything is driven by out_track_id.
+        """
+        out_track_id = uuid.uuid4().hex
+        is_group = str(incoming_message.conversation_type) == '2'
+        if is_group:
+            open_space_id = f'dtv1.card//IM_GROUP.{incoming_message.conversation_id}'
         else:
-            card_data['query'] = '...'
+            open_space_id = f'dtv1.card//IM_ROBOT.{incoming_message.sender_staff_id}'
 
-        card_instance = dingtalk_stream.AICardReplier(self.client, incoming_message)
+        card_param_map = {'content': ''}
-        # print(card_instance)
+        if incoming_message.message_type == 'text':
-        # 先投放卡片: https://open.dingtalk.com/document/orgapp/create-and-deliver-cards
+            card_param_map['query'] = incoming_message.get_text_list()[0]
-        card_instance_id = await card_instance.async_create_and_deliver_card(
+        else:
-            temp_card_id,
+            card_param_map['query'] = '...'
-            card_data,
+
+        await self.create_and_deliver_card(
+            card_template_id=temp_card_id,
+            out_track_id=out_track_id,
+            open_space_id=open_space_id,
+            is_group=is_group,
+            card_param_map=card_param_map,
+            card_data_config={'autoLayout': card_auto_layout},
         )
-        return card_instance, card_instance_id
+        return None, out_track_id
 
     async def send_card_message(self, card_instance, card_instance_id: str, content: str, is_final: bool):
-        content_key = 'content'
+        """Stream a single chunk into an existing card's `content` field."""
         try:
-            await card_instance.async_streaming(
+            await self.streaming_update_card(
-                card_instance_id,
+                out_track_id=card_instance_id,
-                content_key=content_key,
+                content_key='content',
                 content_value=content,
                 append=False,
                 finished=is_final,
                 failed=False,
             )
         except Exception as e:
-            self.logger.exception(e)
+            if self.logger:
-            await card_instance.async_streaming(
+                self.logger.exception(e)
-                card_instance_id,
+            await self.streaming_update_card(
-                content_key=content_key,
+                out_track_id=card_instance_id,
+                content_key='content',
                 content_value='',
                 append=False,
                 finished=is_final,
                 failed=True,
             )
 
+    async def create_and_deliver_card(
+        self,
+        *,
+        card_template_id: str,
+        out_track_id: str,
+        open_space_id: str,
+        is_group: bool,
+        card_param_map: Optional[dict] = None,
+        callback_type: str = 'STREAM',
+        callback_route_key: Optional[str] = None,
+        support_forward: bool = True,
+        dynamic_data_source_configs: Optional[list] = None,
+        card_data_config: Optional[dict] = None,
+        at_user_ids: Optional[dict] = None,
+        recipients: Optional[list] = None,
+    ) -> bool:
+        """POST /v1.0/card/instances/createAndDeliver.
+
+        Mirrors the SDK's `async_create_and_deliver_card` shape but exposes
+        the dynamic-data-source config slot so we can register a pull URL
+        for variable-length button lists.
+        """
+        if not await self.check_access_token():
+            await self.get_access_token()
+
+        cardData: dict = {'cardParamMap': _stringify_card_param_map(card_param_map)}
+        if card_data_config is not None:
+            cardData['config'] = json.dumps(card_data_config)
+
+        body: dict = {
+            'cardTemplateId': card_template_id,
+            'outTrackId': out_track_id,
+            'cardData': cardData,
+            'callbackType': callback_type,
+            'openSpaceId': open_space_id,
+            'imGroupOpenSpaceModel': {'supportForward': support_forward},
+            'imRobotOpenSpaceModel': {'supportForward': support_forward},
+        }
+        if callback_type == 'HTTP' and callback_route_key:
+            body['callbackRouteKey'] = callback_route_key
+
+        if is_group:
+            deliver: dict = {'robotCode': self.robot_code or self.key}
+            if at_user_ids:
+                deliver['atUserIds'] = at_user_ids
+            if recipients is not None:
+                deliver['recipients'] = recipients
+            body['imGroupOpenDeliverModel'] = deliver
+        else:
+            body['imRobotOpenDeliverModel'] = {'spaceType': 'IM_ROBOT'}
+
+        if dynamic_data_source_configs:
+            body['openDynamicDataConfig'] = {'dynamicDataSourceConfigs': dynamic_data_source_configs}
+
+        url = f'{DINGTALK_OPENAPI_BASE}/v1.0/card/instances/createAndDeliver'
+        headers = {
+            'x-acs-dingtalk-access-token': self.access_token,
+            'Content-Type': 'application/json',
+        }
+        try:
+            _stdout_logger.info(
+                'DingTalk createAndDeliver request body: %s',
+                json.dumps(body, ensure_ascii=False)[:1500],
+            )
+            async with httpx.AsyncClient() as client:
+                response = await client.post(url, headers=headers, json=body, timeout=30.0)
+                if response.status_code == 200:
+                    _stdout_logger.info(
+                        'DingTalk createAndDeliver response: %s',
+                        response.text[:500],
+                    )
+                    return True
+                _stdout_logger.error(
+                    'DingTalk createAndDeliver failed: status=%s body=%s',
+                    response.status_code,
+                    response.text,
+                )
+                if self.logger:
+                    await self.logger.error(
+                        f'DingTalk createAndDeliver failed: status={response.status_code} body={response.text}'
+                    )
+                return False
+        except Exception:
+            _stdout_logger.exception('DingTalk createAndDeliver error')
+            if self.logger:
+                await self.logger.error(f'DingTalk createAndDeliver error: {traceback.format_exc()}')
+            return False
+
+    async def streaming_update_card(
+        self,
+        *,
+        out_track_id: str,
+        content_key: str,
+        content_value: str,
+        append: bool,
+        finished: bool,
+        failed: bool = False,
+    ) -> bool:
+        """PUT /v1.0/card/streaming.
+
+        Replaces `dingtalk_stream.AICardReplier.async_streaming` — same body
+        shape (outTrackId / guid / key / content / isFull / isFinalize /
+        isError) per the SDK source.
+        """
+        if not await self.check_access_token():
+            await self.get_access_token()
+
+        body = {
+            'outTrackId': out_track_id,
+            'guid': uuid.uuid4().hex,
+            'key': content_key,
+            'content': content_value,
+            'isFull': not append,
+            'isFinalize': finished,
+            'isError': failed,
+        }
+        url = f'{DINGTALK_OPENAPI_BASE}/v1.0/card/streaming'
+        headers = {
+            'x-acs-dingtalk-access-token': self.access_token,
+            'Content-Type': 'application/json',
+        }
+        try:
+            async with httpx.AsyncClient() as client:
+                response = await client.put(url, headers=headers, json=body, timeout=30.0)
+                if response.status_code == 200:
+                    return True
+                if self.logger:
+                    await self.logger.error(
+                        f'DingTalk card streaming failed: status={response.status_code} body={response.text}'
+                    )
+                return False
+        except Exception:
+            if self.logger:
+                await self.logger.error(f'DingTalk card streaming error: {traceback.format_exc()}')
+            return False
+
+    async def update_card_data(
+        self,
+        *,
+        out_track_id: str,
+        card_param_map: Optional[dict] = None,
+        private_data: Optional[dict] = None,
+    ) -> bool:
+        """PUT /v1.0/card/instances — non-streaming card content update."""
+        if not await self.check_access_token():
+            await self.get_access_token()
+
+        body: dict = {
+            'outTrackId': out_track_id,
+            'cardData': {'cardParamMap': _stringify_card_param_map(card_param_map)},
+        }
+        if private_data:
+            body['privateData'] = private_data
+
+        url = f'{DINGTALK_OPENAPI_BASE}/v1.0/card/instances'
+        headers = {
+            'x-acs-dingtalk-access-token': self.access_token,
+            'Content-Type': 'application/json',
+        }
+        try:
+            _stdout_logger.info(
+                'DingTalk update_card_data request: out_track_id=%s body=%s',
+                out_track_id,
+                json.dumps(body, ensure_ascii=False)[:1500],
+            )
+            async with httpx.AsyncClient() as client:
+                response = await client.put(url, headers=headers, json=body, timeout=30.0)
+                _stdout_logger.info(
+                    'DingTalk update_card_data response: status=%d body=%s',
+                    response.status_code,
+                    response.text[:300],
+                )
+                if response.status_code == 200:
+                    return True
+                if self.logger:
+                    await self.logger.error(
+                        f'DingTalk update card failed: status={response.status_code} body={response.text}'
+                    )
+                return False
+        except Exception:
+            _stdout_logger.exception('DingTalk update_card_data error')
+            if self.logger:
+                await self.logger.error(f'DingTalk update card error: {traceback.format_exc()}')
+            return False
+
+    async def get_legacy_access_token(self) -> Optional[str]:
+        """Fetch the LEGACY (oapi.dingtalk.com) access_token. This is a
+        different auth domain from the v1.0 token cached in
+        ``self.access_token`` — only the legacy token authorises the
+        ``/media/upload`` endpoint that returns an ``@xxx`` media_id
+        consumable by card components like Avatar.imageUrl.
+
+        Returns the token string on success, None on failure. Caches
+        with a 60s safety margin before the documented 7200s expiry.
+        """
+        now = time.time()
+        if (
+            self.legacy_access_token
+            and self.legacy_access_token_expiry_time
+            and now < self.legacy_access_token_expiry_time
+        ):
+            return self.legacy_access_token
+
+        url = 'https://oapi.dingtalk.com/gettoken'
+        try:
+            async with httpx.AsyncClient() as client:
+                response = await client.get(url, params={'appkey': self.key, 'appsecret': self.secret}, timeout=15.0)
+            data = response.json() if response.status_code == 200 else {}
+            if data.get('errcode') == 0 and data.get('access_token'):
+                self.legacy_access_token = data['access_token']
+                expires_in = int(data.get('expires_in', 7200))
+                self.legacy_access_token_expiry_time = now + expires_in - 60
+                return self.legacy_access_token
+            if self.logger:
+                await self.logger.error(
+                    f'DingTalk legacy gettoken failed: status={response.status_code} body={response.text[:200]}'
+                )
+        except Exception:
+            _stdout_logger.exception('DingTalk legacy gettoken error')
+            if self.logger:
+                await self.logger.error(f'DingTalk legacy gettoken error: {traceback.format_exc()}')
+        return None
+
+    async def upload_image_media(self, file_path: str) -> Optional[str]:
+        """Upload an image file to DingTalk media storage and return the
+        ``@xxx`` media_id, which can be passed straight into card variables
+        like Avatar.imageUrl. Endpoint:
+
+            POST https://oapi.dingtalk.com/media/upload?access_token=…&type=image
+
+        Returns the media_id on success, None on any failure (caller
+        should handle a None gracefully — DingTalk falls back to a
+        default avatar when imageUrl is empty/unknown).
+        """
+        if not os.path.exists(file_path):
+            if self.logger:
+                await self.logger.error(f'DingTalk upload_image_media: file not found {file_path}')
+            return None
+
+        token = await self.get_legacy_access_token()
+        if not token:
+            return None
+
+        url = 'https://oapi.dingtalk.com/media/upload'
+        try:
+            with open(file_path, 'rb') as f:
+                file_bytes = f.read()
+            file_name = os.path.basename(file_path)
+            # Best-effort content-type guess; DingTalk accepts the major image
+            # mime types and otherwise infers from the bytes.
+            ext = os.path.splitext(file_name)[1].lower().lstrip('.')
+            mime = {'png': 'image/png', 'jpg': 'image/jpeg', 'jpeg': 'image/jpeg', 'gif': 'image/gif'}.get(
+                ext, 'application/octet-stream'
+            )
+            async with httpx.AsyncClient() as client:
+                response = await client.post(
+                    url,
+                    params={'access_token': token, 'type': 'image'},
+                    files={'media': (file_name, file_bytes, mime)},
+                    timeout=30.0,
+                )
+            data = response.json() if response.status_code == 200 else {}
+            if data.get('errcode') == 0 and data.get('media_id'):
+                _stdout_logger.info('DingTalk upload_image_media OK: media_id=%s', data['media_id'])
+                return data['media_id']
+            if self.logger:
+                await self.logger.error(
+                    f'DingTalk upload_image_media failed: status={response.status_code} body={response.text[:300]}'
+                )
+        except Exception:
+            _stdout_logger.exception('DingTalk upload_image_media error')
+            if self.logger:
+                await self.logger.error(f'DingTalk upload_image_media error: {traceback.format_exc()}')
+        return None
+
     async def start(self):
         """启动 WebSocket 连接，监听消息"""
         self._stopped = False

src/langbot/libs/dingtalk_api/card_callback.py

@@ -0,0 +1,106 @@
+"""STREAM-mode handler for DingTalk card action button clicks.
+
+DingTalk delivers card-action callbacks over the same WebSocket stream used
+for chatbot messages, under the topic `/v1.0/card/instances/callback`. This
+module subclasses `dingtalk_stream.CallbackHandler` and forwards the parsed
+payload to a coroutine the adapter registers, so the resume-paused-workflow
+logic stays in the platform adapter where it belongs.
+
+The `CardCallbackMessage` returned by `from_dict` exposes:
+
+* `card_instance_id` (from `outTrackId`) — the card whose button was clicked
+* `user_id`  — the clicker's userId
+* `content`  — parsed JSON; the click params live here. Where exactly inside
+               `content` they sit depends on the template binding. We probe
+               the common paths.
+* `extension` — parsed JSON; any extra data we set when delivering the card.
+"""
+
+from __future__ import annotations
+
+from typing import Awaitable, Callable, Optional
+
+import dingtalk_stream  # type: ignore
+from dingtalk_stream import AckMessage
+from dingtalk_stream.card_callback import CardCallbackMessage
+
+
+_PARAM_PATHS = (
+    ('params',),
+    ('cardPrivateData', 'params'),
+    ('userPrivateData', 'params'),
+    ('actionData', 'cardPrivateData', 'params'),
+)
+
+
+def _extract_params(content: dict) -> dict:
+    """Return the action params dict regardless of where the template put it."""
+    for path in _PARAM_PATHS:
+        node = content
+        for key in path:
+            if not isinstance(node, dict):
+                node = None
+                break
+            node = node.get(key)
+            if node is None:
+                break
+        if isinstance(node, dict) and node:
+            return node
+    return {}
+
+
+def _merge_params(*sources: dict) -> dict:
+    merged = {}
+    for source in sources:
+        if isinstance(source, dict):
+            merged.update(source)
+    return merged
+
+
+class DingTalkCardActionHandler(dingtalk_stream.CallbackHandler):
+    def __init__(
+        self,
+        dingtalk_stream_client,
+        on_action: Optional[Callable[[dict], Awaitable[None]]] = None,
+    ):
+        super().__init__()
+        self.dingtalk_client = dingtalk_stream_client
+        self.on_action = on_action
+
+    async def process(self, callback: dingtalk_stream.CallbackMessage):
+        try:
+            message = CardCallbackMessage.from_dict(callback.data)
+            content = message.content if isinstance(message.content, dict) else {}
+
+            # `CardCallbackMessage.from_dict` does not surface `actionId` (the
+            # top-level field that ButtonGroup's sendCardRequest event puts
+            # there). Pull it from the raw callback.data instead.
+            raw = callback.data if isinstance(callback.data, dict) else {}
+            params = _merge_params(_extract_params(content), _extract_params(raw))
+            action_id = raw.get('actionId') or ''
+            if not action_id:
+                # Some templates nest it under actionData / cardPrivateData.
+                action_data = raw.get('actionData') or {}
+                if isinstance(action_data, dict):
+                    action_id = action_data.get('actionId') or action_id
+                    if not action_id:
+                        cpd = action_data.get('cardPrivateData') or {}
+                        if isinstance(cpd, dict):
+                            ids = cpd.get('actionIds')
+                            if isinstance(ids, list) and ids:
+                                action_id = str(ids[0])
+
+            payload = {
+                'out_track_id': message.card_instance_id,
+                'user_id': message.user_id,
+                'corp_id': message.corp_id,
+                'action_id': action_id,
+                'params': params,
+                'raw_content': message.content,
+                'extension': message.extension if isinstance(message.extension, dict) else {},
+            }
+            if self.on_action is not None:
+                await self.on_action(payload)
+        except Exception as e:
+            self.logger.error(f'DingTalkCardActionHandler.process error: {e}')
+        return AckMessage.STATUS_OK, 'OK'

src/langbot/libs/qq_official_api/api.py

@@ -12,6 +12,142 @@ import traceback
 from cryptography.hazmat.primitives.asymmetric import ed25519
 
 
+QQ_SELECT_ACTION_PREFIX = '__langbot_select__:'
+
+
+def get_select_field_options(form_data: dict) -> tuple[str, list[str]]:
+    """Return the active select field name and its display/submission values."""
+    field_name = str(form_data.get('_current_input_field') or '').strip()
+    if not field_name:
+        return '', []
+
+    field = next(
+        (
+            item
+            for item in form_data.get('input_defs') or []
+            if str(item.get('output_variable_name') or '').strip() == field_name
+        ),
+        None,
+    )
+    if not field or str(field.get('type') or '').strip().lower() != 'select':
+        return '', []
+
+    source = field.get('option_source') or {}
+    source_value = source.get('value') if isinstance(source, dict) else None
+    if isinstance(source_value, list):
+        return field_name, [str(item) for item in source_value]
+    if isinstance(source_value, str):
+        return field_name, [part.strip() for part in source_value.splitlines() if part.strip()]
+
+    options = field.get('options')
+    if not isinstance(options, list):
+        return field_name, []
+    values = []
+    for item in options:
+        if isinstance(item, dict):
+            values.append(str(item.get('label') or item.get('value') or ''))
+        else:
+            values.append(str(item))
+    return field_name, [value for value in values if value]
+
+
+def build_keyboard_from_select_field(form_data: dict, *, buttons_per_row: int | None = None) -> dict:
+    """Build callback buttons for the currently active Dify select field."""
+    _, options = get_select_field_options(form_data)
+    visible_options = options[:25]
+    if buttons_per_row is None:
+        # Keep small choices readable while fitting up to QQ's 5x5 limit.
+        buttons_per_row = min(5, max(2, (len(visible_options) + 4) // 5))
+    selection_actions = [
+        {
+            'id': f'{QQ_SELECT_ACTION_PREFIX}{idx}',
+            'title': option,
+            'button_style': 'secondary',
+        }
+        for idx, option in enumerate(visible_options)
+    ]
+    return build_keyboard_from_form({'actions': selection_actions}, buttons_per_row=buttons_per_row)
+
+
+def resolve_select_button_action(form_data: dict, action_id: str) -> tuple[str, str] | None:
+    """Resolve a select-button callback to ``(field_name, option_value)``."""
+    if not action_id.startswith(QQ_SELECT_ACTION_PREFIX):
+        return None
+    try:
+        option_index = int(action_id[len(QQ_SELECT_ACTION_PREFIX) :])
+    except ValueError:
+        return None
+
+    field_name, options = get_select_field_options(form_data)
+    if not field_name or option_index < 0 or option_index >= len(options) or option_index >= 25:
+        return None
+    return field_name, options[option_index]
+
+
+def build_keyboard_from_form(form_data: dict, *, buttons_per_row: int = 2) -> dict:
+    """Build a QQ keyboard JSON payload from a Dify human-input form_data.
+
+    Each Dify ``action`` becomes a callback button (``action.type=1``)
+    whose ``data`` is set directly to the Dify ``action_id``. The
+    INTERACTION_CREATE event carries this back as
+    ``data.resolved.button_data`` so the adapter can match the click to
+    the originating form.
+
+    Layout limits per spec: max 5 rows, max 5 buttons per row. We default
+    to 2 buttons per row for legibility; oversized button lists wrap
+    onto additional rows and overflow gets dropped (max 25 visible).
+
+    Args:
+        form_data: Dify ``{"actions": [{"id", "title", "button_style"}, ...]}``.
+        buttons_per_row: 1..5. Mobile UI looks best at 2.
+
+    Returns:
+        ``{"content": {"rows": [{"buttons": [...]}]}}``.
+    """
+    actions = list(form_data.get('actions') or [])[:25]  # 5×5 hard cap
+    buttons_per_row = max(1, min(5, buttons_per_row))
+
+    def _button(idx: int, action: dict) -> dict:
+        action_id = str(action.get('id') or '')
+        label = str(action.get('title') or action_id or f'选项 {idx + 1}')
+        style_raw = (action.get('button_style') or '').lower()
+        # QQ: 0 灰色线框, 1 蓝色线框. Highlight the primary / first action.
+        if style_raw == 'primary' or (style_raw == '' and idx == 0):
+            style = 1
+        else:
+            style = 0
+        return {
+            'id': str(idx + 1),
+            'render_data': {
+                'label': label,
+                # Shown after the user clicks — gives local "已选择" feedback
+                # without a follow-up message. Style mimics DingTalk/Lark's
+                # in-card selection state.
+                'visited_label': f'✓ {label}',
+                'style': style,
+            },
+            'action': {
+                'type': 1,  # callback button
+                'permission': {'type': 2},  # everyone can click
+                'data': action_id,
+                'unsupport_tips': '当前客户端版本不支持此按钮，请升级 QQ',
+            },
+        }
+
+    rows = []
+    for row_start in range(0, len(actions), buttons_per_row):
+        row_actions = actions[row_start : row_start + buttons_per_row]
+        rows.append(
+            {
+                'buttons': [_button(row_start + j, a) for j, a in enumerate(row_actions)],
+            }
+        )
+        if len(rows) >= 5:
+            break
+
+    return {'content': {'rows': rows}}
+
+
 class QQOfficialClient:
     def __init__(self, secret: str, token: str, app_id: str, logger: None, unified_mode: bool = False):
         self.unified_mode = unified_mode
@@ -30,6 +166,10 @@ class QQOfficialClient:
         self.token = token
         self.app_id = app_id
         self._message_handlers = {}
+        # Single optional handler for INTERACTION_CREATE (button click). We
+        # don't multiplex like message handlers — only the adapter cares,
+        # and the click<->resume path needs a single source of truth.
+        self._interaction_handler: Optional[Callable[[Dict[str, Any], Optional[str]], Any]] = None
         self.base_url = 'https://api.sgroup.qq.com'
         self.access_token = ''
         self.access_token_expiry_time = None
@@ -107,6 +247,23 @@ class QQOfficialClient:
                 return response, 200
 
             if payload.get('op') == 0:
+                # INTERACTION_CREATE (button click) skips ``get_message`` —
+                # that helper only flattens message-event fields and would
+                # drop ``data.resolved.button_data`` / ``data.button_id``.
+                if payload.get('t') == 'INTERACTION_CREATE':
+                    if self._interaction_handler:
+                        try:
+                            d = payload.get('d') or {}
+                            # Top-level ``id`` is the ws/event id used as
+                            # ``event_id`` for passive replies. ``d.id``
+                            # is the interaction id used for ACK. Do not
+                            # confuse the two — QQ rejects misuse with
+                            # 40034025.
+                            ws_event_id = payload.get('id')
+                            await self._interaction_handler(d, ws_event_id)
+                        except Exception:
+                            await self.logger.error(f'Error in interaction handler: {traceback.format_exc()}')
+                    return {'code': 0, 'message': 'success'}
                 message_data = await self.get_message(payload)
                 if message_data:
                     event = QQOfficialEvent.from_payload(message_data)
@@ -133,6 +290,21 @@ class QQOfficialClient:
 
         return decorator
 
+    def on_interaction(self):
+        """Register a single handler for INTERACTION_CREATE events.
+
+        The handler receives ``(data_dict, interaction_id)`` — the raw
+        ``d`` payload plus the top-level ``id`` field (the interaction
+        id, needed for the PUT /interactions/{id} ack and for reuse as
+        an ``event_id`` on the resumed reply within 30 minutes).
+        """
+
+        def decorator(func: Callable[[Dict[str, Any], Optional[str]], Any]):
+            self._interaction_handler = func
+            return func
+
+        return decorator
+
     async def _handle_message(self, event: QQOfficialEvent):
         """处理消息事件"""
         msg_type = event.t
@@ -177,8 +349,20 @@ class QQOfficialClient:
         content_type = attachment.get('content_type', '')
         return content_type.startswith('image/')
 
-    async def send_private_text_msg(self, user_openid: str, content: str, msg_id: str):
+    async def send_private_text_msg(
-        """发送私聊消息"""
+        self,
+        user_openid: str,
+        content: str,
+        msg_id: Optional[str] = None,
+        event_id: Optional[str] = None,
+        msg_seq: int = 1,
+    ):
+        """Send a c2c text message.
+
+        Either ``msg_id`` (inbound user msg, free passive reply) or
+        ``event_id`` (e.g. INTERACTION_CREATE id, valid 30 min) is
+        required. Without either, the call costs the proactive-send quota.
+        """
         if not await self.check_access_token():
             await self.get_access_token()
 
@@ -188,11 +372,15 @@ class QQOfficialClient:
                 'Authorization': f'QQBot {self.access_token}',
                 'Content-Type': 'application/json',
             }
-            data = {
+            data: dict[str, Any] = {
                 'content': content,
                 'msg_type': 0,
-                'msg_id': msg_id,
+                'msg_seq': msg_seq,
             }
+            if msg_id:
+                data['msg_id'] = msg_id
+            if event_id:
+                data['event_id'] = event_id
             response = await client.post(url, headers=headers, json=data)
             response_data = response.json()
             if response.status_code == 200:
@@ -201,8 +389,19 @@ class QQOfficialClient:
                 await self.logger.error(f'Failed to send private message: {response_data}')
                 raise ValueError(response)
 
-    async def send_group_text_msg(self, group_openid: str, content: str, msg_id: str):
+    async def send_group_text_msg(
-        """发送群聊消息"""
+        self,
+        group_openid: str,
+        content: str,
+        msg_id: Optional[str] = None,
+        event_id: Optional[str] = None,
+        msg_seq: int = 1,
+    ):
+        """Send a group text message.
+
+        Either ``msg_id`` or ``event_id`` is required (see
+        :meth:`send_private_text_msg` for the distinction).
+        """
         if not await self.check_access_token():
             await self.get_access_token()
 
@@ -212,11 +411,15 @@ class QQOfficialClient:
                 'Authorization': f'QQBot {self.access_token}',
                 'Content-Type': 'application/json',
             }
-            data = {
+            data: dict[str, Any] = {
                 'content': content,
                 'msg_type': 0,
-                'msg_id': msg_id,
+                'msg_seq': msg_seq,
             }
+            if msg_id:
+                data['msg_id'] = msg_id
+            if event_id:
+                data['event_id'] = event_id
             response = await client.post(url, headers=headers, json=data)
             if response.status_code == 200:
                 return
@@ -485,6 +688,107 @@ class QQOfficialClient:
                 raise Exception(f'Failed to send stream message: HTTP {response.status_code} {response.text}')
             return response.json()
 
+    async def send_markdown_keyboard(
+        self,
+        target_type: str,
+        target_id: str,
+        markdown_content: str,
+        keyboard: Optional[dict] = None,
+        msg_id: Optional[str] = None,
+        event_id: Optional[str] = None,
+        msg_seq: int = 1,
+    ) -> dict:
+        """Send a ``msg_type=2`` (markdown) message carrying a keyboard.
+
+        The keyboard ride-along is the only documented way to attach
+        buttons in QQ official; pure keyboard-only messages are not
+        accepted by the server (markdown content is required).
+
+        Args:
+            target_type: 'c2c' (single chat), 'group', 'channel' (text
+                channel — uses POST /channels/{id}/messages instead of v2).
+            target_id: openid for c2c/group, channel_id for channel.
+            markdown_content: Plain markdown text shown above the buttons.
+            keyboard: ``{'content': {'rows': [{'buttons': [...]}]}}`` per
+                the official spec. Use :func:`build_keyboard_from_form`
+                to construct from Dify form_data.
+            msg_id: Inbound user message id; turns this into a passive
+                reply (preferred — no monthly quota cost).
+            event_id: Use ``INTERACTION_CREATE`` event id from a prior
+                button click to keep within the 30-minute passive window
+                without an inbound msg_id.
+            msg_seq: De-dup counter when reusing msg_id.
+        """
+        if not await self.check_access_token():
+            await self.get_access_token()
+
+        if target_type == 'c2c':
+            url = f'{self.base_url}/v2/users/{target_id}/messages'
+        elif target_type == 'group':
+            url = f'{self.base_url}/v2/groups/{target_id}/messages'
+        elif target_type == 'channel':
+            url = f'{self.base_url}/channels/{target_id}/messages'
+        else:
+            raise ValueError(f'Unsupported target_type for markdown+keyboard: {target_type}')
+
+        body: dict[str, Any] = {
+            'msg_type': 2,
+            'markdown': {'content': markdown_content},
+            'msg_seq': msg_seq,
+        }
+        if keyboard and keyboard.get('content', {}).get('rows'):
+            body['keyboard'] = keyboard
+        if msg_id:
+            body['msg_id'] = msg_id
+        if event_id:
+            body['event_id'] = event_id
+
+        async with httpx.AsyncClient(timeout=30) as client:
+            headers = {
+                'Authorization': f'QQBot {self.access_token}',
+                'Content-Type': 'application/json',
+            }
+            response = await client.post(url, headers=headers, json=body)
+            if response.status_code != 200:
+                await self.logger.error(
+                    f'Failed to send markdown+keyboard: HTTP {response.status_code} {response.text}'
+                )
+                raise Exception(f'Failed to send markdown+keyboard: HTTP {response.status_code} {response.text}')
+            return response.json()
+
+    async def ack_interaction(self, interaction_id: str, code: int = 0) -> None:
+        """Acknowledge a button-click INTERACTION_CREATE event.
+
+        QQ keeps the client in a loading spinner until this ack is
+        received. Should be called as soon as the click is parsed, before
+        any heavier downstream work (the actual workflow resume can run
+        async).
+
+        Args:
+            interaction_id: The ``id`` field from the INTERACTION_CREATE event.
+            code: 0=success, 1=fail, 2=rate-limited, 3=duplicate, 4=no
+                permission, 5=admin only. Default 0.
+        """
+        if not interaction_id:
+            return
+        if not await self.check_access_token():
+            await self.get_access_token()
+
+        url = f'{self.base_url}/interactions/{interaction_id}'
+        async with httpx.AsyncClient(timeout=10) as client:
+            headers = {
+                'Authorization': f'QQBot {self.access_token}',
+                'Content-Type': 'application/json',
+            }
+            try:
+                response = await client.put(url, headers=headers, json={'code': code})
+                if response.status_code >= 400:
+                    await self.logger.warning(
+                        f'ack_interaction non-success: HTTP {response.status_code} {response.text}'
+                    )
+            except Exception as e:
+                await self.logger.warning(f'ack_interaction error (non-fatal): {e}')
+
     async def is_token_expired(self):
         """检查token是否过期"""
         if self.access_token_expiry_time is None:
@@ -653,6 +957,12 @@ class QQOfficialClient:
                     d = payload.get('d', {})
                     s = payload.get('s')
                     t = payload.get('t')
+                    # Top-level event id, distinct from `d.id`. Per QQ
+                    # spec this is the only value accepted as ``event_id``
+                    # in subsequent passive-reply send-message calls
+                    # (``d.id`` for INTERACTION_CREATE is the interaction
+                    # id, used solely for PUT /interactions/{id} ack).
+                    ws_event_id = payload.get('id')
 
                     if not isinstance(d, dict):
                         d = {}
@@ -731,7 +1041,22 @@ class QQOfficialClient:
 
                         else:
                             await self.logger.debug(f'Received event: {t}, seq={s}')
-                            if on_event:
+                            # INTERACTION_CREATE bypasses the regular
+                            # on_event dispatcher so the adapter sees the
+                            # top-level ws_event_id (needed as event_id
+                            # for the resumed reply) — same shape as the
+                            # webhook handler.
+                            if t == 'INTERACTION_CREATE':
+                                if self._interaction_handler:
+                                    try:
+                                        result = self._interaction_handler(d, ws_event_id)
+                                        if asyncio.iscoroutine(result):
+                                            await result
+                                    except Exception:
+                                        await self.logger.error(
+                                            f'Error in interaction handler (ws): {traceback.format_exc()}'
+                                        )
+                            elif on_event:
                                 try:
                                     result = on_event(t, d)
                                     if asyncio.iscoroutine(result):

src/langbot/libs/wecom_ai_bot_api/ws_client.py

@@ -20,7 +20,19 @@ from typing import Any, Callable, Optional
 import aiohttp
 
 from langbot.libs.wecom_ai_bot_api import wecombotevent
-from langbot.libs.wecom_ai_bot_api.api import parse_wecom_bot_message, StreamSession
+from langbot.libs.wecom_ai_bot_api.api import (
+    parse_wecom_bot_message,
+    StreamSession,
+    build_human_input_template_card_payload,
+    build_human_input_text_prompt,
+    build_button_interaction_update_card,
+    build_multiple_interaction_update_card,
+    extract_template_card_action,
+    extract_template_card_event_payload,
+    extract_template_card_selections,
+    extract_wecom_event_type,
+    parse_select_button_action,
+)
 from langbot.pkg.platform.logger import EventLogger
 
 DEFAULT_WS_URL = 'wss://openws.work.weixin.qq.com'
@@ -43,6 +55,10 @@ def _generate_req_id(prefix: str) -> str:
     return f'{prefix}_{ts}_{rand}'
 
 
+def _frame_snippet(frame: dict, limit: int = 1000) -> str:
+    return json.dumps(frame, ensure_ascii=False, default=str)[:limit]
+
+
 class WecomBotWsClient:
     """WeChat Work AI Bot WebSocket long connection client.
 
@@ -103,6 +119,22 @@ class WecomBotWsClient:
         # msg_id -> feedback_id (for associating feedback with message)
         self._msg_feedback_ids: dict[str, str] = {}  # msg_id -> feedback_id
 
+        # Dify human-input pause state for ws mode. Keys are task_id (echoed
+        # back in template_card_event.TaskId so we can rebuild the session
+        # context on click).
+        # task_id -> {form_data, msg_id, user_id, chat_id, stream_id, req_id}
+        self._pending_forms_by_task: dict[str, dict] = {}
+        # Reverse: msg_id -> task_id (for cleanup when stream finishes).
+        self._task_id_by_msg: dict[str, str] = {}
+        # Optional card-action callback registered by the adapter.
+        # Signature mirrors the http-mode WecomBotClient:
+        #   async def callback(session, action_id, task_id, raw_event) -> None
+        self._card_action_callback: Optional[Callable] = None
+        # Optional `source` block injected into every interactive
+        # template_card the client builds via `push_form_pause`. Set via
+        # `set_card_source` from the adapter after reading config.
+        self.card_source: Optional[dict] = None
+
     # ── Public API ──────────────────────────────────────────────────
 
     async def connect(self):
@@ -236,6 +268,132 @@ class WecomBotWsClient:
         }
         return await self._send_reply(req_id, body)
 
+    async def reply_template_card(self, req_id: str, card_payload: dict[str, Any]) -> Optional[dict]:
+        """Send a template_card (button_interaction etc.) reply.
+
+        Args:
+            req_id: The req_id from the original message frame.
+            card_payload: Body produced by ``build_button_interaction_payload``;
+                must contain ``msgtype`` and ``template_card`` keys.
+
+        Returns:
+            ACK frame dict, or None on failure.
+        """
+        return await self._send_reply(req_id, card_payload)
+
+    async def update_template_card(
+        self,
+        req_id: str,
+        template_card: dict[str, Any],
+    ) -> Optional[dict]:
+        """Update an existing template_card via WebSocket.
+
+        Uses the ``aibot_respond_update_msg`` command.  Must be called
+        within 5 seconds of receiving the ``template_card_event`` callback,
+        using the **same req_id** from that callback.
+
+        The ``template_card`` dict should contain ``card_type`` and the
+        new content fields (e.g. ``main_title``, ``button_list`` with
+        disabled buttons and ``replace_text``).
+
+        Returns:
+            ACK frame dict, or None on failure.
+        """
+        body: dict[str, Any] = {
+            'response_type': 'update_template_card',
+            'template_card': template_card,
+        }
+        return await self._send_reply(req_id, body, cmd=CMD_RESPOND_UPDATE)
+
+    def set_card_action_callback(self, callback: Callable) -> None:
+        """Register the button-click handler.
+
+        ``async def callback(session, action_id, task_id, raw_event) -> None``
+        — same signature as the http-mode WecomBotClient version so the
+        adapter can hand both off to the same coroutine.
+        """
+        self._card_action_callback = callback
+
+    def set_card_source(self, source: Optional[dict]) -> None:
+        """Set the `source` block injected into every interactive
+        template_card pushed via `push_form_pause`. Pass None to clear."""
+        self.card_source = source
+
+    async def push_form_pause(
+        self, msg_id: str, form_data: dict, task_id: Optional[str] = None
+    ) -> tuple[bool, Optional[str], Optional[str]]:
+        """Attach a Dify human-input pause to the active stream and send
+        the button_interaction card immediately.
+
+        ws mode has no notion of polled "followup" responses — each reply
+        is a one-shot frame send. So unlike the http path (which defers
+        card delivery to the next followup), here we just craft the card
+        and reply with it on the original req_id. The corresponding stream
+        session is then torn down so subsequent chunks don't re-send.
+
+        Returns:
+            ``(ok, stream_id, task_id)``. ``ok=False`` if no active stream
+            for this msg_id (e.g. message arrived in non-stream mode).
+        """
+        key = self._stream_ids.get(msg_id)
+        if not key:
+            return False, None, None
+        req_id, stream_id = key.split('|', 1)
+
+        if not task_id:
+            task_id = f'dify-{secrets.token_hex(12)}'
+
+        session_info = self._stream_sessions.get(msg_id) or {}
+        text_prompt = build_human_input_text_prompt(form_data)
+        if text_prompt:
+            try:
+                ack = await self.reply_text(req_id, text_prompt)
+                if ack is None:
+                    return False, stream_id, None
+            except Exception:
+                await self.logger.error(f'Failed to send human-input text prompt: {traceback.format_exc()}')
+                return False, stream_id, None
+
+            self._stream_ids.pop(msg_id, None)
+            self._stream_last_content.pop(msg_id, None)
+            self._stream_sessions.pop(msg_id, None)
+            return True, stream_id, None
+
+        self._pending_forms_by_task[task_id] = {
+            'form_data': form_data,
+            'msg_id': msg_id,
+            'user_id': session_info.get('user_id', ''),
+            'chat_id': session_info.get('chat_id', ''),
+            'stream_id': stream_id,
+            'req_id': req_id,
+        }
+        self._task_id_by_msg[msg_id] = task_id
+
+        card_payload = build_human_input_template_card_payload(
+            form_data,
+            task_id,
+            source=self.card_source,
+            select_as_buttons=True,
+        )
+        try:
+            await self.reply_template_card(req_id, card_payload)
+        except Exception:
+            await self.logger.error(f'Failed to send button_interaction card: {traceback.format_exc()}')
+            # Roll back the bookkeeping so the next attempt isn't blocked.
+            self._pending_forms_by_task.pop(task_id, None)
+            self._task_id_by_msg.pop(msg_id, None)
+            return False, stream_id, None
+
+        # Tear down the stream — WeCom expects either stream chunks OR a
+        # template_card, not both on the same req_id. Subsequent
+        # push_stream_chunk calls for this msg_id become no-ops.
+        self._stream_ids.pop(msg_id, None)
+        self._stream_last_content.pop(msg_id, None)
+        # Keep _stream_sessions so the button callback can still resolve
+        # user/chat context; it gets cleaned up when the click fires.
+
+        return True, stream_id, task_id
+
     async def send_message(self, chat_id: str, content: str, msgtype: str = 'markdown') -> Optional[dict]:
         """Proactively send a message to a specified chat.
 
@@ -258,6 +416,23 @@ class WecomBotWsClient:
             body['text'] = {'content': content}
         return await self._send_reply(req_id, body, cmd=CMD_SEND_MSG)
 
+    async def send_template_card(self, chat_id: str, card_payload: dict[str, Any]) -> Optional[dict]:
+        """Proactively push a template_card to a chat.
+
+        Used for the resumed-workflow path (button click → new query):
+        synthetic events have no inbound req_id to reply against, so we
+        fall back to proactive ``aibot_send_msg`` instead of reply mode.
+
+        Args:
+            chat_id: userid (single chat) or chatid (group chat).
+            card_payload: ``{"msgtype": "template_card", "template_card": {...}}``
+                as produced by :func:`build_button_interaction_payload`.
+        """
+        req_id = _generate_req_id(CMD_SEND_MSG)
+        body = dict(card_payload)
+        body['chatid'] = chat_id
+        return await self._send_reply(req_id, body, cmd=CMD_SEND_MSG)
+
     async def push_stream_chunk(self, msg_id: str, content: str, is_final: bool = False) -> bool:
         """Push a streaming chunk for a given message ID.
 
@@ -276,10 +451,31 @@ class WecomBotWsClient:
             return False
         req_id, stream_id = key.split('|', 1)
         try:
+            previous_content = self._stream_last_content.get(msg_id, '')
+            if previous_content and content.startswith(previous_content):
+                next_content = content
+            elif previous_content and not content:
+                next_content = previous_content
+            else:
+                next_content = previous_content + content if previous_content else content
+
             # Skip sending if content hasn't changed (e.g. during tool call argument streaming)
-            if not is_final and content == self._stream_last_content.get(msg_id):
+            if not is_final and next_content == previous_content:
                 return True
 
+            # Skip empty/whitespace-only snapshots — the runner injects a
+            # zero-width space ('​') as a pass-through when workflow_paused
+            # fires without any preceding LLM output.  WeCom renders that
+            # as an empty bubble that sits before the form card; skip it.
+            # NOTE: Python str.strip() does NOT strip ​, so we use
+            # a regex that treats any character with Unicode category Zs
+            # (separator space) or Cf (format char like ZWS) as blank.
+            if not is_final:
+                import re as _re
+
+                if not _re.sub(r'[\s​‌‍]', '', next_content):
+                    return True
+
             # Generate feedback_id for final chunk
             feedback_id = ''
             if is_final:
@@ -290,8 +486,10 @@ class WecomBotWsClient:
                 if session_info:
                     self._feedback_sessions[feedback_id] = session_info
 
-            await self.reply_stream(req_id, stream_id, content, finish=is_final, feedback_id=feedback_id)
+            # WeCom replaces the displayed stream content on each refresh, so
-            self._stream_last_content[msg_id] = content
+            # every frame must contain the complete snapshot, not only a delta.
+            await self.reply_stream(req_id, stream_id, next_content, finish=is_final, feedback_id=feedback_id)
+            self._stream_last_content[msg_id] = next_content
             if is_final:
                 self._stream_ids.pop(msg_id, None)
                 self._stream_last_content.pop(msg_id, None)
@@ -465,7 +663,7 @@ class WecomBotWsClient:
             return
 
         # Unknown frame
-        await self.logger.warning(f'Unknown frame: {json.dumps(frame, ensure_ascii=False)[:200]}')
+        await self.logger.warning(f'Unknown frame: {_frame_snippet(frame)}')
 
     async def _handle_message_callback(self, frame: dict):
         """Handle an incoming message callback frame."""
@@ -473,6 +671,13 @@ class WecomBotWsClient:
             body = frame.get('body', {})
             req_id = frame.get('headers', {}).get('req_id', '')
 
+            event_type = extract_wecom_event_type(body)
+            if event_type == 'template_card_event':
+                await self._handle_template_card_event_frame(frame, body)
+                return
+            if event_type:
+                await self.logger.debug(f'Received msg_callback event_type={event_type}: {_frame_snippet(frame)}')
+
             # Parse message using shared logic
             message_data = await parse_wecom_bot_message(body, self.encoding_aes_key, self.logger)
             if not message_data:
@@ -506,8 +711,12 @@ class WecomBotWsClient:
             body = frame.get('body', {})
             req_id = frame.get('headers', {}).get('req_id', '')
 
-            event_info = body.get('event', {})
+            event_info = body.get('event', {}) if isinstance(body.get('event'), dict) else body
-            event_type = event_info.get('eventtype', '')
+            event_type = extract_wecom_event_type(body)
+            if not event_type:
+                await self.logger.warning(f'Received event_callback without event_type: {_frame_snippet(frame)}')
+            else:
+                await self.logger.debug(f'Received event_callback event_type={event_type}')
 
             message_data = {
                 'msgtype': 'event',
@@ -568,6 +777,10 @@ class WecomBotWsClient:
                         await self.logger.error(f'Error in feedback handler: {traceback.format_exc()}')
                 return
 
+            if event_type == 'template_card_event':
+                await self._handle_template_card_event_frame(frame, body)
+                return
+
             event = wecombotevent.WecomBotEvent(message_data)
 
             if event_type in self._message_handlers:
@@ -581,6 +794,72 @@ class WecomBotWsClient:
         except Exception:
             await self.logger.error(f'Error in event callback: {traceback.format_exc()}')
 
+    async def _handle_template_card_event_frame(self, frame: dict, body: dict):
+        """Handle template_card_event frames from event_callback or msg_callback."""
+        tce = extract_template_card_event_payload(body)
+        task_id, event_key, card_type = extract_template_card_action(tce)
+        await self.logger.info(
+            f'Received template_card_event (ws): task_id={task_id} event_key={event_key!r} card_type={card_type}'
+        )
+
+        pending = self._pending_forms_by_task.get(task_id)
+        if pending is None:
+            await self.logger.warning(f'No pending_form found for task_id={task_id} (ws); card event ignored')
+            return
+
+        req_id_for_update = frame.get('headers', {}).get('req_id', '')
+        form_data = pending.get('form_data', {}) or {}
+        selections = extract_template_card_selections(tce, form_data)
+        if not selections:
+            selections = parse_select_button_action(event_key, form_data)
+        if card_type == 'multiple_interaction' and not selections:
+            await self.logger.warning(
+                f'multiple_interaction callback has no parseable selections (ws): raw={str(tce)[:1000]}'
+            )
+            self._drop_pending_form_task(task_id, pending)
+            return
+
+        update_card = build_button_interaction_update_card(
+            form_data,
+            task_id,
+            event_key,
+            source=self.card_source,
+        )
+        if card_type == 'multiple_interaction' or selections:
+            update_card = build_multiple_interaction_update_card(
+                form_data,
+                task_id,
+                selections,
+                source=self.card_source,
+            )
+        try:
+            await self.update_template_card(req_id_for_update, update_card)
+        except Exception:
+            await self.logger.warning(f'Failed to update template card (ws): {traceback.format_exc()}')
+
+        if self._card_action_callback is not None:
+            try:
+                session = StreamSession(
+                    stream_id=pending.get('stream_id', ''),
+                    msg_id=pending.get('msg_id', ''),
+                    chat_id=pending.get('chat_id') or None,
+                    user_id=pending.get('user_id') or None,
+                )
+                session.pending_form = pending.get('form_data')
+                session.pending_form_task_id = task_id
+                await self._card_action_callback(session, event_key, task_id, body)
+            except Exception:
+                await self.logger.error(f'card action callback raised (ws): {traceback.format_exc()}')
+
+        self._drop_pending_form_task(task_id, pending)
+
+    def _drop_pending_form_task(self, task_id: str, pending: dict) -> None:
+        self._pending_forms_by_task.pop(task_id, None)
+        msg_id = pending.get('msg_id', '')
+        if msg_id:
+            self._task_id_by_msg.pop(msg_id, None)
+            self._stream_sessions.pop(msg_id, None)
+
     async def _dispatch_event(self, event: wecombotevent.WecomBotEvent):
         """Dispatch a message event to registered handlers with deduplication."""
         try:

src/langbot/pkg/api/http/controller/groups/platform/adapters.py

@@ -5,6 +5,29 @@ from ... import group
 from langbot.pkg.utils import importutil
 
 
+def _decrypt_qqofficial_secret(encrypted_b64: str, key: bytes) -> str:
+    """Decrypt the AppSecret returned by the QQ Official QR binding endpoint.
+
+    The base64 payload is laid out as `nonce (12 B) | ciphertext | tag (16 B)`.
+    `key` is the 32-byte AES-256 key locally generated when the bind task
+    was created and submitted as `key` to `q.qq.com/lite/create_bind_task`.
+    """
+    import base64
+    from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+    try:
+        raw = base64.b64decode(encrypted_b64)
+    except Exception as exc:
+        raise ValueError('Malformed encrypted credential') from exc
+    if len(key) != 32 or len(raw) <= 28:
+        raise ValueError('Invalid encrypted credential layout')
+    nonce, ciphertext, tag = raw[:12], raw[12:-16], raw[-16:]
+    try:
+        return AESGCM(key).decrypt(nonce, ciphertext + tag, None).decode('utf-8')
+    except Exception as exc:
+        raise ValueError('Failed to decrypt credential') from exc
+
+
 @group.group_class('adapters', '/api/v1/platform/adapters')
 class AdaptersRouterGroup(group.RouterGroup):
     async def initialize(self) -> None:
@@ -37,6 +60,15 @@ class AdaptersRouterGroup(group.RouterGroup):
                 importutil.read_resource_file_bytes(icon_path), mimetype=mimetypes.guess_type(icon_path)[0]
             )
 
+        @self.route('/dingtalk/human-input-card-template', methods=['GET'], auth_type=group.AuthType.NONE)
+        async def _() -> quart.Response:
+            filename = 'dingtalk_human_input_card.json'
+            response = quart.Response(
+                importutil.read_resource_file_bytes(f'templates/{filename}'), mimetype='application/json'
+            )
+            response.headers['Content-Disposition'] = f'attachment; filename={filename}'
+            return response
+
         # In-memory session store for active registrations
         _create_app_sessions: dict = {}
         _SESSION_TTL = 900  # 15 minutes
@@ -650,3 +682,220 @@ class AdaptersRouterGroup(group.RouterGroup):
             if session and session.get('task') and not session['task'].done():
                 session['task'].cancel()
             return self.success(data={})
+
+        # -----------------------------------------------------------------------
+        # QQ Official QR Binding
+        # -----------------------------------------------------------------------
+
+        _qqofficial_sessions: dict = {}
+        _QQOFFICIAL_SESSION_TTL = 300  # 5 minutes (QQ bind QR validity window)
+
+        def _cleanup_expired_qqofficial_sessions():
+            import time
+
+            now = time.time()
+            expired = [
+                sid for sid, s in _qqofficial_sessions.items() if now - s.get('created_at', 0) > _QQOFFICIAL_SESSION_TTL
+            ]
+            for sid in expired:
+                session = _qqofficial_sessions.pop(sid, None)
+                if session and session.get('task') and not session['task'].done():
+                    session['task'].cancel()
+
+        @self.route('/qqofficial/bind', methods=['POST'])
+        async def _() -> str:
+            """Start QQ Official QR binding. Returns session_id + QR URL.
+
+            Flow: generate a local AES-256 key, register it with
+            `q.qq.com/lite/create_bind_task`, then poll
+            `q.qq.com/lite/poll_bind_result` until the user authorizes the
+            bind inside the QQ Bot Assistant on mobile QQ. The encrypted
+            AppSecret returned by the poll endpoint is decrypted with the
+            same key. The key never leaves this process.
+            """
+            import uuid
+            import time
+            import secrets
+            import base64
+            import aiohttp
+
+            QQ_BIND_BASE = 'https://q.qq.com'
+            _cleanup_expired_qqofficial_sessions()
+
+            bind_key_bytes = secrets.token_bytes(32)
+            bind_key = base64.b64encode(bind_key_bytes).decode('ascii')
+
+            session_id = str(uuid.uuid4())
+            session = {
+                'status': 'pending',
+                'qr_url': None,
+                'expire_at': None,
+                'appid': None,
+                'secret': None,
+                'user_openid': None,
+                'error': None,
+                'created_at': time.time(),
+                'task_id': None,
+                'bind_key_bytes': bind_key_bytes,
+                'interval': 2,
+            }
+            _qqofficial_sessions[session_id] = session
+
+            async def run_qr_binding():
+                try:
+                    timeout = aiohttp.ClientTimeout(total=10)
+                    async with aiohttp.ClientSession(timeout=timeout) as http:
+                        # Step 1: create_bind_task — register our AES key, get task_id
+                        async with http.post(
+                            f'{QQ_BIND_BASE}/lite/create_bind_task',
+                            json={'key': bind_key},
+                            headers={'Accept': 'application/json'},
+                        ) as resp:
+                            try:
+                                data = await resp.json(content_type=None)
+                            except (aiohttp.ContentTypeError, ValueError):
+                                session['status'] = 'error'
+                                session['error'] = 'Invalid response from QQ bind service'
+                                return
+                            if int(data.get('retcode', -1)) != 0:
+                                session['status'] = 'error'
+                                session['error'] = (
+                                    data.get('msg') or data.get('message') or 'Failed to create bind task'
+                                )
+                                return
+                            task_id = str((data.get('data') or {}).get('task_id') or '').strip()
+                            if not task_id:
+                                session['status'] = 'error'
+                                session['error'] = 'Missing task_id in QQ response'
+                                return
+
+                        # The QR encodes a URL that mobile QQ opens inside the QQ Bot Assistant.
+                        # `source=langbot` is a courtesy attribution parameter so Tencent
+                        # can see LangBot adoption metrics, matching the convention used by
+                        # other third-party integrations (e.g. hermes-agent uses `source=hermes`).
+                        qr_url = f'{QQ_BIND_BASE}/qqbot/openclaw/connect.html?task_id={task_id}&_wv=2&source=langbot'
+                        session['task_id'] = task_id
+                        session['qr_url'] = qr_url
+                        session['expire_at'] = time.time() + _QQOFFICIAL_SESSION_TTL
+                        session['status'] = 'waiting'
+
+                        # Step 2: poll_bind_result until completed (status=2) or expired (3).
+                        deadline = time.time() + _QQOFFICIAL_SESSION_TTL
+                        while time.time() < deadline:
+                            await asyncio.sleep(session['interval'])
+
+                            async with http.post(
+                                f'{QQ_BIND_BASE}/lite/poll_bind_result',
+                                json={'task_id': task_id},
+                                headers={'Accept': 'application/json'},
+                            ) as poll_resp:
+                                try:
+                                    poll_data = await poll_resp.json(content_type=None)
+                                except (aiohttp.ContentTypeError, ValueError):
+                                    continue
+
+                            if int(poll_data.get('retcode', -1)) != 0:
+                                session['status'] = 'error'
+                                session['error'] = poll_data.get('msg') or poll_data.get('message') or 'Poll failed'
+                                return
+
+                            payload = poll_data.get('data') or {}
+                            try:
+                                raw_status = int(payload.get('status', 0))
+                            except (TypeError, ValueError):
+                                raw_status = 0
+
+                            if raw_status == 2:
+                                appid = str(payload.get('bot_appid') or '').strip()
+                                encrypted = str(payload.get('bot_encrypt_secret') or '').strip()
+                                if not appid or not encrypted:
+                                    session['status'] = 'error'
+                                    session['error'] = 'Incomplete credential payload'
+                                    return
+                                try:
+                                    session['secret'] = _decrypt_qqofficial_secret(
+                                        encrypted,
+                                        bind_key_bytes,
+                                    )
+                                except ValueError as exc:
+                                    session['status'] = 'error'
+                                    session['error'] = str(exc)
+                                    return
+                                session['appid'] = appid
+                                # The scanner's OpenID is returned alongside the credentials —
+                                # surfaced to the dashboard for audit / "bound by" display.
+                                session['user_openid'] = str(payload.get('user_openid') or '').strip() or None
+                                session['status'] = 'success'
+                                return
+
+                            if raw_status == 3:
+                                session['status'] = 'expired'
+                                session['error'] = 'QR code expired'
+                                return
+                            # status 0 / 1: still pending, continue polling
+
+                        session['status'] = 'expired'
+                        session['error'] = 'QR code expired'
+
+                except asyncio.CancelledError:
+                    return
+                except Exception as e:
+                    session['status'] = 'error'
+                    session['error'] = str(e)
+
+            task = asyncio.create_task(run_qr_binding())
+            session['task'] = task
+
+            # Wait up to 10s for the QR URL to be ready before responding.
+            for _ in range(20):
+                if session['qr_url'] or session['error']:
+                    break
+                await asyncio.sleep(0.5)
+
+            if session['error']:
+                task.cancel()
+                return self.http_status(502, -1, session['error'])
+
+            if not session['qr_url']:
+                task.cancel()
+                session['status'] = 'error'
+                session['error'] = 'Timeout waiting for QR code'
+                return self.http_status(504, -1, 'Timeout waiting for QR code')
+
+            return self.success(
+                data={
+                    'session_id': session_id,
+                    'qr_url': session['qr_url'],
+                    'expire_at': session['expire_at'],
+                }
+            )
+
+        @self.route('/qqofficial/bind/status/<session_id>', methods=['GET'])
+        async def _(session_id: str) -> str:
+            """Poll QQ Official QR binding status."""
+            _cleanup_expired_qqofficial_sessions()
+            session = _qqofficial_sessions.get(session_id)
+            if not session:
+                return self.http_status(404, -1, 'Session not found')
+
+            data = {'status': session['status']}
+
+            if session['status'] == 'success':
+                data['appid'] = session['appid']
+                data['secret'] = session['secret']
+                if session.get('user_openid'):
+                    data['user_openid'] = session['user_openid']
+                _qqofficial_sessions.pop(session_id, None)
+            elif session['status'] in ('error', 'expired'):
+                data['error'] = session['error']
+                _qqofficial_sessions.pop(session_id, None)
+
+            return self.success(data=data)
+
+        @self.route('/qqofficial/bind/<session_id>', methods=['DELETE'])
+        async def _(session_id: str) -> str:
+            """Cancel and clean up a QQ Official QR binding session."""
+            session = _qqofficial_sessions.pop(session_id, None)
+            if session and session.get('task') and not session['task'].done():
+                session['task'].cancel()
+            return self.success(data={})

src/langbot/pkg/api/http/controller/groups/resources/mcp.py

@@ -96,6 +96,19 @@ class MCPRouterGroup(group.RouterGroup):
             except Exception as e:
                 return self.http_status(500, -1, f'Failed to get resource templates: {str(e)}')
 
+        @self.route('/servers/<path:server_name>/logs', methods=['GET'], auth_type=group.AuthType.USER_TOKEN)
+        async def _(server_name: str) -> str:
+            """Get logs from an MCP server"""
+            server_name = unquote(server_name)
+            try:
+                limit = int(quart.request.args.get('limit', 200))
+            except (TypeError, ValueError):
+                limit = 200
+            limit = min(limit, 500)
+            level = quart.request.args.get('level') or None
+            logs = await self.ap.mcp_service.get_mcp_server_logs(server_name, limit=limit, level=level)
+            return self.success(data={'logs': logs})
+
         @self.route('/servers/<path:server_name>/resources/read', methods=['POST'], auth_type=group.AuthType.USER_TOKEN)
         async def _(server_name: str) -> str:
             """Read a resource from an MCP server"""

src/langbot/pkg/api/http/service/mcp.py

@@ -188,10 +188,22 @@ class MCPService:
             persisted_session = runtime_mcp_session
 
             async def _refresh_and_report() -> None:
-                if persisted_session.status == MCPSessionStatus.ERROR:
+                # Testing a persisted server should REUSE its live shared-session
+                # process, not rebuild it. Try a lightweight refresh (a real
+                # list_tools probe over the existing connection) first; only fall
+                # back to a full start() when the session has no live connection
+                # to probe (never connected, or the process is actually gone).
+                needs_start = persisted_session.status == MCPSessionStatus.ERROR or persisted_session.session is None
+                if needs_start:
                     await persisted_session.start()
                 else:
-                    await persisted_session.refresh()
+                    try:
+                        await persisted_session.refresh()
+                    except Exception:
+                        # The live connection was stale/dropped: reconnect once
+                        # (reusing the live managed process where possible) and
+                        # re-probe, instead of reporting a false failure.
+                        await persisted_session.start()
                 # Surface the discovered tools so the config page can render them
                 # even for an already-hosted server.
                 ctx.metadata['runtime_info'] = persisted_session.get_runtime_info_dict()
@@ -232,3 +244,19 @@ class MCPService:
             context=ctx,
         )
         return wrapper.id
+
+    async def get_mcp_server_logs(self, server_name: str, limit: int = 200, level: str | None = None) -> list[dict]:
+        """Get recent log lines captured from the MCP server's stderr."""
+        session = self.ap.tool_mgr.mcp_tool_loader.get_session(server_name)
+        if not session:
+            return []
+
+        # Get logs from the session's buffer
+        logs = list(session._log_buffer)
+
+        # Filter by level if specified
+        if level:
+            logs = [log for log in logs if log.get('level') == level]
+
+        # Return the most recent 'limit' logs
+        return logs[-limit:]

src/langbot/pkg/pipeline/pipelinemgr.py

@@ -168,7 +168,7 @@ class RuntimePipeline:
                     bot_message=query.resp_messages[-1],
                     message=result.user_notice,
                     quote_origin=query.pipeline_config['output']['misc']['quote-origin'],
-                    is_final=[msg.is_final for msg in query.resp_messages][0],
+                    is_final=[msg.is_final for msg in query.resp_messages][-1],
                 )
             else:
                 await query.adapter.reply_message(

src/langbot/pkg/pipeline/pool.py

@@ -42,9 +42,13 @@ class QueryPool:
         adapter: abstract_platform_adapter.AbstractMessagePlatformAdapter,
         pipeline_uuid: typing.Optional[str] = None,
         routed_by_rule: bool = False,
+        variables: typing.Optional[dict[str, typing.Any]] = None,
     ) -> pipeline_query.Query:
         async with self.condition:
             query_id = self.query_id_counter
+            initial_variables: dict[str, typing.Any] = {'_routed_by_rule': routed_by_rule}
+            if variables:
+                initial_variables.update(variables)
             query = pipeline_query.Query(
                 bot_uuid=bot_uuid,
                 query_id=query_id,
@@ -53,7 +57,7 @@ class QueryPool:
                 sender_id=sender_id,
                 message_event=message_event,
                 message_chain=message_chain,
-                variables={'_routed_by_rule': routed_by_rule},
+                variables=initial_variables,
                 resp_messages=[],
                 resp_message_chain=[],
                 adapter=adapter,

src/langbot/pkg/pipeline/respback/respback.py

@@ -45,7 +45,7 @@ class SendResponseBackStage(stage.PipelineStage):
 
         try:
             if await query.adapter.is_stream_output_supported() and has_chunks:
-                is_final = [msg.is_final for msg in query.resp_messages][0]
+                is_final = [msg.is_final for msg in query.resp_messages][-1]
                 await query.adapter.reply_message_chunk(
                     message_source=query.message_event,
                     bot_message=query.resp_messages[-1],

src/langbot/pkg/platform/botmgr.py

@@ -501,6 +501,8 @@ class PlatformManager:
             bot_entity.adapter_config,
             logger,
         )
+        if hasattr(adapter_inst, 'ap'):
+            adapter_inst.ap = self.ap
 
         # 如果 adapter 支持 set_bot_uuid 方法，设置 bot_uuid（用于统一 webhook）
         if hasattr(adapter_inst, 'set_bot_uuid'):

src/langbot/pkg/platform/sources/dingtalk.yaml

@@ -103,6 +103,41 @@ spec:
       type: string
       required: true
       default: "填写你的卡片template_id"
+    - name: human_input_card_template_download
+      label:
+        en_US: Download Human Input Card Template
+        zh_Hans: 下载人工输入卡片模板
+        zh_Hant: 下載人工輸入卡片範本
+      description:
+        en_US: "Used as the only card template ID for the whole conversation turn. Download the built-in template, then import the JSON in DingTalk Open Platform > Card Platform / Card Template Management. After DingTalk creates the template, copy its template ID into the field below. The template already wires `content` (MarkdownBlock) and `btns` (ButtonGroup). Leave empty to fall back to the legacy two-card behavior."
+        zh_Hans: "用作整个对话回合唯一卡片的模板 ID。先下载内置模板，再到钉钉开放平台 > 卡片平台 / 卡片模板管理中导入该 JSON；钉钉生成模板后，将模板 ID 填到这里。模板已预先连好 `content` (MarkdownBlock) 与 `btns` (ButtonGroup)。留空则降级为旧的双卡行为。"
+        zh_Hant: "用作整個對話回合唯一卡片的範本 ID。先下載內建範本，再到釘釘開放平台 > 卡片平台 / 卡片範本管理中匯入該 JSON；釘釘產生範本後，將範本 ID 填到這裡。範本已預先連好 `content` (MarkdownBlock) 與 `btns` (ButtonGroup)。留空則降級為舊的雙卡行為。"
+      type: download-link
+      required: false
+      default: ""
+      url: /api/v1/platform/adapters/dingtalk/human-input-card-template
+      download_filename: dingtalk_human_input_card.json
+      help_links:
+        zh: https://open-dev.dingtalk.com/fe/card
+        en: https://open-dev.dingtalk.com/fe/card
+        ja: https://open-dev.dingtalk.com/fe/card
+      help_label:
+        en_US: Import Guide
+        zh_Hans: 导入指引
+        zh_Hant: 匯入指引
+        ja_JP: インポート手順
+    - name: human_input_card_template_id
+      label:
+        en_US: Human Input Card Template ID
+        zh_Hans: 人工输入卡片模板ID
+        zh_Hant: 人工輸入卡片範本ID
+      description:
+        en_US: "Paste the template ID generated after importing the human input card template."
+        zh_Hans: "填写导入人工输入卡片模板后生成的模板 ID。"
+        zh_Hant: "填寫匯入人工輸入卡片範本後產生的範本 ID。"
+      type: string
+      required: false
+      default: ""
 execution:
   python:
     path: ./dingtalk.py

src/langbot/pkg/platform/sources/discord.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import discord
+from discord import ui as discord_ui
 
 import typing
 import re
@@ -8,6 +9,8 @@ import base64
 import uuid
 import os
 import datetime
+import time
+import traceback
 
 # 使用BytesIO创建文件对象，避免路径问题
 import io
@@ -824,6 +827,69 @@ class DiscordEventConverter(abstract_platform_adapter.AbstractEventConverter):
             )
 
 
+class DiscordFormView(discord_ui.View):
+    """Discord ``ui.View`` that renders one button per Dify form action.
+
+    Each button's click triggers ``adapter._on_form_button_click`` which
+    acks the interaction, locks the buttons in place, and enqueues a
+    synthetic ``_dify_form_action`` query so the runner resumes the
+    workflow.
+    """
+
+    # Discord button style mapping for Dify ``button_style`` values.
+    _STYLE_MAP: typing.ClassVar[dict] = {
+        'primary': discord.ButtonStyle.primary,
+        'danger': discord.ButtonStyle.danger,
+        'warning': discord.ButtonStyle.danger,
+        'success': discord.ButtonStyle.success,
+        'default': discord.ButtonStyle.secondary,
+        '': discord.ButtonStyle.secondary,
+    }
+
+    def __init__(
+        self,
+        adapter: 'DiscordAdapter',
+        session_key: str,
+        actions: list,
+        timeout: float = 1800,
+    ):
+        super().__init__(timeout=timeout)
+        self._adapter = adapter
+        self._session_key = session_key
+        # Discord caps a view at 25 children (5 rows × 5 buttons). Trim
+        # silently — most Dify forms have ≤10 actions in practice.
+        for idx, action in enumerate(actions[:25]):
+            action_id = str(action.get('id') or '')
+            label = str(action.get('title') or action_id or f'Option {idx + 1}')
+            style = self._STYLE_MAP.get(
+                str(action.get('button_style') or '').lower(),
+                discord.ButtonStyle.secondary,
+            )
+            # custom_id must be unique within the view and ≤100 chars.
+            # Encode (session, idx) so we can recover the action even
+            # if Dify ids contain unsafe characters.
+            custom_id = f'lb_form:{idx}:{action_id[:80]}'[:100]
+            button = discord_ui.Button(
+                label=label[:80],  # Discord label limit
+                style=style,
+                custom_id=custom_id,
+            )
+            button.callback = self._make_callback(action_id, label)
+            self.add_item(button)
+
+    def _make_callback(self, action_id: str, action_title: str):
+        async def _cb(interaction: discord.Interaction):
+            await self._adapter._on_form_button_click(
+                interaction=interaction,
+                session_key=self._session_key,
+                action_id=action_id,
+                action_title=action_title,
+                view=self,
+            )
+
+        return _cb
+
+
 class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
     bot: discord.Client = pydantic.Field(exclude=True)
 
@@ -837,6 +903,10 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
 
     voice_manager: VoiceConnectionManager | None = pydantic.Field(exclude=True, default=None)
 
+    # Injected by botmgr at construction so the form-button callback can
+    # enqueue a synthetic resume query (`_dify_form_action`) on the pool.
+    ap: typing.Any = pydantic.Field(exclude=True, default=None)
+
     def __init__(self, config: dict, logger: abstract_platform_logger.AbstractEventLogger, **kwargs):
         bot_account_id = config['client_id']
 
@@ -860,8 +930,18 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
 
         args = {}
 
-        if os.getenv('http_proxy'):
+        # Proxy: config > env var > auto-detect.
-            args['proxy'] = os.getenv('http_proxy')
+        # discord.py uses aiohttp which does NOT respect http_proxy env
+        # vars by default — we must pass proxy= explicitly.
+        proxy = (
+            config.get('proxy')
+            or os.getenv('http_proxy')
+            or os.getenv('HTTP_PROXY')
+            or os.getenv('https_proxy')
+            or os.getenv('HTTPS_PROXY')
+        )
+        if proxy:
+            args['proxy'] = proxy
 
         bot = MyClient(intents=intents, **args)
 
@@ -875,6 +955,19 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
             **kwargs,
         )
 
+        # Per-resp-message-id buffer for the accumulated text yielded by
+        # the runner. Discord's edit-message ratelimit (5/5s) makes true
+        # progressive streaming impractical, so we collect chunks and
+        # render once on is_final. ``_form_data`` on the final chunk
+        # diverts to the button-view path.
+        self._stream_buffer: dict[str, str] = {}
+        # session_key -> {form_data, channel_id, thread_id, sender_id,
+        #                 posted_at, view_message_id}
+        # Populated when we send a form view; consumed when the user
+        # clicks a button so we know which workflow_run / form_token to
+        # resume.
+        self._pending_forms: dict[str, dict] = {}
+
     # Voice functionality methods
     async def join_voice_channel(self, guild_id: int, channel_id: int, user_id: int = None) -> discord.VoiceClient:
         """
@@ -1068,7 +1161,12 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
     ):
         msg_to_send, files = await self.message_converter.yiri2target(message)
 
-        assert isinstance(message_source.source_platform_object, discord.Message)
+        # Synthetic events (button-click resume) have no inbound discord
+        # Message. Route via the channel we cached when the user clicked.
+        source = message_source.source_platform_object
+        if not isinstance(source, discord.Message):
+            await self._reply_synthetic(message_source, msg_to_send, files)
+            return
 
         args = {
             'content': msg_to_send,
@@ -1078,7 +1176,7 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
             args['files'] = files
 
         if quote_origin:
-            args['reference'] = message_source.source_platform_object
+            args['reference'] = source
 
         has_at = False
 
@@ -1090,7 +1188,422 @@ class DiscordAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         if has_at:
             args['mention_author'] = True
 
-        await message_source.source_platform_object.channel.send(**args)
+        await source.channel.send(**args)
+
+    async def _reply_synthetic(
+        self,
+        message_source: platform_events.MessageEvent,
+        msg_to_send: str,
+        files: list,
+    ) -> None:
+        """Deliver a reply for a button-click-resumed (synthetic) event.
+
+        We don't have an inbound discord.Message to anchor to; instead
+        look up the channel cached in ``_pending_forms[session_key +
+        '__last_channel']`` from the most recent button click.
+        """
+        if isinstance(message_source, platform_events.GroupMessage):
+            # _handle_form_chunk uses channel_id alone as the session
+            # scope, and launcher_id was set to channel_id when
+            # synthesizing the event.
+            session_key = f'c:{message_source.group.id}'
+        else:
+            session_key = f'p:{message_source.sender.id}'
+
+        cached = self._pending_forms.get(session_key + '__last_channel') or {}
+        channel = cached.get('channel')
+        if channel is None:
+            if self.ap is not None:
+                self.ap.logger.warning(
+                    f'Discord: synthetic reply has no cached channel for '
+                    f'{session_key}; dropping content (len={len(msg_to_send)})'
+                )
+            return
+
+        args: dict[str, typing.Any] = {'content': msg_to_send}
+        if files:
+            args['files'] = files
+        try:
+            await channel.send(**args)
+        except Exception:
+            if self.ap is not None:
+                self.ap.logger.error(f'Discord: synthetic reply send failed: {traceback.format_exc()}')
+
+    # Discord allows 5 edits per 5 seconds per message. We throttle
+    # to one edit per 8 runner-chunks (runner already yields every 8
+    # text_chunks internally), which stays comfortably within limits.
+    _STREAM_EDIT_INTERVAL = 8
+
+    async def is_stream_output_supported(self) -> bool:
+        return True
+
+    async def create_message_card(self, message_id: str, event: platform_events.MessageEvent) -> bool:
+        """Set up a stream context for progressive editing.
+
+        The first non-empty reply_message_chunk will send the initial
+        message; subsequent chunks edit it in place.
+        """
+        source = event.source_platform_object
+        if not isinstance(source, discord.Message):
+            return False
+        self._stream_buffer[message_id] = {
+            'channel': source.channel,
+            'sent_message': None,  # discord.Message set on first send
+            'last_content': '',
+            'chunk_count': 0,
+        }
+        return True
+
+    async def reply_message_chunk(
+        self,
+        message_source: platform_events.MessageEvent,
+        bot_message: typing.Any,
+        message: platform_message.MessageChain,
+        quote_origin: bool = False,
+        is_final: bool = False,
+    ):
+        msg_id = (
+            bot_message.get('resp_message_id')
+            if isinstance(bot_message, dict)
+            else getattr(bot_message, 'resp_message_id', None)
+        )
+
+        text_parts = [m.text for m in message if isinstance(m, platform_message.Plain)]
+        chunk_text = '\n\n'.join(t for t in text_parts if t)
+
+        form_data = getattr(bot_message, '_form_data', None) if not isinstance(bot_message, dict) else None
+
+        ctx = self._stream_buffer.get(msg_id) if msg_id else None
+
+        # If the stream ctx was not set up (create_message_card wasn't
+        # called, e.g. synthetic event), or the final chunk carries a
+        # form, skip progressive editing entirely.
+        if ctx is None or form_data:
+            try:
+                if form_data and is_final:
+                    await self._handle_form_chunk(message_source, form_data)
+                elif is_final and chunk_text:
+                    await self.reply_message(
+                        message_source,
+                        platform_message.MessageChain([platform_message.Plain(text=chunk_text)]),
+                        quote_origin,
+                    )
+            finally:
+                self._stream_buffer.pop(msg_id, None)
+            return
+
+        # Progressive streaming path: send first chunk, edit subsequent.
+        ctx['chunk_count'] += 1
+
+        # Runner yields the full accumulated text on each chunk, so we
+        # always replace (not append).
+        if chunk_text:
+            ctx['last_content'] = chunk_text
+
+        sent = ctx['sent_message']
+
+        if sent is None:
+            # First non-empty chunk — send the initial message.
+            if not ctx['last_content']:
+                return  # No content yet, wait for next chunk.
+            try:
+                sent = await ctx['channel'].send(ctx['last_content'])
+                ctx['sent_message'] = sent
+            except Exception:
+                if self.ap is not None:
+                    self.ap.logger.error(f'Discord stream send failed: {traceback.format_exc()}')
+                self._stream_buffer.pop(msg_id, None)
+                return
+
+        if is_final:
+            # Final chunk — edit to the full content, then clean up.
+            if ctx['last_content'] and ctx['last_content'] != sent.content:
+                try:
+                    await sent.edit(content=ctx['last_content'][:2000])
+                except Exception:
+                    pass  # Best-effort
+            self._stream_buffer.pop(msg_id, None)
+        elif (ctx['chunk_count'] % self._STREAM_EDIT_INTERVAL) == 0:
+            # Intermediate edit — throttle to avoid rate limits.
+            if ctx['last_content'] and ctx['last_content'] != sent.content:
+                try:
+                    await sent.edit(content=ctx['last_content'][:2000])
+                except Exception:
+                    pass  # Rate-limited or deleted — ignore.
+
+    async def _handle_form_chunk(
+        self,
+        message_source: platform_events.MessageEvent,
+        form_data: dict,
+    ) -> None:
+        """Render a Dify form pause as a Discord embed + button View.
+
+        Mirrors the QQ / Telegram / Lark form path: the button's click
+        callback synthesizes a ``_dify_form_action`` query so the runner's
+        ``_merge_pending_form_action`` resumes the workflow.
+        """
+        source = message_source.source_platform_object
+
+        actions = form_data.get('actions') or []
+        if not actions:
+            # Nothing clickable — fall back to plain text.
+            if source is not None:
+                await self.reply_message(
+                    message_source,
+                    platform_message.MessageChain(
+                        [platform_message.Plain(text=str(form_data.get('node_title') or ''))]
+                    ),
+                )
+            return
+
+        node_title = str(form_data.get('node_title') or 'Confirmation needed')
+        form_content = str(form_data.get('form_content') or '').strip()
+
+        # Two paths:
+        #   (a) Real message — extract channel from source.
+        #   (b) Synthetic event (button-click resume) — no
+        #       source_platform_object; recover the channel we cached
+        #       when the user clicked.
+        if isinstance(source, discord.Message):
+            channel = source.channel
+            guild_id = str(source.guild.id) if source.guild else ''
+            sender_id = str(source.author.id)
+            channel_id = str(source.channel.id)
+            session_key = f'c:{channel_id}' if guild_id else f'p:{sender_id}'
+        else:
+            # Synthetic event — resolve session_key from event shape,
+            # then look up the cached channel from the click.
+            if isinstance(message_source, platform_events.GroupMessage):
+                # launcher_id was set to channel_id when we synthesized.
+                channel_id = str(message_source.group.id)
+                session_key = f'c:{channel_id}'
+            else:
+                session_key = f'p:{message_source.sender.id}'
+                channel_id = ''
+
+            cached = self._pending_forms.get(session_key + '__last_channel')
+            channel = cached.get('channel') if cached else None
+            guild_id = (cached or {}).get('guild_id', '')
+            sender_id = str(message_source.sender.id) if message_source.sender else ''
+            if channel is None:
+                if self.ap is not None:
+                    self.ap.logger.warning(
+                        f'Discord: synthetic form chunk has no cached channel for '
+                        f'{session_key}; cannot render form buttons'
+                    )
+                return
+
+        body_parts: list[str] = []
+        if form_content:
+            body_parts.append(form_content)
+        embed_body = '\n\n'.join(body_parts)
+        # Discord embed.description has a 4096 char limit — defensive trim.
+        if len(embed_body) > 4000:
+            embed_body = embed_body[:3990] + '\n\n…(truncated)'
+
+        embed = discord.Embed(
+            title=node_title[:256],
+            description=embed_body,
+            color=discord.Color.blurple(),
+        )
+
+        view = DiscordFormView(
+            adapter=self,
+            session_key=session_key,
+            actions=actions,
+            timeout=1800,  # 30 min — matches Dify form_token TTL
+        )
+
+        try:
+            sent_msg = await channel.send(embed=embed, view=view)
+        except Exception:
+            if self.ap is not None:
+                self.ap.logger.error(f'Discord: form view send failed: {traceback.format_exc()}')
+            return
+
+        self._pending_forms[session_key] = {
+            'form_data': form_data,
+            'channel_id': channel_id,
+            'guild_id': guild_id,
+            'sender_id': sender_id,
+            'view_message_id': str(sent_msg.id),
+            'posted_at': time.time(),
+        }
+
+        if self.ap is not None:
+            self.ap.logger.info(f'Discord: form view posted session={session_key} actions={len(actions)}')
+
+    async def _on_form_button_click(
+        self,
+        interaction: discord.Interaction,
+        session_key: str,
+        action_id: str,
+        action_title: str,
+        view: DiscordFormView,
+    ) -> None:
+        """Handle a click on a form button — ack, resume the workflow,
+        and disable the View buttons so the choice is visually locked in."""
+        import langbot_plugin.api.entities.builtin.provider.session as provider_session
+
+        # ACK first (3-second deadline before Discord shows "interaction failed").
+        try:
+            await interaction.response.defer()
+        except discord.HTTPException:
+            # Already responded somehow — proceed regardless.
+            pass
+
+        pending = self._pending_forms.get(session_key)
+        if not pending:
+            if self.ap is not None:
+                self.ap.logger.warning(
+                    f'Discord: button click on stale session {session_key}; ignoring (action_id={action_id!r})'
+                )
+            await self._lock_view_message(interaction, view, action_title, stale=True)
+            return
+
+        form_data: dict = pending.get('form_data') or {}
+        guild_id = pending.get('guild_id', '')
+        channel_id = pending.get('channel_id', '')
+        initiator_id = str(pending.get('sender_id', '') or '')
+        actor_id = str(interaction.user.id) if interaction.user is not None else initiator_id
+        if not guild_id and initiator_id and actor_id != initiator_id:
+            if self.ap is not None:
+                self.ap.logger.warning(
+                    f'Discord: user {actor_id} cannot act on private form created for {initiator_id}'
+                )
+            await self._lock_view_message(interaction, view, action_title, stale=True)
+            return
+
+        self._pending_forms.pop(session_key, None)
+
+        # Lock the buttons in place: disable everything, mark chosen one.
+        await self._lock_view_message(interaction, view, action_title)
+
+        # In group context the launcher remains the channel so Dify resumes
+        # the original group session. The synthetic sender is still the real
+        # clicker, preserving actor identity for auditing and routing rules.
+        if guild_id:
+            launcher_type = provider_session.LauncherTypes.GROUP
+            launcher_id = channel_id
+        else:
+            launcher_type = provider_session.LauncherTypes.PERSON
+            launcher_id = initiator_id or actor_id
+
+        form_action_data = {
+            'form_token': form_data.get('form_token', ''),
+            'workflow_run_id': form_data.get('workflow_run_id', ''),
+            'action_id': action_id,
+            'action_title': action_title,
+            'node_title': form_data.get('node_title', ''),
+            'user': f'{launcher_type.value}_{launcher_id}',
+            'inputs': {},
+        }
+
+        message_chain = platform_message.MessageChain([platform_message.Plain(text=f'[Form Action: {action_title}]')])
+
+        # Synthesize a platform event so the pipeline can run the resume
+        # query. source_platform_object=None signals "no inbound discord
+        # message" — reply_message must tolerate this (it falls through
+        # to channel.send via the cached interaction.channel below).
+        if launcher_type == provider_session.LauncherTypes.GROUP:
+            synthetic_event: platform_events.MessageEvent = platform_events.GroupMessage(
+                sender=platform_entities.GroupMember(
+                    id=actor_id,
+                    member_name=interaction.user.display_name if interaction.user else '',
+                    permission='MEMBER',
+                    group=platform_entities.Group(
+                        id=launcher_id,
+                        name=channel_id,
+                        permission=platform_entities.Permission.Member,
+                    ),
+                    special_title='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+        else:
+            synthetic_event = platform_events.FriendMessage(
+                sender=platform_entities.Friend(
+                    id=actor_id,
+                    nickname=interaction.user.display_name if interaction.user else '',
+                    remark='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+
+        if self.ap is None:
+            if self.logger:
+                await self.logger.error('Discord: ap not injected; cannot enqueue button-click query')
+            return
+
+        bot_uuid = ''
+        pipeline_uuid = form_data.get('pipeline_uuid') or None
+        for bot in self.ap.platform_mgr.bots:
+            if bot.adapter is self:
+                bot_uuid = bot.bot_entity.uuid
+                pipeline_uuid = pipeline_uuid or bot.bot_entity.use_pipeline_uuid
+                break
+
+        # Remember the channel so _reply_synthetic and _handle_form_chunk
+        # (synthetic-event path) can find a target. guild_id is needed
+        # to reconstruct the launcher_type on subsequent form pauses.
+        self._pending_forms[session_key + '__last_channel'] = {
+            'channel': interaction.channel,
+            'guild_id': guild_id,
+            'posted_at': time.time(),
+        }
+
+        try:
+            await self.ap.query_pool.add_query(
+                bot_uuid=bot_uuid,
+                launcher_type=launcher_type,
+                launcher_id=launcher_id,
+                sender_id=actor_id,
+                message_event=synthetic_event,
+                message_chain=message_chain,
+                adapter=self,
+                pipeline_uuid=pipeline_uuid,
+                variables={
+                    '_dify_form_action': form_action_data,
+                    '_routed_by_rule': True,
+                },
+            )
+            if self.ap is not None:
+                self.ap.logger.info(
+                    f'Discord: button-click query enqueued action_id={action_id!r} '
+                    f'session={session_key} actor_id={actor_id}'
+                )
+        except Exception:
+            if self.ap is not None:
+                self.ap.logger.error(f'Discord: enqueue button-click query failed: {traceback.format_exc()}')
+
+    async def _lock_view_message(
+        self,
+        interaction: discord.Interaction,
+        view: DiscordFormView,
+        chosen_title: str,
+        stale: bool = False,
+    ) -> None:
+        """Disable all buttons on the form view and annotate the chosen
+        one — mirrors DingTalk/Lark's in-card selection feedback."""
+        try:
+            for child in view.children:
+                if not isinstance(child, discord_ui.Button):
+                    continue
+                child.disabled = True
+                if not stale and child.label == chosen_title:
+                    child.style = discord.ButtonStyle.success
+                    if not (child.label or '').startswith('✓ '):
+                        child.label = f'✓ {child.label}'
+            view.stop()
+            if interaction.message is not None:
+                await interaction.message.edit(view=view)
+        except Exception:
+            if self.ap is not None:
+                self.ap.logger.warning(f'Discord: lock-view-message failed (non-fatal): {traceback.format_exc()}')
 
     async def is_muted(self, group_id: int) -> bool:
         return False

src/langbot/pkg/platform/sources/qqofficial.py

@@ -11,7 +11,13 @@ import langbot_plugin.api.definition.abstract.platform.adapter as abstract_platf
 import langbot_plugin.api.entities.builtin.platform.message as platform_message
 import langbot_plugin.api.entities.builtin.platform.events as platform_events
 import langbot_plugin.api.entities.builtin.platform.entities as platform_entities
-from langbot.libs.qq_official_api.api import QQOfficialClient
+from langbot.libs.qq_official_api.api import (
+    QQ_SELECT_ACTION_PREFIX,
+    QQOfficialClient,
+    build_keyboard_from_form,
+    build_keyboard_from_select_field,
+    resolve_select_button_action,
+)
 from langbot.libs.qq_official_api.qqofficialevent import QQOfficialEvent
 from ...utils import image
 from ..logger import EventLogger
@@ -191,6 +197,7 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
     enable_webhook: bool = False
     message_converter: QQOfficialMessageConverter = QQOfficialMessageConverter()
     event_converter: QQOfficialEventConverter = QQOfficialEventConverter()
+    ap: typing.Any = None
 
     def __init__(self, config: dict, logger: EventLogger):
         enable_webhook = config.get('enable-webhook', False)
@@ -216,6 +223,31 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
         self._stream_ctx_ts: dict[str, float] = {}
         self._fallback_text: dict[str, str] = {}
         self._fallback_text_ts: dict[str, float] = {}
+        # Dify form-action bookkeeping for the human-input button flow.
+        # session_key = "<scene>_<id>" where scene is c2c/group/channel and
+        #   id is user_openid / group_openid / channel_id.
+        # session_key -> {form_data, msg_id, event_id, scene, target_id,
+        #                 sender_id, posted_at}
+        # Set when we send a markdown+keyboard card and consulted when:
+        #   (a) INTERACTION_CREATE fires — we look up the form by
+        #       session_key (button's `data` carries the action_id),
+        #   (b) the resumed-workflow query needs to find a passive-reply
+        #       event_id (INTERACTION_CREATE id, 30-min validity).
+        self._pending_forms: dict[str, dict] = {}
+        # session_key -> most recent ``INTERACTION_CREATE`` event_id, used
+        # as the passive event_id for the resumed query's LLM output.
+        self._session_event_ids: dict[str, dict] = {}
+        # Per-anchor msg_seq counter. QQ accepts up to 5 passive replies
+        # per (msg_id|event_id) within 60 min, but each reuse needs a
+        # fresh ``msg_seq`` — re-sending with msg_seq=1 is silently dedup'd.
+        self._anchor_msg_seq: dict[str, int] = {}
+
+        # Wire button-click handler so webhook mode catches INTERACTION_CREATE.
+        # (ws mode is wired separately via on_event in _run_websocket so the
+        # raw payload bypasses get_message's message-only flattening.)
+        @self.bot.on_interaction()
+        async def _on_interaction(event_data: dict, interaction_id: typing.Optional[str]):
+            await self._handle_interaction_create(event_data, interaction_id)
 
     async def reply_message(
         self,
@@ -227,6 +259,13 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
             message_source,
         )
 
+        # Synthetic event (button-click resume): no inbound platform
+        # object → no msg_id. Route via the cached INTERACTION_CREATE
+        # event_id (valid 30 min, no quota cost).
+        if qq_official_event is None:
+            await self._reply_synthetic(message_source, message)
+            return
+
         content_list = await QQOfficialMessageConverter.yiri2target(message)
 
         # 确定 target_type 和 target_id
@@ -376,6 +415,9 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
             await self.logger.info('QQ Official WebSocket connected and ready')
 
         async def on_event(event_type: str, event_data: dict):
+            # INTERACTION_CREATE is dispatched via bot.on_interaction()
+            # (registered in __init__) so we get the top-level ws_event_id
+            # — needed as the passive-reply event_id. It never reaches here.
             # 只处理消息事件，忽略 READY/RESUMED 等系统事件
             message_event_types = {
                 'C2C_MESSAGE_CREATE',
@@ -437,12 +479,36 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
     async def is_stream_output_supported(self) -> bool:
         return self.config.get('enable-stream-reply', False)
 
+    @staticmethod
+    def _is_form_placeholder_chunk(text: str) -> bool:
+        """Return True for invisible placeholder chunks used to carry forms."""
+
+        if not text:
+            return False
+
+        cleaned = text.replace('\u200b', '').replace('\u200c', '').replace('\u200d', '').replace('\ufeff', '').strip()
+        # Some Windows consoles/logs display the zero-width placeholder as
+        # mojibake. Treat those variants as the same non-user-facing marker.
+        return cleaned in {'', '鈥?', '​'}
+
     async def create_message_card(self, message_id: str, event: platform_events.MessageEvent) -> bool:
         source = event.source_platform_object
+        # Synthetic events (button-click resume) have no source object —
+        # they ride a cached INTERACTION_CREATE event_id, not a streamable
+        # msg_id. Skip stream setup; reply_message handles the one-shot
+        # send at is_final.
+        if source is None:
+            return False
         # Streaming API only supports C2C private chat
         if source.t != 'C2C_MESSAGE_CREATE':
             return False
 
+        # The stream endpoint still consumes msg_seq for this inbound msg_id.
+        # Keep the passive-reply counter in sync so a follow-up form card uses
+        # msg_seq=2 instead of being deduplicated by QQ as another seq=1 send.
+        if source.d_id:
+            self._anchor_msg_seq[source.d_id] = max(self._anchor_msg_seq.get(source.d_id, 0), 1)
+
         ctx = {
             'user_openid': source.user_openid,
             'msg_id': source.d_id,
@@ -469,12 +535,38 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
     ):
         # Periodically clean up stale stream contexts
         await self._cleanup_stale_streams()
+
+        # Dify human-input pause: when the runner attaches `_form_data` to
+        # the final chunk, finalize any in-flight stream session and send
+        # a markdown + keyboard message instead. Plain-text content from
+        # earlier chunks is already on the stream; we close it cleanly
+        # and the buttons land as a separate reply.
+        form_data = getattr(bot_message, '_form_data', None) if not isinstance(bot_message, dict) else None
+        if is_final:
+            _resume = getattr(bot_message, '_resume_from_form', None) if not isinstance(bot_message, dict) else None
+            _open_new = getattr(bot_message, '_open_new_card', None) if not isinstance(bot_message, dict) else None
+            if self.ap is not None:
+                self.ap.logger.info(
+                    f'QQ Official reply_message_chunk final: '
+                    f'type={type(bot_message).__name__} '
+                    f'is_final={is_final} '
+                    f'form_data_present={form_data is not None} '
+                    f'resume_from_form={_resume} open_new_card={_open_new} '
+                    f'content_len={len(getattr(bot_message, "content", "") or "")}'
+                )
+        if form_data and is_final:
+            await self._handle_form_chunk(message_source, message, form_data)
+            return
+
         # 提取纯文本内容（当前 chunk 的文本）
         text_parts = []
         for msg in message:
             if type(msg) is platform_message.Plain:
                 text_parts.append(msg.text)
         chunk_text = '\n\n'.join(text_parts)
+        if self._is_form_placeholder_chunk(chunk_text):
+            await self.logger.debug('QQ Official: skipped invisible form placeholder chunk')
+            return
 
         message_id = (
             bot_message.get('resp_message_id')
@@ -484,7 +576,8 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
         if not message_id or message_id not in self._stream_ctx:
             # 非流式场景（如群聊不支持流式），累积文本后一次性回复
             if chunk_text:
-                self._fallback_text[message_id] = self._fallback_text.get(message_id, '') + chunk_text
+                # Chunks carry the latest full snapshot, not a text delta.
+                self._fallback_text[message_id] = chunk_text
                 self._fallback_text_ts[message_id] = time.time()
             if is_final:
                 full_text = self._fallback_text.pop(message_id, '')
@@ -497,7 +590,7 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
 
         # 累积文本
         if chunk_text:
-            ctx['accumulated_text'] += chunk_text
+            ctx['accumulated_text'] = chunk_text
 
         # 未启动会话时，等第一个有内容的 chunk 来建立会话
         if not ctx['session_started']:
@@ -557,3 +650,489 @@ class QQOfficialAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter
         ],
     ):
         return super().unregister_listener(event_type, callback)
+
+    # ------------------------------------------------------------------
+    # Dify human-input button-interaction support
+    # ------------------------------------------------------------------
+
+    _PENDING_FORM_TTL = 1800  # 30 min — matches QQ passive-reply window.
+    _MAX_REPLIES_PER_ANCHOR = 5  # QQ hard limit per msg_id / event_id.
+
+    def _next_msg_seq(self, anchor: str) -> typing.Optional[int]:
+        """Return the next msg_seq for an anchor, or ``None`` if the
+        anchor has already been used 5 times (further sends would be
+        silently dropped by QQ)."""
+        if not anchor:
+            return 1
+        used = self._anchor_msg_seq.get(anchor, 0)
+        if used >= self._MAX_REPLIES_PER_ANCHOR:
+            return None
+        self._anchor_msg_seq[anchor] = used + 1
+        return used + 1
+
+    async def _reply_synthetic(
+        self,
+        message_source: platform_events.MessageEvent,
+        message: platform_message.MessageChain,
+    ) -> None:
+        """Deliver a reply for a synthetic (button-click-resume) event.
+
+        Synthetic events have ``source_platform_object=None`` and no
+        fresh inbound msg_id. The previous INTERACTION_CREATE id we
+        cached in :attr:`_session_event_ids` is a valid passive-reply
+        anchor (``event_id``) for up to 30 minutes — use it.
+        """
+        if isinstance(message_source, platform_events.GroupMessage):
+            target_type = 'group'
+            group = getattr(message_source, 'group', None) or (
+                message_source.sender.group if hasattr(message_source.sender, 'group') else None
+            )
+            target_id = str(group.id) if group else None
+        else:
+            target_type = 'c2c'
+            target_id = str(message_source.sender.id) if message_source.sender else None
+
+        if not target_id:
+            await self.logger.warning('QQ Official: synthetic reply has no target_id; dropping')
+            return
+
+        session_key = f'{target_type}_{target_id}'
+        cached = self._session_event_ids.get(session_key)
+        event_id = cached.get('event_id') if cached else None
+        if cached and (time.time() - cached.get('posted_at', 0)) > self._PENDING_FORM_TTL:
+            event_id = None
+
+        if not event_id:
+            await self.logger.warning(
+                f'QQ Official: no cached event_id for {session_key}; '
+                f'cannot deliver synthetic reply within passive-reply window'
+            )
+            return
+
+        content_list = await QQOfficialMessageConverter.yiri2target(message)
+        text_parts = [c['content'] for c in content_list if c.get('type') == 'text' and c.get('content')]
+        if not text_parts:
+            await self.logger.info('QQ Official: synthetic reply has no text content; skipping')
+            return
+        text = '\n\n'.join(text_parts)
+
+        msg_seq = self._next_msg_seq(event_id)
+        if msg_seq is None:
+            await self.logger.warning(
+                f'QQ Official: anchor {event_id!r} exhausted (>5 passive replies); '
+                f'cannot deliver synthetic reply for {session_key}'
+            )
+            return
+
+        try:
+            if target_type == 'c2c':
+                await self.bot.send_private_text_msg(
+                    user_openid=target_id,
+                    content=text,
+                    event_id=event_id,
+                    msg_seq=msg_seq,
+                )
+            elif target_type == 'group':
+                await self.bot.send_group_text_msg(
+                    group_openid=target_id,
+                    content=text,
+                    event_id=event_id,
+                    msg_seq=msg_seq,
+                )
+        except Exception:
+            await self.logger.error(f'QQ Official: synthetic reply delivery failed: {traceback.format_exc()}')
+
+    def _resolve_target_from_source(self, source: QQOfficialEvent) -> typing.Optional[tuple[str, str]]:
+        """Return ``(target_type, target_id)`` for sending a reply, or
+        ``None`` if the scene cannot host a markdown+keyboard message."""
+        if source is None:
+            return None
+        if source.t == 'C2C_MESSAGE_CREATE':
+            return 'c2c', source.user_openid
+        if source.t == 'GROUP_AT_MESSAGE_CREATE':
+            return 'group', source.group_openid
+        if source.t == 'AT_MESSAGE_CREATE':
+            return 'channel', source.channel_id
+        # DIRECT_MESSAGE_CREATE uses the guild DM API which does not accept
+        # markdown+keyboard at the time of writing — caller falls back to text.
+        return None
+
+    def _resolve_target_from_event(
+        self, message_source: platform_events.MessageEvent
+    ) -> typing.Optional[tuple[str, str]]:
+        """Resolve ``(target_type, target_id)`` from the public event.
+
+        Prefers the platform-native source when present; falls back to
+        the synthesized event's sender/group fields so button-click
+        resume queries can still find a destination.
+        """
+        source = message_source.source_platform_object
+        if source is not None:
+            return self._resolve_target_from_source(source)
+        if isinstance(message_source, platform_events.GroupMessage):
+            group = getattr(message_source, 'group', None) or (
+                message_source.sender.group
+                if message_source.sender and hasattr(message_source.sender, 'group')
+                else None
+            )
+            if group and getattr(group, 'id', None):
+                return 'group', str(group.id)
+        if isinstance(message_source, platform_events.FriendMessage):
+            if message_source.sender and getattr(message_source.sender, 'id', None):
+                return 'c2c', str(message_source.sender.id)
+        return None
+
+    def _prune_pending_forms(self) -> None:
+        now = time.time()
+        stale = [k for k, v in self._pending_forms.items() if now - v.get('posted_at', 0) > self._PENDING_FORM_TTL]
+        for k in stale:
+            self._pending_forms.pop(k, None)
+        stale_e = [
+            k for k, v in self._session_event_ids.items() if now - v.get('posted_at', 0) > self._PENDING_FORM_TTL
+        ]
+        for k in stale_e:
+            self._session_event_ids.pop(k, None)
+
+    async def _handle_form_chunk(
+        self,
+        message_source: platform_events.MessageEvent,
+        message: platform_message.MessageChain,
+        form_data: dict,
+    ) -> None:
+        """Send the markdown + keyboard form prompt for a Dify pause.
+
+        Called from ``reply_message_chunk`` when the runner attaches
+        ``_form_data`` to the final chunk. Replaces what would otherwise
+        be a plain-text numbered-list fallback.
+        """
+        if self.ap is not None:
+            self.ap.logger.info(
+                f'QQ Official _handle_form_chunk entered; '
+                f'source_present={message_source.source_platform_object is not None} '
+                f'form_actions={len(form_data.get("actions") or [])}'
+            )
+        self._prune_pending_forms()
+
+        source = message_source.source_platform_object
+        scene_target = self._resolve_target_from_event(message_source)
+        if scene_target is None:
+            # No rich-UI fit — fall through to existing text path.
+            await self.logger.info('QQ Official: form chunk on unsupported scene; falling back to text')
+            text_parts = [m.text for m in message if type(m) is platform_message.Plain]
+            fallback_msg = platform_message.MessageChain([platform_message.Plain(text='\n\n'.join(text_parts))])
+            try:
+                await self.reply_message(message_source, fallback_msg)
+            except Exception:
+                await self.logger.error(f'QQ Official: form fallback text send failed: {traceback.format_exc()}')
+            return
+
+        target_type, target_id = scene_target
+        session_key = f'{target_type}_{target_id}'
+
+        # Cancel any in-flight stream / fallback ctx so plain-text prefix
+        # doesn't continue alongside the keyboard message.
+        msg_id = getattr(source, 'd_id', '') or '' if source is not None else ''
+        if msg_id:
+            self._stream_ctx.pop(msg_id, None)
+            self._stream_ctx_ts.pop(msg_id, None)
+            self._fallback_text.pop(msg_id, None)
+            self._fallback_text_ts.pop(msg_id, None)
+
+        node_title = form_data.get('node_title') or 'Confirmation needed'
+        form_content = form_data.get('form_content') or ''
+        is_field_step = bool(form_data.get('_current_input_field')) and not form_data.get('_action_select_only')
+        parts = [f'### {node_title}']
+        plain_parts = [node_title]
+        if form_content.strip():
+            parts.append(form_content.strip())
+            plain_parts.append(form_content.strip())
+        markdown_content = '\n\n'.join(parts)
+        plain_content = '\n\n'.join(plain_parts)
+
+        keyboard = build_keyboard_from_select_field(form_data) if is_field_step else None
+        is_text_field_step = is_field_step and not keyboard.get('content', {}).get('rows')
+        if is_text_field_step:
+            keyboard = None
+        if keyboard is None and not is_text_field_step:
+            keyboard = build_keyboard_from_form(form_data, buttons_per_row=2)
+        if keyboard is not None and not keyboard.get('content', {}).get('rows') and not is_text_field_step:
+            # No actions to render — fall back to plain text.
+            text_msg = platform_message.MessageChain([platform_message.Plain(text=plain_content)])
+            try:
+                await self.reply_message(message_source, text_msg)
+            except Exception:
+                await self.logger.error(f'QQ Official: empty-keyboard fallback send failed: {traceback.format_exc()}')
+            return
+
+        # Prefer the inbound msg_id (no quota cost). If the source is a
+        # synthetic event from a prior click, the cached interaction id
+        # serves as event_id for up to 30 min.
+        event_id = None
+        if not msg_id:
+            cached = self._session_event_ids.get(session_key)
+            if cached and (time.time() - cached.get('posted_at', 0)) < self._PENDING_FORM_TTL:
+                event_id = cached.get('event_id')
+
+        anchor = msg_id or event_id or ''
+        msg_seq = self._next_msg_seq(anchor)
+        if msg_seq is None:
+            await self.logger.warning(
+                f'QQ Official: anchor {anchor!r} exhausted (>5 passive replies); '
+                f'cannot deliver form card for session={session_key}'
+            )
+            return
+
+        try:
+            await self.bot.send_markdown_keyboard(
+                target_type=target_type,
+                target_id=target_id,
+                markdown_content=markdown_content,
+                keyboard=keyboard,
+                msg_id=msg_id if (msg_id and not event_id) else None,
+                event_id=event_id,
+                msg_seq=msg_seq,
+            )
+            if self.ap is not None:
+                self.ap.logger.info(
+                    f'QQ Official: form card sent '
+                    f'target={target_type}/{target_id} '
+                    f'msg_id={msg_id!r} event_id={event_id!r} msg_seq={msg_seq}'
+                )
+        except Exception:
+            if self.ap is not None:
+                self.ap.logger.error(
+                    f'QQ Official: send_markdown_keyboard failed, falling back to text: {traceback.format_exc()}'
+                )
+            await self.logger.error(
+                f'QQ Official: send_markdown_keyboard failed, falling back to text: {traceback.format_exc()}'
+            )
+            text_msg = platform_message.MessageChain([platform_message.Plain(text=plain_content)])
+            try:
+                await self.reply_message(message_source, text_msg)
+            except Exception:
+                pass
+            return
+
+        sender_id = ''
+        if source is not None:
+            sender_id = (
+                getattr(source, 'user_openid', None)
+                or getattr(source, 'member_openid', None)
+                or getattr(source, 'd_author_id', None)
+                or ''
+            )
+        if not sender_id and message_source.sender is not None:
+            sender_id = str(getattr(message_source.sender, 'id', '') or '')
+        self._pending_forms[session_key] = {
+            'form_data': form_data,
+            'msg_id': msg_id,
+            'sender_id': sender_id,
+            'target_type': target_type,
+            'target_id': target_id,
+            'source_event_t': source.t if source is not None else None,
+            'posted_at': time.time(),
+        }
+        await self.logger.info(
+            f'QQ Official: form posted session={session_key} actions={len(form_data.get("actions") or [])}'
+        )
+
+    async def _handle_interaction_create(
+        self,
+        event_data: dict,
+        ws_event_id: typing.Optional[str] = None,
+    ) -> None:
+        """Handle a button-click INTERACTION_CREATE event.
+
+        Two IDs at play (QQ keeps them separate):
+            ws_event_id   top-level payload ``id`` (or webhook ``X-Bot-
+                          Event-Id``). The ONLY value accepted as
+                          ``event_id`` for subsequent passive replies.
+            d['id']       the interaction id — used for PUT
+                          /interactions/{id} ack. Cannot be reused as
+                          event_id (QQ returns 40034025 if you try).
+
+        Layout (https://bot.q.qq.com/.../msg-btn.html):
+            chat_type     0 channel / 1 group / 2 c2c
+            data.resolved.button_data  what we set as ``action.data``
+            data.resolved.button_id    ``id`` field on the button row
+        """
+        import langbot_plugin.api.entities.builtin.provider.session as provider_session
+
+        if self.ap is not None:
+            self.ap.logger.info(
+                f'QQ Official _handle_interaction_create entered; '
+                f'ws_event_id={ws_event_id!r} '
+                f'interaction_id={(event_data.get("id") if isinstance(event_data, dict) else None)!r} '
+                f'chat_type={event_data.get("chat_type") if isinstance(event_data, dict) else None}'
+            )
+
+        if not isinstance(event_data, dict):
+            await self.logger.warning(f'QQ Official: INTERACTION_CREATE event_data is not dict: {type(event_data)}')
+            return
+
+        # ACK uses the interaction id, NOT the ws event id.
+        interaction_id = event_data.get('id') or ''
+        if interaction_id:
+            asyncio.create_task(self.bot.ack_interaction(interaction_id, code=0))
+
+        resolved = (event_data.get('data') or {}).get('resolved') or {}
+        action_id = str(resolved.get('button_data') or resolved.get('button_id') or '').strip()
+        if not action_id:
+            await self.logger.warning('QQ Official: INTERACTION_CREATE missing button_data/button_id; ignoring')
+            return
+
+        chat_type = event_data.get('chat_type')
+        scene_target: typing.Optional[tuple[str, str]] = None
+        if chat_type == 2 or event_data.get('user_openid'):
+            scene_target = ('c2c', event_data.get('user_openid') or '')
+        elif chat_type == 1 or event_data.get('group_openid'):
+            scene_target = ('group', event_data.get('group_openid') or '')
+        elif chat_type == 0 or event_data.get('channel_id'):
+            scene_target = ('channel', event_data.get('channel_id') or '')
+
+        if not scene_target or not scene_target[1]:
+            await self.logger.warning(f'QQ Official: INTERACTION_CREATE missing scene/target; raw={event_data}')
+            return
+
+        target_type, target_id = scene_target
+        session_key = f'{target_type}_{target_id}'
+
+        self._prune_pending_forms()
+        pending = self._pending_forms.get(session_key)
+        if not pending:
+            await self.logger.warning(
+                f'QQ Official: no pending form for session {session_key}; click ignored (action_id={action_id!r})'
+            )
+            return
+
+        # Cache ws_event_id so a follow-up pause / text reply can use it
+        # as event_id for passive delivery (30-min window). Falls back to
+        # the interaction_id only if no ws_event_id was provided (e.g.
+        # tests / older payload shape) — QQ will reject that value but
+        # we log so the mismatch is debuggable.
+        cached_event_id = ws_event_id or interaction_id
+        if cached_event_id:
+            self._session_event_ids[session_key] = {
+                'event_id': cached_event_id,
+                'posted_at': time.time(),
+            }
+            # New anchor → fresh 5-reply budget.
+            self._anchor_msg_seq[cached_event_id] = 0
+            if self.ap is not None and not ws_event_id:
+                self.ap.logger.warning(
+                    'QQ Official: INTERACTION_CREATE lacked ws_event_id; '
+                    'falling back to interaction_id (passive reply may be rejected)'
+                )
+
+        form_data: dict = pending.get('form_data') or {}
+        actions = form_data.get('actions') or []
+        select_choice = resolve_select_button_action(form_data, action_id)
+        if action_id.startswith(QQ_SELECT_ACTION_PREFIX) and select_choice is None:
+            await self.logger.warning(f'QQ Official: invalid select action_id={action_id!r} for {session_key}')
+            return
+
+        matched = None
+        if select_choice is None:
+            matched = next(
+                (a for a in actions if str(a.get('id', '')) == action_id),
+                None,
+            )
+            if matched is None:
+                await self.logger.warning(
+                    f'QQ Official: action_id={action_id!r} is not present on pending form for {session_key}'
+                )
+                return
+        self._pending_forms.pop(session_key, None)
+        action_title = select_choice[1] if select_choice else matched.get('title') or action_id
+
+        initiator_id = str(pending.get('sender_id') or '')
+        actor_id = str(event_data.get('member_openid') or event_data.get('user_openid') or initiator_id)
+
+        # Build resume payload matching the shape every other adapter uses
+        # (DingTalk / Lark / Telegram / WeCom). The runner's
+        # _merge_pending_form_action consumes this verbatim.
+        if target_type == 'group' or target_type == 'channel':
+            launcher_type = provider_session.LauncherTypes.GROUP
+            launcher_id = target_id
+        else:
+            launcher_type = provider_session.LauncherTypes.PERSON
+            launcher_id = target_id
+
+        form_action_data = {
+            'form_token': form_data.get('form_token', ''),
+            'workflow_run_id': form_data.get('workflow_run_id', ''),
+            'action_id': '' if select_choice else action_id,
+            'action_title': action_title,
+            'node_title': form_data.get('node_title', ''),
+            'user': f'{launcher_type.value}_{launcher_id}',
+            'inputs': {'select': select_choice[1]} if select_choice else {},
+        }
+        if select_choice:
+            form_action_data['_current_input_field'] = select_choice[0]
+            form_action_data['_input_progress'] = True
+
+        event_label = 'Form Select' if select_choice else 'Form Action'
+        message_chain = platform_message.MessageChain([platform_message.Plain(text=f'[{event_label}: {action_title}]')])
+
+        if launcher_type == provider_session.LauncherTypes.GROUP:
+            synthetic_event: platform_events.MessageEvent = platform_events.GroupMessage(
+                sender=platform_entities.GroupMember(
+                    id=actor_id or launcher_id,
+                    member_name='',
+                    permission='MEMBER',
+                    group=platform_entities.Group(
+                        id=launcher_id,
+                        name='',
+                        permission=platform_entities.Permission.Member,
+                    ),
+                    special_title='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+        else:
+            synthetic_event = platform_events.FriendMessage(
+                sender=platform_entities.Friend(
+                    id=actor_id or launcher_id,
+                    nickname='',
+                    remark='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+
+        if self.ap is None:
+            await self.logger.error('QQ Official: ap not injected; cannot enqueue button-click query')
+            return
+
+        bot_uuid = ''
+        pipeline_uuid = form_data.get('pipeline_uuid') or None
+        for bot in self.ap.platform_mgr.bots:
+            if bot.adapter is self:
+                bot_uuid = bot.bot_entity.uuid
+                pipeline_uuid = pipeline_uuid or bot.bot_entity.use_pipeline_uuid
+                break
+
+        try:
+            await self.ap.query_pool.add_query(
+                bot_uuid=bot_uuid,
+                launcher_type=launcher_type,
+                launcher_id=launcher_id,
+                sender_id=actor_id or launcher_id,
+                message_event=synthetic_event,
+                message_chain=message_chain,
+                adapter=self,
+                pipeline_uuid=pipeline_uuid,
+                variables={
+                    '_dify_form_action': form_action_data,
+                    '_routed_by_rule': True,
+                },
+            )
+            await self.logger.info(
+                f'QQ Official: button-click query enqueued action_id={action_id!r} '
+                f'session={session_key} actor_id={actor_id}'
+            )
+        except Exception:
+            await self.logger.error(f'QQ Official: enqueue button-click query failed: {traceback.format_exc()}')

src/langbot/pkg/platform/sources/qqofficial.yaml

@@ -31,6 +31,18 @@ spec:
       type: array[string]
       required: false
       default: []
+    - name: one-click-bind
+      label:
+        en_US: One-Click QR Binding
+        zh_Hans: 一键扫码绑定
+        zh_Hant: 一鍵掃碼綁定
+      description:
+        en_US: Scan QR code with mobile QQ to auto-fill AppID and Secret (Token is not used and can be left blank)
+        zh_Hans: 使用手机 QQ 扫码绑定，自动填写 AppID 和密钥（当前未使用 Token，可留空）
+        zh_Hant: 使用手機 QQ 掃碼綁定，自動填寫 AppID 和密鑰（目前未使用 Token，可留空）
+      type: qr-code-login
+      login_platform: qqofficial
+      required: false
     - name: appid
       label:
         en_US: App ID
@@ -52,8 +64,12 @@ spec:
         en_US: Token
         zh_Hans: 令牌
         zh_Hant: 令牌
+      description:
+        en_US: Optional. The QR binding cannot return this value; the current adapter implementation does not use it either, so it can be safely left blank.
+        zh_Hans: 可选。扫码绑定无法获取该字段，当前适配器实现也未使用该字段，留空即可。
+        zh_Hant: 可選。掃碼綁定無法取得此欄位，目前介面卡實作亦未使用，留空即可。
       type: string
-      required: true
+      required: false
       default: ""
     - name: enable-webhook
       label:

src/langbot/pkg/platform/sources/telegram.py

@@ -1,15 +1,17 @@
 from __future__ import annotations
-import time
 
 
 import telegram
 import telegram.ext
-from telegram import Update
+from telegram import ForceReply, InlineKeyboardButton, InlineKeyboardMarkup, Update
-from telegram.ext import ApplicationBuilder, ContextTypes, MessageHandler, filters
+from telegram.ext import ApplicationBuilder, ContextTypes, MessageHandler, CallbackQueryHandler, filters
 import telegramify_markdown
 import typing
 import traceback
+import json
 import base64
+import time
+import uuid
 import pydantic
 
 from langbot.pkg.utils import httpclient
@@ -20,6 +22,61 @@ import langbot_plugin.api.entities.builtin.platform.entities as platform_entitie
 import langbot_plugin.api.definition.abstract.platform.event_logger as abstract_platform_logger
 
 
+def _telegram_select_field_options(form_data: dict) -> tuple[str, list[str]]:
+    """Return the active select field and its option values."""
+    field_name = str(form_data.get('_current_input_field') or '').strip()
+    if not field_name:
+        return '', []
+    field = next(
+        (
+            item
+            for item in form_data.get('input_defs') or []
+            if str(item.get('output_variable_name') or '').strip() == field_name
+        ),
+        None,
+    )
+    if not field or str(field.get('type') or '').strip().lower() != 'select':
+        return '', []
+
+    source = field.get('option_source') or {}
+    source_value = source.get('value') if isinstance(source, dict) else None
+    if isinstance(source_value, list):
+        return field_name, [str(item) for item in source_value]
+    if isinstance(source_value, str):
+        return field_name, [part.strip() for part in source_value.splitlines() if part.strip()]
+
+    options = field.get('options')
+    if not isinstance(options, list):
+        return field_name, []
+    values = []
+    for item in options:
+        if isinstance(item, dict):
+            values.append(str(item.get('label') or item.get('value') or ''))
+        else:
+            values.append(str(item))
+    return field_name, [value for value in values if value]
+
+
+def _telegram_form_action_from_callback(data: dict) -> dict | None:
+    """Translate compact Telegram callback data into a runner form action."""
+    if 'x' not in data:
+        return {
+            'action_id': str(data.get('action_id') or data.get('a') or ''),
+            'inputs': {},
+        }
+    try:
+        option_index = int(data['x'])
+    except (TypeError, ValueError):
+        return None
+    if option_index < 0:
+        return None
+    return {
+        'action_id': '',
+        'inputs': {'select': {'index': option_index}},
+        '_input_progress': True,
+    }
+
+
 class TelegramMessageConverter(abstract_platform_adapter.AbstractMessageConverter):
     @staticmethod
     async def yiri2target(message_chain: platform_message.MessageChain, bot: telegram.Bot) -> list[dict]:
@@ -167,7 +224,7 @@ class TelegramEventConverter(abstract_platform_adapter.AbstractEventConverter):
                 time=event.message.date.timestamp(),
                 source_platform_object=event,
             )
-        elif event.effective_chat.type == 'group' or 'supergroup':
+        elif event.effective_chat.type in ('group', 'supergroup'):
             return platform_events.GroupMessage(
                 sender=platform_entities.GroupMember(
                     id=event.effective_chat.id,
@@ -189,6 +246,7 @@ class TelegramEventConverter(abstract_platform_adapter.AbstractEventConverter):
 class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
     bot: telegram.Bot = pydantic.Field(exclude=True)
     application: telegram.ext.Application = pydantic.Field(exclude=True)
+    ap: typing.Any = pydantic.Field(exclude=True, default=None)
 
     message_converter: TelegramMessageConverter = TelegramMessageConverter()
     event_converter: TelegramEventConverter = TelegramEventConverter()
@@ -204,6 +262,48 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         typing.Callable[[platform_events.Event, abstract_platform_adapter.AbstractMessagePlatformAdapter], None],
     ] = {}
 
+    _FORM_ACTION_CACHE_TTL = 30 * 60
+    # callback_data -> (display title, pipeline UUID, expiration time, form group id)
+    _form_action_titles: typing.Dict[str, tuple[str, str, float, str]] = {}
+
+    def _prune_form_action_titles(self, now: float | None = None) -> None:
+        now = time.monotonic() if now is None else now
+        expired = [key for key, (_, _, expires_at, _) in self._form_action_titles.items() if expires_at <= now]
+        for key in expired:
+            self._form_action_titles.pop(key, None)
+
+    def _cache_form_action_titles(
+        self,
+        mappings: dict[str, str],
+        pipeline_uuid: str = '',
+        now: float | None = None,
+    ) -> None:
+        now = time.monotonic() if now is None else now
+        self._prune_form_action_titles(now)
+        group_id = uuid.uuid4().hex
+        expires_at = now + self._FORM_ACTION_CACHE_TTL
+        self._form_action_titles.update(
+            {callback_data: (title, pipeline_uuid, expires_at, group_id) for callback_data, title in mappings.items()}
+        )
+
+    def _take_form_action_context(self, callback_data: str, now: float | None = None) -> tuple[str, str] | None:
+        """Consume a callback and invalidate every button from the same form."""
+        self._prune_form_action_titles(now)
+        entry = self._form_action_titles.get(callback_data)
+        if entry is None:
+            return None
+        title, pipeline_uuid, _, group_id = entry
+        group_keys = [
+            key for key, (_, _, _, cached_group_id) in self._form_action_titles.items() if cached_group_id == group_id
+        ]
+        for key in group_keys:
+            self._form_action_titles.pop(key, None)
+        return title, pipeline_uuid
+
+    def _take_form_action_title(self, callback_data: str, now: float | None = None) -> str | None:
+        context = self._take_form_action_context(callback_data, now)
+        return context[0] if context else None
+
     def __init__(self, config: dict, logger: abstract_platform_logger.AbstractEventLogger):
         async def telegram_callback(update: Update, context: ContextTypes.DEFAULT_TYPE):
             if update.message.from_user.is_bot:
@@ -224,6 +324,117 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
                 telegram_callback,
             )
         )
+
+        async def callback_query_handler(update: Update, context: ContextTypes.DEFAULT_TYPE):
+            query = update.callback_query
+            await query.answer()
+            try:
+                data = json.loads(query.data)
+                if data.get('form_action') or data.get('f'):
+                    import langbot_plugin.api.entities.builtin.provider.session as provider_session
+
+                    # workflow_run_id is not in the callback payload (too large
+                    # for Telegram's 64-byte limit). Only w_suffix is sent;
+                    # the runner resolves the full run id from _PENDING_FORMS.
+                    w_suffix = data.get('w', '')
+                    session_key = data.get('session_key') or data.get('s', '')
+                    callback_action = _telegram_form_action_from_callback(data)
+                    action_context = self._take_form_action_context(query.data) if callback_action is not None else None
+                    if callback_action is None or action_context is None:
+                        await self.logger.warning(f'Invalid or stale Telegram form callback: {query.data!r}')
+                        return
+                    action_title, pipeline_uuid = action_context
+                    # Show selected action feedback by editing the original message
+                    try:
+                        original_text = query.message.text or ''
+                        selected_text = f'{original_text}\n\n✅ {action_title}'
+                        await query.edit_message_text(text=selected_text, reply_markup=None)
+                    except Exception:
+                        # If edit fails (e.g. message too long), just pass
+                        pass
+
+                    if session_key.startswith('group_') or session_key.startswith('g:'):
+                        launcher_type = provider_session.LauncherTypes.GROUP
+                        launcher_id = (
+                            session_key.split(':', 1)[1]
+                            if session_key.startswith('g:')
+                            else session_key[len('group_') :]
+                        )
+                    else:
+                        launcher_type = provider_session.LauncherTypes.PERSON
+                        launcher_id = (
+                            session_key.split(':', 1)[1]
+                            if session_key.startswith('p:')
+                            else session_key[len('person_') :]
+                        )
+
+                    user_id = str(query.from_user.id)
+
+                    # Find bot_uuid and pipeline_uuid
+                    bot_uuid = ''
+                    for b in self.ap.platform_mgr.bots:
+                        if b.adapter is self:
+                            bot_uuid = b.bot_entity.uuid
+                            pipeline_uuid = pipeline_uuid or b.bot_entity.use_pipeline_uuid
+                            break
+
+                    form_action_data = {
+                        # workflow_run_id is intentionally omitted; the runner
+                        # resolves it from w_suffix via _PENDING_FORMS.
+                        'w_suffix': w_suffix,
+                        'user': f'{launcher_type.value}_{launcher_id}',
+                        **callback_action,
+                    }
+
+                    event_label = 'Form Select' if callback_action.get('_input_progress') else 'Form Action'
+                    message_chain = platform_message.MessageChain(
+                        [platform_message.Plain(text=f'[{event_label}: {action_title}]')]
+                    )
+
+                    if launcher_type == provider_session.LauncherTypes.GROUP:
+                        synthetic_event = platform_events.GroupMessage(
+                            sender=platform_entities.GroupMember(
+                                id=user_id,
+                                member_name='',
+                                permission=platform_entities.Permission.Member,
+                                group=platform_entities.Group(
+                                    id=launcher_id,
+                                    name='',
+                                    permission=platform_entities.Permission.Member,
+                                ),
+                            ),
+                            message_chain=message_chain,
+                            source_platform_object=update,
+                        )
+                    else:
+                        synthetic_event = platform_events.FriendMessage(
+                            sender=platform_entities.Friend(
+                                id=user_id,
+                                nickname='',
+                                remark='',
+                            ),
+                            message_chain=message_chain,
+                            source_platform_object=update,
+                        )
+
+                    await self.ap.query_pool.add_query(
+                        bot_uuid=bot_uuid,
+                        launcher_type=launcher_type,
+                        launcher_id=launcher_id,
+                        sender_id=user_id,
+                        message_event=synthetic_event,
+                        message_chain=message_chain,
+                        adapter=self,
+                        pipeline_uuid=pipeline_uuid,
+                        variables={
+                            '_dify_form_action': form_action_data,
+                            '_routed_by_rule': True,
+                        },
+                    )
+            except Exception:
+                await self.logger.error(f'Error in telegram callback query: {traceback.format_exc()}')
+
+        application.add_handler(CallbackQueryHandler(callback_query_handler))
         super().__init__(
             config=config,
             logger=logger,
@@ -314,23 +525,34 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
             args['parse_mode'] = 'MarkdownV2'
         return args
 
+    async def _delete_group_stream_message(self, chat_mode: str, chat_id: int, stream_id: int | None):
+        if chat_mode != 'group' or stream_id is None:
+            return
+        try:
+            await self.bot.delete_message(chat_id=chat_id, message_id=stream_id)
+        except telegram.error.TelegramError:
+            pass
+
+    @staticmethod
+    def _is_form_placeholder_chunk(text: str) -> bool:
+        """Return True for invisible placeholder chunks used to carry forms."""
+
+        if not text:
+            return True
+
+        cleaned = text.replace('\u200b', '').replace('\u200c', '').replace('\u200d', '').replace('\ufeff', '').strip()
+        return cleaned == ''
+
     async def create_message_card(self, message_id, event):
         assert isinstance(event.source_platform_object, Update)
         update = event.source_platform_object
         chat_id = update.effective_chat.id
-        chat_type = update.effective_chat.type
+        effective_message = update.effective_message
-        message_thread_id = update.message.message_thread_id
+        message_thread_id = getattr(effective_message, 'message_thread_id', None) if effective_message else None
 
-        if chat_type == 'private':
+        args = self._build_message_args(chat_id, 'Thinking...', message_thread_id)
-            draft_id = int(time.time() * 1000)
+        send_msg = await self.bot.send_message(**args)
-            self.msg_stream_id[message_id] = ('private', draft_id)
+        self.msg_stream_id[message_id] = ('message', send_msg.message_id, False)
-
-            args = self._build_message_args(chat_id, 'Thinking...', message_thread_id, draft_id=draft_id)
-            await self.bot.send_message_draft(**args)
-        else:
-            args = self._build_message_args(chat_id, 'Thinking...', message_thread_id)
-            send_msg = await self.bot.send_message(**args)
-            self.msg_stream_id[message_id] = ('group', send_msg.message_id)
 
         return True
 
@@ -347,12 +569,15 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         assert isinstance(message_source.source_platform_object, Update)
         update = message_source.source_platform_object
         chat_id = update.effective_chat.id
-        message_thread_id = update.message.message_thread_id
+        effective_message = update.effective_message
+        message_thread_id = getattr(effective_message, 'message_thread_id', None) if effective_message else None
 
         if message_id not in self.msg_stream_id:
             return
 
-        chat_mode, draft_id = self.msg_stream_id[message_id]
+        stream_state = self.msg_stream_id[message_id]
+        chat_mode, stream_id = stream_state[:2]
+        has_visible_content = len(stream_state) > 2 and stream_state[2]
         components = await TelegramMessageConverter.yiri2target(message, self.bot)
 
         if not components or components[0]['type'] != 'text':
@@ -361,17 +586,68 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
             return
 
         content = components[0]['text']
+        form_data = getattr(bot_message, '_form_data', None)
+
+        if form_data and is_final:
+            if not has_visible_content:
+                await self._send_form_action_buttons(message_source, form_data, edit_message_id=stream_id)
+            else:
+                await self._send_form_action_buttons(message_source, form_data)
+            self.msg_stream_id.pop(message_id, None)
+            return
+
+        if self._is_form_placeholder_chunk(content):
+            if is_final and bot_message.tool_calls is None and not has_visible_content:
+                await self._delete_group_stream_message(chat_mode, chat_id, stream_id)
+                self.msg_stream_id.pop(message_id, None)
+            return
 
         if chat_mode == 'private':
-            args = self._build_message_args(chat_id, content, message_thread_id, draft_id=draft_id)
+            # Streaming via draft (ephemeral preview in the chat input area)
-            await self.bot.send_message_draft(**args)
+            if (msg_seq - 1) % 8 == 0 or is_final:
+                args = self._build_message_args(chat_id, content, message_thread_id, draft_id=stream_id)
+                try:
+                    await self.bot.send_message_draft(**args)
+                except telegram.error.BadRequest as exc:
+                    if 'Message_too_long' in str(exc):
+                        args['text'] = content[:4000] + '\n\n… (truncated)'
+                        try:
+                            await self.bot.send_message_draft(**args)
+                        except telegram.error.RetryAfter:
+                            pass
+                    else:
+                        pass  # Ignore other draft errors (cosmetic)
+                self.msg_stream_id[message_id] = (chat_mode, stream_id, True)
             if is_final and bot_message.tool_calls is None:
-                del args['draft_id']
+                # Finalise: send the real message, discard the draft
-                await self.bot.send_message(**args)
+                args = self._build_message_args(chat_id, content, message_thread_id)
+                try:
+                    await self.bot.send_message(**args)
+                except telegram.error.BadRequest as exc:
+                    if 'Message_too_long' in str(exc):
+                        args['text'] = content[:4000] + '\n\n… (truncated)'
+                        await self.bot.send_message(**args)
+                    else:
+                        raise
                 self.msg_stream_id.pop(message_id)
         else:
-            stream_id = draft_id
+            # Streaming via edit_message_text (persistent message)
-            if (msg_seq - 1) % 8 == 0 or is_final:
+            if stream_id is None:
+                args = self._build_message_args(chat_id, content, message_thread_id)
+                try:
+                    send_msg = await self.bot.send_message(**args)
+                except telegram.error.BadRequest as exc:
+                    if 'Message_too_long' in str(exc):
+                        args['text'] = self._process_markdown(content[:4000] + '\n\n鈥?(truncated)')
+                        send_msg = await self.bot.send_message(**args)
+                    else:
+                        raise
+                self.msg_stream_id[message_id] = (chat_mode, send_msg.message_id, True)
+                if is_final and bot_message.tool_calls is None:
+                    self.msg_stream_id.pop(message_id, None)
+                return
+
+            if not has_visible_content or (msg_seq - 1) % 8 == 0 or is_final:
                 args = {
                     'message_id': stream_id,
                     'chat_id': chat_id,
@@ -379,11 +655,137 @@ class TelegramAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
                 }
                 if self.config.get('markdown_card', False):
                     args['parse_mode'] = 'MarkdownV2'
-                await self.bot.edit_message_text(**args)
+                try:
+                    await self.bot.edit_message_text(**args)
+                except telegram.error.BadRequest as exc:
+                    if 'Message_too_long' in str(exc):
+                        args['text'] = self._process_markdown(content[:4000] + '\n\n… (truncated)')
+                        await self.bot.edit_message_text(**args)
+                    else:
+                        raise
+                self.msg_stream_id[message_id] = (chat_mode, stream_id, True)
 
             if is_final and bot_message.tool_calls is None:
                 self.msg_stream_id.pop(message_id)
 
+    async def _send_form_action_buttons(
+        self,
+        message_source: platform_events.MessageEvent,
+        form_data: dict,
+        edit_message_id: int | None = None,
+    ):
+        """Send inline keyboard buttons for Dify form fields or actions."""
+        actions = form_data.get('actions', [])
+        node_title = form_data.get('node_title', '')
+        form_content = form_data.get('form_content', '')
+        workflow_run_id = form_data.get('workflow_run_id', '')
+        # Telegram callback_data is capped at 64 bytes, so we identify the
+        # paused workflow by the last 8 chars of workflow_run_id (unique
+        # within a session with overwhelming probability).
+        w_suffix = workflow_run_id[-8:] if workflow_run_id else ''
+
+        if isinstance(message_source, platform_events.GroupMessage):
+            session_key = f'g:{message_source.group.id}'
+        else:
+            session_key = f'p:{message_source.sender.id}'
+
+        current_field = str(form_data.get('_current_input_field') or '').strip()
+        is_field_step = bool(current_field) and not form_data.get('_action_select_only')
+        select_field, select_options = _telegram_select_field_options(form_data)
+        is_select_field = bool(select_field and select_options)
+        if is_select_field:
+            choices = [(option, {'x': idx}) for idx, option in enumerate(select_options)]
+        elif is_field_step:
+            choices = []
+        else:
+            choices = [(action.get('title', action.get('id', '')), {'a': action.get('id', '')}) for action in actions]
+
+        keyboard = []
+        pending_title_mappings: dict[str, str] = {}
+        oversized = False
+        buttons_per_row = 2 if is_select_field else 1
+        current_row = []
+        for title, choice_data in choices:
+            callback_payload = {'f': 1, **choice_data, 's': session_key}
+            if w_suffix:
+                callback_payload['w'] = w_suffix
+            callback_data = json.dumps(callback_payload, separators=(',', ':'))
+            if len(callback_data.encode('utf-8')) > 64:
+                oversized = True
+                break
+            pending_title_mappings[callback_data] = str(title)
+            current_row.append(InlineKeyboardButton(str(title), callback_data=callback_data))
+            if len(current_row) == buttons_per_row:
+                keyboard.append(current_row)
+                current_row = []
+        if current_row and not oversized:
+            keyboard.append(current_row)
+
+        update = message_source.source_platform_object
+        chat_id = update.effective_chat.id
+        effective_message = update.effective_message
+        message_thread_id = getattr(effective_message, 'message_thread_id', None) if effective_message else None
+
+        heading = f'[{node_title}]'
+        text_lines = [heading]
+        if form_content:
+            text_lines.append(form_content)
+
+        if oversized:
+            # callback_data exceeds Telegram's 64-byte limit — fall back to
+            # a plain-text numbered list so the user can reply by number.
+            for idx, (title, _) in enumerate(choices, start=1):
+                text_lines.append(f'  {idx}. {title}')
+            args = {
+                'chat_id': chat_id,
+                'text': '\n\n'.join(text_lines),
+            }
+        elif keyboard:
+            self._cache_form_action_titles(
+                pending_title_mappings,
+                str(form_data.get('pipeline_uuid') or ''),
+            )
+            reply_markup = InlineKeyboardMarkup(keyboard)
+            args = {
+                'chat_id': chat_id,
+                'text': '\n\n'.join(text_lines),
+                'reply_markup': reply_markup,
+            }
+        elif is_field_step:
+            args = {
+                'chat_id': chat_id,
+                'text': '\n\n'.join(text_lines),
+                # Telegram privacy-mode bots receive replies to ForceReply
+                # prompts even when they cannot read ordinary group messages.
+                'reply_markup': ForceReply(
+                    selective=False,
+                    input_field_placeholder=current_field,
+                ),
+            }
+        else:
+            args = {
+                'chat_id': chat_id,
+                'text': '\n\n'.join(text_lines),
+            }
+
+        if message_thread_id:
+            args['message_thread_id'] = message_thread_id
+
+        if edit_message_id is not None:
+            edit_args = {
+                'chat_id': chat_id,
+                'message_id': edit_message_id,
+                'text': args['text'],
+            }
+            edit_args['reply_markup'] = args.get('reply_markup')
+            try:
+                await self.bot.edit_message_text(**edit_args)
+                return
+            except telegram.error.TelegramError:
+                await self._delete_group_stream_message('group', chat_id, edit_message_id)
+
+        await self.bot.send_message(**args)
+
     def get_launcher_id(self, event: platform_events.MessageEvent) -> str | None:
         if not isinstance(event.source_platform_object, Update):
             return None

src/langbot/pkg/platform/sources/wecombot.py

@@ -11,7 +11,13 @@ import langbot_plugin.api.entities.builtin.platform.events as platform_events
 import langbot_plugin.api.entities.builtin.platform.entities as platform_entities
 from ..logger import EventLogger
 from langbot.libs.wecom_ai_bot_api.wecombotevent import WecomBotEvent
-from langbot.libs.wecom_ai_bot_api.api import WecomBotClient
+from langbot.libs.wecom_ai_bot_api.api import (
+    WecomBotClient,
+    extract_template_card_action,
+    extract_template_card_event_payload,
+    extract_template_card_selections,
+    parse_select_button_action,
+)
 from langbot.libs.wecom_ai_bot_api.ws_client import WecomBotWsClient
 
 
@@ -296,6 +302,7 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
     listeners: dict = {}
     _stream_to_monitoring_msg: dict = {}  # Maps stream_id to (monitoring_message_id, timestamp)
     _STREAM_MAPPING_TTL = 600  # 10 minutes
+    ap: typing.Any = None
 
     def __init__(self, config: dict, logger: EventLogger):
         enable_webhook = config.get('enable-webhook', False)
@@ -336,6 +343,25 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
             _stream_to_monitoring_msg={},
         )
 
+        # Both WecomBotClient (webhook) and WecomBotWsClient (ws long-conn)
+        # expose ``set_card_action_callback``. Wire the click handler so
+        # Dify human-input button taps resume the workflow on either mode.
+        if hasattr(self.bot, 'set_card_action_callback'):
+            self.bot.set_card_action_callback(self._on_card_action)
+
+        # Hand the client a `source` block so every interactive
+        # template_card it emits carries the LangBot logo + name at the
+        # top — the WeCom analogue of DingTalk's Avatar header.
+        # Always on; icon_url accepts plain HTTPS URLs (no upload needed).
+        if hasattr(self.bot, 'set_card_source'):
+            self.bot.set_card_source(
+                {
+                    'icon_url': 'https://raw.githubusercontent.com/RockChinQ/LangBot/master/res/logo-blue.png',
+                    'desc': 'LangBot',
+                    'desc_color': 0,
+                }
+            )
+
     async def reply_message(
         self,
         message_source: platform_events.MessageEvent,
@@ -345,15 +371,37 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         content = await self.message_converter.yiri2target(message)
         _ws_mode = not self.config.get('enable-webhook', False)
 
+        event = message_source.source_platform_object
+        # Synthetic events (button-click resume queries) have no inbound
+        # platform object. Fall back to a proactive send so error
+        # messages and one-shot replies still reach the user.
+        if event is None:
+            if _ws_mode:
+                if isinstance(message_source, platform_events.GroupMessage):
+                    chat_id = str(message_source.group.id)
+                else:
+                    chat_id = str(message_source.sender.id)
+                try:
+                    await self.bot.send_message(chat_id, content)
+                except Exception:
+                    await self.logger.error(
+                        f'WeComBot: proactive reply for synthetic event failed: {traceback.format_exc()}'
+                    )
+            else:
+                await self.logger.warning(
+                    'WeComBot webhook mode cannot reply to a synthetic event '
+                    '(no req_id and no proactive-send credentials); dropping.'
+                )
+            return
+
         if _ws_mode:
-            event = message_source.source_platform_object
+            req_id = event.get('req_id', '') if isinstance(event, dict) else getattr(event, 'req_id', '')
-            req_id = event.get('req_id', '')
             if req_id:
                 await self.bot.reply_text(req_id, content)
             else:
                 await self.bot.set_message(event.message_id, content)
         else:
-            await self.bot.set_message(message_source.source_platform_object.message_id, content)
+            await self.bot.set_message(event.message_id, content)
 
     async def reply_message_chunk(
         self,
@@ -364,9 +412,56 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         is_final: bool = False,
     ):
         content = await self.message_converter.yiri2target(message)
-        msg_id = message_source.source_platform_object.message_id
         _ws_mode = not self.config.get('enable-webhook', False)
 
+        # Synthetic events (e.g. button-click triggered form resume) have
+        # no inbound platform message — no msg_id, no req_id, no stream
+        # session. The output must go via the proactive-send path instead
+        # of the stream/reply path.
+        spo = message_source.source_platform_object
+        if spo is None:
+            return await self._handle_synthetic_chunk(message_source, bot_message, content, is_final, _ws_mode)
+
+        msg_id = spo.message_id
+
+        # Dify human-input pause: when the runner attaches `_form_data` to
+        # the final chunk, hand the button_interaction card off to the
+        # underlying client. In webhook mode the card is queued for the
+        # next followup poll; in ws mode it's sent as a reply frame
+        # immediately. Falls back to plain text when the bot has no active
+        # stream session for this msg_id (rare).
+        form_data = getattr(bot_message, '_form_data', None)
+        if form_data and is_final:
+            if hasattr(self.bot, 'push_form_pause'):
+                ok, stream_id, task_id = await self.bot.push_form_pause(msg_id, form_data)
+                if ok:
+                    await self.logger.info(
+                        f'WeComBot: pending button_interaction registered '
+                        f'stream_id={stream_id} task_id={task_id} ws_mode={_ws_mode}'
+                    )
+                    return {'stream': True, 'form': True, 'task_id': task_id}
+            await self.logger.warning(
+                'WeComBot: cannot register form pause (no active stream session); falling back to plain text'
+            )
+            try:
+                from langbot.pkg.provider.runners.difysvapi import _format_human_input_text
+
+                fallback = _format_human_input_text(
+                    form_data.get('node_title', ''),
+                    form_data.get('form_content', ''),
+                    form_data.get('actions', []) or [],
+                )
+            except Exception:
+                fallback = content or '（人工输入）'
+            if _ws_mode:
+                event = message_source.source_platform_object
+                req_id = event.get('req_id', '') if isinstance(event, dict) else getattr(event, 'req_id', '')
+                if req_id:
+                    await self.bot.reply_text(req_id, fallback)
+            else:
+                await self.bot.set_message(msg_id, fallback)
+            return {'stream': False, 'form': True, 'fallback': True}
+
         if _ws_mode:
             success = await self.bot.push_stream_chunk(msg_id, content, is_final=is_final)
             if not success and is_final:
@@ -385,6 +480,142 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
         """Whether streaming output is enabled for this bot instance."""
         return self.config.get('enable-stream-reply', True)
 
+    async def _handle_synthetic_chunk(
+        self,
+        message_source: platform_events.MessageEvent,
+        bot_message,
+        content: str,
+        is_final: bool,
+        ws_mode: bool,
+    ) -> dict:
+        """Handle reply_message_chunk for synthetic events (button clicks).
+
+        Synthetic events have no inbound message → no msg_id, no req_id,
+        no stream session. We can't do incremental streaming, so we
+        buffer chunks per-conversation and flush on ``is_final`` via the
+        proactive send path.
+
+        Buffer keyed by ``(launcher_type, launcher_id)`` from the
+        synthetic event itself. Only ws mode has a usable proactive-send
+        path right now (``ws_client.send_message`` /
+        ``ws_client.send_template_card``); webhook mode requires a
+        corpid/secret we don't have, so it logs and drops.
+        """
+        if isinstance(message_source, platform_events.GroupMessage):
+            chat_id = str(message_source.group.id)
+        else:
+            chat_id = str(message_source.sender.id)
+
+        form_data = getattr(bot_message, '_form_data', None)
+
+        # Buffer streaming content until is_final.
+        buf_key = chat_id
+        if not hasattr(self, '_synthetic_buffers'):
+            # Attribute-not-declared trick: pydantic forbids dynamic attrs
+            # on the model, but plain instance dicts via object.__setattr__
+            # do work. Lazy-create on first call.
+            object.__setattr__(self, '_synthetic_buffers', {})
+        buffers: dict[str, str] = self._synthetic_buffers
+        if content and not form_data:
+            previous = buffers.get(buf_key, '')
+            if previous and content.startswith(previous):
+                buffers[buf_key] = content
+            elif previous and previous.endswith(content):
+                buffers[buf_key] = previous
+            else:
+                buffers[buf_key] = previous + content
+
+        if not is_final:
+            return {'stream': True, 'synthetic': True, 'buffered': True}
+
+        final_content = buffers.pop(buf_key, '')
+        if content:
+            if final_content and content.startswith(final_content):
+                final_content = content
+            elif final_content and final_content.endswith(content):
+                pass
+            else:
+                final_content = final_content + content
+
+        if not ws_mode:
+            await self.logger.warning(
+                'WeComBot webhook mode cannot proactively push synthetic-event '
+                'output (no corpid/secret); the resume reply is dropped. '
+                f'content_len={len(final_content)} form_data_present={form_data is not None}'
+            )
+            return {'stream': False, 'synthetic': True, 'dropped': True}
+
+        # ws mode: proactive send.
+        try:
+            if form_data:
+                # Determine user_id / chat_id for the routing context of any
+                # subsequent click on this card.
+                if isinstance(message_source, platform_events.GroupMessage):
+                    routing_chat_id = str(message_source.group.id)
+                    routing_user_id = str(message_source.sender.id)
+                else:
+                    routing_chat_id = ''
+                    routing_user_id = str(message_source.sender.id)
+                payload = self._build_button_interaction_payload_from_form(
+                    form_data,
+                    user_id=routing_user_id,
+                    chat_id=routing_chat_id,
+                )
+                await self.bot.send_template_card(chat_id, payload)
+                await self.logger.info(
+                    f'WeComBot ws: proactively sent template_card for synthetic event '
+                    f'chat_id={chat_id} form_token={form_data.get("form_token")!r} '
+                    f'workflow_run_id={form_data.get("workflow_run_id")!r}'
+                )
+            elif final_content:
+                await self.bot.send_message(chat_id, final_content)
+                await self.logger.info(
+                    f'WeComBot ws: proactively sent text for synthetic event chat_id={chat_id} len={len(final_content)}'
+                )
+        except Exception:
+            await self.logger.error(f'WeComBot: synthetic event proactive send failed: {traceback.format_exc()}')
+            return {'stream': False, 'synthetic': True, 'error': True}
+
+        return {'stream': True, 'synthetic': True}
+
+    def _build_button_interaction_payload_from_form(
+        self, form_data: dict, *, user_id: str = '', chat_id: str = ''
+    ) -> dict:
+        """Build a button_interaction payload + track task_id for click resolution.
+
+        Unlike the inbound-event path (where push_form_pause registers the
+        task_id with the active stream session), proactive sends still
+        need the task_id registered so button clicks find pending_form.
+        For ws mode we stash it directly on the ws_client's pending dict.
+        """
+        from langbot.libs.wecom_ai_bot_api.api import build_human_input_template_card_payload
+        import secrets as _secrets
+
+        task_id = f'dify-{_secrets.token_hex(12)}'
+        source = getattr(self.bot, 'card_source', None)
+        payload = build_human_input_template_card_payload(
+            form_data,
+            task_id,
+            source=source,
+            select_as_buttons=not self.config.get('enable-webhook', False),
+        )
+
+        # Register task_id → form_data so the click callback can find it.
+        # user_id / chat_id are required so _on_card_action can route the
+        # resulting synthetic query back to the right user. msg_id / req_id
+        # / stream_id are intentionally empty — synthetic cards have no
+        # inbound message to anchor on.
+        if hasattr(self.bot, '_pending_forms_by_task'):
+            self.bot._pending_forms_by_task[task_id] = {
+                'form_data': form_data,
+                'msg_id': '',
+                'user_id': user_id,
+                'chat_id': chat_id,
+                'stream_id': '',
+                'req_id': '',
+            }
+        return payload
+
     async def send_message(self, target_type, target_id, message):
         _ws_mode = not self.config.get('enable-webhook', False)
         if _ws_mode:
@@ -531,3 +762,191 @@ class WecomBotAdapter(abstract_platform_adapter.AbstractMessagePlatformAdapter):
 
     async def is_muted(self, group_id: int) -> bool:
         pass
+
+    # ------------------------------------------------------------------
+    # Dify human-input button-interaction click handling
+    # ------------------------------------------------------------------
+
+    async def _on_card_action(self, session, action_id: str, task_id: str, raw_event: dict) -> None:
+        """Translate a button click on a button_interaction card into a
+        synthetic ``_dify_form_action`` query enqueued on the pool.
+
+        Pattern mirrors DingTalk / Lark / Telegram so the runner's
+        ``_merge_pending_form_action`` path resumes the workflow.
+        """
+        import langbot_plugin.api.entities.builtin.provider.session as provider_session
+
+        form = session.pending_form or {}
+        await self.logger.info(
+            f'WeComBot _on_card_action: task_id={task_id} action_id={action_id!r} '
+            f'form_token={form.get("form_token")!r} workflow_run_id={form.get("workflow_run_id")!r} '
+            f'session.user_id={session.user_id!r} session.chat_id={session.chat_id!r}'
+        )
+
+        actions = form.get('actions') or []
+        tce = extract_template_card_event_payload(raw_event) if isinstance(raw_event, dict) else {}
+        _, _, card_type = extract_template_card_action(tce)
+        selections = extract_template_card_selections(tce, form)
+        if not selections:
+            selections = parse_select_button_action(action_id, form)
+        await self.logger.info(
+            f'WeComBot template_card selections: task_id={task_id} card_type={card_type} selections={selections}'
+        )
+        if card_type == 'multiple_interaction' and not selections:
+            await self.logger.warning(
+                f'WeComBot: multiple_interaction callback has no parseable selections; raw={str(tce)[:1000]}'
+            )
+            return
+        is_select_submit = card_type == 'multiple_interaction' or bool(selections)
+
+        clean_action_id = '' if is_select_submit else (action_id or '').strip()
+        action_title = clean_action_id
+        for a in actions:
+            if str(a.get('id', '')) == clean_action_id:
+                action_title = a.get('title') or clean_action_id
+                break
+
+        inputs = dict(form.get('inputs') or {})
+        inputs.update(selections)
+
+        def _missing_fields_after_select() -> list[str]:
+            missing: list[str] = []
+            for field in form.get('input_defs') or form.get('all_input_defs') or []:
+                field_name = str(field.get('output_variable_name') or '').strip()
+                if not field_name:
+                    continue
+                if inputs.get(field_name) in (None, '', []):
+                    missing.append(field_name)
+            return missing
+
+        input_progress = False
+        if is_select_submit:
+            missing_fields = _missing_fields_after_select()
+            if not missing_fields and len(actions) == 1:
+                action = actions[0]
+                clean_action_id = str(action.get('id') or '').strip()
+                action_title = action.get('title') or clean_action_id
+            elif not missing_fields and len(actions) > 1:
+                if not self.config.get('enable-webhook', False):
+                    action_form_data = {
+                        'form_content': form.get('raw_form_content') or form.get('form_content') or '',
+                        'raw_form_content': form.get('raw_form_content') or form.get('form_content') or '',
+                        'input_defs': [],
+                        'all_input_defs': form.get('all_input_defs') or form.get('input_defs') or [],
+                        'inputs': inputs,
+                        'actions': actions,
+                        'node_title': form.get('node_title', ''),
+                        'workflow_run_id': form.get('workflow_run_id', ''),
+                        'form_token': form.get('form_token', ''),
+                        'pipeline_uuid': form.get('pipeline_uuid', ''),
+                        '_action_select_only': True,
+                    }
+                    target_chat_id = session.chat_id or session.user_id or ''
+                    try:
+                        payload = self._build_button_interaction_payload_from_form(
+                            action_form_data,
+                            user_id=session.user_id or '',
+                            chat_id=session.chat_id or '',
+                        )
+                        await self.bot.send_template_card(target_chat_id, payload)
+                        await self.logger.info(
+                            f'WeComBot: sent action-select button card after select submit '
+                            f'task_id={task_id} action_count={len(actions)}'
+                        )
+                    except Exception:
+                        await self.logger.error(
+                            f'WeComBot: failed to send action-select button card: {traceback.format_exc()}'
+                        )
+                    return
+                await self.logger.warning(
+                    'WeComBot webhook mode cannot proactively send action-select button card after select submit'
+                )
+                return
+            else:
+                input_progress = True
+                action_title = 'Submit'
+
+        launcher_id = session.user_id or session.chat_id or ''
+        sender_user_id = session.user_id or launcher_id
+        # WeCom AI bot has both single-chat and group-chat; chat_id present
+        # indicates group context.
+        if session.chat_id:
+            launcher_type = provider_session.LauncherTypes.GROUP
+            launcher_id = session.chat_id
+        else:
+            launcher_type = provider_session.LauncherTypes.PERSON
+            launcher_id = session.user_id or ''
+
+        form_action_data = {
+            'form_token': form.get('form_token', ''),
+            'workflow_run_id': form.get('workflow_run_id', ''),
+            'action_id': clean_action_id,
+            'action_title': action_title,
+            'node_title': form.get('node_title', ''),
+            'user': f'{launcher_type.value}_{launcher_id}',
+            'inputs': inputs,
+        }
+        if input_progress:
+            form_action_data['_input_progress'] = True
+
+        message_chain = platform_message.MessageChain([platform_message.Plain(text=f'[Form Action: {action_title}]')])
+
+        if launcher_type == provider_session.LauncherTypes.GROUP:
+            synthetic_event = platform_events.GroupMessage(
+                sender=platform_entities.GroupMember(
+                    id=sender_user_id,
+                    member_name='',
+                    permission=platform_entities.Permission.Member,
+                    group=platform_entities.Group(
+                        id=launcher_id,
+                        name='',
+                        permission=platform_entities.Permission.Member,
+                    ),
+                    special_title='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+        else:
+            synthetic_event = platform_events.FriendMessage(
+                sender=platform_entities.Friend(
+                    id=sender_user_id,
+                    nickname='',
+                    remark='',
+                ),
+                message_chain=message_chain,
+                time=int(time.time()),
+                source_platform_object=None,
+            )
+
+        if self.ap is None:
+            await self.logger.error('WeComBot: ap not injected; cannot enqueue button-click query')
+            return
+
+        bot_uuid = ''
+        pipeline_uuid = form.get('pipeline_uuid') or None
+        for bot in self.ap.platform_mgr.bots:
+            if bot.adapter is self:
+                bot_uuid = bot.bot_entity.uuid
+                pipeline_uuid = pipeline_uuid or bot.bot_entity.use_pipeline_uuid
+                break
+
+        try:
+            await self.ap.query_pool.add_query(
+                bot_uuid=bot_uuid,
+                launcher_type=launcher_type,
+                launcher_id=launcher_id,
+                sender_id=sender_user_id,
+                message_event=synthetic_event,
+                message_chain=message_chain,
+                adapter=self,
+                pipeline_uuid=pipeline_uuid,
+                variables={
+                    '_dify_form_action': form_action_data,
+                    '_routed_by_rule': True,
+                },
+            )
+            await self.logger.info(f'WeComBot: button-click query enqueued action_id={clean_action_id!r}')
+        except Exception:
+            await self.logger.error(f'WeComBot: enqueue button-click query failed: {traceback.format_exc()}')

src/langbot/pkg/provider/tools/loaders/mcp.py

@@ -6,7 +6,7 @@ import json
 import re
 import time
 import typing
-from contextlib import AsyncExitStack
+from contextlib import AsyncExitStack, asynccontextmanager
 import traceback
 from langbot_plugin.api.entities.events import pipeline_query
 import sqlalchemy
@@ -18,6 +18,7 @@ from mcp import ClientSession, StdioServerParameters, types as mcp_types
 from mcp.client.stdio import stdio_client
 from mcp.client.sse import sse_client
 from mcp.client.streamable_http import streamable_http_client
+from mcp.shared.exceptions import McpError
 from pydantic import AnyUrl
 
 from .. import loader
@@ -25,7 +26,7 @@ from ....core import app
 import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
 import langbot_plugin.api.entities.builtin.provider.message as provider_message
 from ....entity.persistence import mcp as persistence_mcp
-from .mcp_stdio import BoxStdioSessionRuntime, MCPServerBoxConfig, MCPSessionErrorPhase  # noqa: F401
+from .mcp_stdio import BoxStdioSessionRuntime, MCPServerBoxConfig, MCPSessionErrorPhase, _ColdStartRetry  # noqa: F401
 
 # Synthesized LLM tools for MCP resources (not from server tools/list).
 # Dispatched in MCPLoader.invoke_tool; placeholder func on LLMTool is never used.
@@ -185,6 +186,16 @@ class MCPSessionStatus(enum.Enum):
     ERROR = 'error'
 
 
+class _TransportReconnect(Exception):
+    """Internal signal: the Box stdio WS transport dropped but the managed
+    process is still alive. Triggers a lightweight transport reconnect that
+    reuses the live process, instead of a full process rebuild.
+
+    Reconnect attempts are NOT counted toward the fatal retry budget, so a
+    long-lived session can survive arbitrarily many transient drops.
+    """
+
+
 class RuntimeMCPSession:
     """运行时 MCP 会话"""
 
@@ -254,6 +265,16 @@ class RuntimeMCPSession:
         self._lifecycle_task = None
         self._shutdown_event = asyncio.Event()
         self._ready_event = asyncio.Event()
+        # Set transiently when a WS transport drop should NOT stop the managed
+        # process (it will be re-attached on the next initialize()).
+        self._preserve_managed_process = False
+
+        # Log buffer for capturing stderr from Box managed process (maxlen=500 keeps
+        # recent lines without unbounded memory growth)
+        import collections as _collections
+
+        self._log_buffer: _collections.deque = _collections.deque(maxlen=500)
+        self._last_stderr_text: str = ''
 
         self._box_stdio_runtime = BoxStdioSessionRuntime(self)
         self.box_config = self._box_stdio_runtime.config
@@ -315,23 +336,34 @@ class RuntimeMCPSession:
 
         await self.session.initialize()
 
-    async def _init_streamable_http_server(self):
+    @asynccontextmanager
-        transport = await self.exit_stack.enter_async_context(
+    async def _streamable_http_session(self) -> typing.AsyncIterator[ClientSession]:
-            streamable_http_client(
+        """Enter a fully initialized Streamable HTTP session as one context.
+
+        Initialization must happen inside the same context manager that owns the
+        MCP transport. The SDK reports request failures by cancelling the host
+        task and raises the real HTTP error from its TaskGroup during context
+        exit. Keeping these nested contexts together guarantees a failed
+        ``__aenter__`` unwinds immediately, so callers see the HTTPStatusError
+        instead of a detached CancelledError. It also owns the injected HTTPX
+        client, which the MCP SDK deliberately does not close for callers.
+        """
+        async with httpx.AsyncClient(
+            headers=self.server_config.get('headers', {}),
+            timeout=self.server_config.get('timeout', 10),
+            follow_redirects=True,
+        ) as http_client:
+            async with streamable_http_client(
                 self.server_config['url'],
-                http_client=httpx.AsyncClient(
+                http_client=http_client,
-                    headers=self.server_config.get('headers', {}),
+            ) as transport:
-                    timeout=self.server_config.get('timeout', 10),
+                read, write, _ = transport
-                    follow_redirects=True,
+                async with ClientSession(read, write) as session:
-                ),
+                    await session.initialize()
-            )
+                    yield session
-        )
 
-        read, write, _ = transport
+    async def _init_streamable_http_server(self):
-
+        self.session = await self.exit_stack.enter_async_context(self._streamable_http_session())
-        self.session = await self.exit_stack.enter_async_context(ClientSession(read, write))
-
-        await self.session.initialize()
 
     async def _init_remote_server(self):
         """Connect to a remote MCP server, auto-detecting the transport.
@@ -346,9 +378,15 @@ class RuntimeMCPSession:
             await self._init_streamable_http_server()
             return
         except Exception as e:
+            if not self._should_fallback_to_sse(e):
+                self.ap.logger.info(
+                    f'MCP server {self.server_name}: Streamable HTTP transport failed '
+                    f'({self._describe_exception(e)}); not falling back to SSE'
+                )
+                raise
             self.ap.logger.info(
-                f'MCP server {self.server_name}: Streamable HTTP transport failed '
+                f'MCP server {self.server_name}: Streamable HTTP initialize failed with a compatible HTTP status '
-                f'({self._describe_exception(e)}), falling back to SSE'
+                f'({self._describe_exception(e)}), falling back to legacy SSE'
             )
 
         # The Streamable HTTP attempt may have partially entered the transport /
@@ -399,11 +437,39 @@ class RuntimeMCPSession:
                     task.cancel()
                 for task in done:
                     if task is monitor_task and not self._shutdown_event.is_set():
+                        # The monitor completed. This is EITHER the managed
+                        # process actually exiting OR just the WS transport
+                        # dropping while the process stays alive in the Box
+                        # runtime. Re-check the real process state so a
+                        # transient transport drop reconnects (reusing the live
+                        # process) instead of tearing the process down and
+                        # running a full rebuild+backoff cycle.
+                        process_still_running = False
+                        try:
+                            process_still_running = await self._box_stdio_runtime._managed_process_is_running()
+                        except Exception:
+                            process_still_running = False
+                        if process_still_running:
+                            self.ap.logger.info(
+                                f'MCP server {self.server_name}: transport dropped but '
+                                f'managed process is still running; reconnecting transport'
+                            )
+                            self.error_phase = MCPSessionErrorPhase.RELAY_CONNECT
+                            # Preserve the live process across the finally-block
+                            # cleanup: only the WS transport should be torn down.
+                            self._preserve_managed_process = True
+                            raise _TransportReconnect('Box managed process transport dropped; reconnecting')
                         self.error_phase = MCPSessionErrorPhase.RUNTIME
                         raise Exception('Box managed process exited unexpectedly')
             else:
                 await self._shutdown_event.wait()
 
+        except _ColdStartRetry:
+            # Cold-start in progress: set the preserve flag BEFORE the finally
+            # block runs so it does not stop the live managed process. The outer
+            # _lifecycle_loop_with_retry will reuse it on the next attempt.
+            self._preserve_managed_process = True
+            raise
         except Exception as e:
             self.status = MCPSessionStatus.ERROR
             self.error_message = str(e)
@@ -424,14 +490,55 @@ class RuntimeMCPSession:
             except Exception as e:
                 self.ap.logger.error(f'Error cleaning up MCP session {self.server_name}: {e}\n{traceback.format_exc()}')
             finally:
-                await self._cleanup_box_stdio_session()
+                # On a transport-only reconnect the managed process is healthy
+                # and will be re-attached on the next initialize(); do NOT stop
+                # it. Any other exit path fully tears the session down.
+                if getattr(self, '_preserve_managed_process', False):
+                    self._preserve_managed_process = False
+                else:
+                    await self._cleanup_box_stdio_session()
 
     async def _lifecycle_loop_with_retry(self):
         """Wrap _lifecycle_loop with retry and exponential backoff."""
-        for attempt in range(self._MAX_RETRIES + 1):
+        attempt = 0
+        while attempt <= self._MAX_RETRIES:
             try:
                 await self._lifecycle_loop()
                 return  # Normal shutdown, don't retry
+            except _TransportReconnect as e:
+                # Transient WS transport drop while the managed process is still
+                # alive. Reconnect promptly WITHOUT consuming the fatal retry
+                # budget and WITHOUT stopping the process — initialize() will
+                # re-attach to the live process. This is what lets a long-lived
+                # stdio MCP survive repeated brief event-loop stalls / pings.
+                if self._shutdown_event.is_set():
+                    return
+                self.ap.logger.info(
+                    f'MCP session {self.server_name}: reconnecting transport ({self._describe_exception(e)})'
+                )
+                self.status = MCPSessionStatus.CONNECTING
+                self.error_message = None
+                self.error_phase = None
+                await asyncio.sleep(1)
+                continue
+            except _ColdStartRetry as e:
+                # The managed process is alive but still cold-starting (e.g.
+                # `npx -y <pkg>` is still installing) and cannot yet answer the
+                # handshake. Reuse the live process and retry the attach WITHOUT
+                # consuming the fatal retry budget or stopping the process, so a
+                # slow cold start is waited out instead of failing. Preserve the
+                # process across the finally-block cleanup.
+                if self._shutdown_event.is_set():
+                    return
+                self._preserve_managed_process = True
+                self.ap.logger.debug(
+                    f'MCP session {self.server_name}: waiting for cold start ({self._describe_exception(e)})'
+                )
+                self.status = MCPSessionStatus.CONNECTING
+                self.error_message = None
+                self.error_phase = None
+                await asyncio.sleep(2)
+                continue
             except Exception as e:
                 self.retry_count = attempt + 1
                 if self._shutdown_event.is_set():
@@ -460,6 +567,7 @@ class RuntimeMCPSession:
                 self.error_message = None
                 self.error_phase = None
                 await asyncio.sleep(delay)
+                attempt += 1
 
     @staticmethod
     def _describe_exception(exc: BaseException) -> str:
@@ -485,6 +593,36 @@ class RuntimeMCPSession:
         unique = [m for m in leaves if not (m in seen or seen.add(m))]
         return '; '.join(unique) if unique else f'{type(exc).__name__}: {exc}'
 
+    @staticmethod
+    def _iter_exception_leaves(exc: BaseException) -> typing.Iterator[BaseException]:
+        sub = getattr(exc, 'exceptions', None)
+        if sub:  # ExceptionGroup / BaseExceptionGroup
+            for child in sub:
+                yield from RuntimeMCPSession._iter_exception_leaves(child)
+        else:
+            yield exc
+
+    @staticmethod
+    def _should_fallback_to_sse(exc: BaseException) -> bool:
+        """Whether a Streamable HTTP failure matches legacy-SSE fallback.
+
+        Only protocol-compatibility responses trigger fallback. Authentication,
+        authorization, throttling, and server failures must remain visible
+        instead of being retried against a different transport.
+
+        MCP SDK 1.26 translates an HTTP 404 initialize response into a synthetic
+        ``McpError(32600, 'Session terminated')`` rather than preserving the
+        HTTPStatusError, so recognize that exact SDK sentinel as 404-compatible.
+        """
+        fallback_statuses = {400, 404, 405}
+        for leaf in RuntimeMCPSession._iter_exception_leaves(exc):
+            if isinstance(leaf, httpx.HTTPStatusError):
+                if leaf.response.status_code in fallback_statuses:
+                    return True
+            elif isinstance(leaf, McpError) and leaf.error.code == 32600 and leaf.error.message == 'Session terminated':
+                return True
+        return False
+
     _MONITOR_POLL_INTERVAL = 5
     _MONITOR_MAX_CONSECUTIVE_ERRORS = 3
 
@@ -927,11 +1065,14 @@ class RuntimeMCPSession:
         return self._box_stdio_runtime.uses_box_stdio()
 
     def _build_box_session_id(self) -> str:
-        # Transient test sessions get their own isolated Box session so a
+        # Both live servers and transient config-page tests share ONE Box
-        # failing/short-lived test can never disturb the shared session that
+        # session ('mcp-shared'). A test therefore reuses the already-running
-        # hosts live, already-connected MCP servers.
+        # container (and, for an existing server, its live managed process)
-        if self.is_transient:
+        # instead of paying a full per-test session cold-start + dependency
-            return f'mcp-test-{self.server_uuid}'
+        # bootstrap. Isolation between a test and the live servers is provided
+        # at the *process* level: each server/test has its own process_id and a
+        # test only ever stops its own process_id (see cleanup_session), so it
+        # never disturbs another server's process or the shared session itself.
         return 'mcp-shared'
 
     def _rewrite_path(self, path: str, host_path: str | None) -> str:

src/langbot/pkg/provider/tools/loaders/mcp_stdio.py

@@ -6,7 +6,7 @@ import os
 import shutil
 import shlex
 import threading
-from contextlib import suppress
+from contextlib import suppress, AsyncExitStack
 from typing import TYPE_CHECKING, Any
 
 import pydantic
@@ -57,6 +57,23 @@ class MCPSessionErrorPhase(enum.Enum):
     BOX_UNAVAILABLE = 'box_unavailable'
 
 
+def _get_default_memory_mb(ap) -> int:
+    """Read box.default_memory_mb from instance config (env: BOX__DEFAULT_MEMORY_MB).
+
+    Falls back to 1536 MB — a safe floor for Node.js V8 + WASM under nsjail.
+    Operators running memory-constrained hosts can lower this; those with large
+    machines can raise it.  Individual MCP servers can still override via their
+    own box.memory_mb setting.
+    """
+    try:
+        data = getattr(getattr(ap, 'instance_config', None), 'data', None)
+        if isinstance(data, dict):
+            return int(data.get('box', {}).get('default_memory_mb', 1536))
+    except (TypeError, ValueError):
+        pass
+    return 1536
+
+
 class MCPServerBoxConfig(pydantic.BaseModel):
     """Structured configuration for running an MCP server inside a Box container."""
 
@@ -74,6 +91,35 @@ class MCPServerBoxConfig(pydantic.BaseModel):
     model_config = pydantic.ConfigDict(extra='ignore')
 
 
+_HANDSHAKE_ATTEMPT_TIMEOUT_SEC = 10.0
+
+
+class _TransferredStack:
+    """Adapts an already-populated AsyncExitStack into an async context manager
+    so ownership of its resources can be transferred into another exit stack.
+    Entering is a no-op; exiting closes the wrapped stack (and thus the live WS
+    transport + ClientSession) when the owning session shuts down."""
+
+    def __init__(self, stack: AsyncExitStack):
+        self._stack = stack
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        await self._stack.aclose()
+        return False
+
+
+class _ColdStartRetry(Exception):
+    """Signal: the managed process is alive but not yet answering the MCP
+    handshake because it is still cold-starting (e.g. `npx -y <pkg>` is still
+    installing). The outer lifecycle retry treats this like a transient
+    reconnect: it reuses the live process and does not count toward the fatal
+    retry budget, so a slow cold start is waited out rather than failing.
+    """
+
+
 class BoxStdioSessionRuntime:
     """Encapsulate Box-backed stdio MCP session orchestration."""
 
@@ -113,7 +159,17 @@ class BoxStdioSessionRuntime:
             read_only_rootfs=self.config.read_only_rootfs if self.config.read_only_rootfs is not None else False,
             image=self.config.image,
             cpus=self.config.cpus,
-            memory_mb=self.config.memory_mb,
+            # Node.js runtimes (npx/bunx) reserve large virtual address space and
+            # load WebAssembly modules (llhttp) on startup; the default 512 MB
+            # cgroup_mem_max is too small and causes OOM kills (return_code=137).
+            # Auto-bump to 1024 MB when the runner is npx/bunx/pnpm dlx.
+            # Per-server override wins; global default comes from
+            # config.yaml box.default_memory_mb (env: BOX__DEFAULT_MEMORY_MB).
+            # Hard floor of 1536 MB: enough for Node.js V8 + WASM without OOM.
+            # Per-server override wins; global default from config.yaml
+            # box.default_memory_mb (env: BOX__DEFAULT_MEMORY_MB), hard floor
+            # of 1536 MB so Node.js V8 + WASM never OOM under nsjail.
+            memory_mb=(self.config.memory_mb or _get_default_memory_mb(self.ap)),
             pids_limit=self.config.pids_limit,
             persistent=True,
         )
@@ -173,28 +229,55 @@ class BoxStdioSessionRuntime:
                     stderr_preview = (result.stderr or '')[:500]
                     raise Exception(f'Dependency install failed (exit code {result.exit_code}): {stderr_preview}')
 
-        try:
+        # Reuse an already-running managed process instead of rebuilding it.
-            process_workspace = (
+        # The Box runtime keeps the managed process alive across a transient
-                self._build_workspace(host_path=host_path, workdir=process_cwd, mount_path=process_cwd)
+        # WebSocket transport drop, so on a reconnect we only need to re-attach
-                if host_path
+        # the WS below. Rebuilding here would needlessly stop a healthy process
-                else workspace
+        # and re-run the (slow, network-touching) dependency bootstrap.
+        if not await self._managed_process_is_running():
+            try:
+                process_workspace = (
+                    self._build_workspace(host_path=host_path, workdir=process_cwd, mount_path=process_cwd)
+                    if host_path
+                    else workspace
+                )
+                payload = process_workspace.build_process_payload(
+                    self.server_config['command'],
+                    self.server_config.get('args', []),
+                    env=self.server_config.get('env', {}),
+                    cwd=process_cwd,
+                )
+                if install_cmd:
+                    payload = self._wrap_process_payload_with_python_env(payload, process_cwd)
+                payload['process_id'] = self.process_id
+                await workspace.box_service.start_managed_process(workspace.session_id, payload)
+            except Exception:
+                self.owner.error_phase = MCPSessionErrorPhase.PROCESS_START
+                raise
+        else:
+            self.ap.logger.info(
+                f'MCP server {self.server_name}: reusing live managed process '
+                f'process_id={self.process_id} (transport reconnect)'
             )
-            payload = process_workspace.build_process_payload(
-                self.server_config['command'],
-                self.server_config.get('args', []),
-                env=self.server_config.get('env', {}),
-                cwd=process_cwd,
-            )
-            if install_cmd:
-                payload = self._wrap_process_payload_with_python_env(payload, process_cwd)
-            payload['process_id'] = self.process_id
-            await workspace.box_service.start_managed_process(workspace.session_id, payload)
-        except Exception:
-            self.owner.error_phase = MCPSessionErrorPhase.PROCESS_START
-            raise
 
+        websocket_url = workspace.get_managed_process_websocket_url(self.process_id)
+
+        # Attach the WS transport + MCP session ONCE, on the owner's exit stack,
+        # in the same task as the serve loop that follows. websocket_client and
+        # ClientSession use anyio task groups whose cancel scope is bound to the
+        # frame/stack that entered them, so they must live on the owner exit
+        # stack (not a deferred/transferred one) or the streams close the moment
+        # initialize() returns and the next request fails with "Connection
+        # closed".
+        #
+        # A slow (`npx -y <pkg>`) cold start makes this single attempt fail
+        # while the process is still alive — the package is still installing and
+        # cannot answer the handshake. We surface that to the outer retry loop
+        # as a _ColdStartRetry: it must NOT stop the process (it is healthy and
+        # will be reused) and must NOT consume the fatal retry budget. The next
+        # attempt re-attaches to the same live process; once it has finished
+        # cold start the handshake succeeds and stays healthy.
         try:
-            websocket_url = workspace.get_managed_process_websocket_url(self.process_id)
             transport = await self.owner.exit_stack.enter_async_context(websocket_client(websocket_url))
             read_stream, write_stream = transport
             self.owner.session = await self.owner.exit_stack.enter_async_context(
@@ -202,12 +285,19 @@ class BoxStdioSessionRuntime:
             )
         except Exception:
             self.owner.error_phase = MCPSessionErrorPhase.RELAY_CONNECT
+            if not await self._managed_process_has_exited():
+                # Process is alive but not yet serving (cold start) — reconnect.
+                raise _ColdStartRetry(f'{self.server_name}: transport not ready during cold start')
             raise
 
         try:
-            await self.owner.session.initialize()
+            await asyncio.wait_for(self.owner.session.initialize(), timeout=_HANDSHAKE_ATTEMPT_TIMEOUT_SEC)
-        except Exception:
+        except Exception as exc:
             self.owner.error_phase = MCPSessionErrorPhase.MCP_INIT
+            if not await self._managed_process_has_exited():
+                raise _ColdStartRetry(
+                    f'{self.server_name}: handshake not ready during cold start ({type(exc).__name__})'
+                )
             raise
 
     async def monitor_process_health(self) -> None:
@@ -234,8 +324,74 @@ class BoxStdioSessionRuntime:
                 )
                 if consecutive_errors >= self.owner._MONITOR_MAX_CONSECUTIVE_ERRORS:
                     return
+
+            # Capture stderr logs from the managed process
+            if isinstance(info, dict):
+                stderr_text = info.get('stderr', '') or info.get('stderr_preview', '')
+            else:
+                stderr_text = getattr(info, 'stderr', '') or getattr(info, 'stderr_preview', '')
+
+            if stderr_text and stderr_text != self.owner._last_stderr_text:
+                # Find new lines not in the previous snapshot
+                old_lines = set(self.owner._last_stderr_text.splitlines()) if self.owner._last_stderr_text else set()
+                new_lines = [l for l in stderr_text.splitlines() if l and l not in old_lines]
+                self.owner._last_stderr_text = stderr_text
+
+                import time as _time
+
+                for line in new_lines:
+                    level = (
+                        'error'
+                        if any(k in line.upper() for k in ('ERROR', 'CRITICAL'))
+                        else 'warning'
+                        if 'WARNING' in line.upper()
+                        else 'debug'
+                        if 'DEBUG' in line.upper()
+                        else 'info'
+                    )
+                    self.owner._log_buffer.append({'ts': _time.time(), 'level': level, 'text': line})
+
             await asyncio.sleep(self.owner._MONITOR_POLL_INTERVAL)
 
+    async def _managed_process_is_running(self) -> bool:
+        """Return True if this server's managed process exists and is running.
+
+        Used to decide whether initialize() must (re)start the process or can
+        simply re-attach the WebSocket transport to a process the Box runtime
+        kept alive across a transient transport drop.
+        """
+        from langbot_plugin.box.models import BoxManagedProcessStatus
+
+        workspace = self._build_workspace()
+        try:
+            info = await workspace.get_managed_process(self.process_id)
+        except Exception:
+            return False
+        status = info.get('status', '') if isinstance(info, dict) else getattr(info, 'status', '')
+        return status in (BoxManagedProcessStatus.RUNNING.value, BoxManagedProcessStatus.RUNNING)
+
+    async def _managed_process_has_exited(self) -> bool:
+        """Return True only if the process is DEFINITIVELY gone (reports EXITED).
+
+        Distinct from ``not _managed_process_is_running()``: a process that has
+        just been spawned may not yet report RUNNING, and a transient query
+        error is not proof of exit. During the cold-start handshake retry we
+        must NOT treat 'not yet running' or 'query failed' as a terminal
+        failure, or we bail out to the outer rebuild path and churn the
+        process (relay then rejects the early re-attach with HTTP 400). Only a
+        successful query that reports EXITED stops the retry loop.
+        """
+        from langbot_plugin.box.models import BoxManagedProcessStatus
+
+        workspace = self._build_workspace()
+        try:
+            info = await workspace.get_managed_process(self.process_id)
+        except Exception:
+            # Unknown — treat as 'still coming up', not exited.
+            return False
+        status = info.get('status', '') if isinstance(info, dict) else getattr(info, 'status', '')
+        return status in (BoxManagedProcessStatus.EXITED.value, BoxManagedProcessStatus.EXITED)
+
     async def _stage_host_path_to_shared_workspace(self, host_path: str) -> str:
         source_path = normalize_host_path(host_path)
         if not source_path:
@@ -342,16 +498,20 @@ class BoxStdioSessionRuntime:
 
         workspace = self._build_workspace(host_path=None)
 
-        # Transient test sessions own their isolated Box session, so tear the
+        # Transient config-page tests now share the same 'mcp-shared' Box
-        # whole session down rather than leaking it. This cannot affect live
+        # session as live servers, so we must NOT tear the session down here —
-        # servers because they live in the separate shared session.
+        # that would kill every other MCP server in the container. A test is
+        # isolated at the process level: it ran under its own process_id, so we
+        # stop only that process, exactly like a live server does below. The
+        # shared session and all other servers' live processes are untouched.
+        # (Staged per-test workspace files are still cleaned up.)
         if getattr(self.owner, 'is_transient', False):
             try:
-                await workspace.cleanup()
+                await workspace.stop_managed_process(self.process_id)
             except Exception as exc:
                 self.ap.logger.warning(
-                    f'MCP server {self.server_name}: failed to delete transient test session '
+                    f'MCP server {self.server_name}: failed to stop transient test process '
-                    f'{self.owner._build_box_session_id()}: {type(exc).__name__}: {exc}'
+                    f'process_id={self.process_id}: {type(exc).__name__}: {exc}'
                 )
             await self._cleanup_staged_workspace()
             return

src/langbot/pkg/vector/mgr.py

@@ -33,6 +33,12 @@ class VectorDBManager:
                 self.vector_db = SeekDBVectorDatabase(self.ap)
                 self.ap.logger.info('Initialized SeekDB vector database backend.')
 
+            elif vdb_type == 'valkey_search':
+                from .vdbs.valkey_search import ValkeySearchVectorDatabase
+
+                self.vector_db = ValkeySearchVectorDatabase(self.ap)
+                self.ap.logger.info('Initialized Valkey Search vector database backend.')
+
             elif vdb_type == 'milvus':
                 from .vdbs.milvus import MilvusVectorDatabase
 

src/langbot/pkg/vector/vdbs/valkey_search.py

@@ -0,0 +1,828 @@
+from __future__ import annotations
+
+import asyncio
+import json
+import struct
+from typing import Any
+
+from langbot.pkg.core import app
+from langbot.pkg.vector.vdb import VectorDatabase, SearchType
+from langbot.pkg.vector.filter_utils import normalize_filter, strip_unsupported_fields
+
+try:
+    from glide import (
+        Batch,
+        GlideClient,
+        GlideClientConfiguration,
+        NodeAddress,
+        RequestError,
+        ServerCredentials,
+        ft,
+        VectorField,
+        VectorFieldAttributesHnsw,
+        VectorFieldAttributesFlat,
+        VectorAlgorithm,
+        VectorType,
+        DistanceMetricType,
+        TagField,
+        TextField,
+        FtCreateOptions,
+        DataType,
+        FtSearchOptions,
+        FtSearchLimit,
+        ReturnField,
+    )
+
+    VALKEY_SEARCH_AVAILABLE = True
+except ImportError:
+    VALKEY_SEARCH_AVAILABLE = False
+
+# Default per-request timeout (ms) for the glide client.  The glide library
+# default is 250ms, which is too low for vector KNN (``FT.SEARCH ... =>[KNN]``)
+# under moderate load or with large indexes and yields spurious TimeoutErrors.
+# Overridable via the ``vdb.valkey_search.request_timeout`` config option.
+_DEFAULT_REQUEST_TIMEOUT_MS = 5000
+
+# Safety cap on the number of SCAN rounds when purging a collection's keys, so
+# a cursor-handling bug or pathological keyspace can never spin forever.
+_MAX_SCAN_ROUNDS = 100000
+
+
+# Mandatory client name for production observability (CLIENT LIST / dashboards).
+VALKEY_CLIENT_NAME = 'langbot_vector_client'
+
+# Fixed, indexed metadata schema.  LangBot's RAG layer stores ``file_id`` on
+# every chunk; it is the only metadata field we promote to a first-class
+# (filterable) index field.  All other metadata is preserved verbatim inside
+# the ``metadata_json`` field so it survives a round-trip, but is NOT
+# filterable (the established Milvus / pgvector pragmatism).
+_INDEXED_TAG_FIELDS = {'file_id'}
+_SUPPORTED_FILTER_FIELDS = set(_INDEXED_TAG_FIELDS)
+
+# Hash field names used for stored documents.
+_FIELD_VECTOR = 'vector'
+_FIELD_DOCUMENT = 'document'
+_FIELD_FILE_ID = 'file_id'
+_FIELD_METADATA = 'metadata_json'
+_VEC_SCORE_ALIAS = '__vec_score'
+
+# Valkey Search has no bare "match everything" token for non-vector queries
+# (a standalone ``*`` is a syntax error).  A negated match on a sentinel tag
+# value that can never exist matches every key, which is the canonical
+# match-all idiom for FT.SEARCH.
+_MATCH_ALL = '-@file_id:{__langbot_match_all_sentinel__}'
+
+# Page size used when enumerating matching keys for deletion.  Deletes
+# paginate through the full result set in batches of this size so that
+# files/filters matching more than one page of chunks are fully removed
+# (no silent truncation / orphaned vectors).
+_DELETE_SCAN_BATCH = 10000
+
+# Characters Valkey Search's TAG query parser cannot handle even when
+# backslash-escaped (the brace delimiters and the wildcard).  file_id TAG
+# values are percent-encoded over this set (plus '%' itself, so the encoding
+# is reversible/unambiguous) before being stored or queried, so an arbitrary
+# file_id round-trips instead of producing an unparseable query.  For normal
+# UUID/hash file_ids none of these characters occur, so the encoding is a
+# no-op and the stored value is unchanged.  The original file_id is always
+# preserved verbatim inside ``metadata_json``.
+_FT_UNSAFE_TAG_CHARS = frozenset('{}*%')
+
+
+class ValkeySearchVectorDatabase(VectorDatabase):
+    """Valkey Search (valkey-bundle) vector database adapter for LangBot.
+
+    Backed by the Valkey Search module shipped in ``valkey/valkey-bundle``,
+    accessed through the official ``valkey-glide`` client's native ``ft``
+    (search) command namespace.  Documents are stored as Valkey HASH keys
+    under a per-collection prefix and indexed by one ``FT.CREATE`` index per
+    collection.
+
+    Supported search types: ``VECTOR``, ``FULL_TEXT`` and ``HYBRID``.
+
+    Hybrid search semantics (IMPORTANT)
+    -----------------------------------
+    Valkey Search hybrid queries follow a *filter-then-KNN* model: the text /
+    metadata filter pre-selects candidate keys and the KNN stage ranks them by
+    vector distance.  This backend does **NOT** implement application-side
+    weighted score fusion.  The ``vector_weight`` argument is therefore
+    accepted for interface compatibility but is **not honored** — passing
+    different weights does not change result ordering.  A one-time warning is
+    emitted the first time a non-default weight is supplied.  App-side score
+    fusion can be layered on later if weighted hybrid ranking is required.
+    """
+
+    @classmethod
+    def supported_search_types(cls) -> list[SearchType]:
+        return [SearchType.VECTOR, SearchType.FULL_TEXT, SearchType.HYBRID]
+
+    def __init__(self, ap: app.Application):
+        if not VALKEY_SEARCH_AVAILABLE:
+            raise ImportError(
+                "valkey-glide is not installed. Install it with: pip install 'valkey-glide>=2.4.1,<3.0.0'"
+            )
+
+        self.ap = ap
+        config = self.ap.instance_config.data['vdb']['valkey_search']
+
+        self._host = config.get('host', 'localhost')
+        self._port = int(config.get('port', 6379))
+        self._db = int(config.get('db', 0))
+        # Auth / TLS are optional (toB / SaaS).  Never logged.
+        self._password = config.get('password', '') or None
+        self._username = config.get('username', '') or None
+        self._tls = bool(config.get('tls', False))
+        self._request_timeout = int(config.get('request_timeout', _DEFAULT_REQUEST_TIMEOUT_MS))
+
+        algorithm = str(config.get('index_algorithm', 'HNSW')).upper()
+        self._algorithm = VectorAlgorithm.FLAT if algorithm == 'FLAT' else VectorAlgorithm.HNSW
+
+        metric = str(config.get('distance_metric', 'COSINE')).upper()
+        self._distance_metric = {
+            'COSINE': DistanceMetricType.COSINE,
+            'L2': DistanceMetricType.L2,
+            'IP': DistanceMetricType.IP,
+        }.get(metric, DistanceMetricType.COSINE)
+
+        # Lazily-created client (created on first use so a down Valkey does not
+        # block LangBot boot).
+        self._client: GlideClient | None = None
+        # Serializes lazy client creation so concurrent first-use callers do not
+        # each construct (and leak) a separate GlideClient.
+        self._client_lock = asyncio.Lock()
+        # Index names we have already ensured this process lifetime.
+        self._ensured_indexes: set[str] = set()
+        # Whether we have already warned about the non-honored vector_weight.
+        self._vector_weight_warned = False
+
+    # ------------------------------------------------------------------ #
+    # Client lifecycle
+    # ------------------------------------------------------------------ #
+    async def _ensure_client(self) -> GlideClient:
+        """Create the glide client on first use (lazy, non-blocking boot)."""
+        if self._client is not None:
+            return self._client
+        # Double-checked locking: serialize creation so two concurrent
+        # first-use callers don't both build a client and leak one.
+        async with self._client_lock:
+            if self._client is not None:
+                return self._client
+
+            credentials = None
+            if self._password is not None:
+                # username is optional alongside a password (ACL "user" vs default user).
+                credentials = ServerCredentials(password=self._password, username=self._username)
+            elif self._username is not None:
+                # A username without a password is not a valid credential pair, and silently
+                # connecting unauthenticated to a potentially shared Valkey instance is a
+                # security footgun (e.g. an env var that failed to resolve). Fail closed.
+                raise ValueError(
+                    'Valkey Search: a username was configured without a password. '
+                    'Set both username and password to use ACL authentication, or remove both.'
+                )
+
+            conf = GlideClientConfiguration(
+                addresses=[NodeAddress(self._host, self._port)],
+                client_name=VALKEY_CLIENT_NAME,
+                database_id=self._db,
+                use_tls=self._tls,
+                lazy_connect=True,
+                credentials=credentials,
+                request_timeout=self._request_timeout,
+            )
+            self._client = await GlideClient.create(conf)
+            self.ap.logger.info(
+                f'Initialized Valkey Search client to {self._host}:{self._port} (db={self._db}, tls={self._tls})'
+            )
+        return self._client
+
+    async def close(self) -> None:
+        """Close the glide client and reset state.
+
+        Safe to call when no client was created. After ``close`` the next
+        operation transparently re-creates the client (``_ensure_client``
+        guards on ``self._client is None``).
+        """
+        if self._client is not None:
+            try:
+                await self._client.close()
+            except Exception:
+                self.ap.logger.warning('Valkey Search: error while closing client (ignored)')
+            finally:
+                self._client = None
+                self._ensured_indexes.clear()
+
+    # ------------------------------------------------------------------ #
+    # Naming helpers
+    # ------------------------------------------------------------------ #
+    @staticmethod
+    def _index_name(collection: str) -> str:
+        return f'idx:{collection}'
+
+    @staticmethod
+    def _key_prefix(collection: str) -> str:
+        return f'kb:{collection}:'
+
+    @staticmethod
+    def _pack_vector(vec: list[float]) -> bytes:
+        """Pack a float vector into little-endian float32 bytes.
+
+        Valkey Search stores and queries vectors as FLOAT32 little-endian
+        blobs (per the search query-language spec).
+        """
+        return struct.pack(f'<{len(vec)}f', *[float(x) for x in vec])
+
+    @staticmethod
+    def _escape_tag(value: str) -> str:
+        """Escape characters that are special inside a TAG ``{...}`` clause.
+
+        The backslash is escaped first so it cannot consume a following
+        escape. This neutralises injection-style values (quotes, parens,
+        ``|``, ``@``, ``:``, spaces, dashes) so a crafted ``file_id`` cannot
+        break out of the clause.
+
+        Note: Valkey Search's TAG query parser cannot handle a literal brace
+        (``{`` / ``}``) or ``*`` even when backslash-escaped.  Callers that pass
+        a ``file_id`` route it through ``_encode_and_escape_tag`` /
+        ``_encode_file_id`` first, which percent-encodes exactly those
+        characters, so an arbitrary ``file_id`` round-trips safely. This raw
+        escaper is still correct for all other special characters.
+        """
+        out = []
+        for ch in str(value):
+            if ch in '\\,.<>{}[]"\':;!@#$%^&*()-+=~| ':
+                out.append('\\')
+            out.append(ch)
+        return ''.join(out)
+
+    @staticmethod
+    def _encode_file_id(value: str) -> str:
+        """Make a ``file_id`` safe to use as an FT TAG token AND query value.
+
+        Percent-encodes the characters Valkey Search's TAG parser cannot handle
+        even when backslash-escaped (``{``, ``}``, ``*``) plus ``%`` itself for
+        reversibility.  Applied identically at write time (the stored TAG field)
+        and query time (filters / ``delete_by_file_id``) so any value matches
+        itself.  For normal UUID/hash ids none of these characters occur, so
+        this is a no-op.  The original value is always kept verbatim in
+        ``metadata_json``; this encoded form is only ever used for the indexed
+        TAG.
+        """
+        out = []
+        for ch in str(value):
+            if ch in _FT_UNSAFE_TAG_CHARS:
+                out.append('%{:02X}'.format(ord(ch)))
+            else:
+                out.append(ch)
+        return ''.join(out)
+
+    def _encode_and_escape_tag(self, value: str) -> str:
+        """Encode an FT-unsafe ``file_id`` then escape TAG special chars."""
+        return self._escape_tag(self._encode_file_id(value))
+
+    # ------------------------------------------------------------------ #
+    # Filter mapping (canonical triples -> FT query fragment)
+    # ------------------------------------------------------------------ #
+    def _triples_to_ft(self, filter: dict[str, Any] | None) -> str:
+        """Translate a canonical filter dict into an FT filter expression.
+
+        Only indexed fields (``file_id``) are filterable; unsupported fields
+        are dropped with a warning (matching the Milvus / pgvector pattern).
+        Returns an empty string when there is no usable filter.
+        """
+        triples = normalize_filter(filter)
+        if not triples:
+            return ''
+        triples = strip_unsupported_fields(triples, _SUPPORTED_FILTER_FIELDS)
+
+        fragments: list[str] = []
+        for field, op, value in triples:
+            # All currently-indexed fields are TAG fields; file_id values are
+            # encoded (FT-unsafe chars) then escaped so any value round-trips.
+            if op == '$eq':
+                fragments.append(f'@{field}:{{{self._encode_and_escape_tag(value)}}}')
+            elif op == '$ne':
+                fragments.append(f'-@{field}:{{{self._encode_and_escape_tag(value)}}}')
+            elif op == '$in':
+                joined = '|'.join(self._encode_and_escape_tag(v) for v in value)
+                fragments.append(f'@{field}:{{{joined}}}')
+            elif op == '$nin':
+                joined = '|'.join(self._encode_and_escape_tag(v) for v in value)
+                fragments.append(f'-@{field}:{{{joined}}}')
+            elif op == '$gt':
+                fragments.append(f'@{field}:[({float(value)} +inf]')
+            elif op == '$gte':
+                fragments.append(f'@{field}:[{float(value)} +inf]')
+            elif op == '$lt':
+                fragments.append(f'@{field}:[-inf ({float(value)}]')
+            elif op == '$lte':
+                fragments.append(f'@{field}:[-inf {float(value)}]')
+            else:
+                # normalize_filter() already rejects unknown operators, so this
+                # only triggers if SUPPORTED_OPS grows without this chain being
+                # updated. Fail closed (rather than silently dropping the
+                # condition, which would widen delete_by_filter's match set).
+                raise ValueError(f'Valkey Search: unhandled filter operator {op!r} on field {field!r}')
+
+        return ' '.join(fragments)
+
+    @staticmethod
+    def _build_text_clause(text: str) -> str:
+        """Build a field-scoped full-text clause for the ``document`` field.
+
+        Each whitespace-delimited word becomes a ``@document:<term>`` term and
+        the terms are AND-ed (space separated).  FT special characters in each
+        term are escaped.  Returns an empty string when *text* has no words.
+        """
+        words = [w for w in str(text).split() if w]
+        if not words:
+            return ''
+        terms = [f'@{_FIELD_DOCUMENT}:{ValkeySearchVectorDatabase._escape_text(w)}' for w in words]
+        return ' '.join(terms)
+
+    @staticmethod
+    def _escape_text(text: str) -> str:
+        """Escape FT full-text special characters in a single term."""
+        out = []
+        for ch in str(text):
+            if ch in '@!{}[]()|-"~*:\\':
+                out.append('\\')
+            out.append(ch)
+        return ''.join(out)
+
+    # ------------------------------------------------------------------ #
+    # Index management
+    # ------------------------------------------------------------------ #
+    async def _ensure_index(self, client: GlideClient, collection: str, dim: int) -> None:
+        index = self._index_name(collection)
+        if index in self._ensured_indexes:
+            return
+
+        # ft.info is O(1) and raises RequestError when the index is absent —
+        # cheaper than ft.list (O(n) over all indexes) and it closes the
+        # check-then-create TOCTOU window.
+        try:
+            await ft.info(client, index)
+            self._ensured_indexes.add(index)
+            return
+        except RequestError:
+            pass
+
+        if self._algorithm == VectorAlgorithm.FLAT:
+            vector_attrs = VectorFieldAttributesFlat(
+                dimensions=dim,
+                distance_metric=self._distance_metric,
+                type=VectorType.FLOAT32,
+            )
+        else:
+            vector_attrs = VectorFieldAttributesHnsw(
+                dimensions=dim,
+                distance_metric=self._distance_metric,
+                type=VectorType.FLOAT32,
+            )
+
+        schema = [
+            VectorField(name=_FIELD_VECTOR, algorithm=self._algorithm, attributes=vector_attrs),
+            TagField(name=_FIELD_FILE_ID),
+            TextField(name=_FIELD_DOCUMENT),
+        ]
+        options = FtCreateOptions(data_type=DataType.HASH, prefixes=[self._key_prefix(collection)])
+        await ft.create(client, index, schema, options)
+        self._ensured_indexes.add(index)
+        self.ap.logger.info(
+            f"Valkey Search index '{index}' created (dim={dim}, algo={self._algorithm.value}, "
+            f'metric={self._distance_metric.value})'
+        )
+
+    @staticmethod
+    def _decode(value: Any) -> str:
+        if isinstance(value, (bytes, bytearray, memoryview)):
+            return bytes(value).decode('utf-8', errors='replace')
+        return str(value)
+
+    # ------------------------------------------------------------------ #
+    # VectorDatabase ABC implementation
+    # ------------------------------------------------------------------ #
+    async def get_or_create_collection(self, collection: str):
+        """Ensure a client exists.
+
+        The index itself requires the vector dimension, which is only known at
+        first ``add_embeddings`` (same constraint as Qdrant / SeekDB), so this
+        is a best-effort no-op when the index does not yet exist.
+        """
+        await self._ensure_client()
+
+    async def add_embeddings(
+        self,
+        collection: str,
+        ids: list[str],
+        embeddings_list: list[list[float]],
+        metadatas: list[dict[str, Any]],
+        documents: list[str] | None = None,
+    ) -> None:
+        if not embeddings_list:
+            return
+
+        client = await self._ensure_client()
+        dim = len(embeddings_list[0])
+        # The index schema is fixed to the first embedding's dimension. A later
+        # embedding of a different length would be packed into a wrong-sized
+        # blob that Valkey stores silently but that yields garbage KNN
+        # distances, so reject mixed dimensions up-front.
+        if any(len(e) != dim for e in embeddings_list[1:]):
+            raise ValueError(f'All embeddings must have dimension {dim}; got mixed lengths')
+        await self._ensure_index(client, collection, dim)
+
+        prefix = self._key_prefix(collection)
+
+        batch = Batch(is_atomic=False)
+        for i, _id in enumerate(ids):
+            key = prefix + str(_id)
+            metadata = metadatas[i] if i < len(metadatas) else {}
+            mapping: dict[str, Any] = {
+                _FIELD_VECTOR: self._pack_vector(embeddings_list[i]),
+                _FIELD_METADATA: json.dumps(metadata, ensure_ascii=False),
+            }
+            file_id = metadata.get('file_id')
+            if file_id is not None:
+                mapping[_FIELD_FILE_ID] = self._encode_file_id(str(file_id))
+            if documents is not None and i < len(documents) and documents[i] is not None:
+                mapping[_FIELD_DOCUMENT] = documents[i]
+
+            batch.hset(key, mapping)
+
+        # Pipeline all HSETs into a single round-trip (non-atomic) instead of
+        # one await per embedding, which is N sequential round-trips for N
+        # chunks.
+        await client.exec(batch, raise_on_error=True)
+
+        self.ap.logger.info(f"Added {len(ids)} embeddings to Valkey Search collection '{collection}'")
+
+    async def search(
+        self,
+        collection: str,
+        query_embedding: list[float],
+        k: int = 5,
+        search_type: str = 'vector',
+        query_text: str = '',
+        filter: dict[str, Any] | None = None,
+        vector_weight: float | None = None,
+    ) -> dict[str, Any]:
+        client = await self._ensure_client()
+        index = self._index_name(collection)
+
+        if not await self._index_exists(client, index):
+            return {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+
+        # vector_weight is accepted for interface parity but NOT honored by this
+        # backend (filter-then-KNN, no weighted fusion).  Warn once.
+        if vector_weight is not None and not self._vector_weight_warned:
+            self.ap.logger.warning(
+                'Valkey Search backend does not honor vector_weight: hybrid search uses '
+                'filter-then-KNN without weighted score fusion. The vector_weight value '
+                'is ignored. See docs/VALKEY_SEARCH_INTEGRATION.md.'
+            )
+            self._vector_weight_warned = True
+
+        filter_expr = self._triples_to_ft(filter)
+
+        if search_type == SearchType.FULL_TEXT:
+            if not query_text:
+                return {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+            text_clause = self._build_text_clause(query_text)
+            if not text_clause:
+                return {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+            query = f'{filter_expr} {text_clause}'.strip() if filter_expr else text_clause
+            return await self._run_text_search(client, index, query, k)
+
+        if search_type == SearchType.HYBRID:
+            # Filter / text pre-selects candidates; KNN ranks. No fusion.
+            pre = filter_expr
+            if query_text:
+                text_clause = self._build_text_clause(query_text)
+                if text_clause:
+                    pre = f'{pre} {text_clause}'.strip() if pre else text_clause
+            pre = pre or '*'
+            query = f'{self._wrap_pre(pre)}=>[KNN {k} @{_FIELD_VECTOR} $BLOB AS {_VEC_SCORE_ALIAS}]'
+            return await self._run_knn_search(client, index, query, query_embedding, k)
+
+        # Default: pure VECTOR search.
+        pre = filter_expr or '*'
+        query = f'{self._wrap_pre(pre)}=>[KNN {k} @{_FIELD_VECTOR} $BLOB AS {_VEC_SCORE_ALIAS}]'
+        return await self._run_knn_search(client, index, query, query_embedding, k)
+
+    @staticmethod
+    def _wrap_pre(pre: str) -> str:
+        """Parenthesize a multi-condition pre-filter before the ``=>`` KNN clause.
+
+        When ``pre`` combines several terms (e.g. ``@file_id:{x} @document:term``)
+        the Valkey Search parser can otherwise mis-associate only the last term
+        with the KNN clause. Wrapping the whole expression forces correct
+        grouping. A bare ``*`` (match-all) and single-term expressions are left
+        untouched.
+        """
+        if pre and pre != '*' and ' ' in pre.strip():
+            return f'({pre})'
+        return pre
+
+    async def _run_knn_search(
+        self,
+        client: GlideClient,
+        index: str,
+        query: str,
+        query_embedding: list[float],
+        k: int,
+    ) -> dict[str, Any]:
+        options = FtSearchOptions(
+            params={'BLOB': self._pack_vector(list(query_embedding))},
+            return_fields=[
+                ReturnField(field_identifier=_VEC_SCORE_ALIAS, alias='distance'),
+                ReturnField(field_identifier=_FIELD_DOCUMENT),
+                ReturnField(field_identifier=_FIELD_METADATA),
+            ],
+            limit=FtSearchLimit(0, k),
+            dialect=2,
+        )
+        try:
+            reply = await ft.search(client, index, query, options)
+        except Exception as exc:
+            if self._is_missing_index_error(exc):
+                return {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+            raise
+        return self._reply_to_chroma(index, reply, has_distance=True)
+
+    async def _run_text_search(
+        self,
+        client: GlideClient,
+        index: str,
+        query: str,
+        k: int,
+    ) -> dict[str, Any]:
+        options = FtSearchOptions(
+            return_fields=[
+                ReturnField(field_identifier=_FIELD_DOCUMENT),
+                ReturnField(field_identifier=_FIELD_METADATA),
+            ],
+            limit=FtSearchLimit(0, k),
+            dialect=2,
+        )
+        try:
+            reply = await ft.search(client, index, query, options)
+        except Exception as exc:
+            if self._is_missing_index_error(exc):
+                return {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+            raise
+        return self._reply_to_chroma(index, reply, has_distance=False)
+
+    @staticmethod
+    def _is_missing_index_error(exc: Exception) -> bool:
+        """Return True if *exc* indicates the FT index does not exist.
+
+        ``FT.DROPINDEX`` is applied eventually, so an index can briefly still
+        appear in ``FT._LIST`` after being dropped; a follow-up search then
+        fails with a "not found" error which we treat as an empty result.
+        """
+        message = str(exc).lower()
+        return 'not found' in message and 'index' in message
+
+    def _iter_reply_docs(self, reply: Any, prefix: str):
+        """Yield ``(doc_id, decoded_fields)`` pairs from an FT.SEARCH reply.
+
+        glide returns ``[total, {key: {field: value}, ...}]``. This shared
+        iterator decodes each key, strips the per-collection prefix to recover
+        the original document id, and decodes the field map — the logic both
+        ``_reply_to_chroma`` and ``list_by_filter`` need.
+        """
+        docs = reply[1] if reply and len(reply) >= 2 and isinstance(reply[1], dict) else {}
+        for key, fields in docs.items():
+            key_str = self._decode(key)
+            doc_id = key_str[len(prefix) :] if prefix and key_str.startswith(prefix) else key_str
+            decoded_fields = {self._decode(fk): fv for fk, fv in fields.items()} if isinstance(fields, dict) else {}
+            yield doc_id, decoded_fields
+
+    def _reply_to_chroma(self, index: str, reply: Any, has_distance: bool) -> dict[str, Any]:
+        """Convert an FT.SEARCH reply into Chroma-style nested lists.
+
+        The KNN score field (aliased ``distance``) is a COSINE/L2 distance
+        directly, so no inversion is needed (unlike Qdrant).
+        """
+        ids: list[str] = []
+        distances: list[float] = []
+        metadatas: list[dict[str, Any]] = []
+
+        if not reply or len(reply) < 2:
+            return {'ids': [ids], 'metadatas': [metadatas], 'distances': [distances]}
+
+        prefix = self._key_prefix(index[len('idx:') :]) if index.startswith('idx:') else ''
+
+        for doc_id, decoded_fields in self._iter_reply_docs(reply, prefix):
+            ids.append(doc_id)
+
+            if has_distance and 'distance' in decoded_fields:
+                try:
+                    distances.append(float(self._decode(decoded_fields['distance'])))
+                except (TypeError, ValueError):
+                    distances.append(0.0)
+            else:
+                distances.append(0.0)
+
+            metadata: dict[str, Any] = {}
+            raw_meta = decoded_fields.get(_FIELD_METADATA)
+            if raw_meta is not None:
+                try:
+                    metadata = json.loads(self._decode(raw_meta))
+                except (TypeError, ValueError):
+                    metadata = {}
+            metadatas.append(metadata)
+
+        return {'ids': [ids], 'metadatas': [metadatas], 'distances': [distances]}
+
+    async def delete_by_file_id(self, collection: str, file_id: str) -> None:
+        client = await self._ensure_client()
+        index = self._index_name(collection)
+        if not await self._index_exists(client, index):
+            self.ap.logger.warning(f"Valkey Search collection '{collection}' not found for deletion")
+            return
+
+        query = f'@{_FIELD_FILE_ID}:{{{self._encode_and_escape_tag(file_id)}}}'
+        keys = await self._search_keys(client, index, query)
+        if keys:
+            await client.delete(keys)
+        self.ap.logger.info(
+            f"Deleted {len(keys)} embeddings from Valkey Search collection '{collection}' with file_id: {file_id}"
+        )
+
+    async def delete_by_filter(self, collection: str, filter: dict[str, Any]) -> int:
+        client = await self._ensure_client()
+        index = self._index_name(collection)
+        if not await self._index_exists(client, index):
+            self.ap.logger.warning(f"Valkey Search collection '{collection}' not found for deletion")
+            return 0
+
+        # Guard against accidental mass deletion: a non-empty filter that maps
+        # to no usable (indexed) conditions must NOT fall back to match-all and
+        # wipe the whole collection.  Skip instead (matching Milvus / pgvector).
+        query = self._triples_to_ft(filter)
+        if not query:
+            self.ap.logger.warning(
+                "Valkey Search delete_by_filter on '%s': filter produced no usable conditions, skipping",
+                collection,
+            )
+            return 0
+        keys = await self._search_keys(client, index, query)
+        if keys:
+            await client.delete(keys)
+        self.ap.logger.info(f"Deleted {len(keys)} embeddings from Valkey Search collection '{collection}' by filter")
+        return len(keys)
+
+    async def list_by_filter(
+        self,
+        collection: str,
+        filter: dict[str, Any] | None = None,
+        limit: int = 20,
+        offset: int = 0,
+    ) -> tuple[list[dict[str, Any]], int]:
+        client = await self._ensure_client()
+        index = self._index_name(collection)
+        if not await self._index_exists(client, index):
+            return [], 0
+
+        query = self._triples_to_ft(filter) or _MATCH_ALL
+        options = FtSearchOptions(
+            return_fields=[
+                ReturnField(field_identifier=_FIELD_DOCUMENT),
+                ReturnField(field_identifier=_FIELD_METADATA),
+            ],
+            limit=FtSearchLimit(offset, limit),
+            dialect=2,
+        )
+        try:
+            reply = await ft.search(client, index, query, options)
+        except Exception as exc:
+            if self._is_missing_index_error(exc):
+                return [], 0
+            raise
+
+        total = 0
+        if reply:
+            try:
+                total = int(reply[0])
+            except (TypeError, ValueError):
+                total = 0
+
+        prefix = self._key_prefix(collection)
+        items: list[dict[str, Any]] = []
+        for doc_id, decoded_fields in self._iter_reply_docs(reply, prefix):
+            document = decoded_fields.get(_FIELD_DOCUMENT)
+            metadata: dict[str, Any] = {}
+            raw_meta = decoded_fields.get(_FIELD_METADATA)
+            if raw_meta is not None:
+                try:
+                    metadata = json.loads(self._decode(raw_meta))
+                except (TypeError, ValueError):
+                    metadata = {}
+
+            items.append(
+                {
+                    'id': doc_id,
+                    'document': self._decode(document) if document is not None else None,
+                    'metadata': metadata,
+                }
+            )
+
+        return items, total
+
+    async def delete_collection(self, collection: str):
+        client = await self._ensure_client()
+        index = self._index_name(collection)
+        self._ensured_indexes.discard(index)
+
+        if await self._index_exists(client, index):
+            try:
+                await ft.dropindex(client, index)
+            except RequestError:
+                # The index was already dropped (e.g. by a concurrent process)
+                # between the existence check and this call — benign. Other
+                # errors (connection / auth) must propagate so the caller knows
+                # the operation failed rather than silently SCAN-deleting next.
+                pass
+
+        # DROPINDEX does not remove the underlying hashes; delete them too.
+        prefix = self._key_prefix(collection)
+        cursor = b'0'
+        deleted = 0
+        for _ in range(_MAX_SCAN_ROUNDS):
+            cursor, keys = await client.scan(cursor, match=f'{prefix}*', count=500)
+            if keys:
+                await client.delete(keys)
+                deleted += len(keys)
+            if cursor in (b'0', '0', 0):
+                break
+        self.ap.logger.info(f"Valkey Search collection '{collection}' deleted ({deleted} keys removed)")
+
+    # ------------------------------------------------------------------ #
+    # Internal search helpers
+    # ------------------------------------------------------------------ #
+    async def _index_exists(self, client: GlideClient, index: str) -> bool:
+        if index in self._ensured_indexes:
+            return True
+        # ft.info is O(1) and raises RequestError when the index does not
+        # exist, vs ft.list which is O(n) over every index on the server and
+        # was being paid on the first query to each collection.
+        try:
+            await ft.info(client, index)
+            self._ensured_indexes.add(index)
+            return True
+        except RequestError:
+            return False
+
+    async def _search_keys(self, client: GlideClient, index: str, query: str) -> list[str]:
+        """Return all matching document keys for a query (NOCONTENT).
+
+        Paginates through the full result set in pages of ``_DELETE_SCAN_BATCH``
+        so that queries matching more than one page of chunks are fully
+        enumerated (avoids silently truncating deletes and leaving orphaned
+        vectors).
+        """
+        keys: list[str] = []
+        offset = 0
+        while True:
+            options = FtSearchOptions(
+                nocontent=True,
+                limit=FtSearchLimit(offset, _DELETE_SCAN_BATCH),
+                dialect=2,
+            )
+            try:
+                reply = await ft.search(client, index, query, options)
+            except Exception as exc:
+                if self._is_missing_index_error(exc):
+                    return keys
+                raise
+
+            if not reply or len(reply) < 2:
+                break
+
+            # reply[0] is the total match count; reply[1] holds this page.
+            total = 0
+            try:
+                total = int(reply[0])
+            except (TypeError, ValueError):
+                total = 0
+
+            docs = reply[1]
+            if isinstance(docs, dict):
+                page = [self._decode(k) for k in docs.keys()]
+            elif isinstance(docs, (list, tuple)):
+                page = [self._decode(k) for k in docs]
+            else:
+                page = []
+
+            if not page:
+                break
+            keys.extend(page)
+
+            offset += len(page)
+            if offset >= total or len(page) < _DELETE_SCAN_BATCH:
+                break
+
+        return keys

src/langbot/templates/config.yaml

@@ -87,6 +87,16 @@ vdb:
         database: 'langbot'
         user: 'postgres'
         password: 'postgres'
+    valkey_search:
+        host: 'localhost'
+        port: 6379            # integration tests use 6380 -> valkey/valkey-bundle:9.1.0
+        db: 0
+        password: ''          # optional (toB auth)
+        username: ''          # optional (ACL user, toB)
+        tls: false            # optional (toB/SaaS)
+        index_algorithm: 'HNSW'   # HNSW | FLAT
+        distance_metric: 'COSINE' # COSINE | L2 | IP
+        request_timeout: 5000     # per-request timeout in ms (glide default 250ms is too low for KNN)
 storage:
     use: local
     cleanup:
@@ -143,6 +153,15 @@ box:
             - './data/box'
             - '/tmp'
         workspace_quota_mb: null  # Optional disk quota override (>= 0). null = profile default.
+    # Default nsjail cgroup memory limit for each MCP stdio server process, in MB.
+    # Node.js MCP servers (npx/bunx) need more memory than Python ones because V8
+    # and WebAssembly modules (e.g. undici llhttp) reserve large virtual address
+    # space at startup. Setting this too low causes processes to be killed with
+    # return_code=137 (OOM kill); the symptom is "Box managed process exited
+    # unexpectedly" in the logs.  Raise on machines with ample RAM; lower only if
+    # you run exclusively Python (uvx) MCP servers.
+    # Can also be set via BOX__DEFAULT_MEMORY_MB. Default: 1536.
+    default_memory_mb: 1536
     docker:
         cpu_limit_enabled: true  # When false, Docker sandbox containers are started without --cpus. Memory and PID limits still apply.
     e2b:

tests/e2e/utils/config_factory.py

@@ -104,6 +104,17 @@ def create_minimal_config(tmpdir: Path, port: int = 15300) -> Path:
                 'user': 'postgres',
                 'password': 'postgres',
             },
+            'valkey_search': {
+                'host': 'localhost',
+                'port': 6379,
+                'db': 0,
+                'password': '',
+                'username': '',
+                'tls': False,
+                'index_algorithm': 'HNSW',
+                'distance_metric': 'COSINE',
+                'request_timeout': 5000,
+            },
         },
         'storage': {
             'use': 'local',

tests/integration/vector/test_valkey_search.py

@@ -0,0 +1,343 @@
+"""Integration tests for the Valkey Search VDB backend.
+
+These are SLOW, real-server tests.  They are gated on ``TEST_VALKEY_URL`` and
+skipped when it is unset (same precedent as the PostgreSQL migration tests).
+
+Run locally against valkey/valkey-bundle:9.1.0::
+
+    podman run -d --name valkey-test-langbot -p 6380:6379 valkey/valkey-bundle:9.1.0
+    TEST_VALKEY_URL=valkey://localhost:6380 \\
+        uv run pytest tests/integration/vector/test_valkey_search.py -m slow -q
+
+The default upstream fast CI lane (``-m "not slow"``) skips these; the local
+supervisor validator MUST run them.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import uuid
+from types import SimpleNamespace
+from urllib.parse import urlparse
+
+import pytest
+
+pytestmark = [pytest.mark.integration, pytest.mark.slow]
+
+
+def _parse_valkey_url(url: str) -> tuple[str, int, int]:
+    """Parse ``valkey://host:port/db`` into ``(host, port, db)``."""
+    parsed = urlparse(url)
+    host = parsed.hostname or 'localhost'
+    port = parsed.port or 6379
+    db = 0
+    if parsed.path and parsed.path.strip('/'):
+        try:
+            db = int(parsed.path.strip('/'))
+        except ValueError:
+            db = 0
+    return host, port, db
+
+
+@pytest.fixture
+def valkey_config():
+    url = os.environ.get('TEST_VALKEY_URL')
+    if not url:
+        pytest.skip('TEST_VALKEY_URL not set')
+    host, port, db = _parse_valkey_url(url)
+    return {
+        'host': host,
+        'port': port,
+        'db': db,
+        'password': '',
+        'username': '',
+        'tls': False,
+        'index_algorithm': 'HNSW',
+        'distance_metric': 'COSINE',
+    }
+
+
+def _make_ap(valkey_config):
+    """Build a minimal fake ``ap`` with the config + a no-op logger."""
+    logger = SimpleNamespace(
+        info=lambda *a, **k: None,
+        warning=lambda *a, **k: None,
+        error=lambda *a, **k: None,
+        debug=lambda *a, **k: None,
+    )
+    instance_config = SimpleNamespace(data={'vdb': {'valkey_search': valkey_config}})
+    return SimpleNamespace(instance_config=instance_config, logger=logger)
+
+
+@pytest.fixture
+async def backend(valkey_config):
+    """Create a Valkey Search backend, skip if module/server unavailable."""
+    from langbot.pkg.vector.vdbs.valkey_search import (
+        ValkeySearchVectorDatabase,
+        VALKEY_SEARCH_AVAILABLE,
+    )
+    from glide import ft
+
+    if not VALKEY_SEARCH_AVAILABLE:
+        pytest.skip('valkey-glide not installed')
+
+    ap = _make_ap(valkey_config)
+    db = ValkeySearchVectorDatabase(ap)
+    client = await db._ensure_client()
+
+    # Module-presence gate: FT.LIST must be available (Search module loaded).
+    try:
+        await ft.list(client)
+    except Exception as exc:  # noqa: BLE001
+        await client.close()
+        pytest.skip(f'Valkey Search module not available: {exc}')
+
+    collection = f'test_{uuid.uuid4().hex[:12]}'
+    yield db, collection
+
+    # Cleanup
+    try:
+        await db.delete_collection(collection)
+    except Exception:
+        pass
+    if db._client is not None:
+        await db._client.close()
+
+
+async def _poll_until(coro_factory, predicate, timeout=5.0, interval=0.2):
+    """Poll an async result until predicate is true (indexer is async)."""
+    deadline = asyncio.get_event_loop().time() + timeout
+    result = await coro_factory()
+    while not predicate(result) and asyncio.get_event_loop().time() < deadline:
+        await asyncio.sleep(interval)
+        result = await coro_factory()
+    return result
+
+
+def _sample_docs():
+    ids = ['d1', 'd2', 'd3']
+    embeddings = [
+        [1.0, 0.0, 0.0, 0.0],
+        [0.0, 1.0, 0.0, 0.0],
+        [0.9, 0.1, 0.0, 0.0],
+    ]
+    metadatas = [
+        {'file_id': 'fileA', 'topic': 'cats'},
+        {'file_id': 'fileB', 'topic': 'dogs'},
+        {'file_id': 'fileA', 'topic': 'cats'},
+    ]
+    documents = [
+        'the quick brown fox',
+        'lazy dogs sleeping',
+        'foxes and cats playing',
+    ]
+    return ids, embeddings, metadatas, documents
+
+
+@pytest.mark.asyncio
+async def test_add_and_vector_search(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+
+    result = await _poll_until(
+        lambda: db.search(collection, [1.0, 0.0, 0.0, 0.0], k=3, search_type='vector'),
+        lambda r: len(r['ids'][0]) >= 1,
+    )
+    assert len(result['ids'][0]) >= 1
+    # Closest to [1,0,0,0] should be d1.
+    assert result['ids'][0][0] == 'd1'
+    assert all(isinstance(d, float) for d in result['distances'][0])
+
+
+@pytest.mark.asyncio
+async def test_full_text_search(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+
+    result = await _poll_until(
+        lambda: db.search(collection, [0.0, 0.0, 0.0, 0.0], k=5, search_type='full_text', query_text='dogs'),
+        lambda r: len(r['ids'][0]) >= 1,
+    )
+    assert 'd2' in result['ids'][0]
+
+
+@pytest.mark.asyncio
+async def test_hybrid_filter_then_knn(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+
+    result = await _poll_until(
+        lambda: db.search(
+            collection,
+            [1.0, 0.0, 0.0, 0.0],
+            k=5,
+            search_type='hybrid',
+            query_text='cats',
+            filter={'file_id': 'fileA'},
+        ),
+        lambda r: len(r['ids'][0]) >= 1,
+    )
+    # Only fileA docs (d1, d3) should be candidates.
+    assert set(result['ids'][0]).issubset({'d1', 'd3'})
+
+
+@pytest.mark.asyncio
+async def test_vector_weight_not_honored(backend):
+    """Passing different vector_weight values must NOT change ranking."""
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+
+    common = dict(
+        collection=collection, query_embedding=[1.0, 0.0, 0.0, 0.0], k=3, search_type='hybrid', query_text='cats'
+    )
+    await _poll_until(lambda: db.search(**common), lambda r: len(r['ids'][0]) >= 1)
+
+    r_low = await db.search(**common, vector_weight=0.1)
+    r_high = await db.search(**common, vector_weight=0.9)
+    assert r_low['ids'][0] == r_high['ids'][0]
+
+
+@pytest.mark.asyncio
+async def test_filter_operators(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+
+    # Wait for indexing.
+    await _poll_until(
+        lambda: db.list_by_filter(collection, limit=10),
+        lambda r: r[1] >= 3,
+    )
+
+    # $eq
+    items, total = await db.list_by_filter(collection, filter={'file_id': 'fileA'})
+    assert total == 2
+    assert {it['id'] for it in items} == {'d1', 'd3'}
+
+    # $ne
+    items, total = await db.list_by_filter(collection, filter={'file_id': {'$ne': 'fileA'}})
+    assert {it['id'] for it in items} == {'d2'}
+
+    # $in
+    items, total = await db.list_by_filter(collection, filter={'file_id': {'$in': ['fileA', 'fileB']}})
+    assert total == 3
+
+    # $nin
+    items, total = await db.list_by_filter(collection, filter={'file_id': {'$nin': ['fileB']}})
+    assert {it['id'] for it in items} == {'d1', 'd3'}
+
+
+@pytest.mark.asyncio
+async def test_delete_by_file_id(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+    await _poll_until(lambda: db.list_by_filter(collection, limit=10), lambda r: r[1] >= 3)
+
+    await db.delete_by_file_id(collection, 'fileA')
+    items, total = await _poll_until(
+        lambda: db.list_by_filter(collection, limit=10),
+        lambda r: r[1] <= 1,
+    )
+    assert {it['id'] for it in items} == {'d2'}
+
+
+@pytest.mark.asyncio
+async def test_delete_by_filter_returns_count(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+    await _poll_until(lambda: db.list_by_filter(collection, limit=10), lambda r: r[1] >= 3)
+
+    deleted = await db.delete_by_filter(collection, filter={'file_id': 'fileA'})
+    assert deleted == 2
+
+
+@pytest.mark.asyncio
+async def test_list_by_filter_pagination(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+    await _poll_until(lambda: db.list_by_filter(collection, limit=10), lambda r: r[1] >= 3)
+
+    page1, total = await db.list_by_filter(collection, limit=2, offset=0)
+    assert total == 3
+    assert len(page1) == 2
+
+    page2, total = await db.list_by_filter(collection, limit=2, offset=2)
+    assert total == 3
+    assert len(page2) == 1
+
+
+@pytest.mark.asyncio
+async def test_delete_collection(backend):
+    db, collection = backend
+    ids, embeddings, metadatas, documents = _sample_docs()
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+    await _poll_until(lambda: db.list_by_filter(collection, limit=10), lambda r: r[1] >= 3)
+
+    await db.delete_collection(collection)
+
+    # After dropping, search on a missing index returns empty.
+    result = await db.search(collection, [1.0, 0.0, 0.0, 0.0], k=3, search_type='vector')
+    assert result['ids'][0] == []
+
+
+@pytest.mark.asyncio
+async def test_adversarial_filter_and_query_input(backend):
+    """Crafted FT special chars in file_id / query_text must not break out.
+
+    Guarantees locked in here:
+    * A file_id full of injection-style chars (quotes, parens, ``|``, ``@``,
+      ``:``, spaces, dashes) only ever matches its own row — the payload is
+      escaped to literal TAG content, never interpreted as extra clauses.
+    * A query_text full of FT operators does not raise and does not widen the
+      result set.
+    * A file_id containing FT-unsafe chars (``{`` / ``}`` / ``*``) is
+      percent-encoded, so it round-trips correctly: an exact match returns ONLY
+      its own row and never widens to an unrelated row, and the query does not
+      raise.
+    """
+    db, collection = backend
+
+    # Injection-style file_id WITHOUT FT-unsafe chars (the realistic surface).
+    injection_fid = 'evil") @file_id (".id|x-y:z'
+    # file_id WITH FT-unsafe chars that previously could not be queried.
+    brace_fid = 'x} @file_id:{*'
+    ids = ['adv1', 'benign2', 'brace3']
+    embeddings = [[1.0, 0.0, 0.0, 0.0], [0.0, 1.0, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0]]
+    metadatas = [{'file_id': injection_fid}, {'file_id': 'plainB'}, {'file_id': brace_fid}]
+    documents = ['payload row content', 'unrelated benign content', 'brace row content']
+    await db.add_embeddings(collection, ids, embeddings, metadatas, documents)
+    await _poll_until(lambda: db.list_by_filter(collection, limit=10), lambda r: r[1] >= 3)
+
+    # Exact-match on the crafted file_id returns ONLY its own row.
+    items, total = await db.list_by_filter(collection, filter={'file_id': injection_fid})
+    assert total == 1
+    assert {it['id'] for it in items} == {'adv1'}
+
+    # A query_text packed with FT operators must not raise and must not match
+    # the benign row (escaped to literal terms, none of which it contains).
+    result = await db.search(
+        collection,
+        [0.0, 0.0, 0.0, 0.0],
+        k=5,
+        search_type='full_text',
+        query_text='@document:{*} | -()~ "evil"',
+    )
+    assert 'benign2' not in result['ids'][0]
+
+    # The brace/star-bearing file_id is encoded, so it round-trips: exact match
+    # returns ONLY its own row and never widens. No RequestError is raised.
+    b_items, b_total = await db.list_by_filter(collection, filter={'file_id': brace_fid})
+    assert b_total == 1
+    assert {it['id'] for it in b_items} == {'brace3'}
+
+    # And deletion by that file_id removes exactly its own row.
+    deleted = await db.delete_by_filter(collection, filter={'file_id': brace_fid})
+    assert deleted == 1

tests/unit_tests/COVERAGE_EXCLUSIONS.md

@@ -27,7 +27,7 @@
 
 ### 4. 向量数据库 (`vector/vdbs/`)
 - **路径**: `src/langbot/pkg/vector/vdbs/`
-- **模块**: chroma, milvus, pgvector, qdrant, seekdb
+- **模块**: chroma, milvus, pgvector, qdrant, seekdb, valkey_search
 - **排除原因**: 需要真实向量数据库实例运行
 - **测试方式**: 需要 Docker 启动测试数据库或 mock
 - **状态**: 后续可补充 mock 测试

tests/unit_tests/platform/test_dingtalk_adapter.py

@@ -0,0 +1,254 @@
+"""Tests for DingTalk adapter helper behavior."""
+
+import asyncio
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from langbot.pkg.platform.sources.dingtalk import (
+    DingTalkAdapter,
+    _dingtalk_card_markdown,
+    _dingtalk_clean_form_content,
+    _dingtalk_completed_input_lines,
+    _dingtalk_extract_component_inputs,
+    _dingtalk_form_component_params,
+    _dingtalk_missing_completed_input_lines,
+    _dingtalk_pending_input_defs,
+)
+
+
+def test_dingtalk_select_component_params_expose_options():
+    params = _dingtalk_form_component_params(
+        {
+            '_current_input_field': 'choice',
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+            'inputs': {},
+        }
+    )
+
+    assert params['select_visible'] == 'true'
+    assert params['select_placeholder'] == 'choice'
+    assert params['select_options'] == ['A', 'B']
+    assert [option['value'] for option in params['index_o']] == ['A', 'B']
+    assert [option['value'] for option in params['test_index']] == ['A', 'B']
+    assert params['index_o'][0]['text']['zh_CN'] == 'A'
+    assert params['index_o'][0]['text']['en_US'] == 'A'
+    assert params['select_index'] == -1
+
+
+def test_dingtalk_extract_select_from_builtin_result_dict():
+    inputs = _dingtalk_extract_component_inputs({'selectResult': {'index': 1, 'value': 'B'}})
+
+    assert inputs == {'select': 'B'}
+
+
+def test_dingtalk_extract_select_from_template_param_string():
+    inputs = _dingtalk_extract_component_inputs({'select': '{"index": 1, "value": "B"}'})
+
+    assert inputs == {'select': '{"index": 1, "value": "B"}'}
+
+
+def test_dingtalk_extract_input_and_select_together():
+    inputs = _dingtalk_extract_component_inputs(
+        {
+            'inputResult': {'value': 'looks good'},
+            '__built_in_selectResult__': {'index': 0, 'value': 'A'},
+        }
+    )
+
+    assert inputs == {'input': 'looks good', 'select': 'A'}
+
+
+def test_dingtalk_extract_component_inputs_strips_card_line_endings():
+    inputs = _dingtalk_extract_component_inputs(
+        {
+            'inputResult': {'value': '回复我测试\r\n'},
+            'selectResult': {'value': '1\r'},
+        }
+    )
+
+    assert inputs == {'input': '回复我测试', 'select': '1'}
+
+
+def test_dingtalk_pending_input_defs_includes_file_fields():
+    pending = _dingtalk_pending_input_defs(
+        {
+            'input_defs': [
+                {'output_variable_name': 'comment', 'type': 'paragraph'},
+                {'output_variable_name': 'files', 'type': 'file-list'},
+            ],
+            'inputs': {'comment': 'ready'},
+        }
+    )
+
+    assert [field['output_variable_name'] for field in pending] == ['files']
+
+
+def test_dingtalk_completed_input_lines_include_text_and_select_values():
+    lines = _dingtalk_completed_input_lines(
+        {
+            'all_input_defs': [
+                {'output_variable_name': 'comment', 'type': 'paragraph'},
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                },
+            ],
+            'inputs': {'comment': 'looks good', 'choice': 'B'},
+        }
+    )
+
+    assert lines == ['✅ comment：looks good', '✅ choice：B']
+
+
+def test_dingtalk_completed_inputs_are_not_repeated_when_already_interleaved():
+    form_data = {
+        'all_input_defs': [
+            {'output_variable_name': 'us_input', 'type': 'paragraph'},
+            {'output_variable_name': 'xiala', 'type': 'select'},
+        ],
+        'inputs': {'us_input': '回复我测试\r', 'xiala': '1'},
+    }
+    form_content = '你好\n请输入你的问题\n✅ us_input：回复我测试\n请选择你的答案\n✅ xiala：1'
+
+    assert _dingtalk_missing_completed_input_lines(form_data, form_content) == []
+
+
+def test_dingtalk_completed_inputs_are_appended_when_template_does_not_render_them():
+    form_data = {
+        'all_input_defs': [{'output_variable_name': 'comment', 'type': 'paragraph'}],
+        'inputs': {'comment': 'ready'},
+    }
+
+    assert _dingtalk_missing_completed_input_lines(form_data, 'Please review') == ['✅ comment：ready']
+
+
+def test_dingtalk_clean_form_content_uses_all_input_defs():
+    content = _dingtalk_clean_form_content(
+        {
+            'raw_form_content': 'Hello\n\n{{#$output.comment#}}\n\n{{#$output.choice#}}\n',
+            'input_defs': [],
+            'all_input_defs': [
+                {'output_variable_name': 'comment', 'type': 'paragraph'},
+                {'output_variable_name': 'choice', 'type': 'select'},
+            ],
+        }
+    )
+
+    assert content == 'Hello'
+
+
+def test_dingtalk_field_stage_keeps_prior_prompts_and_completed_values():
+    content = _dingtalk_clean_form_content(
+        {
+            '_current_input_field': 'choice',
+            'raw_form_content': ('Question\n{{#$output.comment#}}\nChoose an answer\n{{#$output.choice#}}'),
+            'form_content': 'Choose an answer',
+            'input_defs': [
+                {'output_variable_name': 'comment', 'type': 'paragraph'},
+                {'output_variable_name': 'choice', 'type': 'select'},
+            ],
+            'inputs': {'comment': 'hello'},
+        }
+    )
+
+    assert '{{#$output.' not in content
+    assert content.index('Question') < content.index('comment')
+    assert content.index('comment') < content.index('Choose an answer')
+
+
+def test_dingtalk_final_action_stage_interleaves_prompts_and_completed_values():
+    content = _dingtalk_clean_form_content(
+        {
+            '_action_select_only': True,
+            'raw_form_content': ('11\nQuestion\n{{#$output.comment#}}\nChoose an answer\n{{#$output.choice#}}'),
+            'all_input_defs': [
+                {'output_variable_name': 'comment', 'type': 'paragraph'},
+                {'output_variable_name': 'choice', 'type': 'select'},
+            ],
+            'inputs': {'comment': 'hello', 'choice': 'B'},
+        }
+    )
+
+    assert '{{#$output.' not in content
+    assert content.startswith('11\nQuestion')
+    assert content.index('Question') < content.index('comment')
+    assert content.index('comment') < content.index('Choose an answer')
+    assert content.index('Choose an answer') < content.index('choice')
+
+
+def test_dingtalk_card_markdown_preserves_internal_line_breaks():
+    assert _dingtalk_card_markdown('11\nQuestion\nCompleted') == '11<br>Question<br>Completed'
+
+
+def _build_card_action_adapter() -> DingTalkAdapter:
+    adapter = DingTalkAdapter.model_construct(
+        card_state={
+            'card-1': {
+                'session_key': 'group_group-1',
+                'launcher_type': 'group',
+                'launcher_id': 'group-1',
+                'sender_user_id': 'initiator-1',
+                'form_token': 'token-1',
+                'workflow_run_id': 'run-1',
+                'actions': [{'id': 'approve', 'title': 'Approve'}],
+                'node_title': 'Review',
+                'form_content': 'Please review',
+                'input_defs': [],
+                'inputs': {},
+            }
+        },
+        active_turn_card={},
+        active_turn_text={},
+    )
+    adapter.logger = AsyncMock()
+    adapter.ap = MagicMock()
+    adapter.ap.platform_mgr.bots = []
+    adapter.ap.query_pool.add_query = AsyncMock()
+    return adapter
+
+
+@pytest.mark.asyncio
+async def test_dingtalk_group_card_action_uses_clicker_as_sender():
+    adapter = _build_card_action_adapter()
+
+    with patch.object(DingTalkAdapter, '_mark_card_resolved', new=AsyncMock()) as mark_resolved:
+        await adapter._on_card_action(
+            {
+                'out_track_id': 'card-1',
+                'user_id': 'reviewer-2',
+                'action_id': 'approve',
+                'params': {},
+            }
+        )
+        await asyncio.sleep(0)
+
+    call = adapter.ap.query_pool.add_query.await_args
+    assert call.kwargs['launcher_id'] == 'group-1'
+    assert call.kwargs['sender_id'] == 'reviewer-2'
+    assert call.kwargs['message_event'].sender.id == 'reviewer-2'
+    mark_resolved.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_dingtalk_unknown_card_action_is_rejected():
+    adapter = _build_card_action_adapter()
+
+    await adapter._on_card_action(
+        {
+            'out_track_id': 'card-1',
+            'user_id': 'reviewer-2',
+            'action_id': 'not-on-card',
+            'params': {},
+        }
+    )
+
+    adapter.ap.query_pool.add_query.assert_not_awaited()
+    adapter.logger.warning.assert_awaited_once()

tests/unit_tests/platform/test_dingtalk_api.py

@@ -0,0 +1,42 @@
+"""Tests for DingTalk API payload helpers."""
+
+import json
+
+from langbot.libs.dingtalk_api.api import _stringify_card_param_map
+
+
+def test_dingtalk_card_param_map_stringifies_select_component_arrays():
+    params = _stringify_card_param_map(
+        {
+            'content': 'Pick one',
+            'btns': json.dumps([{'text': 'OK'}], ensure_ascii=False),
+            'select_options': ['A', 'B'],
+            'index_o': [
+                {
+                    'value': 'A',
+                    'text': {'zh_CN': 'A', 'en_US': 'A'},
+                }
+            ],
+            'test_index': [
+                {
+                    'value': 'A',
+                    'text': {'zh_CN': 'A', 'en_US': 'A'},
+                }
+            ],
+            'select_index': -1,
+        }
+    )
+
+    assert params['content'] == 'Pick one'
+    assert params['btns'] == '[{"text": "OK"}]'
+    assert params['select_options'] == '["A", "B"]'
+    assert json.loads(params['index_o'])[0]['value'] == 'A'
+    assert json.loads(params['test_index'])[0]['value'] == 'A'
+    assert params['select_index'] == '-1'
+
+
+def test_dingtalk_card_param_map_stringifies_unregistered_structures():
+    params = _stringify_card_param_map({'other': ['A'], 'empty': None})
+
+    assert params['other'] == '["A"]'
+    assert params['empty'] == ''

tests/unit_tests/platform/test_lark_adapter.py

@@ -0,0 +1,196 @@
+"""Tests for Lark adapter helper behavior."""
+
+from langbot.pkg.platform.sources.lark import (
+    LarkAdapter,
+    _lark_clean_form_content,
+    _lark_completed_input_lines,
+    _lark_current_input_defs,
+    _lark_extract_action_form_inputs,
+    _lark_should_update_stream_element,
+    _lark_visible_form_content,
+)
+
+
+def test_lark_current_input_defs_only_returns_active_stage():
+    input_defs = [
+        {'output_variable_name': 'us_input', 'type': 'paragraph'},
+        {'output_variable_name': 'xiala', 'type': 'select'},
+    ]
+
+    assert _lark_current_input_defs(
+        {
+            '_current_input_field': 'xiala',
+            'input_defs': input_defs,
+        }
+    ) == [input_defs[1]]
+    assert (
+        _lark_current_input_defs(
+            {
+                '_action_select_only': True,
+                'input_defs': input_defs,
+            }
+        )
+        == []
+    )
+
+
+def test_lark_form_field_elements_only_render_active_stage():
+    adapter = LarkAdapter.model_construct()
+    form_data = {
+        '_current_input_field': 'xiala',
+        'input_defs': [
+            {'output_variable_name': 'us_input', 'type': 'paragraph'},
+            {
+                'output_variable_name': 'xiala',
+                'type': 'select',
+                'option_source': {'type': 'constant', 'value': ['1', '2']},
+            },
+        ],
+    }
+
+    elements, input_name_map, file_help_lines = adapter._build_lark_form_field_elements(form_data)
+
+    assert len(elements) == 1
+    assert elements[0]['tag'] == 'select_static'
+    assert elements[0]['label']['content'] == 'xiala'
+    assert list(input_name_map.values()) == ['xiala']
+    assert file_help_lines == []
+
+
+def test_lark_form_stage_skips_closed_streaming_element_update():
+    assert not _lark_should_update_stream_element(
+        resume_from=False,
+        form_data={'_current_input_field': 'xiala'},
+        msg_seq=1,
+        is_final=True,
+    )
+    assert _lark_should_update_stream_element(
+        resume_from=False,
+        form_data=None,
+        msg_seq=1,
+        is_final=True,
+    )
+
+
+def test_lark_final_action_stage_interleaves_prompts_and_completed_values():
+    form_content = _lark_visible_form_content(
+        {
+            '_action_select_only': True,
+            'raw_form_content': ('11\nQuestion\n{{#$output.us_input#}}\nChoose an answer\n{{#$output.xiala#}}\n'),
+            'all_input_defs': [
+                {'output_variable_name': 'us_input', 'type': 'paragraph'},
+                {'output_variable_name': 'xiala', 'type': 'select'},
+            ],
+            'inputs': {'us_input': 'hello', 'xiala': '2'},
+        }
+    )
+
+    assert '{{#$output.' not in form_content
+    assert form_content.startswith('11\nQuestion')
+    assert form_content.index('Question') < form_content.index('us_input')
+    assert form_content.index('us_input') < form_content.index('Choose an answer')
+    assert form_content.index('Choose an answer') < form_content.index('xiala')
+
+
+def test_lark_completed_input_lines_include_text_select_and_files():
+    lines = _lark_completed_input_lines(
+        {
+            'all_input_defs': [
+                {'output_variable_name': 'us_input', 'type': 'paragraph'},
+                {'output_variable_name': 'xiala', 'type': 'select'},
+                {'output_variable_name': 'files', 'type': 'file-list'},
+            ],
+            'inputs': {
+                'us_input': '你好',
+                'xiala': 'or',
+                'files': [{'upload_file_id': 'file-1'}, {'upload_file_id': 'file-2'}],
+            },
+        }
+    )
+
+    assert lines == [
+        '✅ us_input：你好',
+        '✅ xiala：or',
+        '✅ files：2 file(s)',
+    ]
+
+
+def test_lark_clean_form_content_removes_all_input_placeholders():
+    content = _lark_clean_form_content(
+        '人工介入\n\n{{#$output.us_input#}}\n\n{{#$output.xiala#}}\n',
+        [
+            {'output_variable_name': 'us_input', 'type': 'paragraph'},
+            {'output_variable_name': 'xiala', 'type': 'select'},
+        ],
+    )
+
+    assert content == '人工介入'
+
+
+def test_lark_extract_action_form_inputs_from_json_form_value():
+    class Action:
+        form_value = '{"Input_1_us_input_abcd12": "hello", "Select_2_xiala_abcd12": "B"}'
+        input_value = None
+        option = None
+        name = None
+
+    inputs = _lark_extract_action_form_inputs(
+        Action(),
+        {
+            'input_name_map': {
+                'Input_1_us_input_abcd12': 'us_input',
+                'Select_2_xiala_abcd12': 'xiala',
+            }
+        },
+    )
+
+    assert inputs == {'us_input': 'hello', 'xiala': 'B'}
+
+
+def test_lark_extract_action_form_inputs_from_webhook_dict_action():
+    inputs = _lark_extract_action_form_inputs(
+        {
+            'form_value': {
+                'Input_1_us_input_abcd12': 'hello',
+                'Select_2_xiala_abcd12': {'value': 'B', 'text': {'content': 'Option B'}},
+            }
+        },
+        {
+            'input_name_map': {
+                'Input_1_us_input_abcd12': 'us_input',
+                'Select_2_xiala_abcd12': 'xiala',
+            }
+        },
+    )
+
+    assert inputs == {'us_input': 'hello', 'xiala': {'value': 'B', 'text': {'content': 'Option B'}}}
+
+
+def test_lark_extract_action_form_inputs_maps_dotted_component_names():
+    inputs = _lark_extract_action_form_inputs(
+        {
+            'form_value': {
+                'Form_1_token_abcd12.Input_1_us_input_abcd12': 'hello',
+            }
+        },
+        {
+            'input_name_map': {
+                'Input_1_us_input_abcd12': 'us_input',
+            }
+        },
+    )
+
+    assert inputs == {'us_input': 'hello'}
+
+
+def test_lark_completed_input_lines_display_select_value_from_object():
+    lines = _lark_completed_input_lines(
+        {
+            'all_input_defs': [
+                {'output_variable_name': 'xiala', 'type': 'select'},
+            ],
+            'inputs': {'xiala': {'value': 'B', 'text': {'content': 'Option B'}}},
+        }
+    )
+
+    assert lines == ['✅ xiala：B']

tests/unit_tests/platform/test_qqofficial_api.py

@@ -0,0 +1,224 @@
+"""Tests for QQ Official keyboard payload helpers."""
+
+import asyncio
+import time
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+import langbot_plugin.api.entities.builtin.platform.message as platform_message
+
+from langbot.libs.qq_official_api.api import (
+    QQ_SELECT_ACTION_PREFIX,
+    build_keyboard_from_select_field,
+    get_select_field_options,
+    resolve_select_button_action,
+)
+
+
+def _select_form_data() -> dict:
+    return {
+        '_current_input_field': 'choice',
+        'input_defs': [
+            {
+                'output_variable_name': 'choice',
+                'type': 'select',
+                'option_source': {'type': 'constant', 'value': ['A', 'B', 'C']},
+            }
+        ],
+    }
+
+
+def test_qq_select_field_builds_callback_buttons():
+    keyboard = build_keyboard_from_select_field(_select_form_data(), buttons_per_row=2)
+
+    rows = keyboard['content']['rows']
+    assert [[button['render_data']['label'] for button in row['buttons']] for row in rows] == [
+        ['A', 'B'],
+        ['C'],
+    ]
+    assert rows[0]['buttons'][0]['action']['data'] == f'{QQ_SELECT_ACTION_PREFIX}0'
+    assert rows[0]['buttons'][1]['action']['data'] == f'{QQ_SELECT_ACTION_PREFIX}1'
+
+
+def test_qq_select_button_resolves_field_and_value():
+    form_data = _select_form_data()
+
+    assert get_select_field_options(form_data) == ('choice', ['A', 'B', 'C'])
+    assert resolve_select_button_action(form_data, f'{QQ_SELECT_ACTION_PREFIX}1') == ('choice', 'B')
+    assert resolve_select_button_action(form_data, f'{QQ_SELECT_ACTION_PREFIX}99') is None
+
+
+def test_qq_select_keyboard_fits_twenty_five_options():
+    form_data = _select_form_data()
+    form_data['input_defs'][0]['option_source']['value'] = [f'Option {idx}' for idx in range(25)]
+
+    rows = build_keyboard_from_select_field(form_data)['content']['rows']
+
+    assert len(rows) == 5
+    assert all(len(row['buttons']) == 5 for row in rows)
+
+
+def test_qq_non_select_field_does_not_build_keyboard():
+    form_data = {
+        '_current_input_field': 'comment',
+        'input_defs': [{'output_variable_name': 'comment', 'type': 'paragraph'}],
+    }
+
+    assert build_keyboard_from_select_field(form_data)['content']['rows'] == []
+
+
+def _stream_test_adapter():
+    from langbot.pkg.platform.sources.qqofficial import QQOfficialAdapter
+
+    adapter = QQOfficialAdapter.model_construct()
+    adapter.logger = AsyncMock()
+    adapter.bot = MagicMock()
+    adapter.bot.send_stream_msg = AsyncMock(return_value={'id': 'stream-1'})
+    adapter.bot.send_markdown_keyboard = AsyncMock(return_value={'id': 'message-1'})
+    adapter.ap = None
+    adapter._stream_ctx = {}
+    adapter._stream_ctx_ts = {}
+    adapter._fallback_text = {}
+    adapter._fallback_text_ts = {}
+    return adapter
+
+
+@pytest.mark.asyncio
+async def test_qq_stream_uses_cumulative_chunks_as_snapshots():
+    adapter = _stream_test_adapter()
+    adapter._stream_ctx['message-1'] = {
+        'user_openid': 'user-1',
+        'msg_id': 'source-1',
+        'stream_msg_id': None,
+        'msg_seq': 1,
+        'index': 0,
+        'last_update_ts': 0,
+        'accumulated_text': '',
+        'sent_length': 0,
+        'session_started': False,
+    }
+    adapter._stream_ctx_ts['message-1'] = time.time()
+    source = MagicMock()
+
+    await adapter.reply_message_chunk(
+        source,
+        {'resp_message_id': 'message-1'},
+        platform_message.MessageChain([platform_message.Plain(text='<think>one')]),
+    )
+    await adapter.reply_message_chunk(
+        source,
+        {'resp_message_id': 'message-1'},
+        platform_message.MessageChain([platform_message.Plain(text='<think>one two')]),
+        is_final=True,
+    )
+
+    assert [call.kwargs['content'] for call in adapter.bot.send_stream_msg.await_args_list] == [
+        '<think>one',
+        ' two',
+    ]
+
+
+@pytest.mark.asyncio
+async def test_qq_non_streaming_fallback_keeps_latest_snapshot_only():
+    from langbot.pkg.platform.sources.qqofficial import QQOfficialAdapter
+
+    adapter = _stream_test_adapter()
+    source = MagicMock()
+
+    with patch.object(QQOfficialAdapter, 'reply_message', new=AsyncMock()) as reply_message:
+        await adapter.reply_message_chunk(
+            source,
+            {'resp_message_id': 'message-1'},
+            platform_message.MessageChain([platform_message.Plain(text='Hel')]),
+        )
+        await adapter.reply_message_chunk(
+            source,
+            {'resp_message_id': 'message-1'},
+            platform_message.MessageChain([platform_message.Plain(text='Hello')]),
+            is_final=True,
+        )
+
+    sent_chain = reply_message.await_args.args[1]
+    assert str(sent_chain) == 'Hello'
+
+
+@pytest.mark.asyncio
+async def test_qq_text_field_prompt_keeps_form_content():
+    from langbot.pkg.platform.sources.qqofficial import QQOfficialAdapter
+
+    adapter = _stream_test_adapter()
+    adapter._pending_forms = {}
+    adapter._session_event_ids = {}
+    adapter._anchor_msg_seq = {}
+    source = MagicMock()
+    source.d_id = 'source-1'
+    source.t = 'C2C_MESSAGE_CREATE'
+    event = MagicMock()
+    event.source_platform_object = source
+    event.sender.id = 'user-1'
+    form_data = {
+        '_current_input_field': 'us_input',
+        'node_title': 'Manual input',
+        'form_content': '1234\nEnter your question',
+        'input_defs': [{'output_variable_name': 'us_input', 'type': 'paragraph'}],
+        'actions': [{'id': 'yes', 'title': 'yes'}],
+    }
+
+    with patch.object(QQOfficialAdapter, '_resolve_target_from_event', return_value=('c2c', 'user-1')):
+        await adapter._handle_form_chunk(event, platform_message.MessageChain([]), form_data)
+
+    send_call = adapter.bot.send_markdown_keyboard.await_args.kwargs
+    assert send_call['markdown_content'] == '### Manual input\n\n1234\nEnter your question'
+    assert send_call['keyboard'] is None
+
+
+@pytest.mark.asyncio
+async def test_qq_select_click_enqueues_input_progress_query():
+    import langbot.pkg.core.app  # noqa: F401
+    from langbot.pkg.platform.sources.qqofficial import QQOfficialAdapter
+
+    adapter = QQOfficialAdapter.model_construct()
+    adapter.logger = AsyncMock()
+    adapter.bot = MagicMock()
+    adapter.bot.ack_interaction = AsyncMock()
+    adapter.ap = MagicMock()
+    adapter.ap.platform_mgr.bots = []
+    adapter.ap.query_pool.add_query = AsyncMock()
+    adapter._pending_forms = {
+        'group_group-1': {
+            'form_data': {
+                **_select_form_data(),
+                'form_token': 'token-1',
+                'workflow_run_id': 'run-1',
+                'node_title': 'Review',
+                'actions': [{'id': 'approve', 'title': 'Approve'}],
+            },
+            'sender_id': 'initiator-1',
+            'posted_at': time.time(),
+        }
+    }
+    adapter._session_event_ids = {}
+    adapter._anchor_msg_seq = {}
+
+    await adapter._handle_interaction_create(
+        {
+            'id': 'interaction-1',
+            'chat_type': 1,
+            'group_openid': 'group-1',
+            'member_openid': 'reviewer-2',
+            'data': {'resolved': {'button_data': f'{QQ_SELECT_ACTION_PREFIX}1'}},
+        },
+        ws_event_id='event-1',
+    )
+    await asyncio.sleep(0)
+
+    call = adapter.ap.query_pool.add_query.await_args
+    form_action = call.kwargs['variables']['_dify_form_action']
+    assert call.kwargs['launcher_id'] == 'group-1'
+    assert call.kwargs['sender_id'] == 'reviewer-2'
+    assert form_action['action_id'] == ''
+    assert form_action['inputs'] == {'select': 'B'}
+    assert form_action['_current_input_field'] == 'choice'
+    assert form_action['_input_progress'] is True
+    adapter.bot.ack_interaction.assert_awaited_once_with('interaction-1', code=0)

tests/unit_tests/platform/test_telegram_adapter.py

@@ -0,0 +1,158 @@
+"""Tests for Telegram Dify form callback helpers."""
+
+import json
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+from telegram import ForceReply
+
+import langbot_plugin.api.entities.builtin.platform.entities as platform_entities
+import langbot_plugin.api.entities.builtin.platform.events as platform_events
+import langbot_plugin.api.entities.builtin.platform.message as platform_message
+from langbot.pkg.platform.sources.telegram import (
+    TelegramAdapter,
+    _telegram_form_action_from_callback,
+    _telegram_select_field_options,
+)
+
+
+def _select_form_data() -> dict:
+    return {
+        '_current_input_field': 'choice',
+        'input_defs': [
+            {
+                'output_variable_name': 'choice',
+                'type': 'select',
+                'option_source': {'type': 'constant', 'value': ['A', 'B', 'C']},
+            }
+        ],
+    }
+
+
+def test_telegram_select_field_options_are_extracted():
+    assert _telegram_select_field_options(_select_form_data()) == ('choice', ['A', 'B', 'C'])
+
+
+def test_telegram_select_callback_becomes_input_progress():
+    assert _telegram_form_action_from_callback({'f': 1, 'x': 1}) == {
+        'action_id': '',
+        'inputs': {'select': {'index': 1}},
+        '_input_progress': True,
+    }
+
+
+def test_telegram_action_callback_remains_final_action():
+    assert _telegram_form_action_from_callback({'f': 1, 'a': 'approve'}) == {
+        'action_id': 'approve',
+        'inputs': {},
+    }
+
+
+def test_telegram_invalid_select_callback_is_rejected():
+    assert _telegram_form_action_from_callback({'f': 1, 'x': -1}) is None
+    assert _telegram_form_action_from_callback({'f': 1, 'x': 'invalid'}) is None
+
+
+def test_telegram_form_callback_cache_consumes_the_whole_form_group():
+    adapter = TelegramAdapter.model_construct()
+    adapter._form_action_titles = {}
+    adapter._cache_form_action_titles({'callback-a': 'A', 'callback-b': 'B'}, now=100.0)
+
+    assert adapter._take_form_action_title('callback-a', now=101.0) == 'A'
+    assert adapter._take_form_action_title('callback-a', now=101.0) is None
+    assert adapter._take_form_action_title('callback-b', now=101.0) is None
+    assert adapter._form_action_titles == {}
+
+
+def test_telegram_form_callback_cache_prunes_expired_entries():
+    adapter = TelegramAdapter.model_construct()
+    adapter._form_action_titles = {}
+    adapter._cache_form_action_titles({'callback-a': 'A'}, now=100.0)
+
+    assert adapter._take_form_action_title('callback-a', now=100.0 + adapter._FORM_ACTION_CACHE_TTL) is None
+    assert adapter._form_action_titles == {}
+
+
+def test_telegram_form_callback_cache_preserves_pipeline_uuid():
+    adapter = TelegramAdapter.model_construct()
+    adapter._form_action_titles = {}
+    adapter._cache_form_action_titles(
+        {'callback-a': 'Approve'},
+        pipeline_uuid='pipeline-routed',
+        now=100.0,
+    )
+
+    assert adapter._take_form_action_context('callback-a', now=101.0) == (
+        'Approve',
+        'pipeline-routed',
+    )
+
+
+@pytest.mark.asyncio
+async def test_telegram_select_field_sends_two_column_inline_keyboard():
+    bot = MagicMock()
+    bot.send_message = AsyncMock()
+    adapter = TelegramAdapter.model_construct(bot=bot, config={}, msg_stream_id={}, seq=1, listeners={})
+    adapter._form_action_titles = {}
+
+    update = MagicMock()
+    update.effective_chat.id = 123
+    update.effective_message.message_thread_id = None
+    event = platform_events.FriendMessage(
+        sender=platform_entities.Friend(id='user-1', nickname='', remark=''),
+        message_chain=platform_message.MessageChain([]),
+        source_platform_object=update,
+    )
+    form_data = {
+        **_select_form_data(),
+        'node_title': 'Review',
+        'form_content': 'Choose one',
+        'workflow_run_id': 'workflow-run-12345678',
+        'actions': [{'id': 'approve', 'title': 'Approve'}],
+    }
+
+    await adapter._send_form_action_buttons(event, form_data)
+
+    args = bot.send_message.await_args.kwargs
+    rows = args['reply_markup'].inline_keyboard
+    assert [[button.text for button in row] for row in rows] == [['A', 'B'], ['C']]
+    callback_data = rows[0][1].callback_data
+    assert len(callback_data.encode('utf-8')) <= 64
+    assert json.loads(callback_data)['x'] == 1
+    assert callback_data in adapter._form_action_titles
+
+
+@pytest.mark.asyncio
+async def test_telegram_text_field_does_not_show_action_buttons():
+    bot = MagicMock()
+    bot.send_message = AsyncMock()
+    adapter = TelegramAdapter.model_construct(bot=bot, config={}, msg_stream_id={}, seq=1, listeners={})
+    adapter._form_action_titles = {}
+
+    update = MagicMock()
+    update.effective_chat.id = 123
+    update.effective_message.message_thread_id = None
+    event = platform_events.FriendMessage(
+        sender=platform_entities.Friend(id='user-1', nickname='', remark=''),
+        message_chain=platform_message.MessageChain([]),
+        source_platform_object=update,
+    )
+    form_data = {
+        '_current_input_field': 'us_input',
+        'input_defs': [{'output_variable_name': 'us_input', 'type': 'paragraph'}],
+        'node_title': '人工介入',
+        'form_content': 'us_input (paragraph): reply "us_input: <value>"',
+        'workflow_run_id': 'workflow-run-12345678',
+        'actions': [{'id': 'yes', 'title': 'yes'}, {'id': 'no', 'title': 'no'}],
+    }
+
+    await adapter._send_form_action_buttons(event, form_data)
+
+    args = bot.send_message.await_args.kwargs
+    assert isinstance(args['reply_markup'], ForceReply)
+    assert args['reply_markup'].selective is False
+    assert args['reply_markup'].input_field_placeholder == 'us_input'
+    assert 'Please reply' not in args['text']
+    assert args['text'].startswith('[人工介入]')
+    assert 'us_input (paragraph)' in args['text']
+    assert adapter._form_action_titles == {}

tests/unit_tests/platform/test_wecombot_template_card.py

@@ -0,0 +1,475 @@
+import sys
+import types
+
+import pytest
+
+
+logger_module = types.ModuleType('langbot.pkg.platform.logger')
+logger_module.EventLogger = object
+sys.modules.setdefault('langbot.pkg.platform.logger', logger_module)
+
+from langbot.libs.wecom_ai_bot_api.api import (  # noqa: E402
+    WecomBotClient,
+    build_button_interaction_payload,
+    build_human_input_template_card_payload,
+    build_button_interaction_update_card,
+    build_multiple_interaction_update_card,
+    extract_template_card_event_payload,
+    extract_template_card_selections,
+    extract_wecom_event_type,
+    extract_template_card_action,
+    build_human_input_text_prompt,
+    parse_select_button_action,
+)
+from langbot.libs.wecom_ai_bot_api.ws_client import WecomBotWsClient  # noqa: E402
+
+
+def test_extract_template_card_action_supports_nested_button_key():
+    task_id, event_key, card_type = extract_template_card_action(
+        {
+            'taskId': 'task-1',
+            'cardType': 'button_interaction',
+            'button': {'key': 'approve'},
+        }
+    )
+
+    assert task_id == 'task-1'
+    assert event_key == 'approve'
+    assert card_type == 'button_interaction'
+
+
+def test_extract_wecom_event_type_supports_top_level_template_card_event():
+    payload = {
+        'eventtype': 'template_card_event',
+        'template_card_event': {
+            'TaskId': 'task-1',
+            'CardType': 'multiple_interaction',
+            'ResponseData': '{"select_list":[{"question_key":"choice","option_id":"opt_2"}]}',
+        },
+    }
+
+    assert extract_wecom_event_type(payload) == 'template_card_event'
+    assert extract_template_card_event_payload(payload)['TaskId'] == 'task-1'
+
+
+def test_extract_wecom_event_type_infers_template_card_event_from_top_level_card_fields():
+    payload = {
+        'TaskId': 'task-1',
+        'CardType': 'button_interaction',
+        'EventKey': 'approve',
+    }
+
+    assert extract_wecom_event_type(payload) == 'template_card_event'
+    assert extract_template_card_event_payload(payload)['EventKey'] == 'approve'
+
+
+def test_build_button_interaction_update_card_marks_clicked_button():
+    card = build_button_interaction_update_card(
+        {
+            'node_title': 'Manual Review',
+            'form_content': 'Please choose one action.',
+            'actions': [
+                {'id': 'approve', 'title': 'Approve', 'button_style': 'primary'},
+                {'id': 'reject', 'title': 'Reject', 'button_style': 'danger'},
+            ],
+        },
+        task_id='task-1',
+        action_id='reject',
+        source={'desc': 'LangBot'},
+    )
+
+    assert card['main_title'] == {'title': 'Manual Review'}
+    assert card['sub_title_text'] == 'Please choose one action.'
+    assert card['button_list'][0] == {'text': 'Approve', 'style': 2, 'key': 'approve'}
+    assert card['button_list'][1] == {
+        'text': '✅ Reject',
+        'style': 1,
+        'key': 'reject',
+        'replace_text': '✅ Reject',
+    }
+    assert card['source'] == {'desc': 'LangBot'}
+
+
+def test_build_button_interaction_payload_uses_preselected_button_styles_before_click():
+    payload = build_button_interaction_payload(
+        {
+            'node_title': 'Manual Review',
+            'actions': [
+                {'id': 'approve', 'title': 'Approve', 'button_style': 'primary'},
+                {'id': 'reject', 'title': 'Reject', 'button_style': 'danger'},
+            ],
+        },
+        task_id='task-1',
+    )
+
+    assert payload['template_card']['button_list'] == [
+        {'text': 'Approve', 'style': 2, 'key': 'approve'},
+        {'text': 'Reject', 'style': 2, 'key': 'reject'},
+    ]
+
+
+def test_build_payload_uses_multiple_interaction_for_pending_select_field():
+    payload = build_human_input_template_card_payload(
+        {
+            'node_title': 'Manual Review',
+            'form_content': 'Choose a label\n\n{{#$output.choice#}}',
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+            'inputs': {},
+            'actions': [{'id': 'yes', 'title': 'Yes'}],
+        },
+        task_id='task-1',
+        source={'desc': 'LangBot'},
+    )
+
+    card = payload['template_card']
+    assert card['card_type'] == 'multiple_interaction'
+    assert card['source'] == {'desc': 'LangBot'}
+    assert card['select_list'] == [
+        {
+            'question_key': 'choice',
+            'title': 'choice',
+            'selected_id': 'opt_1',
+            'option_list': [
+                {'id': 'opt_1', 'text': 'A'},
+                {'id': 'opt_2', 'text': 'B'},
+            ],
+        }
+    ]
+    assert card['submit_button'] == {'text': 'Submit', 'key': 'submit_human_input'}
+
+
+def test_build_payload_can_emulate_select_as_buttons_for_wecombot_ws():
+    form_data = {
+        'node_title': 'Manual Review',
+        'form_content': 'Choose a label\n\n{{#$output.choice#}}',
+        'input_defs': [
+            {
+                'output_variable_name': 'choice',
+                'type': 'select',
+                'option_source': {'type': 'constant', 'value': ['A', 'B']},
+            }
+        ],
+        'inputs': {},
+        'actions': [{'id': 'yes', 'title': 'Yes'}],
+    }
+    payload = build_human_input_template_card_payload(
+        form_data,
+        task_id='task-1',
+        select_as_buttons=True,
+    )
+
+    card = payload['template_card']
+    assert card['card_type'] == 'button_interaction'
+    assert card['button_list'][0]['text'] == 'A'
+    assert parse_select_button_action(card['button_list'][1]['key'], form_data) == {'choice': 'B'}
+
+
+def test_text_input_card_uses_current_stage_content_without_direct_reply_prompt():
+    payload = build_human_input_template_card_payload(
+        {
+            'node_title': '人工介入',
+            'form_content': '11\n请输入你的问题\n\n{{#$output.us_input#}}',
+            'raw_form_content': ('11\n请输入你的问题\n{{#$output.us_input#}}\n请选择你的答案\n{{#$output.xiala#}}'),
+            'input_defs': [
+                {
+                    'output_variable_name': 'us_input',
+                    'type': 'paragraph',
+                    'label': '请输入你的问题',
+                }
+            ],
+            'inputs': {},
+            'actions': [{'id': 'yes', 'title': 'yes'}],
+            '_current_input_field': 'us_input',
+        },
+        task_id='task-1',
+    )
+
+    card = payload['template_card']
+    assert 'desc' not in card['main_title']
+    assert card['sub_title_text'] == '11\n请输入你的问题'
+    assert card['button_list'] == []
+
+
+def test_build_human_input_text_prompt_for_current_text_field():
+    prompt = build_human_input_text_prompt(
+        {
+            'node_title': '人工介入',
+            'form_content': '11\n请输入你的问题\n{{#$output.us_input#}}',
+            'raw_form_content': ('11\n请输入你的问题\n{{#$output.us_input#}}\n请选择你的答案\n{{#$output.xiala#}}'),
+            'input_defs': [
+                {
+                    'output_variable_name': 'us_input',
+                    'type': 'paragraph',
+                    'label': '请输入你的问题',
+                }
+            ],
+            '_current_input_field': 'us_input',
+        }
+    )
+
+    assert prompt == '人工介入\n\n11\n请输入你的问题'
+
+
+@pytest.mark.asyncio
+async def test_ws_push_form_pause_sends_text_prompt_without_empty_card():
+    client = WecomBotWsClient('bot-id', 'secret', object())
+    client._stream_ids['msg-1'] = 'req-1|stream-1'
+    client._stream_sessions['msg-1'] = {'user_id': 'user-1'}
+    sent = []
+
+    async def fake_reply_text(req_id, content):
+        sent.append((req_id, content))
+        return {}
+
+    client.reply_text = fake_reply_text
+
+    ok, stream_id, task_id = await client.push_form_pause(
+        'msg-1',
+        {
+            'node_title': '人工介入',
+            'form_content': '11\n请输入你的问题\n{{#$output.us_input#}}',
+            'raw_form_content': ('11\n请输入你的问题\n{{#$output.us_input#}}\n请选择你的答案\n{{#$output.xiala#}}'),
+            'input_defs': [
+                {
+                    'output_variable_name': 'us_input',
+                    'type': 'paragraph',
+                    'label': '请输入你的问题',
+                }
+            ],
+            '_current_input_field': 'us_input',
+        },
+    )
+
+    assert ok is True
+    assert stream_id == 'stream-1'
+    assert task_id is None
+    assert sent == [('req-1', '人工介入\n\n11\n请输入你的问题')]
+    assert client._pending_forms_by_task == {}
+    assert 'msg-1' not in client._stream_ids
+
+
+@pytest.mark.asyncio
+async def test_ws_stream_sends_cumulative_snapshots_to_wecom():
+    client = WecomBotWsClient('bot-id', 'secret', object())
+    client._stream_ids['msg-1'] = 'req-1|stream-1'
+    client._stream_sessions['msg-1'] = {}
+    sent = []
+
+    async def fake_reply_stream(req_id, stream_id, content, finish=False, feedback_id=''):
+        sent.append((req_id, stream_id, content, finish))
+        return {}
+
+    client.reply_stream = fake_reply_stream
+
+    assert await client.push_stream_chunk('msg-1', '你', is_final=False)
+    assert await client.push_stream_chunk('msg-1', '你好', is_final=False)
+    assert await client.push_stream_chunk('msg-1', '你好', is_final=True)
+
+    assert sent == [
+        ('req-1', 'stream-1', '你', False),
+        ('req-1', 'stream-1', '你好', False),
+        ('req-1', 'stream-1', '你好', True),
+    ]
+
+
+@pytest.mark.asyncio
+async def test_webhook_stream_queues_cumulative_snapshots_for_followups():
+    client = WecomBotClient('', '', '', object(), unified_mode=True)
+    session, _ = client.stream_sessions.create_or_get({'msgid': 'msg-1', 'chatid': '', 'from': {'userid': 'user-1'}})
+
+    assert await client.push_stream_chunk('msg-1', '你', is_final=False)
+    assert await client.push_stream_chunk('msg-1', '你好', is_final=False)
+    assert await client.push_stream_chunk('msg-1', '你好', is_final=True)
+
+    chunks = [
+        await client.stream_sessions.consume(session.stream_id),
+        await client.stream_sessions.consume(session.stream_id),
+        await client.stream_sessions.consume(session.stream_id),
+    ]
+    assert [(chunk.content, chunk.is_final) for chunk in chunks] == [
+        ('你', False),
+        ('你好', False),
+        ('你好', True),
+    ]
+
+
+def test_human_input_payload_keeps_action_select_stage_as_buttons():
+    payload = build_human_input_template_card_payload(
+        {
+            'node_title': 'Manual Review',
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+            'inputs': {'choice': 'B'},
+            'actions': [
+                {'id': 'approve', 'title': 'Approve'},
+                {'id': 'reject', 'title': 'Reject'},
+            ],
+            '_action_select_only': True,
+        },
+        task_id='task-1',
+    )
+
+    card = payload['template_card']
+    assert card['card_type'] == 'button_interaction'
+    assert card['button_list'] == [
+        {'text': 'Approve', 'style': 2, 'key': 'approve'},
+        {'text': 'Reject', 'style': 2, 'key': 'reject'},
+    ]
+
+
+def test_extract_template_card_selections_maps_selected_id_to_option_text():
+    selections = extract_template_card_selections(
+        {
+            'SelectedItems': [
+                {'QuestionKey': 'choice', 'SelectedId': 'opt_2'},
+            ],
+        },
+        {
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+        },
+    )
+
+    assert selections == {'choice': 'B'}
+
+
+def test_extract_template_card_selections_reads_nested_response_data_json():
+    selections = extract_template_card_selections(
+        {
+            'CardType': 'multiple_interaction',
+            'ResponseData': '{"select_list":[{"question_key":"choice","option_id":"opt_2"}]}',
+        },
+        {
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+        },
+    )
+
+    assert selections == {'choice': 'B'}
+
+
+def test_extract_template_card_selections_reads_response_data_direct_mapping():
+    selections = extract_template_card_selections(
+        {
+            'CardType': 'multiple_interaction',
+            'EventKey': 'submit_human_input',
+            'ResponseData': '{"choice":"opt_2"}',
+        },
+        {
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+        },
+    )
+
+    assert selections == {'choice': 'B'}
+
+
+def test_build_multiple_interaction_update_card_disables_selected_value_without_submitted_text():
+    card = build_multiple_interaction_update_card(
+        {
+            'node_title': 'Manual Review',
+            'form_content': 'Choose a label\n{{#$output.choice#}}',
+            'raw_form_content': 'Choose a label\n{{#$output.choice#}}',
+            'input_defs': [
+                {
+                    'output_variable_name': 'choice',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['A', 'B']},
+                }
+            ],
+            '_current_input_field': 'choice',
+        },
+        task_id='task-1',
+        selections={'choice': 'B'},
+    )
+
+    assert card['card_type'] == 'multiple_interaction'
+    assert card['main_title']['desc'] == 'Choose a label\n✅ choice：B'
+    assert card['submit_button']['text'] == '✅'
+    assert card['select_list'][0]['disable'] is True
+    assert card['select_list'][0]['selected_id'] == 'opt_2'
+
+
+def test_select_stage_only_shows_current_prompt_in_a_separate_message():
+    raw_content = '11\n请输入你的问题\n{{#$output.us_input#}}\n请选择你的答案\n{{#$output.xiala#}}'
+    payload = build_human_input_template_card_payload(
+        {
+            'node_title': '人工介入',
+            'form_content': '请选择你的答案\n{{#$output.xiala#}}',
+            'raw_form_content': raw_content,
+            'input_defs': [
+                {
+                    'output_variable_name': 'xiala',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['1', '2']},
+                }
+            ],
+            'all_input_defs': [
+                {'output_variable_name': 'us_input', 'type': 'paragraph'},
+                {
+                    'output_variable_name': 'xiala',
+                    'type': 'select',
+                    'option_source': {'type': 'constant', 'value': ['1', '2']},
+                },
+            ],
+            'inputs': {'us_input': '你叫啥'},
+            '_current_input_field': 'xiala',
+        },
+        task_id='task-1',
+    )
+
+    assert payload['template_card']['main_title']['desc'] == '请选择你的答案'
+
+
+def test_action_stage_only_shows_content_after_fields_without_placeholders():
+    raw_content = '11\n请输入你的问题\n{{#$output.us_input#}}\n请选择你的答案\n{{#$output.xiala#}}\n请选择操作'
+    payload = build_human_input_template_card_payload(
+        {
+            'node_title': '人工介入',
+            'form_content': raw_content,
+            'raw_form_content': raw_content,
+            'input_defs': [],
+            'all_input_defs': [
+                {'output_variable_name': 'us_input', 'type': 'paragraph'},
+                {'output_variable_name': 'xiala', 'type': 'select'},
+            ],
+            'inputs': {'us_input': '你叫啥', 'xiala': '2'},
+            'actions': [
+                {'id': 'yes', 'title': 'yes'},
+                {'id': 'no', 'title': 'no'},
+            ],
+            '_action_select_only': True,
+        },
+        task_id='task-1',
+    )
+
+    card = payload['template_card']
+    assert card['sub_title_text'] == '请选择操作'
+    assert '{{#$output.' not in card['sub_title_text']
+    assert [button['text'] for button in card['button_list']] == ['yes', 'no']

tests/unit_tests/provider/test_mcp_box_integration.py

@@ -417,7 +417,7 @@ class TestBuildBoxSessionPayload:
         payload = s._build_box_session_payload('session-123')
         assert payload['image'] == 'node:20'
         assert payload['cpus'] == 2.0
-        assert payload['memory_mb'] == 1024
+        assert payload["memory_mb"] == 1024
         assert payload['pids_limit'] == 256
 
     def test_none_fields_excluded(self, mcp_module):
@@ -639,10 +639,13 @@ class TestGetRuntimeInfoDict:
         assert info['box_session_id'] == 'mcp-shared'
         assert info['box_enabled'] is True
 
-    def test_transient_test_session_is_isolated_from_shared(self, mcp_module):
+    def test_transient_test_shares_session_but_isolated_by_process(self, mcp_module):
-        """A transient test session (config-page "test", no persisted UUID)
+        """A transient config-page "test" now shares the same 'mcp-shared' Box
-        must NOT share the live "mcp-shared" Box session. Regression: a failing
+        session as live servers (so a test reuses the running container / live
-        test churned the shared session and tore down healthy live servers."""
+        process instead of a cold per-test session bootstrap). Isolation is at
+        the PROCESS level: the test runs under its own process_id and only ever
+        stops that process_id, so it cannot disturb another server's live
+        process or the shared session itself."""
         ap = _make_ap()
         ap.box_service.available = True
         transient = _make_session(
@@ -670,10 +673,12 @@ class TestGetRuntimeInfoDict:
         )
         assert transient.is_transient is True
         assert live.is_transient is False
-        # Isolated session id for the test, shared for the live server.
+        # Both share ONE Box session ...
-        assert transient._build_box_session_id() == 'mcp-test-gen-uuid-123'
+        assert transient._build_box_session_id() == 'mcp-shared'
         assert live._build_box_session_id() == 'mcp-shared'
-        assert transient._build_box_session_id() != live._build_box_session_id()
+        assert transient._build_box_session_id() == live._build_box_session_id()
+        # ... but are isolated by distinct process_ids within that session.
+        assert transient._box_stdio_runtime.process_id != live._box_stdio_runtime.process_id
 
     def test_stdio_session_refuses_when_box_unavailable(self, mcp_module):
         """Policy: when Box is configured but unavailable (disabled in config
@@ -824,3 +829,129 @@ async def test_init_box_stdio_server_stages_host_path_in_shared_workspace(mcp_mo
     assert process_payload['command'] == 'python'
     assert process_payload['args'] == ['/workspace/.mcp/u1/workspace/server.py']
     assert process_payload['cwd'] == '/workspace/.mcp/u1/workspace'
+
+
+@pytest.mark.asyncio
+async def test_stdio_handshake_raises_coldstart_retry_while_process_alive(mcp_module, tmp_path, monkeypatch):
+    """During a slow (npx) cold start the handshake fails while the managed
+    process is still alive. initialize() must raise _ColdStartRetry (so the
+    outer lifecycle loop reuses the live process and retries without stopping it
+    or consuming the fatal budget), NOT a fatal error."""
+    from contextlib import asynccontextmanager
+
+    mcp_stdio_module = sys.modules['langbot.pkg.provider.tools.loaders.mcp_stdio']
+
+    class ColdClientSession:
+        def __init__(self, *_args):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, exc_type, exc, tb):
+            return False
+
+        async def initialize(self):
+            # Process still cold-starting: handshake fails.
+            raise Exception('Connection closed')
+
+    @asynccontextmanager
+    async def fake_websocket_client(_url: str):
+        yield ('read-stream', 'write-stream')
+
+    monkeypatch.setattr(mcp_stdio_module, 'ClientSession', ColdClientSession)
+    monkeypatch.setattr(mcp_stdio_module, 'websocket_client', fake_websocket_client)
+    monkeypatch.setattr(mcp_stdio_module, '_HANDSHAKE_ATTEMPT_TIMEOUT_SEC', 1.0, raising=False)
+
+    ap = _make_ap()
+    ap.box_service.available = True
+    ap.box_service.create_session = AsyncMock(return_value={})
+    ap.box_service.start_managed_process = AsyncMock(return_value={})
+    ap.box_service.get_managed_process_websocket_url = Mock(return_value='ws://box/p')
+
+    session = _make_session(
+        mcp_module,
+        {
+            'name': 'slow',
+            'uuid': 'slow-uuid',
+            'mode': 'stdio',
+            'command': 'npx',
+            'args': ['-y', 'some-mcp'],
+        },
+        ap=ap,
+    )
+
+    # Process is NOT exited (still cold-starting) and not yet running for reuse.
+    async def _not_exited():
+        return False
+
+    session._box_stdio_runtime._managed_process_has_exited = _not_exited
+
+    async def _not_running():
+        return False
+
+    session._box_stdio_runtime._managed_process_is_running = _not_running
+
+    with pytest.raises(mcp_stdio_module._ColdStartRetry):
+        await session._init_box_stdio_server()
+
+    # Process was started exactly once (the retry will reuse it, not rebuild).
+    assert ap.box_service.start_managed_process.await_count == 1
+    await session.exit_stack.aclose()
+
+
+@pytest.mark.asyncio
+async def test_stdio_handshake_raises_fatal_when_process_exited(mcp_module, tmp_path, monkeypatch):
+    """If the handshake fails AND the process has definitively exited, that is a
+    real failure — initialize() must NOT swallow it as a cold-start retry."""
+    from contextlib import asynccontextmanager
+
+    mcp_stdio_module = sys.modules['langbot.pkg.provider.tools.loaders.mcp_stdio']
+
+    class DeadClientSession:
+        def __init__(self, *_args):
+            pass
+
+        async def __aenter__(self):
+            return self
+
+        async def __aexit__(self, exc_type, exc, tb):
+            return False
+
+        async def initialize(self):
+            raise Exception('Connection closed')
+
+    @asynccontextmanager
+    async def fake_websocket_client(_url: str):
+        yield ('read-stream', 'write-stream')
+
+    monkeypatch.setattr(mcp_stdio_module, 'ClientSession', DeadClientSession)
+    monkeypatch.setattr(mcp_stdio_module, 'websocket_client', fake_websocket_client)
+    monkeypatch.setattr(mcp_stdio_module, '_HANDSHAKE_ATTEMPT_TIMEOUT_SEC', 1.0, raising=False)
+
+    ap = _make_ap()
+    ap.box_service.available = True
+    ap.box_service.create_session = AsyncMock(return_value={})
+    ap.box_service.start_managed_process = AsyncMock(return_value={})
+    ap.box_service.get_managed_process_websocket_url = Mock(return_value='ws://box/p')
+
+    session = _make_session(
+        mcp_module,
+        {'name': 'dead', 'uuid': 'dead-uuid', 'mode': 'stdio', 'command': 'npx', 'args': ['-y', 'x']},
+        ap=ap,
+    )
+
+    async def _exited():
+        return True
+
+    session._box_stdio_runtime._managed_process_has_exited = _exited
+
+    async def _not_running():
+        return False
+
+    session._box_stdio_runtime._managed_process_is_running = _not_running
+
+    with pytest.raises(Exception) as ei:
+        await session._init_box_stdio_server()
+    assert not isinstance(ei.value, mcp_stdio_module._ColdStartRetry)
+    await session.exit_stack.aclose()

tests/unit_tests/provider/test_mcp_remote_transport.py

@@ -0,0 +1,244 @@
+from __future__ import annotations
+
+import asyncio
+import json
+from contextlib import asynccontextmanager
+from types import SimpleNamespace
+from typing import Any, cast
+from unittest.mock import AsyncMock, Mock
+
+import httpx
+import pytest
+from aiohttp import web
+from mcp import types as mcp_types
+
+from langbot.pkg.provider.tools.loaders.mcp import RuntimeMCPSession
+
+
+class _TransportProbe:
+    def __init__(self, streamable_status: int | None) -> None:
+        self.streamable_status = streamable_status
+        self.streamable_posts = 0
+        self.streamable_messages: list[str] = []
+        self.sse_gets = 0
+        self.sse_messages: list[str] = []
+        self.streamable_request_started = asyncio.Event()
+        self.release_streamable_request = asyncio.Event()
+        self._sse_response: web.StreamResponse | None = None
+
+    async def handle_mcp_endpoint(self, request: web.Request) -> web.StreamResponse:
+        if request.method == 'POST':
+            self.streamable_posts += 1
+            self.streamable_request_started.set()
+            if self.streamable_status is None:
+                await self.release_streamable_request.wait()
+                return web.Response(status=204)
+            if self.streamable_status == 200:
+                message = await request.json()
+                method = message.get('method', '')
+                self.streamable_messages.append(method)
+                if method == 'initialize':
+                    return web.json_response(
+                        {
+                            'jsonrpc': '2.0',
+                            'id': message['id'],
+                            'result': {
+                                'protocolVersion': mcp_types.LATEST_PROTOCOL_VERSION,
+                                'capabilities': {'tools': {}},
+                                'serverInfo': {'name': 'streamable-test', 'version': '1.0.0'},
+                            },
+                        }
+                    )
+                if method == 'tools/list':
+                    return web.json_response(
+                        {
+                            'jsonrpc': '2.0',
+                            'id': message['id'],
+                            'result': {
+                                'tools': [
+                                    {
+                                        'name': 'echo',
+                                        'description': 'Echo test input',
+                                        'inputSchema': {'type': 'object'},
+                                    }
+                                ]
+                            },
+                        }
+                    )
+                return web.Response(status=202)
+            return web.Response(status=self.streamable_status)
+
+        self.sse_gets += 1
+        response = web.StreamResponse(
+            status=200,
+            headers={
+                'Content-Type': 'text/event-stream',
+                'Cache-Control': 'no-cache',
+            },
+        )
+        await response.prepare(request)
+        self._sse_response = response
+        await response.write(b'event: endpoint\ndata: /messages?session_id=test-session\n\n')
+        try:
+            while request.transport is not None and not request.transport.is_closing():
+                await asyncio.sleep(0.05)
+        except asyncio.CancelledError:
+            raise
+        return response
+
+    async def handle_sse_message(self, request: web.Request) -> web.Response:
+        message = await request.json()
+        method = message.get('method', '')
+        self.sse_messages.append(method)
+
+        if method == 'initialize':
+            response_message = {
+                'jsonrpc': '2.0',
+                'id': message['id'],
+                'result': {
+                    'protocolVersion': mcp_types.LATEST_PROTOCOL_VERSION,
+                    'capabilities': {},
+                    'serverInfo': {'name': 'legacy-sse-test', 'version': '1.0.0'},
+                },
+            }
+            assert self._sse_response is not None
+            payload = json.dumps(response_message, separators=(',', ':'))
+            await self._sse_response.write(f'event: message\ndata: {payload}\n\n'.encode())
+
+        return web.Response(status=202)
+
+
+@asynccontextmanager
+async def _transport_server(streamable_status: int | None):
+    probe = _TransportProbe(streamable_status)
+    application = web.Application()
+    application.router.add_route('*', '/mcp', probe.handle_mcp_endpoint)
+    application.router.add_post('/messages', probe.handle_sse_message)
+    runner = web.AppRunner(application, shutdown_timeout=0.1)
+    await runner.setup()
+    site = web.TCPSite(runner, '127.0.0.1', 0)
+    await site.start()
+    server = cast(asyncio.Server, site._server)
+    port = server.sockets[0].getsockname()[1]
+    try:
+        yield probe, f'http://127.0.0.1:{port}/mcp'
+    finally:
+        await runner.cleanup()
+
+
+def _session(url: str, *, timeout: float = 2) -> RuntimeMCPSession:
+    app = cast(Any, SimpleNamespace(logger=Mock()))
+    return RuntimeMCPSession(
+        'remote-transport-test',
+        {'uuid': 'srv-1', 'mode': 'remote', 'url': url, 'timeout': timeout},
+        True,
+        app,
+    )
+
+
+def _contains_http_status(exc: BaseException, status_code: int) -> bool:
+    return any(
+        isinstance(leaf, httpx.HTTPStatusError) and leaf.response.status_code == status_code
+        for leaf in RuntimeMCPSession._iter_exception_leaves(exc)
+    )
+
+
+async def _close_session(session: RuntimeMCPSession) -> None:
+    await session.exit_stack.aclose()
+
+
+@pytest.mark.asyncio
+async def test_remote_transport_real_streamable_http_success_keeps_session_usable():
+    async with _transport_server(200) as (probe, url):
+        session = _session(url)
+        try:
+            await session._init_remote_server()
+            assert session.session is not None
+            tools = await session.session.list_tools()
+            assert [tool.name for tool in tools.tools] == ['echo']
+            assert probe.streamable_posts >= 2
+            assert probe.streamable_messages[:2] == ['initialize', 'notifications/initialized']
+            assert 'tools/list' in probe.streamable_messages
+            assert probe.sse_gets == 0
+        finally:
+            await _close_session(session)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize('status_code', [400, 404, 405])
+async def test_remote_transport_real_streamable_http_error_falls_back_to_legacy_sse(status_code: int):
+    async with _transport_server(status_code) as (probe, url):
+        session = _session(url)
+        try:
+            await session._init_remote_server()
+            assert session.session is not None
+            assert probe.streamable_posts == 1
+            assert probe.sse_gets == 1
+            assert 'initialize' in probe.sse_messages
+        finally:
+            await _close_session(session)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize('status_code', [401, 403, 406, 415, 429, 500])
+async def test_remote_transport_real_non_compatibility_error_does_not_fallback(status_code: int):
+    async with _transport_server(status_code) as (probe, url):
+        session = _session(url)
+        try:
+            with pytest.raises(BaseException) as exc_info:
+                await session._init_remote_server()
+            assert _contains_http_status(exc_info.value, status_code)
+            assert probe.streamable_posts == 1
+            assert probe.sse_gets == 0
+        finally:
+            await _close_session(session)
+
+
+@pytest.mark.asyncio
+async def test_remote_transport_real_timeout_does_not_fallback():
+    async with _transport_server(None) as (probe, url):
+        session = _session(url, timeout=0.05)
+        try:
+            with pytest.raises(BaseException) as exc_info:
+                await session._init_remote_server()
+            assert any(
+                isinstance(leaf, httpx.TimeoutException)
+                for leaf in RuntimeMCPSession._iter_exception_leaves(exc_info.value)
+            )
+            assert probe.streamable_posts == 1
+            assert probe.sse_gets == 0
+        finally:
+            probe.release_streamable_request.set()
+            await _close_session(session)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize('error_type', [httpx.ConnectError, httpx.ConnectTimeout])
+async def test_remote_transport_connection_errors_do_not_fallback(error_type: type[httpx.RequestError]):
+    request = httpx.Request('POST', 'https://unreachable.invalid/mcp')
+    error = error_type('connection failed', request=request)
+    session = _session(str(request.url))
+    session._init_streamable_http_server = AsyncMock(side_effect=error)
+    session._init_sse_server = AsyncMock()
+
+    with pytest.raises(type(error)) as exc_info:
+        await session._init_remote_server()
+
+    assert exc_info.value is error
+    session._init_sse_server.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_remote_transport_external_cancellation_is_not_converted_to_sse_fallback():
+    async with _transport_server(None) as (probe, url):
+        session = _session(url)
+        task = asyncio.create_task(session._init_remote_server())
+        await asyncio.wait_for(probe.streamable_request_started.wait(), timeout=2)
+        task.cancel()
+        try:
+            with pytest.raises(asyncio.CancelledError):
+                await task
+            assert probe.sse_gets == 0
+        finally:
+            probe.release_streamable_request.set()
+            await _close_session(session)

tests/unit_tests/provider/test_mcp_resources.py

@@ -4,6 +4,7 @@ import base64
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, Mock
 
+import httpx
 import pytest
 from mcp import types as mcp_types
 
@@ -54,6 +55,50 @@ def _query() -> SimpleNamespace:
     return SimpleNamespace(variables={})
 
 
+def _http_status_error(status_code: int) -> httpx.HTTPStatusError:
+    request = httpx.Request('POST', 'https://example.com/mcp')
+    response = httpx.Response(status_code, request=request)
+    return httpx.HTTPStatusError(f'HTTP {status_code}', request=request, response=response)
+
+
+@pytest.mark.asyncio
+async def test_remote_transport_falls_back_to_sse_for_compatible_http_status_in_exception_group():
+    session = RuntimeMCPSession(
+        'remote',
+        {'uuid': 'srv-1', 'mode': 'remote', 'url': 'https://example.com/mcp'},
+        True,
+        _app(),
+    )
+    session._init_streamable_http_server = AsyncMock(
+        side_effect=ExceptionGroup('transport failed', [_http_status_error(405)])
+    )
+    session._init_sse_server = AsyncMock()
+
+    await session._init_remote_server()
+
+    session._init_streamable_http_server.assert_awaited_once()
+    session._init_sse_server.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_remote_transport_does_not_fallback_for_auth_http_status():
+    session = RuntimeMCPSession(
+        'remote',
+        {'uuid': 'srv-1', 'mode': 'remote', 'url': 'https://example.com/mcp'},
+        True,
+        _app(),
+    )
+    error = _http_status_error(403)
+    session._init_streamable_http_server = AsyncMock(side_effect=error)
+    session._init_sse_server = AsyncMock()
+
+    with pytest.raises(httpx.HTTPStatusError):
+        await session._init_remote_server()
+
+    session._init_streamable_http_server.assert_awaited_once()
+    session._init_sse_server.assert_not_awaited()
+
+
 @pytest.mark.asyncio
 async def test_read_resource_envelope_truncates_caches_and_records_trace():
     session = _connected_session()

tests/unit_tests/vector/test_mgr.py

@@ -33,7 +33,7 @@ class TestVectorDBManagerInitialization:
         mocks['langbot.pkg.core.app'] = MagicMock()
 
         # Mock all VDB backend implementations
-        for backend in ['chroma', 'qdrant', 'seekdb', 'milvus', 'pgvector_db']:
+        for backend in ['chroma', 'qdrant', 'seekdb', 'milvus', 'pgvector_db', 'valkey_search']:
             mocks[f'langbot.pkg.vector.vdbs.{backend}'] = MagicMock()
 
         return mocks
@@ -123,6 +123,25 @@ class TestVectorDBManagerInitialization:
 
             mock_seekdb_class.assert_called_once_with(mock_app)
 
+    def test_initialize_valkey_search_backend(self):
+        """Valkey Search config uses ValkeySearchVectorDatabase backend."""
+        vdb_config = {'use': 'valkey_search'}
+        mock_app = self._create_mock_app(vdb_config)
+
+        mocks = self._make_vector_import_mocks()
+        mock_valkey_class = MagicMock()
+        mocks['langbot.pkg.vector.vdbs.valkey_search'].ValkeySearchVectorDatabase = mock_valkey_class
+
+        with isolated_sys_modules(mocks):
+            from langbot.pkg.vector.mgr import VectorDBManager
+
+            mgr = VectorDBManager(mock_app)
+
+            import asyncio
+            asyncio.get_event_loop().run_until_complete(mgr.initialize())
+
+            mock_valkey_class.assert_called_once_with(mock_app)
+
     def test_initialize_milvus_backend_with_uri(self):
         """Milvus config with custom URI."""
         vdb_config = {

tests/unit_tests/vector/test_valkey_search_filter.py

@@ -0,0 +1,388 @@
+"""Unit tests for the Valkey Search VDB backend's pure helpers.
+
+These tests exercise the filter-to-FT mapping, float32 packing, tag/text
+escaping, FT.SEARCH reply parsing and the import guard.  They run in the fast
+CI lane and require NO running Valkey server.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import struct
+from importlib import import_module
+from unittest.mock import AsyncMock
+
+import pytest
+
+
+def get_valkey_module():
+    """Lazy import of the valkey_search backend module."""
+    return import_module('langbot.pkg.vector.vdbs.valkey_search')
+
+
+def make_backend():
+    """Construct a backend instance without running its __init__.
+
+    The constructor needs a live ``ap`` + config; for pure-helper tests we
+    only need a bare instance with the attributes the helpers touch.
+    """
+    mod = get_valkey_module()
+    backend = object.__new__(mod.ValkeySearchVectorDatabase)
+    # _ensure_client serializes creation through this lock; set it here since
+    # __init__ (which normally creates it) is bypassed.
+    backend._client_lock = asyncio.Lock()
+    return backend
+
+
+class TestFloat32Packing:
+    """Tests for _pack_vector little-endian float32 packing."""
+
+    def test_pack_round_trips(self):
+        mod = get_valkey_module()
+        vec = [0.1, -2.5, 3.0, 4.25]
+        packed = mod.ValkeySearchVectorDatabase._pack_vector(vec)
+        assert isinstance(packed, bytes)
+        assert len(packed) == 4 * len(vec)
+        unpacked = list(struct.unpack(f'<{len(vec)}f', packed))
+        for original, restored in zip(vec, unpacked):
+            assert restored == pytest.approx(original, rel=1e-6)
+
+    def test_pack_is_little_endian(self):
+        mod = get_valkey_module()
+        packed = mod.ValkeySearchVectorDatabase._pack_vector([1.0])
+        assert packed == struct.pack('<f', 1.0)
+
+
+class TestTagEscaping:
+    """Tests for _escape_tag."""
+
+    def test_escapes_special_chars(self):
+        mod = get_valkey_module()
+        escaped = mod.ValkeySearchVectorDatabase._escape_tag('a-b c.d')
+        assert '\\-' in escaped
+        assert '\\ ' in escaped
+        assert '\\.' in escaped
+
+    def test_plain_value_unchanged(self):
+        mod = get_valkey_module()
+        assert mod.ValkeySearchVectorDatabase._escape_tag('abc123') == 'abc123'
+
+
+class TestFileIdEncoding:
+    """Tests for _encode_file_id (FT-unsafe char percent-encoding)."""
+
+    def test_uuid_is_noop(self):
+        mod = get_valkey_module()
+        fid = '550e8400-e29b-41d4-a716-446655440000'
+        assert mod.ValkeySearchVectorDatabase._encode_file_id(fid) == fid
+
+    def test_encodes_braces_star_and_percent(self):
+        mod = get_valkey_module()
+        enc = mod.ValkeySearchVectorDatabase._encode_file_id('a{b}c*d%e')
+        # '{'=7B '}'=7D '*'=2A '%'=25
+        assert enc == 'a%7Bb%7Dc%2Ad%25e'
+        # No raw FT-unsafe char survives.
+        assert all(ch not in enc for ch in '{}*') or '%' in enc
+
+    def test_encoding_is_deterministic_and_collision_safe(self):
+        mod = get_valkey_module()
+        enc = mod.ValkeySearchVectorDatabase._encode_file_id
+        # A literal "%7B" must not collide with an encoded "{".
+        assert enc('{') != enc('%7B')
+        assert enc('{') == '%7B'
+        assert enc('%7B') == '%257B'
+
+    def test_filter_encodes_unsafe_chars_in_tag_query(self):
+        backend = make_backend()
+        # The emitted TAG query must contain the encoded form, never raw braces.
+        frag = backend._triples_to_ft({'file_id': 'x}y{z*'})
+        assert '7D' in frag and '7B' in frag and '2A' in frag
+        # No raw '*' from the value, and exactly one opening/closing brace (the
+        # TAG-clause delimiters) — the value's own braces were encoded away.
+        assert '*' not in frag
+        assert frag.count('{') == 1 and frag.count('}') == 1
+        assert frag.startswith('@file_id:{') and frag.endswith('}')
+
+    def test_filter_in_operator_encodes_each_value(self):
+        backend = make_backend()
+        frag = backend._triples_to_ft({'file_id': {'$in': ['a*b', 'c}d']}})
+        assert '2A' in frag and '7D' in frag
+        assert '*' not in frag
+
+
+class TestFilterToFt:
+    """Tests for _triples_to_ft filter mapping (all 8 operators)."""
+
+    def test_empty_filter_returns_empty_string(self):
+        backend = make_backend()
+        assert backend._triples_to_ft(None) == ''
+        assert backend._triples_to_ft({}) == ''
+
+    def test_eq_tag(self):
+        backend = make_backend()
+        assert backend._triples_to_ft({'file_id': 'abc'}) == '@file_id:{abc}'
+
+    def test_explicit_eq_tag(self):
+        backend = make_backend()
+        assert backend._triples_to_ft({'file_id': {'$eq': 'abc'}}) == '@file_id:{abc}'
+
+    def test_ne_tag(self):
+        backend = make_backend()
+        assert backend._triples_to_ft({'file_id': {'$ne': 'abc'}}) == '-@file_id:{abc}'
+
+    def test_in_tag(self):
+        backend = make_backend()
+        assert backend._triples_to_ft({'file_id': {'$in': ['a', 'b']}}) == '@file_id:{a|b}'
+
+    def test_nin_tag(self):
+        backend = make_backend()
+        assert backend._triples_to_ft({'file_id': {'$nin': ['a', 'b']}}) == '-@file_id:{a|b}'
+
+    def test_numeric_range_operators(self):
+        backend = make_backend()
+        # file_id is the only indexed field; numeric ops still render via the
+        # generic range fragment, so use file_id to keep the field supported.
+        # Values are cast to float (defensive against non-numeric input and a
+        # future NUMERIC field becoming an injection surface).
+        assert backend._triples_to_ft({'file_id': {'$gt': 5}}) == '@file_id:[(5.0 +inf]'
+        assert backend._triples_to_ft({'file_id': {'$gte': 5}}) == '@file_id:[5.0 +inf]'
+        assert backend._triples_to_ft({'file_id': {'$lt': 5}}) == '@file_id:[-inf (5.0]'
+        assert backend._triples_to_ft({'file_id': {'$lte': 5}}) == '@file_id:[-inf 5.0]'
+
+    def test_numeric_range_rejects_non_numeric(self):
+        backend = make_backend()
+        # A non-numeric range value fails closed rather than interpolating raw.
+        with pytest.raises((ValueError, TypeError)):
+            backend._triples_to_ft({'file_id': {'$gt': 'not-a-number'}})
+
+    def test_unsupported_field_dropped(self):
+        backend = make_backend()
+        # Non-indexed fields are dropped (returns empty expression).
+        assert backend._triples_to_ft({'some_other_field': 'x'}) == ''
+
+    def test_multiple_supported_keys_anded(self):
+        backend = make_backend()
+        # Two conditions on the same indexed field are joined with a space (AND).
+        result = backend._triples_to_ft({'file_id': {'$in': ['a', 'b']}})
+        assert result == '@file_id:{a|b}'
+
+
+class TestTextEscaping:
+    """Tests for _escape_text full-text escaping."""
+
+    def test_escapes_ft_special_chars(self):
+        mod = get_valkey_module()
+        escaped = mod.ValkeySearchVectorDatabase._escape_text('hello@world|test')
+        assert '\\@' in escaped
+        assert '\\|' in escaped
+
+
+class TestReplyToChroma:
+    """Tests for _reply_to_chroma FT.SEARCH reply parsing."""
+
+    def test_parses_knn_reply(self):
+        backend = make_backend()
+        # glide returns [total, {key: {field: value}}]
+        reply = [
+            2,
+            {
+                b'kb:col1:id1': {
+                    b'distance': b'0.10',
+                    b'document': b'hello',
+                    b'metadata_json': b'{"file_id": "f1"}',
+                },
+                b'kb:col1:id2': {
+                    b'distance': b'0.25',
+                    b'document': b'world',
+                    b'metadata_json': b'{"file_id": "f2"}',
+                },
+            },
+        ]
+        result = backend._reply_to_chroma('idx:col1', reply, has_distance=True)
+        assert result['ids'][0] == ['id1', 'id2']
+        assert result['distances'][0] == [pytest.approx(0.10), pytest.approx(0.25)]
+        assert result['metadatas'][0][0] == {'file_id': 'f1'}
+        assert result['metadatas'][0][1] == {'file_id': 'f2'}
+
+    def test_empty_reply(self):
+        backend = make_backend()
+        result = backend._reply_to_chroma('idx:col1', [0, {}], has_distance=True)
+        assert result == {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+
+    def test_malformed_reply(self):
+        backend = make_backend()
+        result = backend._reply_to_chroma('idx:col1', [], has_distance=True)
+        assert result == {'ids': [[]], 'metadatas': [[]], 'distances': [[]]}
+
+    def test_text_search_reply_no_distance(self):
+        backend = make_backend()
+        reply = [
+            1,
+            {
+                b'kb:col1:id1': {
+                    b'document': b'hello',
+                    b'metadata_json': b'{"file_id": "f1"}',
+                },
+            },
+        ]
+        result = backend._reply_to_chroma('idx:col1', reply, has_distance=False)
+        assert result['ids'][0] == ['id1']
+        assert result['distances'][0] == [0.0]
+
+
+class TestImportGuard:
+    """Tests for the ImportError guard when glide is unavailable."""
+
+    def test_constructor_raises_when_unavailable(self, monkeypatch):
+        mod = get_valkey_module()
+        monkeypatch.setattr(mod, 'VALKEY_SEARCH_AVAILABLE', False)
+        with pytest.raises(ImportError, match='valkey-glide'):
+            mod.ValkeySearchVectorDatabase(ap=None)
+
+
+class TestSupportedSearchTypes:
+    """Tests for supported_search_types."""
+
+    def test_supports_vector_full_text_hybrid(self):
+        mod = get_valkey_module()
+        from langbot.pkg.vector.vdb import SearchType
+
+        types = mod.ValkeySearchVectorDatabase.supported_search_types()
+        assert SearchType.VECTOR in types
+        assert SearchType.FULL_TEXT in types
+        assert SearchType.HYBRID in types
+
+
+class TestDeleteByFilterGuard:
+    """Regression tests for the delete_by_filter mass-deletion guard.
+
+    A non-empty filter referencing only non-indexed fields must NOT fall back
+    to match-all and wipe the whole collection: it must skip and return 0.
+    """
+
+    async def test_unsupported_only_filter_skips_and_returns_zero(self):
+        backend = make_backend()
+        # Make the client/index lookups succeed without a real server.
+        backend._client = AsyncMock()
+        backend.ap = type('Ap', (), {'logger': AsyncMock()})()
+        backend._ensure_client = AsyncMock(return_value=backend._client)
+        backend._index_exists = AsyncMock(return_value=True)
+        # _search_keys must never be reached for an unusable filter.
+        backend._search_keys = AsyncMock(
+            side_effect=AssertionError('_search_keys must not be called for an unusable filter')
+        )
+
+        # Filter references only a non-indexed field -> maps to no FT conditions.
+        deleted = await backend.delete_by_filter('col1', {'some_other_field': 'x'})
+
+        assert deleted == 0
+        backend._client.delete.assert_not_called()
+
+    async def test_supported_filter_deletes_matching_keys(self):
+        backend = make_backend()
+        backend._client = AsyncMock()
+        backend.ap = type('Ap', (), {'logger': AsyncMock()})()
+        backend._ensure_client = AsyncMock(return_value=backend._client)
+        backend._index_exists = AsyncMock(return_value=True)
+        backend._search_keys = AsyncMock(return_value=['kb:col1:id1', 'kb:col1:id2'])
+
+        deleted = await backend.delete_by_filter('col1', {'file_id': 'f1'})
+
+        assert deleted == 2
+        backend._client.delete.assert_awaited_once_with(['kb:col1:id1', 'kb:col1:id2'])
+
+
+class TestClose:
+    """Tests for the close() teardown."""
+
+    async def test_close_resets_client_and_indexes(self):
+        backend = make_backend()
+        client = AsyncMock()
+        backend._client = client
+        backend.ap = type('Ap', (), {'logger': AsyncMock()})()
+        backend._ensured_indexes = {'idx:col1'}
+
+        await backend.close()
+
+        client.close.assert_awaited_once()
+        assert backend._client is None
+        assert backend._ensured_indexes == set()
+
+    async def test_close_is_noop_when_no_client(self):
+        backend = make_backend()
+        backend._client = None
+        backend.ap = type('Ap', (), {'logger': AsyncMock()})()
+        backend._ensured_indexes = set()
+        # Should not raise.
+        await backend.close()
+        assert backend._client is None
+
+
+class TestCredentialsBuild:
+    """Tests for the auth-credential construction in _ensure_client."""
+
+    def _prep_backend(self, mod, monkeypatch, *, username, password):
+        backend = make_backend()
+        backend._client = None
+        backend._host = 'localhost'
+        backend._port = 6379
+        backend._db = 0
+        backend._tls = False
+        backend._username = username
+        backend._password = password
+        backend._request_timeout = 5000
+        backend._ensured_indexes = set()
+        warnings: list[str] = []
+        backend.ap = type(
+            'Ap',
+            (),
+            {
+                'logger': type(
+                    'L', (), {'info': lambda self, *a, **k: None, 'warning': lambda s, m, *a, **k: warnings.append(m)}
+                )()
+            },
+        )()
+
+        created = {}
+
+        class _FakeClient:
+            @staticmethod
+            async def create(conf):
+                created['conf'] = conf
+                return AsyncMock()
+
+        cred_calls: list[dict] = []
+
+        def _fake_credentials(**kwargs):
+            cred_calls.append(kwargs)
+            return ('CRED', kwargs)
+
+        monkeypatch.setattr(mod, 'GlideClient', _FakeClient)
+        monkeypatch.setattr(mod, 'ServerCredentials', _fake_credentials)
+        monkeypatch.setattr(mod, 'GlideClientConfiguration', lambda **kw: kw)
+        monkeypatch.setattr(mod, 'NodeAddress', lambda *a, **k: ('node', a, k))
+        return backend, created, cred_calls, warnings
+
+    async def test_username_without_password_fails_closed(self, monkeypatch):
+        mod = get_valkey_module()
+        backend, created, cred_calls, warnings = self._prep_backend(mod, monkeypatch, username='acluser', password=None)
+
+        # A username without a password must fail closed rather than silently
+        # connecting unauthenticated to a (potentially shared) Valkey instance.
+        with pytest.raises(ValueError, match='without a password'):
+            await backend._ensure_client()
+
+        assert cred_calls == []  # ServerCredentials NOT constructed
+        assert 'conf' not in created  # client never created
+
+    async def test_password_builds_credentials(self, monkeypatch):
+        mod = get_valkey_module()
+        backend, created, cred_calls, warnings = self._prep_backend(
+            mod, monkeypatch, username='acluser', password='secret'
+        )
+
+        await backend._ensure_client()
+
+        assert len(cred_calls) == 1
+        assert cred_calls[0] == {'password': 'secret', 'username': 'acluser'}
+        assert created['conf']['credentials'] == ('CRED', {'password': 'secret', 'username': 'acluser'})

uv.lock

@@ -2084,6 +2084,7 @@ dependencies = [
     { name = "tiktoken" },
     { name = "urllib3" },
     { name = "uv" },
+    { name = "valkey-glide" },
     { name = "websockets" },
 ]
 
@@ -2123,7 +2124,7 @@ requires-dist = [
     { name = "ebooklib", specifier = ">=0.18" },
     { name = "gewechat-client", specifier = ">=0.1.5" },
     { name = "html2text", specifier = ">=2024.2.26" },
-    { name = "langbot-plugin", specifier = "==0.4.7" },
+    { name = "langbot-plugin", specifier = "==0.4.13" },
     { name = "langchain", specifier = ">=1.3.9" },
     { name = "langchain-core", specifier = ">=1.3.3" },
     { name = "langchain-text-splitters", specifier = ">=1.1.2" },
@@ -2172,6 +2173,7 @@ requires-dist = [
     { name = "tiktoken", specifier = ">=0.9.0" },
     { name = "urllib3", specifier = ">=2.7.0" },
     { name = "uv", specifier = ">=0.11.15" },
+    { name = "valkey-glide", specifier = ">=2.4.1,<3.0.0" },
     { name = "websockets", specifier = ">=15.0.1" },
 ]
 
@@ -2187,7 +2189,7 @@ dev = [
 
 [[package]]
 name = "langbot-plugin"
-version = "0.4.7"
+version = "0.4.13"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "aiofiles" },
@@ -2208,9 +2210,9 @@ dependencies = [
     { name = "watchdog" },
     { name = "websockets" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/00/6c/850a92cc05583ab71e6ac527ff50fa5f3db43cc16e86f4f6d5ac684da9a8/langbot_plugin-0.4.7.tar.gz", hash = "sha256:16b24d79fc55c0a6b15d901b1134de9e298e2cadab9a9b25c9231f731333ebe9", size = 333867, upload-time = "2026-07-02T16:49:46.258Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/40/a6/1eaf77c3b81e9de3390c504c5f627dc41f43bff6df9aff0e1e31d796b6f0/langbot_plugin-0.4.13.tar.gz", hash = "sha256:f936340e67679c21f1e7e7f1447339f31a0a2c965db060ecfbd9d0c51bb0d6fe", size = 334887, upload-time = "2026-07-04T05:38:59.942Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c6/f5/a00a929ffb4dcbd958e5f5249516f19ecc720f25f072c9bfda8bbb56851c/langbot_plugin-0.4.7-py3-none-any.whl", hash = "sha256:1364f80fcf448f4503e87d1367af60438abd2cf16e78be6807032cbdb4710a56", size = 220999, upload-time = "2026-07-02T16:49:44.868Z" },
+    { url = "https://files.pythonhosted.org/packages/e9/bf/fc9671a7afbd933440c38403c84d918c1022fdeed16e22a6ab3b2aec83ff/langbot_plugin-0.4.13-py3-none-any.whl", hash = "sha256:9d45ebc7a7ee0413d6db9baa009fcbf0ad07e2e1753a6f0a27f37b8b665cd1ee", size = 221884, upload-time = "2026-07-04T05:38:58.525Z" },
 ]
 
 [[package]]
@@ -5984,6 +5986,39 @@ wheels = [
     { url = "https://files.pythonhosted.org/packages/e4/16/c1fd27e9549f3c4baf1dc9c20c456cd2f822dbf8de9f463824b0c0357e06/uvloop-0.22.1-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:6cde23eeda1a25c75b2e07d39970f3374105d5eafbaab2a4482be82f272d5a5e", size = 4296730, upload-time = "2025-10-16T22:17:00.744Z" },
 ]
 
+[[package]]
+name = "valkey-glide"
+version = "2.4.1"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+    { name = "protobuf" },
+    { name = "sniffio" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/72/a2/582b34c6acc8dc857c537f6007459cba48dfa0dc404789a657e5c1a998c0/valkey_glide-2.4.1.tar.gz", hash = "sha256:f1155d84156d11b90488aa67e90102f0bf98a45314f5b99308ac9074c05f7241", size = 898030, upload-time = "2026-05-28T21:41:55.881Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/60/60/961ce40492a56ef831a905dfe03df4a81c0705152f6a8e49c541c634f49e/valkey_glide-2.4.1-cp311-cp311-macosx_10_7_x86_64.whl", hash = "sha256:d7285d03c2df040f26874b7f4ae96f040da2daecc9a34fa99da6f4e6ce5149c8", size = 7482152, upload-time = "2026-05-28T21:41:02.205Z" },
+    { url = "https://files.pythonhosted.org/packages/a4/b2/5a05567f0fc385dcbbbf6ab1061f0bc00443d51c2996e95eed45feaedda9/valkey_glide-2.4.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:5d2e82b74127897ccb7a957ad455787816a75fdc8c60a5e8004aef65ea93e99c", size = 6928601, upload-time = "2026-05-28T21:41:04.543Z" },
+    { url = "https://files.pythonhosted.org/packages/c5/d9/7ea2b47cff0a2f99921eb0db404215f828ced7814bd09ede9c93b65d20bc/valkey_glide-2.4.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4094128cb07e06e87013b7afab1e9388f8f5aeebe48ea6cbd54de15bd772e644", size = 7236977, upload-time = "2026-05-28T21:41:06.055Z" },
+    { url = "https://files.pythonhosted.org/packages/00/7a/6cda6b42156ed260e765e4ad2d6ab831607775e218a00fbb0d93411c4e8f/valkey_glide-2.4.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9f8dc0f3a36adb1cbe4e167972ca4758acdfed6baf58a4db94bbb713df56c8f5", size = 7691446, upload-time = "2026-05-28T21:41:07.833Z" },
+    { url = "https://files.pythonhosted.org/packages/c5/b4/da8c058baaee414a6bb2450742359f3b3b6993b23281bf227c5089f0099c/valkey_glide-2.4.1-cp312-cp312-macosx_10_7_x86_64.whl", hash = "sha256:5f8df64f6a4f0fd7203113103101fdf0aaa7ff0e7557312611de11ab89c6db75", size = 7472646, upload-time = "2026-05-28T21:41:09.451Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/94/e1e311cb56597272b9cb69afb3fe8e2e7dd3371f88c92836015deddc6f49/valkey_glide-2.4.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:b45e35f44c17e88f8cd8082f8d8061a9763238c44ef20b11b615f6d87235864a", size = 6943375, upload-time = "2026-05-28T21:41:11.079Z" },
+    { url = "https://files.pythonhosted.org/packages/76/00/0e42e2f6866ebf0de552e076dc585a487b488b5b818c52460d28b50de65b/valkey_glide-2.4.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cf812b498925a30abab6e1a9f82f5eb821e967904fe7724729b2c82c47e29edf", size = 7237469, upload-time = "2026-05-28T21:41:12.733Z" },
+    { url = "https://files.pythonhosted.org/packages/f5/4c/c5dd9a1ed995453b0d9ca75a5af87e881c14e6eebdbf5a5fa78c3bae23fc/valkey_glide-2.4.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:214e2faca98966eea3eaf9e09de616862423815a5059843a9884125e2427a344", size = 7678744, upload-time = "2026-05-28T21:41:14.634Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/2f/3df5702fc68684cef3e09f9cb6ed85578ddb08dc43593b1694c977f396fa/valkey_glide-2.4.1-cp313-cp313-macosx_10_7_x86_64.whl", hash = "sha256:c18976553ba663c03f7cc18c7e6075f4cbd2236c18b051e3d55bb213c6c44cb4", size = 7472972, upload-time = "2026-05-28T21:41:16.063Z" },
+    { url = "https://files.pythonhosted.org/packages/54/a3/6a74c6f996fa9e411e66b6f0e645fead2e0a341f1371e4cf3212efa54412/valkey_glide-2.4.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:43006e19cd63d66051263fa34a8ad47ba7d08a199585689b3f12f56ed6c9a005", size = 6943012, upload-time = "2026-05-28T21:41:17.492Z" },
+    { url = "https://files.pythonhosted.org/packages/fc/e7/d10ec41dca703f8c5dcbcba2b905e660c1cf56be53c4d5e368d7aa23d220/valkey_glide-2.4.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:b652a2a62aad87738e8f0e0aa5bf660ba91449c9fdb88550ccbc42e5fec08fe7", size = 7237842, upload-time = "2026-05-28T21:41:18.995Z" },
+    { url = "https://files.pythonhosted.org/packages/0a/a3/8916a9ed9e871686db444c86e601773245852ba1ad451ce1bb06f7aed91d/valkey_glide-2.4.1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fbd27d26947fd9f1b6e9eaf0abce4bccfde779c1e618b310c4d725424b609793", size = 7678919, upload-time = "2026-05-28T21:41:20.502Z" },
+    { url = "https://files.pythonhosted.org/packages/05/35/6d39ec3cbd24d85ad8e1051e29e6509c0999f760aff5af7851c1a1981471/valkey_glide-2.4.1-cp314-cp314-macosx_10_7_x86_64.whl", hash = "sha256:91fb7ff97acdabc8f641255b548a48627bb731e65037b1126745bf8a0022e87d", size = 7471906, upload-time = "2026-05-28T21:41:22.135Z" },
+    { url = "https://files.pythonhosted.org/packages/ab/fc/3c28f794b7d35e13101598669c1d249c0a9f0408c545c87212e364c6ee4e/valkey_glide-2.4.1-cp314-cp314-macosx_11_0_arm64.whl", hash = "sha256:d49a2537c2de44b0fc57691b1ae6c3d6f481e6f7f7eb879c0d28921d0aaec67d", size = 6943495, upload-time = "2026-05-28T21:41:23.783Z" },
+    { url = "https://files.pythonhosted.org/packages/2e/15/fb884631f5df78dc538c56bca9391165e40906b9b63ca65633d1be5bf980/valkey_glide-2.4.1-cp314-cp314-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:cded9f14e448da5a96f61c066395f2c7e2846f2afe74cacc8634da0ae0c3425f", size = 7257720, upload-time = "2026-05-28T21:41:25.361Z" },
+    { url = "https://files.pythonhosted.org/packages/73/79/0b881017194386d21812b929a81dd8afd51d6b8d92280895b45913854785/valkey_glide-2.4.1-cp314-cp314-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5f249ab5bd0d69befe35897cf51a8fc9e01e9c8c9fe03087a68e6fe6d3e31d0d", size = 7682318, upload-time = "2026-05-28T21:41:26.996Z" },
+    { url = "https://files.pythonhosted.org/packages/7a/4d/f2b4e508692fcd21e76c7cbdc4f988bec7f4675e60f4f35ef482a826f6ae/valkey_glide-2.4.1-pp311-pypy311_pp73-macosx_10_7_x86_64.whl", hash = "sha256:775df9c7421a187c41caf003e4af5f073ed7e4b8abe50f8b9bec712cb03e12bf", size = 7479155, upload-time = "2026-05-28T21:41:42.399Z" },
+    { url = "https://files.pythonhosted.org/packages/52/d8/8a3495f5582dccb4c8e7faf6a73baf3dbc4580701923f06d8abf210ff22d/valkey_glide-2.4.1-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:0d87f21c77004240189cc3c5aab156966487afd81ffdee04225a52c7bd7132e4", size = 6938571, upload-time = "2026-05-28T21:41:44.078Z" },
+    { url = "https://files.pythonhosted.org/packages/f3/5a/a70077f76c2f18e94ec4309857b248beb7a8c7a3a50e30242abde2c3827d/valkey_glide-2.4.1-pp311-pypy311_pp73-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:44376ef5fe7a25287095b073d8abde510a50b1ead0143662394b3da9717863ef", size = 7260021, upload-time = "2026-05-28T21:41:45.837Z" },
+    { url = "https://files.pythonhosted.org/packages/aa/12/72d31522e06fcc9b391118c1f69a09002224e78114b1db0d01b96008dc59/valkey_glide-2.4.1-pp311-pypy311_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a59cc0a21d7a8b1b3caeb299f23817429b5fe6579bd4cb016382e6b7a10de984", size = 7693093, upload-time = "2026-05-28T21:41:47.617Z" },
+]
+
 [[package]]
 name = "virtualenv"
 version = "20.36.1"

web/src/app/home/bots/components/bot-form/BotForm.tsx

@@ -269,6 +269,10 @@ export default function BotForm({
               options: item.options,
               show_if: item.show_if,
               login_platform: item.login_platform,
+              url: item.url,
+              download_filename: item.download_filename,
+              help_links: item.help_links,
+              help_label: item.help_label,
             }),
         ),
       );

web/src/app/home/components/dynamic-form/DynamicFormComponent.tsx

@@ -24,7 +24,15 @@ import { useTranslation } from 'react-i18next';
 import { cn } from '@/lib/utils';
 import { Input } from '@/components/ui/input';
 import { Button } from '@/components/ui/button';
-import { Copy, Check, Globe, Info, QrCode } from 'lucide-react';
+import {
+  Copy,
+  Check,
+  Globe,
+  Info,
+  QrCode,
+  Download,
+  ExternalLink,
+} from 'lucide-react';
 import { copyToClipboard } from '@/app/utils/clipboard';
 import {
   Tooltip,
@@ -33,6 +41,7 @@ import {
   TooltipTrigger,
 } from '@/components/ui/tooltip';
 import { systemInfo } from '@/app/infra/http';
+import { getAdapterDocUrl } from '@/app/infra/entities/adapter-docs';
 
 /**
  * Resolve the value referenced by a `show_if.field` string.
@@ -291,6 +300,52 @@ function WebhookUrlField({
   );
 }
 
+function DownloadLinkField({
+  label,
+  description,
+  url,
+  filename,
+  helpUrl,
+  helpLabel,
+}: {
+  label: string;
+  description?: string;
+  url: string;
+  filename?: string;
+  helpUrl?: string | null;
+  helpLabel: string;
+}) {
+  const baseUrl = import.meta.env.VITE_API_BASE_URL || window.location.origin;
+  const downloadUrl = url.startsWith('http') ? url : `${baseUrl}${url}`;
+
+  return (
+    <FormItem className="min-w-0">
+      <FormLabel className="break-words">{label}</FormLabel>
+      <div className="flex min-w-0 flex-wrap items-center gap-2">
+        <Button asChild variant="outline" size="sm">
+          <a href={downloadUrl} download={filename}>
+            <Download className="h-4 w-4" />
+            {label}
+          </a>
+        </Button>
+        {helpUrl && (
+          <Button asChild variant="ghost" size="sm">
+            <a href={helpUrl} target="_blank" rel="noopener noreferrer">
+              <ExternalLink className="h-4 w-4" />
+              {helpLabel}
+            </a>
+          </Button>
+        )}
+      </div>
+      {description && (
+        <p className="max-w-2xl text-sm break-words text-muted-foreground">
+          {description}
+        </p>
+      )}
+    </FormItem>
+  );
+}
+
 /**
  * Display-only component for `__system.*` fields (e.g. the deployment's
  * outbound IPs that the operator must add to a platform's trusted-IP list).
@@ -405,7 +460,7 @@ export default function DynamicFormComponent({
 }) {
   const isInitialMount = useRef(true);
   const previousInitialValues = useRef(initialValues);
-  const { t } = useTranslation();
+  const { t, i18n } = useTranslation();
 
   // Normalize a form value according to its field type.
   // This ensures legacy/malformed data (e.g. a plain string for
@@ -460,6 +515,7 @@ export default function DynamicFormComponent({
           item.type !== 'webhook-url' &&
           item.type !== 'embed-code' &&
           item.type !== 'qr-code-login' &&
+          item.type !== 'download-link' &&
           !item.name.startsWith(SYSTEM_FIELD_PREFIX),
       ),
     [itemConfigList],
@@ -777,6 +833,30 @@ export default function DynamicFormComponent({
             );
           }
 
+          if (config.type === 'download-link') {
+            if (!config.url) return null;
+
+            return (
+              <DownloadLinkField
+                key={config.id}
+                label={extractI18nObject(config.label)}
+                description={
+                  config.description
+                    ? extractI18nObject(config.description)
+                    : undefined
+                }
+                url={config.url}
+                filename={config.download_filename}
+                helpUrl={getAdapterDocUrl(config.help_links, i18n.language)}
+                helpLabel={
+                  config.help_label
+                    ? extractI18nObject(config.help_label)
+                    : t('bots.viewAdapterDocs')
+                }
+              />
+            );
+          }
+
           // QR code login button (e.g. Feishu one-click create, WeChat scan login)
           if (config.type === 'qr-code-login') {
             return (

web/src/app/home/components/dynamic-form/DynamicFormItemConfig.ts

@@ -18,6 +18,10 @@ export class DynamicFormItemConfig implements IDynamicFormItemSchema {
   options?: IDynamicFormItemOption[];
   show_if?: IShowIfCondition;
   login_platform?: string;
+  url?: string;
+  download_filename?: string;
+  help_links?: Record<string, string>;
+  help_label?: I18nObject;
 
   constructor(params: IDynamicFormItemSchema) {
     this.id = params.id;
@@ -30,6 +34,10 @@ export class DynamicFormItemConfig implements IDynamicFormItemSchema {
     this.options = params.options;
     this.show_if = params.show_if;
     this.login_platform = params.login_platform;
+    this.url = params.url;
+    this.download_filename = params.download_filename;
+    this.help_links = params.help_links;
+    this.help_label = params.help_label;
   }
 }
 

web/src/app/home/components/home-sidebar/HomeSidebar.tsx

@@ -264,6 +264,48 @@ function saveListExpansionState(state: SidebarListExpansionState) {
 
 // Maximum number of entity sub-items visible before "More" toggle
 const MAX_VISIBLE_ITEMS = 5;
+const MCP_REFRESH_POLL_INTERVAL_MS = 1000;
+const MCP_REFRESH_TIMEOUT_MS = 60000;
+
+function sleep(ms: number) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+async function waitForMCPRefreshTask(taskId: number) {
+  const deadline = Date.now() + MCP_REFRESH_TIMEOUT_MS;
+
+  while (Date.now() < deadline) {
+    const task = await httpClient.getAsyncTask(taskId);
+    if (task.runtime.done) return task;
+    await sleep(MCP_REFRESH_POLL_INTERVAL_MS);
+  }
+
+  throw new Error(`Timed out waiting for MCP refresh task ${taskId}`);
+}
+
+async function refreshEnabledMCPConnections() {
+  const resp = await httpClient.getMCPServers();
+  const enabledServers = resp.servers.filter((server) => server.enable);
+  if (enabledServers.length === 0) return;
+
+  const taskResults = await Promise.allSettled(
+    enabledServers.map((server) => httpClient.testMCPServer(server.name, {})),
+  );
+  const taskIds: number[] = [];
+
+  for (const result of taskResults) {
+    if (
+      result.status === 'fulfilled' &&
+      typeof result.value.task_id === 'number'
+    ) {
+      taskIds.push(result.value.task_id);
+    } else if (result.status === 'rejected') {
+      console.error('Failed to start MCP refresh task:', result.reason);
+    }
+  }
+
+  await Promise.allSettled(taskIds.map(waitForMCPRefreshTask));
+}
 
 // Sort entity items by updatedAt descending (most recent first), items without updatedAt go last
 function sortByRecent(items: SidebarEntityItem[]): SidebarEntityItem[] {
@@ -352,11 +394,19 @@ function NavItems({
     if (extRefreshing) return;
     setExtRefreshing(true);
     try {
-      await Promise.all([
+      const results = await Promise.allSettled([
         sidebarData.refreshPlugins(),
-        sidebarData.refreshMCPServers(),
         sidebarData.refreshSkills(),
+        refreshEnabledMCPConnections(),
       ]);
+      const mcpRefreshResult = results[2];
+      if (mcpRefreshResult.status === 'rejected') {
+        console.error(
+          'Failed to refresh MCP connections:',
+          mcpRefreshResult.reason,
+        );
+      }
+      await sidebarData.refreshMCPServers();
     } finally {
       setExtRefreshing(false);
     }

web/src/app/home/components/qrcode-login/QrCodeLoginDialog.tsx

@@ -16,7 +16,12 @@ import {
 } from 'lucide-react';
 import QRCode from 'qrcode';
 
-export type QrLoginPlatform = 'feishu' | 'weixin' | 'dingtalk' | 'wecombot';
+export type QrLoginPlatform =
+  | 'feishu'
+  | 'weixin'
+  | 'dingtalk'
+  | 'wecombot'
+  | 'qqofficial';
 
 interface PlatformConfig {
   titleKey: string;
@@ -29,6 +34,7 @@ interface PlatformConfig {
   apiBase: string;
   extractSuccess: (data: Record<string, string>) => Record<string, string>;
   successNoteKey?: string;
+  boundByKey?: string;
 }
 
 const PLATFORM_CONFIGS: Record<QrLoginPlatform, PlatformConfig> = {
@@ -92,6 +98,22 @@ const PLATFORM_CONFIGS: Record<QrLoginPlatform, PlatformConfig> = {
     }),
     successNoteKey: 'wecombot.robotNameNote',
   },
+  qqofficial: {
+    titleKey: 'qqofficial.createBinding',
+    connectingKey: 'qqofficial.connecting',
+    scanQRCodeKey: 'qqofficial.scanQRCode',
+    waitingKey: 'qqofficial.waitingForScan',
+    successKey: 'qqofficial.bindSuccess',
+    failedKey: 'qqofficial.bindFailed',
+    retryKey: 'qqofficial.retry',
+    apiBase: '/api/v1/platform/adapters/qqofficial/bind',
+    extractSuccess: (data) => ({
+      appid: data.appid,
+      secret: data.secret,
+    }),
+    successNoteKey: 'qqofficial.tokenNote',
+    boundByKey: 'qqofficial.boundBy',
+  },
 };
 
 interface QrCodeLoginDialogProps {
@@ -118,6 +140,7 @@ export default function QrCodeLoginDialog({
   const [qrDataUrl, setQrDataUrl] = useState('');
   const [expireIn, setExpireIn] = useState(0);
   const [errorMessage, setErrorMessage] = useState('');
+  const [successMeta, setSuccessMeta] = useState('');
   const pollTimerRef = useRef<ReturnType<typeof setInterval> | null>(null);
   const countdownRef = useRef<ReturnType<typeof setInterval> | null>(null);
   const checkExpiredRef = useRef<ReturnType<typeof setInterval> | null>(null);
@@ -178,6 +201,7 @@ export default function QrCodeLoginDialog({
     setQrDataUrl('');
     setExpireIn(0);
     setErrorMessage('');
+    setSuccessMeta('');
 
     const token = localStorage.getItem('token');
     const baseUrl = import.meta.env.VITE_API_BASE_URL || window.location.origin;
@@ -275,6 +299,13 @@ export default function QrCodeLoginDialog({
             sessionIdRef.current = null;
             cleanup();
             setState('success');
+            // Platform may return extra audit metadata (e.g. QQ Official returns
+            // the scanner's user_openid) — surface it briefly before the dialog closes.
+            if (rest.user_openid && cfg.boundByKey) {
+              setSuccessMeta(
+                tRef.current(cfg.boundByKey, { openid: rest.user_openid }),
+              );
+            }
             setTimeout(() => {
               onSuccessRef.current(cfg.extractSuccess(rest));
               onOpenChangeRef.current(false);
@@ -395,6 +426,11 @@ export default function QrCodeLoginDialog({
               <p className="text-sm text-green-600 font-medium">
                 {t(platformConfig.successKey)}
               </p>
+              {successMeta && (
+                <p className="text-xs text-muted-foreground text-center max-w-xs break-all">
+                  {successMeta}
+                </p>
+              )}
               {platformConfig.successNoteKey && (
                 <p className="text-xs text-muted-foreground text-center max-w-xs">
                   {t(platformConfig.successNoteKey)}

web/src/app/home/mcp/MCPDetailContent.tsx

@@ -157,6 +157,10 @@ export default function MCPDetailContent({ id }: { id: string }) {
     navigate(`/home/mcp?id=${encodeURIComponent(serverName)}`);
   }
 
+  const handlePersistedTestComplete = useCallback(async () => {
+    await refreshMCPServers();
+  }, [refreshMCPServers]);
+
   function confirmDelete() {
     httpClient
       .deleteMCPServer(id)
@@ -364,6 +368,7 @@ export default function MCPDetailContent({ id }: { id: string }) {
             onRuntimeInfoChange={(runtimeInfo) =>
               setDetailRuntimeStatus(runtimeInfo?.status ?? null)
             }
+            onPersistedTestComplete={handlePersistedTestComplete}
           />
         </div>
       </div>

web/src/app/home/mcp/components/mcp-form/MCPForm.tsx

@@ -41,6 +41,7 @@ import {
 } from '@/components/ui/card';
 import { httpClient } from '@/app/infra/http/HttpClient';
 import { Tabs, TabsContent, TabsList, TabsTrigger } from '@/components/ui/tabs';
+import MCPLogs from '@/app/home/mcp/components/mcp-form/MCPLogs';
 import MCPReadme from '@/app/home/mcp/components/mcp-form/MCPReadme';
 import {
   MCPServerRuntimeInfo,
@@ -487,6 +488,7 @@ interface MCPFormProps {
   onDirtyChange?: (dirty: boolean) => void;
   onTestingChange?: (testing: boolean) => void;
   onRuntimeInfoChange?: (runtimeInfo: MCPServerRuntimeInfo | null) => void;
+  onPersistedTestComplete?: (serverName: string) => void | Promise<void>;
   /** Reported when the form cannot be saved because the current mode is
    * ``stdio`` and the Box sandbox is disabled/unavailable. Parents that
    * render the Save button outside this component should disable it. */
@@ -511,6 +513,7 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
     onDirtyChange,
     onTestingChange,
     onRuntimeInfoChange,
+    onPersistedTestComplete,
     onSaveBlockedChange,
     layout = 'stacked',
     sideHeader,
@@ -822,6 +825,8 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
       // are always current.
       const serverName =
         isEditMode && initServerName ? initServerName : form.getValues('name');
+      const shouldTestPersistedServer =
+        isEditMode && !!initServerName && !form.formState.isDirty;
       const formExtraArgs = form.getValues('extra_args') ?? [];
       const formStdioArgs = form.getValues('args') ?? [];
       let extraArgsData: MCPServerExtraArgsRemote | MCPServerExtraArgsStdio;
@@ -844,12 +849,20 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
         };
       }
 
-      const { task_id } = await httpClient.testMCPServer('_', {
+      const testTarget = shouldTestPersistedServer ? serverName : '_';
-        name: serverName,
+      const testPayload = shouldTestPersistedServer
-        mode,
+        ? {}
-        enable: true,
+        : ({
-        extra_args: extraArgsData,
+            name: serverName,
-      } as MCPServer);
+            mode,
+            enable: true,
+            extra_args: extraArgsData,
+          } as MCPServer);
+
+      const { task_id } = await httpClient.testMCPServer(
+        testTarget,
+        testPayload,
+      );
 
       if (!task_id) {
         throw new Error(t('mcp.noTaskId'));
@@ -875,14 +888,18 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
                 resource_count: 0,
                 resources: [],
               });
+              if (shouldTestPersistedServer) {
+                await onPersistedTestComplete?.(serverName);
+              }
             } else {
-              if (isEditMode) {
+              if (shouldTestPersistedServer) {
                 await loadServerForEdit(serverName);
+                await onPersistedTestComplete?.(serverName);
               } else {
-                // Create mode has no persisted server to reload tools from.
+                // Transient tests have no persisted server to reload tools from.
                 // The backend stashes the discovered runtime info (status +
-                // tools) in the test task's metadata before tearing the
+                // tools) in the task metadata before tearing the transient
-                // transient session down — surface it so a successful test
+                // session down — surface it so a successful test
                 // shows the tool list instead of "no tools found".
                 const runtimeInfoFromTest = taskResp.task_context?.metadata
                   ?.runtime_info as MCPServerRuntimeInfo | undefined;
@@ -1207,6 +1224,9 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
         <TabsTrigger value="resources" className="flex-none px-4">
           {resourcesTabLabel}
         </TabsTrigger>
+        <TabsTrigger value="logs" className="flex-none px-4">
+          {t('mcp.tabLogs')}
+        </TabsTrigger>
       </TabsList>
       <TabsContent value="docs" className="mt-4 min-h-0 flex-1 overflow-y-auto">
         <MCPReadme readme={readme} />
@@ -1235,6 +1255,9 @@ const MCPForm = forwardRef<MCPFormHandle, MCPFormProps>(function MCPForm(
           t={t}
         />
       </TabsContent>
+      <TabsContent value="logs" className="mt-4 min-h-0 flex-1 overflow-y-auto">
+        {persistedServerName && <MCPLogs serverName={persistedServerName} />}
+      </TabsContent>
     </Tabs>
   ) : (
     runtimePanel

web/src/app/home/mcp/components/mcp-form/MCPLogs.tsx

@@ -0,0 +1,149 @@
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { httpClient } from '@/app/infra/http/HttpClient';
+import { useTranslation } from 'react-i18next';
+import { PluginLogEntry } from '@/app/infra/entities/plugin';
+import { Button } from '@/components/ui/button';
+import { Switch } from '@/components/ui/switch';
+import { Label } from '@/components/ui/label';
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from '@/components/ui/select';
+import { RefreshCw } from 'lucide-react';
+
+const LEVEL_OPTIONS = ['ALL', 'DEBUG', 'INFO', 'WARNING', 'ERROR'] as const;
+
+function levelClassName(level: string): string {
+  switch (level) {
+    case 'ERROR':
+    case 'CRITICAL':
+      return 'text-red-500';
+    case 'WARNING':
+      return 'text-amber-500';
+    case 'DEBUG':
+      return 'text-gray-400 dark:text-gray-500';
+    default:
+      return 'text-gray-700 dark:text-gray-300';
+  }
+}
+
+export default function MCPLogs({ serverName }: { serverName: string }) {
+  const { t } = useTranslation();
+  const [logs, setLogs] = useState<PluginLogEntry[]>([]);
+  const [isLoading, setIsLoading] = useState(false);
+  const [level, setLevel] = useState<string>('ALL');
+  const [autoRefresh, setAutoRefresh] = useState(true);
+  const scrollRef = useRef<HTMLDivElement>(null);
+  const atBottomRef = useRef(true);
+
+  const fetchLogs = useCallback(() => {
+    setIsLoading(true);
+    httpClient
+      .getMcpServerLogs(serverName, 500, level === 'ALL' ? undefined : level)
+      .then((res) => {
+        setLogs(res.logs ?? []);
+      })
+      .catch(() => {
+        setLogs([]);
+      })
+      .finally(() => {
+        setIsLoading(false);
+      });
+  }, [serverName, level]);
+
+  useEffect(() => {
+    fetchLogs();
+  }, [fetchLogs]);
+
+  // Auto-refresh poll loop.
+  useEffect(() => {
+    if (!autoRefresh) return;
+    const timer = setInterval(fetchLogs, 3000);
+    return () => clearInterval(timer);
+  }, [autoRefresh, fetchLogs]);
+
+  // Keep view pinned to bottom when the user is already at the bottom.
+  useEffect(() => {
+    const el = scrollRef.current;
+    if (el && atBottomRef.current) {
+      el.scrollTop = el.scrollHeight;
+    }
+  }, [logs]);
+
+  function handleScroll() {
+    const el = scrollRef.current;
+    if (!el) return;
+    atBottomRef.current = el.scrollHeight - el.scrollTop - el.clientHeight < 40;
+  }
+
+  return (
+    <div className="flex h-full flex-col">
+      <div className="flex shrink-0 flex-wrap items-center gap-2 px-1 pb-3 sm:px-6">
+        <Select value={level} onValueChange={setLevel}>
+          <SelectTrigger className="h-8 w-[130px]">
+            <SelectValue />
+          </SelectTrigger>
+          <SelectContent>
+            {LEVEL_OPTIONS.map((opt) => (
+              <SelectItem key={opt} value={opt}>
+                {opt === 'ALL' ? t('mcp.logsLevelAll') : opt}
+              </SelectItem>
+            ))}
+          </SelectContent>
+        </Select>
+        <Button
+          type="button"
+          variant="outline"
+          size="sm"
+          className="h-8"
+          onClick={fetchLogs}
+          disabled={isLoading}
+        >
+          <RefreshCw
+            className={`mr-1.5 size-3.5 ${isLoading ? 'animate-spin' : ''}`}
+          />
+          {t('mcp.logsRefresh')}
+        </Button>
+        <div className="flex items-center gap-2">
+          <Switch
+            id="mcp-logs-auto-refresh"
+            checked={autoRefresh}
+            onCheckedChange={setAutoRefresh}
+          />
+          <Label
+            htmlFor="mcp-logs-auto-refresh"
+            className="cursor-pointer text-sm font-normal text-muted-foreground"
+          >
+            {t('mcp.logsAutoRefresh')}
+          </Label>
+        </div>
+      </div>
+
+      <div
+        ref={scrollRef}
+        onScroll={handleScroll}
+        className="min-h-0 flex-1 overflow-auto bg-gray-50 px-3 py-3 font-mono text-xs leading-relaxed dark:bg-gray-900/40 sm:px-6"
+      >
+        {logs.length === 0 ? (
+          <div className="py-8 text-center text-sm text-gray-500 dark:text-gray-400">
+            {t('mcp.logsEmpty')}
+          </div>
+        ) : (
+          logs.map((entry, idx) => (
+            <div
+              key={`${entry.ts}-${idx}`}
+              className={`whitespace-pre-wrap break-all ${levelClassName(
+                entry.level,
+              )}`}
+            >
+              {entry.text}
+            </div>
+          ))
+        )}
+      </div>
+    </div>
+  );
+}

web/src/app/infra/entities/form/dynamic.ts

@@ -44,6 +44,10 @@ export interface IDynamicFormItemSchema {
   scopes?: string[];
   accept?: string; // For file type: accepted MIME types
   login_platform?: string; // For qr-code-login type: platform identifier (e.g. 'feishu', 'weixin')
+  url?: string; // For download-link type: relative or absolute download URL
+  download_filename?: string; // Optional filename for download-link type
+  help_links?: Record<string, string>; // Optional docs links for display-only fields
+  help_label?: I18nObject; // Optional label for help_links
 }
 
 export enum DynamicFormItemType {
@@ -72,6 +76,7 @@ export enum DynamicFormItemType {
   WEBHOOK_URL = 'webhook-url',
   EMBED_CODE = 'embed-code',
   QR_CODE_LOGIN = 'qr-code-login',
+  DOWNLOAD_LINK = 'download-link',
 }
 
 export interface IFileConfig {

web/src/app/infra/http/BackendClient.ts

@@ -671,6 +671,21 @@ export class BackendClient extends BaseHttpClient {
     );
   }
 
+  public getMcpServerLogs(
+    serverName: string,
+    limit: number = 200,
+    level?: string,
+  ): Promise<{ logs: PluginLogEntry[] }> {
+    const params = new URLSearchParams();
+    params.set('limit', String(limit));
+    if (level) {
+      params.set('level', level);
+    }
+    return this.get(
+      `/api/v1/mcp/servers/${encodeURIComponent(serverName)}/logs?${params.toString()}`,
+    );
+  }
+
   public getPluginAssetURL(
     author: string,
     name: string,

web/src/app/wizard/page.tsx

@@ -229,6 +229,10 @@ export default function WizardPage() {
           options: item.options,
           show_if: item.show_if,
           login_platform: item.login_platform,
+          url: item.url,
+          download_filename: item.download_filename,
+          help_links: item.help_links,
+          help_label: item.help_label,
         }),
     );
   }, [adapters, selectedAdapter]);
@@ -249,6 +253,10 @@ export default function WizardPage() {
           options: item.options,
           show_if: item.show_if,
           login_platform: item.login_platform,
+          url: item.url,
+          download_filename: item.download_filename,
+          help_links: item.help_links,
+          help_label: item.help_label,
         }),
     );
   }, [selectedRunnerConfigStage]);

web/src/i18n/locales/en-US.ts

@@ -828,6 +828,12 @@ const enUS = {
     tabTools: 'Tools',
     tabResources: 'Resources',
     tabDocs: 'Docs',
+    tabLogs: 'Logs',
+    logsLevelAll: 'All levels',
+    logsRefresh: 'Refresh',
+    logsAutoRefresh: 'Auto refresh',
+    logsEmpty:
+      'No logs yet. Runtime logs from the MCP server will appear here.',
     noReadme: 'No documentation available',
     parseResultFailed: 'Failed to parse test result',
     noResultReturned: 'Test returned no result',
@@ -1816,6 +1822,19 @@ const enUS = {
     robotNameNote:
       'Robot Name cannot be obtained automatically. Please fill it in manually.',
   },
+  qqofficial: {
+    createBinding: 'One-Click QR Binding for QQ Official Bot',
+    scanQRCode:
+      'Scan the QR code below with mobile QQ and authorize the binding in QQ Bot Assistant',
+    waitingForScan: 'Waiting for scan',
+    bindSuccess: 'Bound successfully! AppID and Secret have been filled in',
+    bindFailed: 'Binding failed',
+    connecting: 'Connecting to QQ service...',
+    retry: 'Retry',
+    tokenNote:
+      'The Token field is not used by the current adapter — you can leave it blank.',
+    boundBy: 'Bound by QQ user {{openid}}',
+  },
   pluginPages: {
     selectFromSidebar: 'Select a plugin page from the sidebar',
     invalidPage: 'Invalid plugin page',

web/src/i18n/locales/es-ES.ts

@@ -842,6 +842,12 @@ const esES = {
     tabTools: 'Herramientas',
     tabResources: 'Recursos',
     tabDocs: 'Documentación',
+    tabLogs: 'Registros',
+    logsLevelAll: 'Todos los niveles',
+    logsRefresh: 'Actualizar',
+    logsAutoRefresh: 'Actualización automática',
+    logsEmpty:
+      'Aún no hay registros. Los registros de ejecución del servidor MCP aparecerán aquí.',
     noReadme: 'No hay documentación disponible',
     parseResultFailed: 'Error al analizar el resultado de la prueba',
     noResultReturned: 'La prueba no devolvió resultados',
@@ -1734,6 +1740,20 @@ const esES = {
     robotNameNote:
       'El nombre del robot no puede obtenerse automáticamente. Introdúcelo manualmente.',
   },
+  qqofficial: {
+    createBinding: 'Vinculación QR con un clic para el bot oficial de QQ',
+    scanQRCode:
+      'Escanea el código QR siguiente con QQ móvil y autoriza la vinculación en «QQ Bot Assistant»',
+    waitingForScan: 'Esperando escaneo',
+    bindSuccess:
+      '¡Vinculación correcta! AppID y Secret se han rellenado automáticamente',
+    bindFailed: 'Error en la vinculación',
+    connecting: 'Conectando con el servicio de QQ...',
+    retry: 'Reintentar',
+    tokenNote:
+      'El campo Token no es utilizado por el adaptador actual; puedes dejarlo vacío.',
+    boundBy: 'Vinculado por el usuario QQ {{openid}}',
+  },
   pluginPages: {
     selectFromSidebar: 'Selecciona una página de plugin en la barra lateral',
     invalidPage: 'Página de plugin no válida',

web/src/i18n/locales/ja-JP.ts

@@ -834,6 +834,11 @@ const jaJP = {
     tabTools: 'ツール',
     tabResources: 'リソース',
     tabDocs: 'ドキュメント',
+    tabLogs: 'ログ',
+    logsLevelAll: 'すべてのレベル',
+    logsRefresh: '更新',
+    logsAutoRefresh: '自動更新',
+    logsEmpty: 'ログはありません。MCPサーバーの実行ログがここに表示されます。',
     noReadme: 'ドキュメントがありません',
     parseResultFailed: 'テスト結果の解析に失敗しました',
     noResultReturned: 'テスト結果が返されませんでした',
@@ -1728,6 +1733,19 @@ const jaJP = {
     retry: '再試行',
     robotNameNote: 'ロボット名は自動取得できません。手動で入力してください。',
   },
+  qqofficial: {
+    createBinding: 'ワンクリックで QQ 公式ボットを QR バインド',
+    scanQRCode:
+      '以下の QR コードをモバイル QQ でスキャンし、「QQ ボットアシスタント」でバインドを承認してください',
+    waitingForScan: 'スキャン待ち',
+    bindSuccess: 'バインド成功！AppID と Secret が自動入力されました',
+    bindFailed: 'バインド失敗',
+    connecting: 'QQ サービスに接続中...',
+    retry: '再試行',
+    tokenNote:
+      'Token フィールドは現行アダプターでは使用しません。空欄のままで構いません。',
+    boundBy: 'QQ ユーザー {{openid}} によりバインドされました',
+  },
   pluginPages: {
     selectFromSidebar: 'サイドバーからプラグインページを選択してください',
     invalidPage: '無効なプラグインページ',

web/src/i18n/locales/ru-RU.ts

@@ -839,6 +839,12 @@ const ruRU = {
     tabTools: 'Инструменты',
     tabResources: 'Ресурсы',
     tabDocs: 'Документация',
+    tabLogs: 'Журнал',
+    logsLevelAll: 'Все уровни',
+    logsRefresh: 'Обновить',
+    logsAutoRefresh: 'Автообновление',
+    logsEmpty:
+      'Журналов пока нет. Здесь будут отображаться журналы выполнения MCP-сервера.',
     noReadme: 'Документация отсутствует',
     parseResultFailed: 'Не удалось разобрать результат теста',
     noResultReturned: 'Тест не вернул результат',
@@ -1705,6 +1711,19 @@ const ruRU = {
     robotNameNote:
       'Имя бота нельзя получить автоматически. Пожалуйста, введите его вручную.',
   },
+  qqofficial: {
+    createBinding: 'Привязка официального бота QQ по QR-коду',
+    scanQRCode:
+      'Отсканируйте QR-код ниже мобильным QQ и подтвердите привязку в «QQ Bot Assistant»',
+    waitingForScan: 'Ожидание сканирования',
+    bindSuccess: 'Привязка успешна! AppID и Secret заполнены автоматически',
+    bindFailed: 'Не удалось выполнить привязку',
+    connecting: 'Подключение к сервису QQ...',
+    retry: 'Повторить',
+    tokenNote:
+      'Поле Token не используется текущим адаптером — его можно оставить пустым.',
+    boundBy: 'Привязано пользователем QQ {{openid}}',
+  },
   pluginPages: {
     selectFromSidebar: 'Выберите страницу плагина на боковой панели',
     invalidPage: 'Недопустимая страница плагина',

web/src/i18n/locales/th-TH.ts

@@ -817,6 +817,11 @@ const thTH = {
     tabTools: 'เครื่องมือ',
     tabResources: 'ทรัพยากร',
     tabDocs: 'เอกสาร',
+    tabLogs: 'บันทึก',
+    logsLevelAll: 'ทุกระดับ',
+    logsRefresh: 'รีเฟรช',
+    logsAutoRefresh: 'รีเฟรชอัตโนมัติ',
+    logsEmpty: 'ยังไม่มีบันทึก บันทึกการทำงานของ MCP Server จะแสดงที่นี่',
     noReadme: 'ไม่มีเอกสาร',
     parseResultFailed: 'ไม่สามารถแยกวิเคราะห์ผลการทดสอบได้',
     noResultReturned: 'การทดสอบไม่ส่งผลลัพธ์กลับมา',
@@ -1667,6 +1672,18 @@ const thTH = {
     retry: 'ลองใหม่',
     robotNameNote: 'ไม่สามารถดึงชื่อบอตได้โดยอัตโนมัติ กรุณากรอกด้วยตนเอง',
   },
+  qqofficial: {
+    createBinding: 'ผูกบอต QQ Official ด้วย QR คลิกเดียว',
+    scanQRCode:
+      'สแกนคิวอาร์โค้ดด้านล่างด้วย QQ มือถือ แล้วอนุญาตการผูกใน «QQ Bot Assistant»',
+    waitingForScan: 'กำลังรอสแกน',
+    bindSuccess: 'ผูกสำเร็จ! AppID และ Secret ถูกกรอกอัตโนมัติแล้ว',
+    bindFailed: 'การผูกล้มเหลว',
+    connecting: 'กำลังเชื่อมต่อบริการ QQ...',
+    retry: 'ลองใหม่',
+    tokenNote: 'อะแดปเตอร์ปัจจุบันไม่ได้ใช้ฟิลด์ Token จึงเว้นว่างไว้ได้',
+    boundBy: 'ผูกโดยผู้ใช้ QQ {{openid}}',
+  },
   pluginPages: {
     selectFromSidebar: 'เลือกหน้าปลั๊กอินจากแถบด้านข้าง',
     invalidPage: 'หน้าปลั๊กอินไม่ถูกต้อง',

web/src/i18n/locales/vi-VN.ts

@@ -832,6 +832,12 @@ const viVN = {
     tabTools: 'Công cụ',
     tabResources: 'Tài nguyên',
     tabDocs: 'Tài liệu',
+    tabLogs: 'Nhật ký',
+    logsLevelAll: 'Tất cả cấp độ',
+    logsRefresh: 'Làm mới',
+    logsAutoRefresh: 'Tự động làm mới',
+    logsEmpty:
+      'Chưa có nhật ký. Nhật ký chạy của MCP Server sẽ hiển thị ở đây.',
     noReadme: 'Không có tài liệu',
     parseResultFailed: 'Phân tích kết quả kiểm tra thất bại',
     noResultReturned: 'Kiểm tra không trả về kết quả',
@@ -1695,6 +1701,19 @@ const viVN = {
     retry: 'Thử lại',
     robotNameNote: 'Không thể tự động lấy tên bot. Vui lòng điền thủ công.',
   },
+  qqofficial: {
+    createBinding: 'Liên kết bot QQ Official bằng QR một chạm',
+    scanQRCode:
+      'Quét mã QR bên dưới bằng QQ trên di động và xác nhận liên kết trong «QQ Bot Assistant»',
+    waitingForScan: 'Đang chờ quét',
+    bindSuccess: 'Liên kết thành công! AppID và Secret đã được điền tự động',
+    bindFailed: 'Liên kết thất bại',
+    connecting: 'Đang kết nối tới dịch vụ QQ...',
+    retry: 'Thử lại',
+    tokenNote:
+      'Bộ chuyển đổi hiện tại không dùng trường Token; có thể để trống.',
+    boundBy: 'Được liên kết bởi người dùng QQ {{openid}}',
+  },
   pluginPages: {
     selectFromSidebar: 'Chọn một trang plugin từ thanh bên',
     invalidPage: 'Trang plugin không hợp lệ',

web/src/i18n/locales/zh-Hans.ts

@@ -794,6 +794,11 @@ const zhHans = {
     tabTools: '工具',
     tabResources: '资源',
     tabDocs: '文档',
+    tabLogs: '日志',
+    logsLevelAll: '全部级别',
+    logsRefresh: '刷新',
+    logsAutoRefresh: '自动刷新',
+    logsEmpty: '暂无日志。MCP 服务器的运行日志会显示在这里。',
     noReadme: '暂无文档',
     parseResultFailed: '解析测试结果失败',
     noResultReturned: '测试未返回结果',
@@ -1736,6 +1741,17 @@ const zhHans = {
     retry: '重试',
     robotNameNote: '机器人名称无法自动获取，请手动填写。',
   },
+  qqofficial: {
+    createBinding: '一键扫码绑定 QQ 机器人',
+    scanQRCode: '请使用手机 QQ 扫描以下二维码，在「QQ 机器人助手」中授权绑定',
+    waitingForScan: '等待扫码中',
+    bindSuccess: '绑定成功！AppID 与密钥已自动填入',
+    bindFailed: '绑定失败',
+    connecting: '正在连接 QQ 服务...',
+    retry: '重试',
+    tokenNote: 'Token 字段当前适配器未使用，留空即可。',
+    boundBy: '由 QQ 用户 {{openid}} 扫码绑定',
+  },
   pluginPages: {
     selectFromSidebar: '从侧边栏选择一个插件页面',
     invalidPage: '无效的插件页面',

web/src/i18n/locales/zh-Hant.ts

@@ -793,6 +793,11 @@ const zhHant = {
     tabTools: '工具',
     tabResources: '資源',
     tabDocs: '文件',
+    tabLogs: '日誌',
+    logsLevelAll: '全部級別',
+    logsRefresh: '重新整理',
+    logsAutoRefresh: '自動重新整理',
+    logsEmpty: '暫無日誌。MCP 服務器的運行日誌會顯示在這裡。',
     noReadme: '暫無文件',
     parseResultFailed: '解析測試結果失敗',
     noResultReturned: '測試未返回結果',
@@ -1645,6 +1650,17 @@ const zhHant = {
     retry: '重試',
     robotNameNote: '機器人名稱無法自動取得，請手動填寫。',
   },
+  qqofficial: {
+    createBinding: '一鍵掃碼綁定 QQ 機器人',
+    scanQRCode: '請使用手機 QQ 掃描以下 QR Code，在「QQ 機器人助手」中授權綁定',
+    waitingForScan: '等待掃碼中',
+    bindSuccess: '綁定成功！AppID 與密鑰已自動填入',
+    bindFailed: '綁定失敗',
+    connecting: '正在連線 QQ 服務...',
+    retry: '重試',
+    tokenNote: 'Token 欄位目前介面卡未使用，留空即可。',
+    boundBy: '由 QQ 用戶 {{openid}} 掃碼綁定',
+  },
   pluginPages: {
     selectFromSidebar: '從側邊欄選擇一個插件頁面',
     invalidPage: '無效的插件頁面',