bot

Create a directory under src/server/services/bot/platforms with definition.ts, schema.ts, client.ts, const.ts, and protocol-spec.md files. Implement a ClientFactory returning a PlatformClient, register it in platforms/index.ts, add the UI icon, and add i18n keys for the schema labels.

A POST to /api/agent/webhooks/[platform]/[appId] hits BotMessageRouter, which lazy-loads the bot, merges schema defaults with provider settings, and registers Chat SDK handlers. Events then flow to AgentBridgeService, which executes the agent via AiAgentService and replies through the platform client.

In-memory mode runs execAgent with stepCallbacks in one process, editing a progress message live with a 30-minute timeout. Queue mode posts step and completion webhooks delivered via QStash to /api/agent/webhooks/bot-callback, returning immediately while BotCallbackService handles updates.

Discord, Slack, and Telegram support both markdown and message editing. Feishu and Lark strip markdown but support edits, while QQ and WeChat support neither, so only final plain-text replies are sent.

The router caches loaded bots in memory, but the TRPC update and delete mutations call BotMessageRouter.invalidateBot to clear the cache. If changes do not take effect, verify the invalidation ran and that the gateway client was stopped via GatewayService.stopClient.

GatewayService pushes connect requests to a Redis-backed BotConnectQueue and marks status queued. A cron hitting GET /api/agent/gateway starts persistent providers within a 10-minute window and polls the queue every 30 seconds for new requests.