流媒体応答

OpenClaw's "streaming" involves three layers of mechanisms that are easy to confuse:

1. LLM token streaming：模型但令牌每个文本返执行（内部实现的详情）。

2. Block streaming: delivers completed text blocks to the channel early (users may see multiple updates).

3. Chunking: splits one reply into multiple messages based on channel character limits (independent of streaming).

块流的目标：在模型完全完成之前向用户显示"部分完成的段落"，尽可能保持段落边界自然。

交换机：

- 全局默认：''agents.defaults.blockStreamingDefault: "on"''

- Channel override: most channels require explicit ''channels.<channel>.blockStreaming: true''

- Telegram is the exception: it also has draft streaming (see below).

調整参数：

- ''agents.defaults.blockStreamingBreak'': when to flush (e.g., ''text_end'').

- ''agents.defaults.blockStreamingChunk''：''minChars''、''maxChars''、''breakPreference''。

- ''agents.defaults.blockStreamingCoalesce'': merges overly dense small blocks to reduce screen flooding.

- ''agents.defaults.humanDelay'': makes delivery look more "human".

Chunking is a "hard split" before delivery:

- Telegram / WhatsApp默认約4000文字

- Discord是2000文字 + 行数限制

- Slack是4000文字

设置：

- ''channels.<channel>.textChunkLimit''

- ''channels.<channel>.chunkMode''（''length'' 或 ''newline''）

分块尝试避免分割围栏代码块，但在极端情况下可能仍需要分割。

Telegram supports "updating the same message" using draft mode in private chat topics/threads. OpenClaw controls this via ''channels.telegram.streamMode'':

- ''"off"'': disables draft streaming, allows block streaming (multiple messages)

- ''"partial"'' (default): uses draft updates in private chat topics (users see one message continuously growing)

- ''"block"'': writes block streaming blocks to draft (reduces multiple messages)

注意：

- Only supported in Telegram private chats + topics/threads.

- Draft streaming and block streaming are mutually exclusive (when enabled, block streaming is disabled).

一般的那理由：

- ''agents.defaults.blockStreamingDefault'' is still ''off''

- You set ''blockStreamingDefault'' but didn't enable ''channels.<channel>.blockStreaming: true'' on the channel

- ''blockStreamingCoalesce'' merged blocks, so they appeared sent at once

- 模型但大机那块一度在输出已执行（flush积分钟没有）

- Telegram draft streaming is enabled (looks like "same message update", not multiple)

- Be careful enabling block streaming in public group chats (screen flooding); prefer larger ''minChars'' or stronger coalesce.

- Block streaming works best in private chats.

- If the channel supports reply threading, block streaming's multiple messages will look more "scattered". Use ''replyToMode'' and channel policy to optimize threading.