How to Configure OpenClaw Streaming and Response Chunking

OpenClaw connects 22+ messaging channels to a single AI gateway, but response delivery is not a single pipeline. The gateway runs two independent streaming layers that solve different problems.

A 2024 Forrester study found that a 2-second delay in chatbot responses can reduce task completion rates by 15 to 20 percent. For AI agents that produce multi-paragraph answers, the delay compounds. Users stare at a blank chat bubble while the model finishes generating, then receive a wall of text all at once. OpenClaw's two-layer approach addresses both sides of this problem.

Block streaming waits for the model to finish a logical chunk (a paragraph, a code block, a sentence) and sends it as a normal channel message. The user's chat history looks clean because each message is a complete, properly formatted unit. Block streaming is the structural layer: it controls how a long response gets split into separate messages.

Preview streaming tackles perceived latency. While the model is still generating, the gateway sends a temporary message and continuously updates it with new tokens. On Telegram, this means editing the same message via editMessageText. On Discord, it uses automation hooks edits. On Slack, it can use the native streaming API. The preview vanishes once the final block-streamed messages land.

These layers are independent. You can run block streaming without preview, preview without block, both together, or neither. Per-channel overrides let you tune each platform separately, which matters because a Discord server with inline embeds behaves differently from a WhatsApp thread with a 4,000-character message cap.

The EmbeddedBlockChunker is the engine behind block streaming. It buffers the model's output and decides where to split it into separate messages. Understanding its algorithm saves you from trial-and-error tuning.

The chunker uses a low/high bound system. It will not emit a chunk until the buffer reaches minChars (the low bound), and it prefers to split before maxChars (the high bound). If the buffer exceeds maxChars with no natural break point, it forces a hard split.

Between those bounds, the chunker looks for the best break point using a preference hierarchy: paragraph break first, then newline, then sentence boundary, then whitespace. You control this order with the breakPreference setting.

Break preference hierarchy (default order):

1. Paragraph (double newline)
2. Newline (single newline)
3. Sentence (period, question mark, exclamation followed by space)
4. Whitespace (any space character)
5. Hard break (forced split at maxChars)

The chunker also handles code fences intelligently. If it must split inside a fenced code block, it closes the fence at the split point and reopens it in the next chunk. This keeps markdown rendering valid across all channels.

One constraint to plan for: maxChars is clamped to the channel's textChunkLimit. If you set maxChars: 6000 but WhatsApp's limit is 4,000 characters, the chunker will use 4,000. This happens silently, so check your channel limits before tuning chunk sizes.

Each messaging platform has different constraints: message length limits, edit support, and native streaming capabilities. OpenClaw lets you override streaming settings per channel through its configuration system, so you can optimize delivery for each platform independently.

Discord: preview plus block streaming

Discord supports message editing, which makes it a strong fit for preview streaming. Users see tokens arriving in a temporary message while the model generates, then block streaming sends the final formatted messages as permanent posts. You can also cap the number of lines per message to prevent enormous embeds from dominating the channel. The combination of both layers gives Discord users the best of both worlds: real-time feedback during generation and clean, readable output when the response is complete.

Telegram: preview streaming with tool progress

Telegram's message-editing API makes preview streaming work well here too. The useful addition for Telegram is tool progress: users see status indicators like "searching web" or "reading file" while the agent works through tool calls. This is especially valuable for agents that run multiple tools per response, because without progress indicators the user just sees a static "typing" state for several seconds.

WhatsApp: block streaming only

WhatsApp does not support editing sent messages, which rules out preview streaming entirely. Block streaming is the only option. The important constraint is WhatsApp's 4,000-character message limit. OpenClaw's chunker respects per-channel character caps, so it will split long responses into separate messages that fit within the limit. Set the break preference to paragraph or newline boundaries for cleaner splits.

Slack: native streaming transport

Slack offers its own streaming API that OpenClaw can use directly, bypassing the edit-based preview approach. Native transport delivers smoother incremental updates with lower latency than editing a message repeatedly. Pair it with block streaming and moderately aggressive coalescing to batch short fragments before sending, which keeps the channel from flooding with tiny updates.

Store and version your OpenClaw streaming configs

Free 50GB workspace with semantic search across your config files, agent logs, and response archives. No credit card, no expiration, MCP-ready for your agents.

Start 14-Day Trial

Without coalescing, a fast model can fire dozens of tiny messages in quick succession. The coalescing layer merges consecutive block chunks before sending, turning a stream of fragments into readable messages.

Three parameters control coalescing behavior. The minimum character threshold prevents tiny one-line fragments from shipping as standalone messages. The maximum character cap forces a send when the buffer gets large enough. The idle gap threshold flushes whatever is buffered when no new content has arrived for a set number of milliseconds.

The coalescer also respects your break preference when joining chunks. If set to paragraph mode, chunks are joined with double newlines. Newline mode joins with single newlines. This keeps the merged output formatted consistently.

You can override coalescing per channel. A WhatsApp channel might use aggressive coalescing (high minimum characters, long idle gap) to send fewer, larger messages. A Discord channel might use lighter coalescing so responses feel more interactive. The OpenClaw streaming docs list the exact config keys and their defaults per platform.

One common mistake: setting the idle gap too low on channels with rate limits. If your agent pauses briefly during tool calls, a short idle threshold triggers premature flushes, and the rapid message bursts can hit platform rate limits. Start with a 2-second idle gap and adjust downward only if responsiveness feels sluggish.

Raw streaming can feel robotic. Messages appear at machine speed with perfectly even timing that no human typist would produce. OpenClaw's human delay feature adds randomized pauses between block replies to make the conversation feel more natural.

Three modes are available. The default is off, with no delay and immediate sends. Natural mode adds randomized pauses between 800ms and 2,500ms, mimicking human typing rhythm. Custom mode lets you define the delay range yourself.

Human delay applies between block-streamed messages, not during preview streaming. This is intentional. Preview streaming should feel fast because it is showing work-in-progress. Block streaming delivers the finished product, and a brief pause between messages gives the reader time to process each one before the next arrives.

This setting is more useful than it might sound. In customer-facing bots, perfectly instant multi-message responses can feel overwhelming or robotic. In internal tools where speed matters more than polish, turn it off. Per-channel overrides let you use natural pacing for customer-facing WhatsApp while keeping instant delivery for your team's Slack channel.

The agent's response files, configuration, and prompt templates all benefit from version control. If you use a shared workspace like Fastio to store your OpenClaw configuration alongside agent outputs, every team member can see which streaming settings were active when a particular response was generated. Enable Intelligence Mode on the workspace and you can search your config history by meaning, asking questions like "when did we change the WhatsApp chunk limit?" Fastio's Business Trial includes 50GB of storage and included credits with no credit card required, which covers config files, agent logs, and the response archives that help you tune streaming over time. Agents can read and write config files through the Fastio MCP server, and ownership transfer lets a developer hand the workspace off to a client's team when the bot is ready for production.

Messages arriving out of order on Telegram

Telegram rate-limits message edits. If preview streaming fires updates too quickly, they can arrive out of sequence. Fix this by increasing the idle gap in your coalescing settings, or switch from preview mode to block-only streaming. Block mode sends chunked updates as new messages without editing a draft, which sidesteps the edit rate limit entirely.

Code blocks splitting mid-line on WhatsApp

WhatsApp's 4,000-character limit can force splits inside code fences. The chunker closes and reopens fences automatically, but some WhatsApp clients render the closing fence oddly. Lower the max character threshold to give the chunker room to find a clean break before hitting the platform limit.

Excessive message spam on Discord

If your agent sends 15+ messages for a single response, your coalescing settings are too loose. Increase the minimum character threshold and idle gap timer so the coalescer batches more content per message. Also check the lines-per-message cap; setting it too low forces more splits than necessary.

Preview updates feel laggy on Slack

Enable native streaming transport if you have not already. The edit-based fallback adds round-trip latency per update. Native transport uses Slack's own streaming infrastructure and delivers smoother incremental updates. If native transport is already on, the bottleneck is likely your model's token generation speed, not OpenClaw's delivery layer.

Agent tool calls cause premature message breaks

When an agent runs tools mid-generation, the model pauses output. If the coalescing idle timer is set too low, the coalescer interprets this pause as idle time and flushes the buffer. The result is an incomplete first message followed by the rest of the response in a second message. Switch to message_end flush mode for agents that frequently use tools, or increase the idle threshold above the typical tool execution time.