feat(telegram): streaming response via sendMessageDraft for private chats by El-Chiang · Pull Request #1433 · HKUDS/nanobot

El-Chiang · 2026-03-02T14:42:09Z

Summary

Add real-time streaming response support for Telegram private chats using the sendMessageDraft API (Bot API 9.3+, opened to all bots in 9.5).

What it does

When a user messages the bot in a private Telegram chat, the bot now streams its response in real-time — the user sees text appearing character by character, similar to ChatGPT's typing effect.

How it works

Detection: _should_stream_private_telegram() checks if the incoming message is from a Telegram private chat
Streaming: During the first LLM turn, tokens are streamed via on_stream_chunk callback
Draft updates: sendMessageDraft is called with throttling (350ms interval) to push partial text to the user
Finalization: When generation completes, draft is cleaned up and final message is sent via normal sendMessage
Fallback: If sendMessageDraft fails, silently falls back to normal non-streaming behavior

Key design decisions

Private chats only: sendMessageDraft is a Telegram API limitation — it only works in private chats, not groups
First turn only: Streaming only applies to the first LLM generation turn; subsequent tool-call turns use normal sending
Throttled updates: 350ms minimum interval between draft pushes to avoid rate limiting
Graceful degradation: Any draft API failure silently falls back to standard message sending

Files changed

File	Changes
`nanobot/agent/loop.py`	Add streaming detection, stream chunk callback, stream_id management
`nanobot/channels/telegram.py`	Full `sendMessageDraft` implementation with draft state management
`nanobot/providers/base.py`	Add `stream` and `on_stream_chunk` params to `chat()` interface
`nanobot/providers/litellm_provider.py`	Implement streaming in LiteLLM provider
`nanobot/providers/custom_provider.py`	Implement streaming in custom provider
`nanobot/providers/openai_codex_provider.py`	Signature update for compatibility

References

Telegram Bot API 9.5 changelog — sendMessageDraft opened to all bots
sendMessageDraft docs

…hat support - Remove global litellm.api_base setting that interfered with provider routing when env vars like ANTHROPIC_BASE_URL were present - Pass api_key explicitly per-request to prevent env var interference - Add group chat support for DingTalk channel (auto-detect conversationType, route to groupMessages/send API for group, oToMessages/batchSend for private) - Pass metadata through OutboundMessage so DingTalk send() knows the chat type - Add providers/ module documentation (PROVIDERS.md) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add conversation_title extraction from DingTalk messages to improve group chat context. Group messages now include a [群:title] prefix and sender name so the agent knows who is speaking in which group. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

- Add MCPManager for connecting external MCP servers and registering tools - Add agent-browser, search-sessions, and send-ding skills - Add subagent support with session management - Update DingTalk channel with conversation title in messages - Fix cron jobs.json to preserve Chinese characters (ensure_ascii=False) - Add MCP server configuration to schema and CLI Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

External devices (e.g. M5Stack) can POST to /api/chat and block until the agent replies, completing one conversation turn per HTTP round-trip. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

DingTalk channel can now upload and send images via OApi media endpoint. Message tool gains a media parameter for attaching local image files and passes through channel metadata for request-response correlation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Track finish_reason through iterations and provide informative messages when the model returns empty content or hits the iteration limit. Increase default max_tool_iterations from 20 to 50. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Log compact previews of model responses including finish_reason, tool call names, and truncated content for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…oding Add 30s timeout to MCP server connections to prevent a slow/unreachable server from blocking the entire gateway startup. Also fix HTTP channel JSON responses to output readable UTF-8 instead of escaped unicode, and default HttpConfig host to 127.0.0.1 for security. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

… misc improvements - Add thinking (enabled/adaptive) and effort (low/medium/high/max) params to AgentLoop and LiteLLM provider - Add configurable memory_daily_subdir for daily notes path - Rewrite CLAUDE.md in Chinese, streamline content - Improve DingTalk group message detection by chat_id prefix fallback - Update system prompt identity and simplify instructions - Add journal skill filtering rules for sensitive content Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Previously, tool calls and results were only kept in memory during a single message processing cycle and discarded when saving to session. This caused the model to "forget" it had used tools when the user asked follow-up questions in the next turn. Now prepends a <tool_use_summary> block to the saved assistant message with truncated args (100 chars) and results (200 chars), so the model retains awareness of its tool usage across conversation turns. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Write daily rotating logs to ~/.nanobot/logs/ via loguru file sink, enabled automatically in both agent and gateway commands. Expose log path in system prompt so the assistant can search logs when needed. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…in assistant content Previously tool use summaries were prepended to assistant content as <tool_use_summary> XML tags, which caused the model to mimic the format in its replies. Now persists them as a synthetic _tool_use_summary tool call + tool result pair so the model retains tool awareness across turns without polluting its output. Also: - Handle empty response after message tool as normal completion - Add tool_use_log tracking to _process_system_message - Update system prompt with restart log path and memory reminder - Minor journal skill wording tweaks Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add meme/sticker support to journal skill with usage guidelines - Add INFO log for tool use summary when saving to session Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add _send_with_media() for sending photos via Telegram Bot API - Support single photo with send_photo and album with send_media_group - Add reply_to_message_id support for all message types - Refactor send logic into _send_text() and _send_with_media() - Add long message splitting (4096 char limit) - Improve incoming message handling with reply context

- Add _contains_silent_marker() to detect [SILENT] in model output - Suppress outbound message when model outputs [SILENT] - Update system prompt with message dedup rules - Prevents double-sending when message tool already sent the reply

- Add allow_bot_messages and allow_bot_from config options - Skip webhook messages and self-bot messages - Support allowlist-based bot message filtering - Track bot's own user ID from READY event

Co-authored-by: Codex <codex@openai.com>

… timeout fix

Co-authored-by: Codex <noreply@openai.com>

…on and heartbeat

Heartbeat and cron notification messages were written to isolated sessions ("heartbeat" / "cron:{job_id}"), so when users replied to a notification, the agent had no context of what it had sent. Changes: - Add _route_session_key() and _cron_execution_session_key() helpers - Heartbeat: use channel:chat_id as session key instead of fixed "heartbeat" - Cron (deliver=True): use target user session instead of cron:{job_id} - Cache heartbeat_target to prevent target drift between execute and notify - Add 4 unit tests for routing logic Closes: specs/2026-02-28_heartbeat-cron-history-injection.md

…tion

…stream-0228

…nking mode, cron improvements

…hats - Add sendMessageDraft support for real-time streaming in Telegram private chats - Implement draft state management with throttling (350ms interval) - Add graceful fallback when sendMessageDraft fails - Stream only first LLM turn (tool calls use normal send) - Provider layer: add stream + on_stream_chunk callback to chat() - Finalize draft before sending final message to avoid duplicates Refs: Telegram Bot API 9.3+ sendMessageDraft method

wenjielei1990 · 2026-03-03T04:49:43Z

Nice feature...Also looking into the same feature, the streaming capability.

But given the large number of file get changed (62 files), and many Chinese word in this commit, I wonder if this will ever get merged, given the limited bandwidth of the reviewer.

El-Chiang and others added 30 commits February 9, 2026 17:36

Merge branch 'main' of https://github.com/HKUDS/nanobot

0ea2a82

feat: add skills

f5a1cc4

feat: add HTTP channel for synchronous request-response communication

5b31d6c

External devices (e.g. M5Stack) can POST to /api/chat and block until the agent replies, completing one conversation turn per HTTP round-trip. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

feat: add LLM response summary logging

a3c219f

Log compact previews of model responses including finish_reason, tool call names, and truncated content for easier debugging. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs: add MCP server documentation to README

84bf5a5

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

chore: add CLAUDE.md and feature specs

bc67597

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Merge branch 'main' of https://github.com/HKUDS/nanobot

0c1741e

feat: add meme support to journal skill and log tool use summary

d7905cd

- Add meme/sticker support to journal skill with usage guidelines - Add INFO log for tool use summary when saving to session Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

🐱 kaguya was here (bot PR)

c1e9462

chore: update gitignore to exclude pem files and ds store

bb1de31

fix: harden litellm fallback and realign session history windows

4e36ab0

feat: add bot message filtering for Discord channel

5b8216c

- Add allow_bot_messages and allow_bot_from config options - Skip webhook messages and self-bot messages - Support allowlist-based bot message filtering - Track bot's own user ID from READY event

docs: add Discord bot message filtering config to README

fdd6257

feat: update ignore

1c5c404

Merge branch 'kaguya-bot-pr'

4de74ae

feat: 支持发送 sticker

18ec325

feat: skills 迁移

33cbbfa

fix: stabilize telegram message_time handling and logging previews

27bf8cb

Co-authored-by: Codex <codex@openai.com>

El-Chiang and others added 24 commits February 20, 2026 11:02

merge upstream/main: pin deps, codex routing fix, feishu media, shell…

e1782eb

… timeout fix

fix: disable consolidation temporarily, minor cleanup

d5d5135

merge: sync upstream/main and resolve conflicts

3fe2d44

fix(loop): treat placeholder empty responses as blank output

ffdb4b2

fix(loop): suppress empty fallback text on external channels

36d49e4

feat: 增加 dingtalk 不走代理配置

62f8983

fix: 修复钉钉消息预览功能

8544d8b

refactor: remove yuque-reply skill and improve telegram error handling

15877a2

refactor: move Current Time section to end of system prompt

1ae4739

feat: 提示词优化

bfc98b3

merge: sync upstream/main and resolve conflicts

4f54418

chore: 更新 spec 格式

10897fb

feat: normalize litellm usage cache metrics

32fbac2

Co-authored-by: Codex <noreply@openai.com>

feat: support heartbeat-specific model and interval config

019b1e5

feat: add dedicated model and provider support for memory consolidati…

0c34cbd

…on and heartbeat

Merge pull request #3 from El-Chiang/fix/heartbeat-cron-history-injec…

2f5520a

…tion

feat: add codex command description

24f2c85

Merge branch 'main' of https://github.com/El-Chiang/nanobot

ed32986

Merge branch 'main' of https://github.com/HKUDS/nanobot into merge-up…

56761f0

…stream-0228

Merge upstream/main (2026-03-01): 39 commits including web proxy, thi…

54d29ff

…nking mode, cron improvements

refactor: add TYPE_CHECKING imports and clean up type annotations

ddc6c6d

fix: delete current_time in user message

32cc5db

github-actions Bot mentioned this pull request Mar 3, 2026

🦞 OpenClaw 生态日报 2026-03-03 duanyytop/agents-radar#46

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(telegram): streaming response via sendMessageDraft for private chats#1433

feat(telegram): streaming response via sendMessageDraft for private chats#1433
El-Chiang wants to merge 69 commits intoHKUDS:mainfrom
El-Chiang:feat/telegram-streaming-draft

El-Chiang commented Mar 2, 2026

Uh oh!

wenjielei1990 commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

El-Chiang commented Mar 2, 2026

Summary

What it does

How it works

Key design decisions

Files changed

References

Uh oh!

wenjielei1990 commented Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants