Problem

Bot API 9.4 introduced DM forum topics, but the upstream Telegram adapter has three issues that make new DM topics behave poorly:

1. Topics are invisible on mobile/desktop clients for minutes after creation. The forum_topic_created service message alone does not force Telegram clients to refresh their topic tree — without a real message in the new topic, Desktop / iOS / Android may not render it until the next sync cycle (or app restart).

2. User-created topics are never seen by Hermes. connect() registers PTB handlers for filters.TEXT, filters.COMMAND, filters.LOCATION, and media — none of which match forum_topic_created service updates. When a user creates a topic in the Telegram UI, Hermes silently drops the event and never learns the topic name. The forum_topic_created check inside _build_message_event is effectively dead code for this path.

3. (Quality-of-life) Telegram clients auto-name new DM topics from the verbatim first user message, producing long sentence-shaped names (e.g. Can we discuss next week's meeting agenda with ABC Company?) instead of short labels.

Fix

A. Seed message — _create_dm_topic and _cache_dm_topic_from_message now post a short 📌 <name> message into the new thread. This forces all clients to sync the topic into their UI immediately.

B. forum_topic_created handler — connect() registers a new handler on filters.StatusUpdate.FORUM_TOPIC_CREATED so user-initiated topics reach _cache_dm_topic_from_message.

C. Opt-in LLM rename — new config flag platforms.telegram.extra.auto_rename_dm_topics (default false). When enabled, the auxiliary title-generation model produces a clean 3-7 word label and editForumTopic is called.

About the rename model

The rename uses call_llm(task="title_generation", ...), the same auxiliary slot that session-title generation already uses. With the default auto routing it follows the main provider — there is no hard dependency on any specific model vendor. Operators who already use Claude / GPT / Gemini / etc. as their main model get rename-via-that-model for free; operators who use a reasoning model that emits <think>...</think> traces (DeepSeek-R1, Qwen QwQ, and similar) get a defensive cleaner that strips closed and unterminated traces plus any residual XML-like markup, so a bad LLM response can never land as a topic name.

Compatibility

• A and B are pure bug fixes — no behaviour change for users who do not use DM topics; for users who do, topics now appear immediately on every client and user-created topics are tracked.
• C is opt-in (auto_rename_dm_topics: false by default), so cost-conscious deployments are unaffected.

Tests

24 new test cases added to tests/gateway/test_dm_topics.py. Full suite: 59 passed in 1.23s. New cases cover:

• _clean_dm_topic_title (11 cases): closed / unterminated <think>, alternate <thinking> tag, residual XML, multi-line scratchpad, quotes / brackets, Title: prefix, trailing punctuation, length cap, residual < / > rejection, empty input. Some fixtures use non-Latin script to exercise the cleaner against e.g. Chinese topic names.
• _send_dm_topic_seed_message (5 cases): name in seed text, long-name truncation, empty-name fallback, no-bot no-op, send-error swallow.
• _create_dm_topic (2 cases): seeds on success, no seed on failure.
• _cache_dm_topic_from_message (4 cases): schedules seed for new entry, idempotent on re-cache, skips rename when flag off, schedules rename when flag on.
• _handle_forum_topic_created (2 cases): caches topic from service message, ignores non-topic messages.
• _maybe_rename_dm_topic (4 cases): applies generated title via editForumTopic, skips on empty / unchanged title, swallows edit errors.

Manual verification

Tested in a live deployment (two Hermes profiles running this code as a startup-time monkey-patch hook before the upstream change). New DM topics:

• Appear immediately on Telegram Desktop, iOS, and Web (previously invisible for several minutes on at least one client per topic).
• Receive a 📌 <name> seed message that survives the rename cycle.
• With auto_rename_dm_topics: true, long sentence-shaped auto-names get renamed to short labels via the configured auxiliary model. Reasoning-model <think> traces never leak into the topic name thanks to the cleaner.

Docs

Updated website/docs/user-guide/messaging/telegram.md:

• Added step 5 to "How it works" describing the seed message
• Updated the "topics created outside the config" tip to note that forum_topic_created is now properly handled
• New "Auto-rename via LLM (opt-in)" section documenting auto_rename_dm_topics and the auxiliary-model routing

Commits

1. dcd2d06 — fix(telegram): seed DM topics, register forum_topic_created handler, optional LLM rename
2. 8932688 — docs(telegram): use English examples and clarify auxiliary-model independence