OpenClaw closed out Sunday with a pair of Telegram delivery fixes aimed at messages that looked sent but were not fully visible to users. PR #97286 preserves literal reasoning-looking tags, while PR #97304 fixes chunked reply delivery loss in Telegram groups.
Both PRs were merged minutes before the nightly cutoff on June 28, and both carried P1 priority plus Telegram-visible proof. Together, they make Telegram replies less surprising when the model output contains angle-bracket text or when OpenClaw has to split a long response into multiple messages.
Literal Text Should Stay Literal
The first fix addresses a parsing boundary. Telegram already sends normal reply text through an HTML-escaped path, but the reasoning-lane coordinator was still guessing that unflagged answer text containing tags such as <think> or <thinking> should be treated as private reasoning.
That created a product bug: ordinary answer text could be removed from the visible answer path before Telegram's HTML renderer had a chance to escape it safely. The new behavior trusts the typed delivery contract. Only payloads explicitly marked as reasoning go through the reasoning lane; unflagged answer text stays answer text.
For users, that means a response that literally discusses <think> tags should show the text instead of disappearing into a channel-local reasoning guess.
Long Replies Keep Their First Chunk
The second fix targets long Telegram group responses with streaming.mode: "off" and replyToMode: "first". In that configuration, Telegram could accept multiple Bot API sends but render the first long formatted native-reply chunk as unsupported in some clients. Users would see later chunks, but not the beginning of the answer.
OpenClaw now centralizes chunk reply policy so multi-part first replies consume the reply target without attaching native reply params to every chunk. Single-chunk replies can still use native reply-to behavior. The fix also keeps regular group reply threads from emitting message_thread_id as if they were forum topics.
That is an important distinction for Telegram operators because group chats, forum topics, and direct messages all have different threading semantics.
What Operators Gain
The practical result is better delivery fidelity:
- Literal reasoning-looking text remains visible when it is ordinary answer text.
- Real reasoning payloads still use the separate reasoning lane.
- Long final answers in Telegram groups should show the beginning, middle, and end.
- Multipart delivery receipts now include the relevant Telegram message IDs for bookkeeping.
Validation
The literal-tag fix included focused tests for Telegram sending, reasoning-lane coordination, formatting, and OpenAI transport streaming. The chunked-delivery fix included 396 passing focused Telegram tests across delivery, helper, bot creation, send, and channel adapter paths.
The real-behavior proof is especially useful here. For the chunked reply case, the PR reports that before the fix the first chunk rendered as unsupported and the BEGIN marker was missing. After the fix, three chunks rendered as normal text with BEGIN, MID1, MID2, and END all present.
For a chat-first assistant, that is the right bar: not just "the API accepted it," but "the user can actually read it."