When streaming was enabled on the gateway, the stream consumer created a single message at the start and kept editing it as tokens arrived. Tool progress messages were sent as separate messages below it. Since edits don't change message position on Telegram/Matrix/Discord, the final response ended up stuck above all tool progress messages — users had to scroll up past potentially dozens of tool call lines to read the answer. The agent already sends stream_delta_callback(None) at tool boundaries (before _execute_tool_calls). The stream consumer was ignoring this signal. Now it treats None as a segment break: finalizes the current message (removes cursor), resets _message_id, and the next text chunk creates a fresh message below the tool progress messages. Timeline before: [msg 1: 'Let me search...' → edits → 'Here is the answer'] ← top [msg 2: tool progress lines] ← bottom Timeline after: [msg 1: 'Let me search...'] ← top [msg 2: tool progress lines] [msg 3: 'Here is the answer'] ← bottom (visible) Reported by SkyLinx on Discord.
14 KiB
14 KiB