← all posts site ↗

stopping the telegram message flood

the out-of-the-box hermes telegram experience is unusable for anything beyond testing. every tool call generates a message. every thinking step generates a message. the agent runs a chain of tool invocations for something simple like "check the weather" and your phone buzzes eight times in three seconds. each buzz is a wall of text showing tool inputs, intermediate reasoning, and partial results that no human needs to see. the worst part is that this makes the agent look broken — the messages arrive out of order because some tools are async, so you get a result before the reasoning that produced it, then another reasoning steps that references a result you haven't seen yet.

the fix lives in the hermes config.yaml under the telegram block. there are two settings that control the noise. hide_transcripts for telegram defaults to false — flip it to true and the agent stops broadcasting every tool execution step. no more "running function: get_weather, parameters: {city: austin}" appearing as a separate message. the second is interim_assistant_messages, also per-platform. set both to false and the agent sends exactly one message per interaction: the final answer. the thinking transcript, the tool chain, the intermediate results — all suppressed. the agent still processes everything the same way internally, it just doesn't dump the intermediate state into your chat.

i went further and tweaked the HTTP bridge settings too. the http bridge in hermes controls how the agent communicates with the telegram API, not just content filtering. i lowered the message split threshold and enabled message chunking with a smaller max size so the final response arrives as digestible bubbles instead of one massive wall of text. the combination of suppressed intermediates and smarter message chunking turned the bot from a spammer into something that feels like a normal conversation partner. i can now ask it to research something, put my phone down, and come back to one well-structured response instead of a notification log that looks like a debug console.

the lesson is that tool-output-as-conversation is a terrible default for any async messaging platform. discord handles the flood better because it has threads and the notifications are less aggressive, but telegram's notification model punishes multi-message responses. the right approach for mobile messaging is output coalescing — buffer all tool results, build the complete response, and send it once. hermes supports this now but you have to explicitly configure it. i'm considering writing a skill that auto-configures these settings when the telegram platform is detected so i don't have to remember them on every setup.