From 041f143abfc074cca81ccd3702a89e196fb6e88c Mon Sep 17 00:00:00 2001 From: Engel Nyst Date: Fri, 30 Oct 2026 07:07:20 +0100 Subject: [PATCH] docs(sdk): explain automatic conversation title prompts Co-authored-by: openhands --- sdk/guides/agent-server/overview.mdx | 31 ++++++++++++++++++++++++++++ 1 file changed, 31 insertions(+) diff --git a/sdk/guides/agent-server/overview.mdx b/sdk/guides/agent-server/overview.mdx index 6e4c40e53..89cb1f8c5 100644 --- a/sdk/guides/agent-server/overview.mdx +++ b/sdk/guides/agent-server/overview.mdx @@ -57,6 +57,37 @@ Same interfaces as local: */} +## Automatic conversation titles + +The Agent Server generates a title from the first user message when `autotitle` +is enabled (the default). A configured `title_llm_profile` selects the title +model; otherwise the server uses the agent's LLM, with message truncation as a +fallback when generation fails. + +To customize the title-generation instructions, set `prompt` in the conversation +creation request. Use `{conversation_content}` for the first user message and +`{max_length}` for the output title-length limit. If the conversation placeholder +is absent, the server appends the user message to the prompt. Empty or unset +prompts retain the default instructions. + +```json +{ + "autotitle": true, + "prompt": "Use an action-first title of at most {max_length} characters. Describe the requested task without implying completion. Return only the title.\n\n{conversation_content}" +} +``` + +These fields are part of the normal conversation creation request; include the +required agent and workspace configuration as well. The title call receives the +complete first user message and accepts prompts longer than 2,000 characters. +Model context-window and provider request limits still apply. The default output +title limit remains 50 characters. A custom prompt replaces the default user +instructions, including their emoji requirement, so include any desired emoji +rule explicitly. + +Title generation uses a separate LLM call. Skills loaded into the main agent's +context do not automatically reach it; include relevant instructions in `prompt`. + ## Architecture Overview Remote Agent Servers follow a simple three-part architecture: