tui: new sessions retain Thinking, User, or Asking as their title #57043

Activity

1. opencode-agent commented on Oct 19, 2026

   

   opencode-agentBot

   on Oct 19, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the clear report, @pgasparini-byte.

   In v2 the session title comes from a background request to a model. OpenCode keeps the first line of what that model writes. The TUI doesn't put labels like Thinking into the title itself, so Thinking, User or Asking were most likely written by the model, and that depends a lot on which model it is. To narrow it down, could you share:

   1. The provider and model you use for sessions, e.g. anthropic/..., openai/..., or a local or self-hosted model. Please say whether it is a reasoning or thinking model.
   2. Any title or small-model setting in your config, e.g. an agents.title override or a small model setting. If you have none, just say so.
   3. One concrete example: the exact first prompt you sent and the title the session ended up with.

   You can reply here with the details, or edit the issue. If you edit it, please also leave a comment saying you've updated it, so we know you're done.

   Related issues, ordered by likelihood that they cover the same problem:

   • Title agent produces conversational replies instead of session titles (V2) #54185: on some models the title agent writes a conversational reply instead of a title
   • Bug: V2 session titles include model commentary #51997: the model's planning or commentary text ends up in the title, because only the first line is kept (fix proposed in fix(core): exclude commentary from generated session titles #52006)
   • session: title generation request has no output limit and runs away on local thinking models #54401: title generation has no output limit and misbehaves on thinking models (fix proposed in fix(core): bound session title generation with a small output limit #54411)
2. opencode-agent

   added

   triaging

   on Oct 19, 2026
3. pgasparini-byte commented on Oct 19, 2026

   

   pgasparini-byte

   on Oct 19, 2026

   Author

   More actions

   Thanks — I checked the local configuration and the persisted session data.

   Configuration

   • Session model: litellm-team/auto
   • Small model: litellm-team/lite
   • Provider: a self-hosted LiteLLM gateway (litellm.turbocrowd.net)
   • agents.title: no override configured
   • Plugins: none configured

   I cannot confirm the exact downstream model selected by the LiteLLM router for every title request, but the affected titles strongly indicate that a reasoning/thinking-capable model is being used for title generation, or that its reasoning output is not being stripped before OpenCode stores the first line.

   Concrete persisted example

   First prompt:

   /mcps
   

   Stored session title:

   <think>The user message is "/mcps" - a command likely related to MCP servers. Need a brief title....
   

   This is not a user-facing title. It is clearly internal reasoning from the title-generation request, persisted verbatim as the session title.

   The same local session history also contains another title beginning with:

   <think>The user is asking in Italian about Qonto (a business banking/finance tool) - wanting help...
   

   This likely explains the user-facing Thinking, User, and Asking labels reported earlier: they appear to be short or first-line fragments emitted by the title model, not TUI activity labels.

   This seems especially close to #51997 and #54401. A title-generation path that strips <think>...</think> / reasoning output and validates the remaining title before persistence would address this case.

   No agents.title override is currently configured. If you can confirm the recommended configuration for a dedicated non-reasoning title model in a LiteLLM-routed setup, I can test it as a workaround.
4. opencode-agent

   removed

   triaging

   on Oct 19, 2026
5. opencode-agent commented on Oct 19, 2026

   

   opencode-agentBot

   on Oct 19, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the extra details, @pgasparini-byte. I can confirm this.

   I used a simulated title model that returns its reasoning inline as plain text, <think>The user message is "/mcps" … Need a brief title.</think> followed by MCP servers list, which is how your LiteLLM gateway appears to pass it through. After the first prompt, the session tab title became <think>The user message is "… instead of MCP servers list. The recording and screenshot below show it, and the script below reproduces it.

   It happens on 2.0.32 and still happens on the latest dev (cec70f3). v2's title code keeps the first non-empty line of the title model's output and doesn't remove <think>…</think> blocks first. v1 used to strip them.

   My simulation stands in for the gateway, so I couldn't confirm that your LiteLLM router sends the reasoning as plain text. Your stored titles suggest it does.

   Until this is addressed, a possible workaround is to point the title agent at a model that doesn't emit reasoning, e.g. a non-reasoning LiteLLM model as small_model, or an agents.title override with such a model. Turning reasoning off for that model on the LiteLLM side should also work.

   Related: #51997 (model commentary taken as the title) and #54401 (no output limit on title requests).

   recording.mp4

   

   Reproduction script (drive.ts)

   // Repro for anomalyco/opencode#57043: title model output containing inline
   // <think>...</think> reasoning (as some OpenAI-compatible gateways such as
   // LiteLLM return it) is saved verbatim as the session title.
   import { Effect } from "effect"
   import { Llm, OpenCodeDriver } from "opencode-drive"
   
   const TITLE_OUTPUT =
     '<think>The user message is "/mcps" - a command likely related to MCP servers. Need a brief title.</think>\nMCP servers list'
   
   export default OpenCodeDriver.use(
     {
       ...(process.env.OC_DEV ? { opencode: { dev: process.env.OC_DEV } } : {}),
       tui: { recording: true },
     },
     ({ ui, llm }) =>
       Effect.gen(function* () {
         // Simulated title model streams reasoning inline in the text channel.
         yield* llm.title(() => Effect.succeed(TITLE_OUTPUT))
         yield* llm.queue(Llm.text("Here are your MCP servers."))
         yield* ui.submit("list my mcp servers")
         yield* ui.waitFor("Here are your MCP servers.")
         yield* Effect.sleep("3 seconds")
         yield* ui.screenshot("after-title")
         yield* ui.waitFor("<think>")
         yield* ui.screenshot("think-title")
       }),
   )

   Run it with opencode-drive: opencode-drive run ./drive.ts. It drives opencode from your PATH; set OC_DEV to the path of an OpenCode checkout to run that from source instead.
6. opencode-agent

   added

   reproduced

   on Oct 19, 2026