session: auto-title saved as raw <think> tag with reasoning small models via ollama (openai-compatible) #53493

Summary

Session auto-titling stores the literal opening think tag (<think>) as the session title whenever the configured small_model streams reasoning into the text channel (ollama @ai-sdk/openai-compatible + a thinking-capable small model). Every new session gets titled <think> — 19 consecutive sessions observed across 2.0.20 and 2.0.24, zero exceptions.

Environment

• opencode version: 2.0.24 (first observed on 2.0.20; v1.18.30 titled the same setup correctly)
• OS: Darwin 27.0.0 arm64 (macOS)
• Terminal: iTerm.app (TERM=xterm-256color, COLORTERM=truecolor)
• Shell: /bin/zsh
• Install/channel: latest, Homebrew tap anomalyco/tap/opencode-v2
• Active plugins: none configured

Reproduction

1. ~/.config/opencode/opencode.json:

   {
     "model": "openrouter/z-ai/glm-5.3-flash",
     "small_model": "ollama/lfm2.5:latest",
     "provider": {
       "ollama": {
         "npm": "@ai-sdk/openai-compatible",
         "options": { "baseURL": "http://localhost:11434/v1" },
         "models": { "lfm2.5:latest": { "name": "LFM2.5 8B MoE" } }
       }
     }
   }

2. Ollama running locally with lfm2.5:latest (LiquidAI LFM2.5, thinking-capable).
3. Start any session and send a first message.

Expected Behavior

The session title is a short summary of the first user message (per the built-in title agent prompt).

Actual Behavior

The session title is the 7-character string <think> — the opening tag of the model's reasoning block. 100% of new sessions affected (19/19 observed over Oct 3–6).

Root cause (from packages/opencode/src/session/prompt.ts, SessionPrompt.ensureTitle):

• The title text is assembled from textDelta events only, then cleaned with

  text.replace(/<think>[\s\S]*?<\/think>\s*/g, "")
      .split("\n").map(trim).find(line => line.length > 0)

• Through the ollama OpenAI-compatible endpoint, LFM2.5 emits its reasoning inline in content, beginning with <think>. In the opencode flow the closing </think> never lands in the text-delta stream, so the closed-pair regex can't strip anything and the first non-empty line — the dangling <think> — becomes the title.
• provider.getSmallModel() correctly resolves small_model from config (confirmed in provider.ts), so the right model is called; only the parsing leaks the tag.

Notes:

• v1.18.30 produced clean titles with the identical model/config, so this is a v2 title-pipeline regression.
• Verified fix on the user side: override the built-in title agent's model in config — "agent": { "title": { "model": "ollama/granite4.1:3b" } } (a non-thinking small model) — which yields clean titles. The built-in agent merge in agent.ts (if (value.model) item.model = Provider.parseModel(value.model)) honors this.

Additional Context

Same parser family but distinct manifestation from #42287 (reasoning-only output -> empty title, silently unset) and #51997 (commentary phases leak): here the closed-pair strip fails entirely because the closing tag never reaches the text-delta stream, so a dangling opener is stored as the title.

Possible fixes: strip a dangling opening <think> when no closing tag is present (e.g. treat text after an unclosed tag opener as the payload, or also accept split("</think>").pop()); or filter reasoning deltas out of the text stream before assembly; or validate the cleaned title against tag-like content and fall back to a truncation of the first user message.

Happy to provide the affected session records or a trace if useful.