Title generation fails from unavailable gpt-5-nano fallback and possible variant leakage into OpenAI small_model #26181

Summary

There appear to be two related title-generation problems in OpenCode:

1. Unavailable default/fallback small model: OpenCode can attempt to use gpt-5-nano for title/small-model work. For some Codex/OpenAI subscribers, gpt-5-nano is not available, so title generation fails unless the user configures another small_model.

2. Possible main-model variant leakage after configuring a replacement: After setting an explicit OpenAI small_model, for example:

   "small_model": "openai/gpt-5.4-mini"

   OpenCode correctly selects that model for title generation:

   providerID=openai modelID=gpt-5.4-mini small=true agent=title
   

   But title generation can still fail when the main/current agent model is also OpenAI GPT and uses high/xhigh/variant-style reasoning. The title remains New session - ....

The second issue looks like possible main-model variant/options leakage, or another OpenAI title-call isolation bug, rather than incorrect title model selection.

Environment

• OpenCode: 1.14.40
• Upgraded from: 1.14.31
• Provider: OpenAI/Codex subscription
• gpt-5-nano: unavailable on this subscription
• Reproduced with external plugins disabled using --pure

Config Used During Failure

{
  "small_model": "openai/gpt-5.4-mini"
}

No agent.title.model override was set during the failing tests.

Actual Behavior

Case 1: Default/fallback gpt-5-nano

Without explicit small_model, OpenCode can try to use gpt-5-nano for title generation.

For Codex/OpenAI subscribers who do not have access to that model, title generation fails or errors.

Case 2: Explicit OpenAI small_model

After setting:

"small_model": "openai/gpt-5.4-mini"

OpenCode correctly selects that model for title generation:

providerID=openai modelID=gpt-5.4-mini small=true agent=title

However, when the main/current agent model is also OpenAI GPT and uses high/xhigh or variant-style reasoning, title generation can still fail and the session title remains default.

This reproduced in --pure mode, so it does not appear to be plugin-specific.

Control Tests

These controls worked:

• Main OpenAI model + non-OpenAI title model:

  {
    "agent": {
      "title": {
        "model": "google/gemini-3-flash-preview"
      }
    }
  }

  Result: title generated successfully.

• Main OpenAI model without high/xhigh variant + OpenAI title model:

  Result: title generated successfully.

• Non-OpenAI main model + OpenAI title model:

  Result: title generated successfully.

A temporary chat.params diagnostic hook saw title options as:

{ "store": false, "reasoningEffort": "low" }

No visible leak of xhigh, provider-level medium, include, reasoningSummary, or textVerbosity was observed at that hook layer.

Expected Behavior

• OpenCode should not hardcode or blindly fall back to gpt-5-nano when that model is unavailable for the active subscription/provider.
• If small_model is explicitly configured, title generation should use it independently of the main model.
• Main model variants/options must not affect small:true title-generation calls.
• If title generation fails, OpenCode should log or surface why instead of silently leaving New session - ....

Related Issues / PRs

This looks related to the existing class of title/small-model isolation bugs:

• fix: session title generation with OpenAI models. #11678: OpenAI title generation needed store:false.
• [BUG] Session title generation fails silently since v1.3.3 — effort parameter leaks into small model call #20269: effort/variant leaking into title small-model calls.
• fix(session): prevent model options from overriding small model defaults #20582: model/agent options overriding small-model defaults.
• fix: prevent main model thinking variant from applying to small model #6839: prevent main-model thinking variant from applying to small-model calls.
• title agent use not allowed param reasoning_effort: minimal #22796 / Bug: Copilot title generation sends invalid reasoning_effort=minimal to gpt-5-mini #22919: GPT-5-mini title generation sending invalid reasoning_effort=minimal.
• Auto-title generation silently fails when provider has no small model #25344 / fix: ensureTitle falls back to main model when getSmallModel fails #25340: silent title-generation failure / fallback handling around unavailable small models.

This issue seems to combine two related failure modes: unavailable gpt-5-nano fallback for some Codex/OpenAI users, and possible main-model variant/options leakage or isolation failure after switching to an explicit OpenAI small_model.

Current Workaround

Use a non-OpenAI model specifically for the native title agent:

{
  "agent": {
    "title": {
      "mode": "primary",
      "model": "google/gemini-3-flash-preview"
    }
  }
}

This avoids the OpenAI title-generation path while keeping the main OpenAI model unchanged.

Suggested Fixes / Questions

• Validate/default around gpt-5-nano availability before using it for title generation.

• Provide a safer fallback chain for Codex/OpenAI subscribers where gpt-5-nano is unavailable.

• Audit title generation for main-model variant/options leakage into small:true calls.

• Add a regression test for: main OpenAI high/xhigh variant + explicit OpenAI small_model.

• Add debug logging for title stream failure, empty output, or missing text-delta.

• Confirm whether all final OpenAI title-call options are resolved after provider/model/agent/variant merging to safe values such as:

  { "store": false, "reasoningEffort": "low" }