Bug: --browser-thinking-time extended can select non-Pro GPT-5.5 instead of GPT-5.5 Pro

[oirehT]

Hi, when setting thinkingTime, with the intention of forcing 5.5 Pro Extended in the browser, the selector picks Extended for the regular 5.5 model instead. When just using the model "gpt-5.5-pro" with no thinkingTime, it respects the choice made on the website and correctly selects the Pro model, but the model selector is broken for 5.5 Pro when trying to explicitly use Extended. If you want I can try to look into this issue myself and submit a pr, if you like. Cheers and thanks.

Below is a minimal bug report from codex:

Summary

When using Oracle browser mode with model: "gpt-5.5-pro" and browser.thinkingTime: "extended", Oracle appears to select
normal GPT-5.5 and then set that model to Extended, instead of selecting GPT-5.5 Pro.

Config

{
  model: "gpt-5.5-pro",
  engine: "browser",
  browser: {
    chatgptUrl: "https://chatgpt.com/g/g-p-.../project",
    manualLogin: true,
    modelStrategy: "select",
    thinkingTime: "extended",
  },
}

Expected

Oracle selects GPT-5.5 Pro and, if possible, sets Pro thinking time to Extended.

Actual

Oracle selects normal GPT-5.5, then applies Extended thinking time there. The resulting ChatGPT UI is not using the Pro
model.

Workaround

Remove browser.thinkingTime and only use:

  {
    model: "gpt-5.5-pro",
    engine: "browser",
    browser: {
      manualLogin: true,
      modelStrategy: "select",
    },
  }

This avoids touching the thinking-time picker and prioritizes selecting the Pro model.

Version

Oracle CLI 0.10.0

[pdurlej]

I think this is now covered by the open browser hardening work, but I would keep the issue open until the relevant PR lands.

Direct fix: #176 scopes --browser-thinking-time extended to the selected GPT-5.5 Pro row and includes live DOM verification that model-switcher-gpt-5-5-pro is checked while regular model-switcher-gpt-5-5-thinking is not.

Related hardening: #177 adds post-selection validation for gpt-5.5-pro so Oracle fails loudly if the UI resolves to Thinking instead of Pro/Extended, plus CI-covered diagnostics for future selector drift.

So the intended close path is: merge #176 and/or #177, then this issue should be safe to close as fixed.

[steipete]

Fixed on main.

The Pro/Extended mis-selection path is covered by 8dc0b411 (#177): Oracle validates that gpt-5.5-pro resolves to a Pro/Extended label and treats gpt-5.5-pro + browserThinkingTime: "extended" as Pro Extended via model selection, avoiding the regular GPT-5.5 thinking-effort picker. ecb2d320 also prevents model selection from treating per-row thinking-effort controls as model options.

Live proof from the current code path: cookie/session attach passed, Model picker: Pro, Thinking time: Pro Extended (via model selection), and the run returned the requested marker. Main CI is green: https://github.com/steipete/oracle/actions/runs/25417634473