Problem:

• /answer can select openai-codex/gpt-5.1-codex-mini or anthropic/claude-haiku-4-5 even when those providers are not actually configured.
• This happens because selectExtractionModel() treats modelRegistry.getApiKeyAndHeaders(model).ok as “model is available”.
• In pi, auth.ok === true only means auth resolution did not error; it does not guarantee that credentials exist.
• As a result, /answer may choose an unconfigured preferred model instead of the current model, and extraction then fails.

Root cause:

• getApiKeyAndHeaders() can return { ok: true, apiKey: undefined, headers: undefined } for an unconfigured provider.
• The extension was using that as the availability check for Codex/Haiku.

Fix:

• In selectExtractionModel(), require modelRegistry.hasConfiguredAuth(model) before considering Codex or Haiku selectable.
• Keep the existing fallback behavior to the current model unchanged.