[AI Agent] Model selector only shows gpt-5.1-mini and ignores the configured model

[trikunai]

Description

When the AI agent is enabled in the OpenCom web UI, the model selector is hardcoded to only show gpt-5.1-mini as an option. This causes two problems: the displayed model does not match the one actually being used by the backend, and users have no way to switch models through the UI.

---

Bug 1 — Model selector is hardcoded

The model dropdown only shows gpt-5.1-mini regardless of what model is configured in Convex environment variables.

• Steps to reproduce: Enable AI agent → open the model selector dropdown
• Expected: Dropdown shows all available models fetched from the OpenAI API, with the currently configured one selected
• Actual: Dropdown only shows gpt-5.1-mini

---

Bug 2 — Selected model does not match what's used

Even though the selector shows gpt-5.1-mini, the actual API calls are being made with a different model (gpt-5-nano, as confirmed via the OpenAI API usage dashboard).

• The UI state is misleading and the Convex env var (OPENAI_MODEL or equivalent) is either being ignored or overridden by a hardcoded default
• Any model selection made in the UI should be persisted and honoured by the backend

---

Proposed fixes

1. On agent enable / settings open, fetch the model list dynamically from GET https://api.openai.com/v1/models and populate the selector with the results (optionally filtered to chat-capable models)
2. Read the currently configured model from the Convex env var and pre-select it in the dropdown
3. Persist the user-selected model back to the Convex env (or equivalent settings store) so the backend uses the correct one

// Example: fetch models and filter to chat-capable ones
const res = await fetch('https://api.openai.com/v1/models', {
  headers: { Authorization: `Bearer ${apiKey}` }
});
const { data } = await res.json();
const chatModels = data
  .filter(m => m.id.startsWith('gpt-'))
  .sort((a, b) => b.created - a.created);

---

Improvement — Support additional / custom OpenAI-compatible providers

Many users bring their own keys from third-party providers that expose an OpenAI-compatible API (e.g. Azure OpenAI, Groq, Together AI, local Ollama). The agent settings should support configuring a custom base URL alongside the API key, so these providers work out of the box without code changes.

Suggested settings fields:

• Provider — dropdown: OpenAI, Custom (OpenAI-compatible)
• Base URL — shown when "Custom" is selected; defaults to https://api.openai.com/v1
• API key — existing field, unchanged
• Model — dynamically fetched from whichever base URL is set

This would make the model selector reuse the same dynamic-fetch logic from Bug 1 fix — just pointed at the custom base URL instead of OpenAI's.

---

Environment

• Interface: OpenCom web
• Model shown in UI: gpt-5.1-mini
• Model used in practice (per OpenAI dashboard): gpt-5-nano
• API key configured via: Convex environment variables