Summary

This PR fixes three bugs that prevent custom providers from working correctly when using the v12+ providers: configuration format with base_url field and list-format models.

Problem

Users configuring custom providers like this:

providers:
  alibaba-coding:
    base_url: https://coding.dashscope.aliyuncs.com/v1
    api_key: sk-xxx
    models:
      - qwen3.5-plus
      - qwen3-max-2026-01-23
      - qwen3-coder-next

Experience three issues:

1. CLI rejects the provider - hermes chat --provider alibaba-coding fails with "invalid choice"
2. Model picker shows duplicate entries - Provider appears twice, one with "0 models"
3. Models not loaded - Custom provider shows "(0 models)" in picker UI

Root Causes

Bug 1: Hardcoded provider choices in CLI (main.py:6427)

The --provider argument used a hardcoded list of builtin providers, rejecting any custom providers defined in config.

Bug 2: Missing base_url field check (model_switch.py:1043)

The user_providers section only checked api and url fields, but the newer config format uses base_url (matching OpenAI-compatible API conventions).

Bug 3: Models field type mismatch (config.py:1831)

The normalization function only accepted dict-format models:

if isinstance(models, dict) and models:

But list-format (which is more intuitive for simple provider configs) was rejected, causing models to be dropped during normalization.

Changes

1. Dynamic provider loading (main.py)

• Added _get_provider_choices() function that loads config at argparse setup time
• Merges builtin + custom providers from config.yaml
• Falls back gracefully if config loading fails

2. Support base_url field (model_switch.py)

• Updated API URL lookup to check base_url first, then api, then url
• Maintains backward compatibility with legacy formats

3. Accept list-format models (config.py)

• Changed type check from dict to (dict, list)
• Both formats now work correctly

Testing

Before fix:

$ hermes chat --provider alibaba-coding
error: argument --provider: invalid choice: 'alibaba-coding'

# Model picker shows:
# - alibaba-coding (8 models)
# - alibaba-coding (0 models)  ← duplicate with no models

After fix:

$ hermes chat --provider alibaba-coding
# Works correctly

# Model picker shows:
# - alibaba-coding (8 models)  ← single entry, all models loaded

Impact

• ✅ Zero breaking changes - all existing configs continue to work
• ✅ Enables proper support for v12+ provider configuration format
• ✅ Consistent field priority across codebase (base_url → api → url)
• ✅ Supports both dict and list formats for models field

Checklist

• Fixes affect only custom provider loading logic
• Backward compatible with legacy config formats
• Tested with real custom provider configuration
• No new dependencies added