Summary

On latest main, the CLI /model picker still shows OpenAI and OpenAI Direct with 0 models in a clean repro, even though those providers have live catalogs available.

This appears to be narrower than the earlier duplicate-row report: the duplicate-row behavior looks addressed on main, but the live count mismatch still reproduces.

Repro environment

• Repo: NousResearch/hermes-agent
• Commit: ce089169
• Context: clean worktree at origin/main
• Config loaded from the user's normal Hermes config/env

Steps to reproduce

1. Check out latest main
2. Load normal Hermes env/config
3. Call list_authenticated_providers() using the current config and compatibility-expanded custom_providers

Repro script:

from hermes_cli.env_loader import load_hermes_dotenv
load_hermes_dotenv()
from hermes_cli.config import load_config, get_compatible_custom_providers
from hermes_cli.model_switch import list_authenticated_providers

cfg = load_config()
providers = cfg.get('providers', {}) if isinstance(cfg, dict) else {}
custom = get_compatible_custom_providers(cfg)
rows = list_authenticated_providers(
    current_provider=(cfg.get('model', {}) or {}).get('provider', ''),
    user_providers=providers,
    custom_providers=custom,
)
for r in rows:
    print(r['name'], r['slug'], r['total_models'], r.get('source', ''))

Actual output

Ollama         ollama-launch   0   user-config
OpenRouter     openrouter      33  built-in
GitHub Copilot copilot         14  built-in
Anthropic      anthropic       8   built-in
OpenAI         openai          0   built-in
OpenAI Direct  openai-direct   0   user-config

Expected behavior

OpenAI and OpenAI Direct should not show 0 when live catalogs are available for those providers.

At minimum, the picker should surface a nonzero count consistent with the provider's available models, rather than a misleading zero.

Notes

• This issue is intentionally scoped only to the live OpenAI/OpenAI Direct count mismatch on latest main.
• It is separate from the earlier duplicate-row investigation/PR (#13190), which was closed as partially redundant.
• In earlier local testing on a fix branch, probing live catalogs for openai / openai-direct resolved the zero-count symptom.