Problem

When users switch from a custom endpoint (e.g. local Ollama/vLLM) to a named provider (e.g. OpenRouter) via hermes model, OPENAI_BASE_URL persists in ~/.hermes/.env. The main inference client uses config.yaml correctly, but auxiliary clients (compression, vision, delegation) with provider:auto pick up the stale env var and route requests to the old custom endpoint.

Fixes #5161

Solution

Two-pronged fix:

1. Startup warning in auxiliary client (agent/auxiliary_client.py)

• In _resolve_auto(), detect when OPENAI_BASE_URL is set but config.yaml model.provider is a named provider (not custom)
• Log a WARNING once per process with actionable advice
• Uses a module-level _stale_base_url_warned flag to avoid spamming

2. Proactive cleanup on provider switch (hermes_cli/main.py)

• New _clear_stale_openai_base_url() helper called after every provider flow in select_provider_and_model()
• When the user switches to any named provider, checks if OPENAI_BASE_URL is still set in .env and clears it
• Skips cleanup for custom, cancel, remove-custom, and custom:* providers
• Prints a confirmation message so the user knows what happened

Tests

4 new tests in tests/agent/test_auxiliary_client.py::TestStaleBaseUrlWarning:

• Warning fires when OPENAI_BASE_URL conflicts with named provider
• No warning when provider is custom (OPENAI_BASE_URL is expected)
• No warning when OPENAI_BASE_URL is not set
• Warning only fires once per process

4 new tests in tests/hermes_cli/test_clear_stale_base_url.py:

• OPENAI_BASE_URL is cleared when provider is a named provider
• OPENAI_BASE_URL is preserved when provider is custom
• No error when OPENAI_BASE_URL is already empty
• No cleanup when provider is not configured in config.yaml

All 8 tests pass. Pre-existing test failures are unrelated.