Bug Description

When switching providers via hermes model, the api_mode field in ~/.hermes/config.yaml is not updated for API-key providers (Z.AI, MiniMax, Kimi). The previous provider's api_mode value persists, causing requests to hit the wrong API endpoint path.

Steps to Reproduce

1. Configure Hermes for Copilot:

   model:
     default: gpt-5.4
     provider: copilot
     base_url: https://api.githubcopilot.com
     api_mode: codex_responses

2. Run hermes model and select Z.AI → GLM-5-turbo
3. Config is updated to:

   model:
     default: glm-5-turbo
     provider: zai
     base_url: https://api.z.ai/api/coding/paas/v4
     api_mode: codex_responses    # stale - should be chat_completions

4. Run any query → Hermes sends POST to /v4/responses → HTTP 404

Expected Behavior

hermes model should set api_mode appropriately for each provider when switching. At minimum, the generic API-key provider flow (_model_flow_api_key_provider) should either set api_mode: chat_completions (the correct default for Z.AI, MiniMax, Kimi) or clear the existing api_mode so Hermes falls back to its default.

Actual Behavior

api_mode is left untouched from the previous provider. Only model.default, model.provider, and model.base_url are written.

Root Cause

In hermes_cli/main.py, the _model_flow_api_key_provider() function (line ~2091) writes:

model["provider"] = provider_id
model["base_url"] = effective_base
save_config(cfg)

But never touches model["api_mode"]. Compare with the Copilot flow (line ~1771) which correctly calls copilot_model_api_mode() to determine the right api_mode.

This affects all providers that go through _model_flow_api_key_provider: Z.AI, MiniMax, Kimi.

Environment

• OS: Ubuntu/WSL2
• Python: 3.11
• Hermes Agent: v0.4.0+

Related

• Fix OpenCode model selection, normalization, and mixed API-surface routing #3014 — Similar api_mode issue but specific to OpenCode providers
• Discovered when debugging Paperclip agent runs that routed GPT-5.4 to Z.AI due to provider mismatch, which then hit /v4/responses because api_mode: codex_responses was left over from a Copilot session