hermes - 💡(How to fix) Fix Add first-class per-turn and per-tool-call model overrides [1 pull requests]

Problem

Hermes supports model switching today, but the paths are split:

• /model can hot-swap the live chat session.
• --global and dashboard model settings persist to config.yaml.
• image_generate and some auxiliary paths read provider/model from config at call time.
• Auxiliary tasks can be pinned through auxiliary.<task>.{provider,model}.

This works, but it is still not a unified dynamic routing model. Users naturally ask things like:

• “Use Claude for this task.”
• “Generate this image with GPT-Image-2.”
• “Try Grok Imagine for this one.”
• “Use a cheaper model for the next summary.”

Today those requests are just normal text unless the user knows the exact slash command or config surface. Tool calls like image_generate also do not have a first-class provider / model override, so changing the backend usually means editing config instead of scoping the choice to one output.

Proposal

Introduce a unified runtime model override mechanism with explicit scope:

• once: applies to the next assistant turn only.
• tool_call: applies to one tool invocation, e.g. image/video generation.
• session: applies to the current live session, equivalent to /model session-only behavior.
• global: persists to config.yaml, equivalent to /model --global.

The override should carry:

provider: openai | anthropic | openrouter | xai | ...
model: gpt-image-2 | claude-sonnet-... | ...
api_mode: optional
base_url: optional
api_key: optional/resolved
scope: once | tool_call | session | global
target: chat | image_generate | video_generate | auxiliary.<task>

Why this matters

This gives Hermes one shared primitive for dynamic model selection instead of each surface solving it differently. It also creates a clean foundation for future natural-language routing and recommendation features without making the first implementation depend on an intelligent router.

Suggested implementation path

1. Extract the live-session /model switching result into a reusable runtime override object.
2. Add a small override store on the session/agent runtime for once and tool_call scoped overrides.
3. Extend generation tool schemas, starting with image_generate, to accept optional provider and model.
4. Route tool-level overrides through the same provider registry/config resolution used by existing image/video providers.
5. Keep natural-language parsing out of the first PR, but make it possible for a later command or router to set the same override object.

Acceptance criteria

• /model <model> --provider <provider> continues to work as today.
• A user can request one image with a specific image model without changing config.yaml.
• A one-turn chat override does not persist into later turns.
• Session-scoped switching still updates the live agent client, context compressor, and _primary_runtime.
• Dashboard/global config writes remain new-session-only unless routed through the live session override path.
• Tests cover chat session switch, one-turn override cleanup, and image-generation provider/model override.