hermes - 💡(How to fix) Fix Add named LLM model entities in config and allow skills/cron to reference them [1 participants]

Problem

Hermes currently spreads model configuration across multiple subsystem-specific knobs (model, fallback_model, smart_model_routing.cheap_model, auxiliary.*, delegation). That works for a single default model, but it does not scale well when different jobs need different cost/capability tradeoffs.

Examples:

• interactive chat should use a strong model
• simple cron jobs should use a cheap fast model
• skill generation or complex reasoning may need a different model profile
• auxiliary workflows like compression, memory flush, or approval may need dedicated settings

Proposal

Add a named model registry to config.yaml so users can define reusable model entities with parameters like:

• provider
• model
• context window
• temperature
• max output tokens
• reasoning effort
• timeout

Then allow subsystems such as skills, cron jobs, delegation, and auxiliary tasks to reference a model entity by name.

Suggested shape

models:
  default:
    provider: opencode-go
    model: kimi-k2.5
    context_window: 128000
    temperature: 0.7
    max_output_tokens: 8192
    reasoning_effort: medium
    timeout: 120

  cheap_cron:
    extends: default
    temperature: 0.1
    max_output_tokens: 1024
    reasoning_effort: low
    timeout: 30

Example usage

cron:
  jobs:
    update-service:
      schedule: "*/15 * * * *"
      model_ref: cheap_cron

delegation:
  model_ref: default

Benefits

• cheaper background work
• clearer intent in config
• fewer duplicated provider/model blocks
• easier to tune per workload
• backward compatible if the current top-level model stays as the default reference

Acceptance criteria

• users can define named model entities in YAML
• subsystems can reference model entities by name
• inheritance/override semantics are documented
• missing or invalid references fail clearly
• existing config continues to work without migration

Notes

Hermes already has some adjacent config concepts (smart_model_routing, fallback_model, auxiliary.*), so this could likely be added as a unifying layer rather than a rewrite.