hermes - 💡(How to fix) Fix Per-agent fallback tier overrides for fleet right-sizing [1 participants]

Problem

fallback_providers in ~/.hermes/config.yaml is a single global chain. In a multi-agent fleet (27+ cortextOS agents), every agent falls back to the same models regardless of task complexity. A spam-classification agent gets opus; an architecture-doc agent gets capped at sonnet. Token spend is wrong-sized in both directions.

Proposal

Add a tier field to per-agent config (or, in hermes-only setups, a CLI flag / env var). Hermes reads the tier and selects from a tiered fallback chain.

Suggested tier mapping (configurable, these are defaults):

Config shape

# ~/.hermes/config.yaml
fallback_tiers:
  cheap:
    - { provider: anthropic, model: claude-haiku-4-5, key_env: ANTHROPIC_API_KEY }
    - { provider: openai,    model: gpt-5-nano,        key_env: OPENAI_API_KEY }
    - { provider: anthropic, model: claude-sonnet-4-5, key_env: ANTHROPIC_API_KEY }
  mid:
    - { provider: anthropic, model: claude-sonnet-4-5, key_env: ANTHROPIC_API_KEY }
    - { provider: openai,    model: gpt-5,              key_env: OPENAI_API_KEY }
    - { provider: anthropic, model: claude-opus-4-7,   key_env: ANTHROPIC_API_KEY }
  hard:
    - { provider: anthropic, model: claude-opus-4-7,   key_env: ANTHROPIC_API_KEY }
    - { provider: openai,    model: gpt-5,              key_env: OPENAI_API_KEY }
    - { provider: anthropic, model: claude-sonnet-4-5, key_env: ANTHROPIC_API_KEY }

Per-agent override (cortextOS pattern):

// ~/cortextos/orgs/gradata/agents/classifier/config.json
{ "runtime": "hermes", "tier": "cheap" }

Or per-invocation:

hermes chat --tier cheap
HERMES_TIER=hard hermes chat

Acceptance criteria

• Agents with tier: cheap use the cheap chain; opus never fires for them
• Agents with tier: hard start on opus directly (skip the sonnet rung)
• Default tier (mid) when unspecified, preserves current behavior for single-user setups
• Falls back to global fallback_providers if fallback_tiers not configured (backward compat)
• 🔄 Fallback chain (N providers): ... banner shows the tier name when applicable

Why this matters

Current fleet behavior on ollama-cloud 429: all 27 agents try sonnet next. If they collectively saturate sonnet too, they all promote to gpt-5 simultaneously. Tier-aware routing spreads load across the model bands instead of stampeding one rung.

Workaround until shipped

User can manually set per-agent model: overrides in each config.json, but that bypasses fallback entirely (no chain for the override). Tier-aware native support is cleaner.