local-model-router

name	local-model-router
description	Route tasks to local Ollama models or cloud APIs based on triage output. Health-check aware with automatic cloud fallback.
version	1.0.0
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["routing","ollama","local-model","cloud","dispatch","cost-optimization"],"related_skills":["complexity-triage","humanizer-gate","skill-discovery"]}}

Local Model Router

Status: paused. Per the standing "no local models for now" directive (current machine too slow; revisit on capable hardware), do not route any work to local models. All dispatches and LLM-backed support passes go through the active cloud coding harness (coding.default_engine, override with --engine). The routing logic below is retained for when local models are re-enabled; until then, treat every task as cloud-routed and ignore local recommendations from triage.

Route coding tasks between local Ollama models and cloud APIs based on complexity triage results.

Available models

Local (Ollama)

• gemma4:latest (8B, Q4_K_M) — formatting, filtering, simple analysis, rewriting, classification
  • Good for: S-sized tasks, humanizer passes, triage classification, commit rewriting
  • Not good for: multi-file reasoning, complex architecture, deep debugging

Cloud

• gemini-3.5-flash (default cloud model) — full reasoning, multi-file, architecture

Health check

Before routing to local, always verify Ollama is healthy:

terminal(
    command="python3 ~/.hermes-coder/scripts/ollama_manager.py health",
    workdir="~/.hermes-coder",
    timeout=10
)

Exit 0 = healthy, proceed with local routing. Exit 1 = unhealthy, route everything to cloud.

Routing rules

1. Triage says S + Ollama healthy → dispatch to local model
2. Triage says M/L/XL → always cloud
3. Triage says S but task mentions security/auth/crypto/migration → cloud override
4. Ollama unhealthy → all tasks to cloud (silent fallback, no error)

Model selection for dispatch

Local model routing applies to:

• Humanizer gateway calls (internal, handled by the script itself)
• Triage LLM calls (internal, handled by the script itself)
• OpenCode dispatches (via --api-base and -m flags)

For claude-code dispatches, the model always uses Anthropic's cloud API. Triage output informs --max-turns and skill injection but does not change the model.

Future extensibility

When larger models are added (gemma4:27b, qwen3-coder:30b), update:

1. config.yaml ollama.models section with new entry
2. triage.routing.local_sizes to include M
3. scripts/ollama_utils.py MODEL_REGISTRY with new capabilities