# Hermes Agent + AIWG — Minimal MCP Configuration
#
# This config connects AIWG as an MCP sidecar.
# AIWG exposes exactly 5 tools: workflow-run, artifact-read, artifact-write,
# template-render, agent-list. Schema overhead: ~3,000 tokens.
#
# Architecture: Hermes (host) → MCP → AIWG (sidecar)
# Hermes owns: conversation, memory, sessions
# AIWG owns: .aiwg/ artifacts, workflows, templates
#
# Usage:
#   1. Copy the mcp_servers and delegation blocks into ~/.hermes/config.yaml
#   2. Config reloads automatically (v0.4.0+) — no restart needed
#   3. Verify: ask Hermes "what AIWG tools are available?"

mcp_servers:
  aiwg:
    command: "aiwg"
    args: ["mcp", "serve"]

# Delegation config — route AIWG workflows to a coding-optimized model
# while the parent stays on hermes3 for conversation.
delegation:
  model: "ollama/qwen2.5-coder:14b"   # Best tool call accuracy for structured output
  max_iterations: 50

# Recommended model strategy — two roles, two models:
#
#   Parent (conversation, soul, memory):
#     hermes3          — best for roleplay/soul features (recommended)
#     mistral:7b       — solid general-purpose alternative
#     llama3.2:3b      — lightweight, fast on low VRAM
#
#   Coding / tool calls (via delegation.model config):
#     qwen2.5-coder:14b — best tool call accuracy, recommended for AIWG workflows
#     qwen2.5-coder:7b  — smaller variant, still excellent tool calls, lower VRAM
#     qwen3.5:9b        — vision + 256K context, strong structured output (8GB VRAM)
#     qwen3:8b          — strong structured output, thinking/non-thinking modes
#     phi4-mini          — compact alternative for low VRAM
#
#   Note: Qwen models have the best tool call accuracy open-weight. Use them
#   for any AIWG workflow involving structured output or function calling.
#
# Pull recommended pair: ollama pull hermes3 && ollama pull qwen2.5-coder:14b

# Compression settings for local hardware (12GB VRAM)
compression:
  enabled: true
  threshold: 0.30              # Lower than default 0.50 — avoids abrupt context cliffs
  summary_model: "ollama/qwen2.5-coder:7b"
  summary_provider: "custom"
  summary_base_url: "http://localhost:11434/v1"

# Context window (auto-detected from /models endpoint, or set manually)
# model:
#   context_length: 32768

max_tool_rounds: 10