config.yaml.example

# NERF Configuration
# Copy to config.yaml (project root) or ~/.config/nerf/config.yaml
#
# Precedence: env vars > project config > user config (~/.config/nerf/config.yaml)
#
# All LLM operations route through pi-mono (@nerf/pi-ai). 26 providers supported.
# Supports ${VAR} and ${VAR:-default} environment variable substitution.
#
# Environment variable overrides:
#   LLM_BACKEND          overrides llm_backend
#   ANTHROPIC_API_KEY    overrides claude.api_key
#   OPENAI_API_KEY       overrides openai.api_key
#   OPENROUTER_API_KEY   overrides openrouter.api_key
#   GEMINI_API_KEY       overrides google.api_key
#   OLLAMA_BASE_URL      overrides ollama.base_url
#   NERF_API_SECRET      overrides auto-generated REST API secret
#   NERF_DEBUG           enable debug logging (set to "1")

# Active backend. Options:
#   claude        — Anthropic Claude (API key or OAuth)
#   ollama        — Local Ollama server (free, no API key required)
#   openai        — OpenAI GPT models
#   openai-compat — Any OpenAI-compatible endpoint: LM Studio, vLLM, Groq, Together, etc.
#   openrouter    — 200+ models via OpenRouter
#   google        — Google Gemini (API key)
#   mistral       — Mistral AI
#   pi-ai         — Direct pi-ai provider spec (most flexible)
llm_backend: claude

# ---------------------------------------------------------------------------
# Backend Configurations
# ---------------------------------------------------------------------------

# -- Anthropic Claude -------------------------------------------------------
# API key OR OAuth (no key needed if authenticated via `nerf setup`).
# OAuth credentials are stored in ~/.nerf/agent/auth.json and auto-refreshed.
claude:
  api_key: "${ANTHROPIC_API_KEY}"   # Leave empty to use OAuth
  model: "claude-sonnet-4-6"
  timeout: 120

# -- Ollama (local, free) ---------------------------------------------------
# Install: https://ollama.ai
# Start:   ollama serve
# Models:  ollama pull qwen2.5:32b
#
# Recommended models:
#   qwen3-coder:30b    — Qwen3 Coder 30B MoE (3.3B activated) — strong code + reasoning
#   qwen2.5:32b        — previous gen, good general purpose baseline
#   qwen2.5-coder:32b  — optimized for code tasks (prev gen)
#   llama3.3:70b       — Meta's latest, requires 48GB+ RAM
#   deepseek-r1:32b    — strong reasoning model
ollama:
  base_url: "${OLLAMA_BASE_URL:-http://127.0.0.1:11434/v1}"
  model: "nerf"          # local NERF persona built via: ollama create nerf -f Modelfile
  timeout: 300

# -- LM Studio --------------------------------------------------------------
# Use for models not available in Ollama (heretic/uncensored models, custom GGUF).
# Install: https://lmstudio.ai/download
#
# Starting the server:
#   GUI:     Local Server tab → Start Server (default port 1234)
#   CLI:     lms server start --port 1234   (requires LM Studio app running)
#   Headless Linux: DISPLAY=:0 lm-studio --no-sandbox & && lms server start
#
# To use LM Studio as backend, set llm_backend: openai-compat and configure below.

# -- OpenAI -----------------------------------------------------------------
openai:
  api_key: "${OPENAI_API_KEY}"
  model: "gpt-4o"
  timeout: 120

# -- OpenRouter (200+ models) -----------------------------------------------
# Models: https://openrouter.ai/models
# Useful for: DeepSeek R1, Qwen, Llama, Claude, GPT-4o — all via one key.
openrouter:
  api_key: "${OPENROUTER_API_KEY}"
  model: "anthropic/claude-3.5-sonnet"
  timeout: 120

# -- Google Gemini ----------------------------------------------------------
google:
  api_key: "${GEMINI_API_KEY}"
  model: "gemini-2.5-flash"
  timeout: 120

# -- Mistral AI -------------------------------------------------------------
mistral:
  api_key: "${MISTRAL_API_KEY}"
  model: "mistral-large-latest"
  timeout: 120

# -- OpenAI-Compatible Endpoint ---------------------------------------------
# Works with: LM Studio, vLLM, Groq, Together AI, Cerebras, Ollama (alt), any local server.
#
# LM Studio example:
#   base_url: "http://127.0.0.1:1234/v1"
#   api_key:  "lm-studio"        # any non-empty string
#   model:    "your-model-name"  # must match the loaded model in LM Studio
#
# Groq example:
#   base_url: "https://api.groq.com/openai/v1"
#   api_key:  "${GROQ_API_KEY}"
#   model:    "llama-3.3-70b-versatile"
#
# vLLM example:
#   base_url: "http://your-server:8000/v1"
#   api_key:  "EMPTY"
#   model:    "meta-llama/Llama-3-70b-instruct"
openai-compat:
  base_url: ""
  api_key: ""
  model: ""
  timeout: 300

# -- Direct pi-ai provider spec (most flexible) ----------------------------
# provider: anthropic | openai | google | mistral | groq | xai | huggingface
#           bedrock | azure | vertex | ollama | lmstudio | openrouter
#           cerebras | together | perplexity | cohere | fireworks | deepseek
pi-ai:
  provider: "anthropic"
  model: "claude-sonnet-4-6"

# ---------------------------------------------------------------------------
# Per-Phase Model Routing
# ---------------------------------------------------------------------------
# Use expensive models where quality matters (planning), fast/cheap for research.
# Mix local and cloud freely. Prefix with provider/ for explicit routing.
#
# Examples:
#   "claude-opus-4-6"                    — Anthropic direct
#   "openrouter/deepseek/deepseek-r1"    — via OpenRouter
#   "ollama/qwen2.5:32b"                 — local Ollama
#   "google/gemini-2.5-pro"              — Google direct

models:
  research: "claude-sonnet-4-6"
  planning:
    model: "claude-opus-4-6"
    fallbacks:
      - "openrouter/anthropic/claude-3.5-sonnet"
      - "google/gemini-2.5-pro"
  execution: "claude-sonnet-4-6"
  completion: "claude-sonnet-4-6"

# ---------------------------------------------------------------------------
# Auto Mode Supervision
# ---------------------------------------------------------------------------

auto_supervisor:
  soft_timeout_minutes: 20     # Warn and suggest compacting
  idle_timeout_minutes: 10     # Stop if no progress detected
  hard_timeout_minutes: 30     # Force stop

budget_ceiling: 50.00          # USD. Auto mode stops when reached.

# ---------------------------------------------------------------------------
# Verification
# ---------------------------------------------------------------------------
# Shell commands run after each task execution. Task fails if any exits non-zero.

verification_commands: []
  # - "npm run lint"
  # - "npm run test"
  # - "pytest tests/"

verification_auto_fix: true    # Retry with auto-fix on failure
verification_max_retries: 2    # Max retries per command

# ---------------------------------------------------------------------------
# Token Optimization
# ---------------------------------------------------------------------------

token_profile: balanced        # budget | balanced | quality

# ---------------------------------------------------------------------------
# Engagement
# ---------------------------------------------------------------------------

unique_engagement_ids: true    # Append random suffix to engagement IDs (prevents collisions)
auto_report: true              # Generate report on engagement completion

# ---------------------------------------------------------------------------
# Signal Bot (optional)
# ---------------------------------------------------------------------------

signal:
  signal_service: "${SIGNAL_SERVICE:-signal-api:8080}"
  phone_number: "${SIGNAL_PHONE_NUMBER}"
  whitelist: []                # Phone numbers allowed to interact with the bot
  session_timeout: 3600        # Seconds before session history is cleared