settings.yaml.example

# Godspeed Settings — copy to ~/.godspeed/settings.yaml
#
# Godspeed works out of the box with free local models via Ollama.
# To use paid models, set the appropriate API key as an env var and
# change the model below.
#
# Model switching:
#   - CLI flag:     godspeed -m claude-sonnet-4-20250514
#   - Env var:      GODSPEED_MODEL=gpt-4o godspeed
#   - This file:    edit the 'model' field below
#   - At runtime:   /model claude-sonnet-4-20250514

# ─── LLM ────────────────────────────────────────────────────────────────
# Default: free local model via Ollama (no API key needed)
# Install: ollama pull qwen3:4b
#
# Recommended production driver: NVIDIA NIM free-tier Qwen3.5-397B —
# winner of the 2026-04 benchmark shootout (0.608 overall, 11/20 pass).
# Requires NVIDIA_NIM_API_KEY. See experiments/benchmark_shootout_2026_04.md.
#   model: nvidia_nim/qwen/qwen3.5-397b-a17b
#   fallback_models:
#     - ollama/qwen3-coder:latest
model: ollama/qwen3:4b

# Fallback chain — tried in order if the primary model fails.
# Useful for mixing local + cloud: local model for speed, cloud for fallback.
fallback_models: []

# ─── Popular model examples (uncomment one) ─────────────────────────────
#
# Free local models (Ollama — no API key):
#   model: ollama/qwen3:4b               # 4B params, fast, good tool calling
#   model: ollama/qwen3:8b               # 8B params, better reasoning
#   model: ollama/qwen3:14b              # 14B params, ~75 tok/s, fits fully in 16GB VRAM
#   model: ollama/qwen3-coder:latest     # MoE 30B total / 3B active, coder-tuned,
#                                        # ~18GB, ~53 tok/s. Tool calls via Godspeed's
#                                        # Qwen3-Coder parser shim (llm/qwen3_coder_parser.py)
#   model: ollama/gemma4:e4b             # Google Gemma 4, excellent coding
#   model: ollama/llama3.3:8b            # Meta Llama 3.3
#   model: ollama/mistral:7b             # Mistral 7B
#   model: ollama/deepseek-r1:8b         # DeepSeek R1, strong reasoning
#
# NVIDIA NIM (free frontier-quality via build.nvidia.com, set NVIDIA_NIM_API_KEY):
#   model: nvidia_nim/deepseek-ai/deepseek-v3                    # strong tool calling
#   model: nvidia_nim/nvidia/llama-3.3-nemotron-super-49b-v1     # NVIDIA-tuned Llama
#   model: nvidia_nim/meta/llama-3.3-70b-instruct                # broad compatibility
#   model: nvidia_nim/qwen/qwen3-coder-480b-a35b-instruct        # if available in catalog
#
# Paid cloud models (set API key as env var):
#   model: claude-sonnet-4-20250514    # ANTHROPIC_API_KEY
#   model: claude-opus-4-20250514      # ANTHROPIC_API_KEY
#   model: gpt-4o                      # OPENAI_API_KEY
#
# Moonshot (Kimi) direct API — for models not yet on NVIDIA NIM free tier.
# Requires MOONSHOT_API_KEY from https://platform.moonshot.ai/. As of
# 2026-04-20, K2.6 is on Moonshot's paid API; K2.5 is also on NIM free.
#   model: moonshot/kimi-k2.6          # flagship, 256K ctx; $0.95/$4.00 per MTok
#   model: moonshot/kimi-k2.5          # mature, generally available
#   fallback_models:
#     - nvidia_nim/moonshotai/kimi-k2.5   # free fallback on rate limits
#
# Ollama Cloud — hosted inference via Ollama (requires `ollama signin`):
# Free tier (1 concurrent, "light usage"):
#   model: ollama/kimi-k2.5:cloud      # K2.5 free on Ollama Cloud, 64K ctx
# Paid (Pro $20/mo = 3 concurrent; Max $100/mo = 10 concurrent):
#   model: ollama/kimi-k2.6:cloud      # K2.6 flagship, requires Ollama Pro
#   model: ollama/gpt-oss:120b-cloud   # GPT-OSS 120B variants
# Usage is qualitative ("light/day-to-day/heavy"), 16K max output tokens.
#   model: gpt-4o-mini                 # OPENAI_API_KEY
#   model: gemini/gemini-2.0-flash     # GEMINI_API_KEY
#   model: deepseek/deepseek-chat      # DEEPSEEK_API_KEY
#
# Mixed strategy — fast local + smart cloud fallback:
#   model: ollama/qwen3:4b
#   fallback_models:
#     - nvidia_nim/deepseek-ai/deepseek-v3
#     - claude-sonnet-4-20250514

# ─── Task-aware model routing ───────────────────────────────────────────
# Use a strong model for fresh planning turns and a cheap one for
# everything else (file edits, reads, shell continuations). The
# classifier in godspeed/llm/router.py decides per-turn — no extra LLM
# call. Empty values are skipped.
#
# Example: NIM Kimi for reasoning, local Ollama for the rest.
#   strong_model: nvidia_nim/moonshotai/kimi-k2.5
#   cheap_model:  ollama/qwen3:4b
#
# Power users can write `routing:` directly — explicit entries always
# win over the shortcuts:
#   routing:
#     plan:      claude-opus-4
#     edit:      ollama/qwen3-coder:latest
#     read:      ollama/qwen3:4b
#     shell:     ollama/qwen3:4b
#     architect: claude-opus-4
strong_model: ""
cheap_model: ""

# ─── Context ────────────────────────────────────────────────────────────
max_context_tokens: 100000
compaction_threshold: 0.8

# ─── Permissions ────────────────────────────────────────────────────────
# Deny rules are checked first and always win. Project configs can only
# ADD deny rules, never remove global ones.
permissions:
  deny:
    # Secrets and credentials
    - "FileRead(.env)"
    - "FileRead(.env.*)"
    - "FileRead(.env.local)"
    - "FileRead(.env.production)"
    - "FileRead(.env.staging)"
    - "FileRead(*.pem)"
    - "FileRead(*.key)"
    - "FileRead(*.p12)"
    - "FileRead(*.pfx)"
    - "FileRead(*.jks)"
    - "FileRead(*credentials*)"
    - "FileRead(*secret*)"
    - "FileRead(.aws/*)"
    - "FileRead(.gcloud/*)"
    - "FileRead(.azure/*)"
    - "FileRead(.ssh/*)"
    - "FileRead(docker-compose*.secret*)"
    - "FileWrite(.env)"
    - "FileWrite(.env.*)"
    - "FileWrite(.ssh/*)"
    - "FileWrite(.aws/*)"
  allow:
    # Safe dev commands — auto-approved
    - "shell(git *)"
    - "shell(ruff *)"
    - "shell(pytest *)"
    - "shell(make *)"
  ask:
    # Everything else requires confirmation
    - "shell(*)"

# ─── Audit ──────────────────────────────────────────────────────────────
audit:
  enabled: true
  retention_days: 30