Configuration Reference

Configuration Reference

Edit ~/.claude/skills/h3/config.json (or %USERPROFILE%\.claude\skills\h3\config.json on Windows):

{
  "model": "z-ai/glm-5",
  "free_model": "nvidia/nemotron-3-nano-30b-a3b:free",
  "council_models": {
    "correctness": "openai/gpt-5.2",
    "performance": "google/gemini-3-pro-preview",
    "security": "x-ai/grok-4"
  },
  "reasoning": "high",
  "docs_folder": "documents",
  "max_context": 200000,
  "enable_web_search": true
}

Options

Web Search (Enabled by Default)

Web search is enabled by default. This adds the :online suffix and Exa search plugin to API calls, allowing models to reference up-to-date documentation, CVEs, and best practices.

To disable, set "enable_web_search": false in config.json.

Notes:

• Web search adds ~$0.005 per request (Exa pricing)
• Free models (:free suffix) do not support web search

Council Models

Customize which models are used for each council role:

{
  "council_models": {
    "correctness": "openai/gpt-5.2",
    "performance": "google/gemini-3-pro-preview",
    "security": "x-ai/grok-4"
  }
}

Notes:

• You can override individual roles (unspecified roles use defaults)
• The same models are used for both code and plan reviews (with different prompts)
• Web search (:online suffix) is automatically added when enable_web_search: true

Example: Use Claude for correctness:

{
  "council_models": {
    "correctness": "anthropic/claude-sonnet-4"
  }
}

Context Limits

The context limit includes: diff + full file contents + documentation + test files. For very large PRs, the skill automatically breaks them into module-by-module reviews.

Available Models (via OpenRouter)

The full council.py is in the public repo at scripts/council.py. You can read, audit, and modify it - it's MIT licensed. Council models are configurable via council_models in config.json.

Free Models

Free tiers rotate on OpenRouter. Use the included utility to find current free models:

python ~/.claude/skills/h3/scripts/list-free-models.py

# Example output:
# FREE MODELS ON OPENROUTER (newest first)
# ------------------------------------------------------------
#  1. nvidia/nemotron-3-nano-30b-a3b:free                             2025-12-09
#  2. deepseek/deepseek-r1:free          [THINKING]            2025-01-20
#  3. google/gemini-2.0-flash:free       [THINKING]            2025-01-10

Or browse manually: openrouter.ai/models

Free Tier Limits (OpenRouter)

Notes:

• Adding $10+ credits unlocks higher limits but keeps free model access
• Some free models may have stricter limits
• Heavy usage can trigger rate limiting

Cost Comparison

For a typical small code review (~10K chars / ~2.5K tokens input):

For a larger review (~50K chars / ~12.5K tokens input): Single ~$0.01, Council ~$0.19

Default is Single (GLM 5) for excellent quality at minimal cost. Use --free for zero cost, or --council for multi-model consensus.