| provider | LLM provider to use: anthropic, openai, litellm, or claude-code | no | anthropic |
| model | Model name (e.g., claude-haiku-4-5-20250414, haiku, sonnet, opus, gpt-4o, or any LiteLLM model) | no | "" |
| anthropic_api_key | Anthropic API key (required when provider is anthropic) | no | — |
| openai_api_key | OpenAI API key (required when provider is openai) | no | — |
| litellm_api_key | API key for LiteLLM provider | no | — |
| litellm_api_base | Base URL for LiteLLM-compatible API | no | — |
| claude_code_oauth_token | Claude Code OAuth token for Pro/Max subscription auth (generate with 'claude setup-token') | no | — |
| claude_code_path | Path to Claude Code CLI binary (for claude-code provider; auto-installed if not found) | no | "" |
| config_path | Path to .skill-eval.yml config file | no | .skill-eval.yml |
| skills_path | Path to skills directory (overrides config file) | no | — |
| github_token | GitHub token for PR comments and check status | no | ${{ github.token }} |
| parallel_evals | Number of evals to run in parallel (1-10, lower for rate-limited APIs) | no | 3 |
| eval_preset | Eval preset: smoke (3 trials), reliable (10 trials), or regression (25 trials). Overridden by explicit eval_trials. | no | — |
| eval_trials | Number of trials per eval test case for pass@k/pass^k statistical metrics (1-50, default 1 for single-run) | no | 1 |
| min_pass_rate | Minimum eval pass rate (0.0-1.0) to consider the run successful. Default 1.0 requires all evals to pass. | no | — |
| fail_on_error | Fail the action if any evaluation errors are found | no | true |