Run an all-in-one discovery workflow for a target executable. Discovery asks an LLM agent to understand the target command, author five complex goal-oriented scenarios, run the generated scenario batch, judge usage quality, and summarize the results.

Real agent execution is disabled unless AX_EVAL_ENABLED=1 is set, because discovery may spend LLM API credits and execute agent-driven CLI commands.

Usage: ax-eval discover [OPTIONS] <TARGET>

Arguments:
  <TARGET>
          Target executable binary or command to discover

Options:
      --tool <TOOL>
          Agent tool to evaluate in generated scenarios
          
          [default: opencode]

      --model <MODEL>
          Model to evaluate in generated scenarios

      --discover-tool <DISCOVER_TOOL>
          Agent tool used for inspect, fixture authoring, and final summary

      --discover-model <DISCOVER_MODEL>
          Agent model used for inspect, fixture authoring, and final summary

      --judge-model <JUDGE_MODEL>
          Judge model for LLM-as-judge evaluation

      --judge-tool <JUDGE_TOOL>
          Tool to use for LLM-as-judge evaluation

      --timeout-secs <TIMEOUT_SECS>
          Maximum execution time in seconds per agent command
          
          [default: 300]

  -v, --verbose
          Enable verbose output (or set RUST_LOG for fine-grained control)

  -h, --help
          Print help (see a summary with '-h')

Example:
  AX_EVAL_ENABLED=1 ax-eval discover qipu --tool opencode

Use --discover-tool/--discover-model when the agent authoring the discovery artifacts should differ from the evaluated scenario-run agent.
