Skip to main content

Inspecting and Exporting Results

Inspect a step

After a run completes (or while it’s running), use inspect to see exactly what happened at any step:
This prints all artifacts for step 7 by default (--show all):
  • Patch — the JSON diff the optimizer proposed
  • Skill diff — a unified diff of the skill document before and after
  • Gate scores — the before/after scores that drove the accept/reject decision
  • Skips — any eval cases that were skipped this step (e.g. due to timeout)

Show only what you need

--show options:

Interpreting gate scores

A typical gate_scores.json looks like this:
If delta < epsilon, the patch is rejected. If the step was rejected, the skill document is rolled back to the skill_before.md state.

Finding a step to inspect

Use status to read the history table first, then pick the step with the largest positive delta or the last accepted step before a regression:

Export the best skill

When you’re happy with the results, export the best skill document to your skills directory:
This copies best_skill.md from the run directory to the destination path. Parent directories are created if they don’t exist. The output is byte-identical to the run directory copy. Note: best_skill.md is the highest-scoring skill document seen across all accepted steps — not necessarily the final step. If the last few steps were rejected, best_skill.md reflects the best earlier checkpoint.

Comparing multiple runs

To compare two runs side by side:

Artifacts reference

Every step directory contains the same set of files: Top-level run files:

See also