jayfarei/opentraces/skill-opt-v1

skill-opt-v1

Scored-rollout projection for the SkillOpt text-space skill optimizer (arXiv 2605.23904; design in kb/br/66-skillopt-text-space-skill-optimizer.md).

scripts/build_rows.py reads each already-captured trace from the local bucket, scores it with opentraces.quality.engine.assess_trace, and emits one row per trace. The trace's overall_utility (in [0,1] after scaling) is the rollout score; the failed and passed quality checks become failure_tags and success_tags, which the optimizer's reflection step and the held-out validation gate consume.

This workflow is the forward-pass-over-captured-evidence layer. It does not run an agent; the live re-rollout of a candidate skill is a later slice. The opentraces workflow optimize consumer reads these rows, splits them into a train and a held-out selection set by trace-id hash, proposes bounded add/delete/replace edits, accepts only strictly-improving candidates, and exports best_skill.md plus edit_apply_report.json.

Row shape

{
  "trace_id": "…",
  "project_slug": "…",
  "session_id": "…",
  "score": 0.62,            // overall_utility / 100, the rollout score
  "overall_utility": 62.0,
  "failure_tags": ["rl.RL1_grounded_outcome", "training.T3_clean_diff"],
  "success_tags": ["conformance.C1_schema_version"],
  "summary": "task description (truncated)",
  "skill_digest": null
}

Scope

The run packet's scope may carry project_slug (or project) to narrow the bucket scan to one project; otherwise every project's traces are projected.