Phase 6 — Metrics + trajectory analysis

Do not reimplement reporting here. Once the run finishes, hand off to the wbbench-report-skills skill (invoke it via the Skill tool). That skill owns metric computation and trajectory analysis and routes to the matching benchmark-specific workflow (WB-Bench-Office / WB-Bench-Web / WB-Bench-Code).

What to pass

A single evaluation run artifact directory RUN_DIR — one model, one dataset, one round. For a run launched here that's:

<bench.jobs_dir>/<slug>/<timestamp>/

(bench jobs_dir defaults to results, so typically results/<slug>/<timestamp>/). Point at the timestamped run dir itself — the one containing config.json / result.json / job.log / trial dirs — not a parent results/ and not a single trial dir.

If several timestamps exist under the slug, pass the exact one; the report skill refuses to guess when a parent contains multiple runs.

Boundary

wbbench-report-skills treats Harbor artifacts as read-only and writes only to a user-specified output directory. Let it produce the metrics report and trajectory analysis; this setup skill's job ends at a completed run.