Run Ledger
Scrub an agent run like a tape: steps, a two-run diff, golden-file assertions.
Try←→scrubSpaceplayTabedit a step
5/7
- Elapsed
- Cost
Golden file
6/6- PASSroles = 19
- PASSboards ≤ URLs
- PASSroles ≤ analysed ≤ postings
- PASSverify before write
- PASSsources cited
- PASScost ≤ $0.40
Edit: Count roles
- Problem
- Agent runs are opaque: a wall of logs and a final answer.
- Decision
- One playhead, typed steps, assertions that re-run on every edit, diffs by step id.
- Cut
- Token-level traces and live streaming.
Three sample agent runs: steps and numbers are real; latency, cost and the re-run are synthetic.