review-runlisted

# Review run The reviewer's attention is the scarce resource in an agentic setup, and it is spent badly by default: an evenly-weighted list of everything that happened, in a viewer that cannot act on the answer. Two rules fix that. **Decisions before evidence**, because only decisions need judgement — evidence needs checking, which is faster. And **the review happens where the answer can be acted on**, so a verdict costs a sentence rather than a copy, a window switch and a paste. ## 1 — Find what needs review With no id given, scan for runs and present them in this order: 1. **Blocked** — `status: stopped` in `.agent/current-run.json`, or a note in `.agent/handoff/`. The human is the blocker here, so it goes first. 2. **Handed over** — evidence present, `pr-package` written. 3. **Failed** — a failure report but no handover. For an overview across many runs at once, `tools/evidence-board.mjs` renders them as columns. This skill is for going *into* one. ## 2 — Present one run One run at a time. A wall of three runs gets one run's worth of attention spread thin. **The headline, one line.** Gates, screenshots, console — the numbers, so the reviewer knows what kind of run this was before reading anything. **Decisions, in reversal-cost order.** For each high or medium entry: what was decided, what the alternative was, and **your own recommendation with its reason**. A decision presented without a recommendation turns into a design discussion the reviewer did not ask for;