/runs and /runs/<id> shipped 2026-08-09 (pantry 6f6cced). Both work and both are a first cut. Two asks from the owner the same day, and they are the two altitudes the surface is currently missing.

The index wants a shape, not a list

Right now it is a card grid ordered by date, which answers "what happened last" and nothing else. The questions it should answer are about the series: is adherence improving, when did the gaps cluster, which run is the one that grew past its scope. A timeline or flow reading of the same data is the natural form, and the repo already has the pieces:

• /timeline already exists and is git-derived plan bars with dependency arrows. Runs are dated, scoped and evidenced, so they are the same kind of series. The first question is whether this is a new view or a layer on the timeline that already ships, and the second answer is cheaper and probably better.
• FIGURES owns the form. Any diagram here follows the flow scaffold and the palette, and mermaid does not reach a published page. A timeline of runs is a data-viz figure, not a decoration.
• The honest limit: four or five reports is not a series. This is worth building when the ledger has enough entries that a shape exists to see, and building it against five will produce a chart that looks like a chart and says nothing.

The detail page is one long scroll

It renders, in order: missing evidence, scope growth, the meta, the gates, the plans, the dirty files, the scope, the touched, the skills, then the whole report body. That order is deliberate (what is missing first, the run's own prose last, because the body is where a run is most able to sound finished) and it should survive whatever the reorganisation is.

What it needs on top of that:

• A summary at the top, before anything else: status, date, diffstat, evidence score, whether scope grew. The one-line answer to "do I need to read this".
• Grouping, so the middle is skimmable. Tabs, a sidebar, or accordions. Accordions are the cheapest and need no client script, which matters because every other PANTRY surface except /decisions and /map is static server-rendered and adding a script here would be the first exception with no reason behind it.
• Gate output has to stay expanded, or expanded by default. Collapsing the verbatim output behind a click is the one move this page must not make: LOOP §4a's whole point is that the evidence is present rather than summarized, and a thing behind a click is a thing nobody opens.

Known rough edges, already seen live

• A long diffstat runs three lines on a card and squashes its neighbour. Truncating it was rejected (this page does not shorten evidence), so the fix is layout. Fixed: its own line on the card.
• The meta definition list wraps into ragged columns at some widths. Fixed: each fact is capped at 60ch and the row aligns to the top, so a long verified-by line stops stranding its neighbours.

Tasks

• [x] Decide: a runs layer on /timeline, or a separate view. A layer on /timeline, and not yet. The two arguments settle it together: the timeline already draws dated, dependency-linked bars from real git commits, and a run is the same kind of dated thing, so a second view would be a fork of a drawing that exists. And the ledger holds five reports. Building either shape against five produces a chart that looks like a chart and says nothing, which is this plan's own warning. The decision is made; the drawing waits for the data.
• [x] The detail summary block first. Five facts — outcome, closed, diffstat, evidence carried, scope held or grown — as a strip above everything else. Label over value, the same way round as the meta list below it.
• [x] Grouping for the middle sections, accordions unless there is a reason to take a script. <details> on Plans claimed, Declared scope, Touched and Skills, each carrying its count in the summary line so a shut section still informs. No script: this stays the static surface it was.
• [x] Keep gate output out of any collapse, and add a test that asserts it. Nothing folds above the body: the gap list, the scope growth, the gates and their results and the report's own prose all stay open. The test asserts the gate output sits after the last </details> and that no fold is open across any of the four sections above it.
• [x] The card layout fix for long diffstats. The diffstat gets its own line rather than trailing the date run; truncating it stays rejected.
• [ ] Revisit the timeline once the ledger has enough entries to have a shape. Checked 2026-08-10: six reports, one more than the five this plan already called not a series, so it stays deferred. Counting again every session and re-arguing it is its own small waste, so the bar is a number now: twelve reports, or a ledger spanning two calendar months, whichever comes first. Twelve is roughly a quarter of sessions at the current rate, which is the point where a run of gaps or a climb in evidence carried would be a trend rather than a coincidence.

What is still rough

The strip's diffstat repeats the one in the meta list a screen below. That is the summary doing its job (it is the fact most likely to decide whether to read on) rather than a duplication to remove, but if the meta list is ever reworked, this is the pair to look at.