← board

Who reads tools/progress.sh check?

Surfaced by frankH and frank-coordinator, 2026-09-01. The tool is not broken and does not need writing. It needs a reader, which is a workflow question.

What it cost, concretely

perf-p-parsefactorcore-walks-a-92-arm-name-chain-per-factor — p60 — sat in blocked/ with its only blocker rejected. ready/next never scan blocked/, so no path could ever surface it, and check had been saying so in two independent classes the whole time. Repaired 2026-09-01.

frankH found the family the same way: bug-b-tlist-has-no-list-property had been unblocked for days behind a blocked-by naming a slug that no longer existed — the Track A defect had been renamed and closed.

Also standing right now: 4 STALE-PARK-HELD addressed to frankA whose instruction is literally "tell the holder", with nothing doing the telling.

The fork — and it is not one decision, it is two

Mechanical classes. STALE-EDGE-HIDDEN, BLOCKED-BY-REJECTED, and a blocked-by naming a nonexistent slug. In every case a ticket is hidden from the ranker for a reason that is provably not real, and confirming it takes no judgement — the blocker is closed, rejected, or absent. This is the ranker being wrong, not a workflow preference.

Judgement classes. STALE-PARK, NEAR-DUP, DEAD-COMMIT, PROSE-EDGE. These cannot be swept, and check's own output says why: a STALE-PARK hit matches SLUGS, not questions — a blocker that settled the OPPOSITE question reads exactly like one that settled yours. Auto-resolving these would manufacture wrong closures at scale.

Recommendation

  1. Wire the mechanical subset into ready/next so ranked work cannot be hidden by a dead edge. A p60 invisible to every path is a ranker defect.
  2. Leave the judgement classes on-demand. Accept that they age. The alternative is worse than aging.
  3. Give STALE-PARK-HELD a router. Its instruction already names its reader — the holder — and the coordinator's relay is the mechanism that exists. It relayed frankA's four on 2026-09-01.

What needs the owner rather than a default

Option (a) as originally posed — run check in the tstate cycle and auto-file what it finds — is the one this ticket will not take on a default. It spends Track T's machine, and "widening your own gate spends the machine that produces the 8" applies to the ranker's hygiene as much as to a test tier. It would also auto-file the judgement classes, which is the one thing the caveat above forbids.

This is the fourth item in one evening where the information existed and the reader did not (frank-coordinator's count): the auto-filed regression heading, the tstate detail block, trackt.py health answering about the wrong host, and this. That pattern is the argument for (1); it is not an argument for (a).