feat(skills): measure how far the bug-pattern digest has fallen behind

[ako]

Step 3a of the fix-issue restructure — the trigger. (#345 repointed diagnosis at the digest; #354 made the findings consumable; this makes the gap visible so the digest stops decaying.)

The number

digest: 607 of 631 findings added since the last bug-pattern sync (2026-05-24, 3 pages)

96%. The three docs-wiki/bug-patterns/ pages were written from the ~24 findings that existed in May. Nothing has re-synced since, because the digest is on demand and no step demands it: the After-Every-Fix checklist feeds the input and nothing consumes it.

Three parts

Findings gain a date, backfilled by git blame over the table they came from — 630 of 631 matched, the last being the one appended after the extraction. It is last touched, not first written: a row later corrected carries the correction's date, and the README says so rather than letting the field imply more than it knows.

scripts/digest-status.sh reports the gap — pages, last sync, findings, how many arrived since, and a per-area breakdown:

  area                    findings     since   in a page
  mdl/executor                 248       238   NO
  cmd/mxcli                     95        95   NO
  mdl/backend                   83        82   NO
  mdl/grammar                   53        51   NO
  mdl/visitor                   28        27   yes
  sdk/mpr                       22        16   yes

The "in a page" check is a substring heuristic on purpose. The alternative is a covers: list in every page's frontmatter, which is one more thing to keep true and would rot silently the first time someone forgot it. It's skipped for a bare first segment like model, which matches almost any prose and would answer a confident yes.

check-findings prints the one-line form on every run. That placement is the actual mechanism, not a convenience: it's the command the checklist already runs, so the number lands in front of whoever just appended a finding. A report nobody invokes is precisely how this went three months without a sync.

Advisory, deliberately

Everything here exits 0. A stale digest must not block an unrelated fix, and neither number is a target to drive to zero — they exist so the decision to sync is made deliberately rather than by default. make digest-status for the breakdown; the checklist line in CLAUDE.md and fix-issue.md says to sync when a class of failure keeps recurring, not when a counter crosses a threshold.

/mxcli-dev:wiki-sync now points at digest-status as the scope question for the bug-patterns/ category.

Next

Step 3b: actually writing the pattern pages. 607 undigested findings, and the areas with the largest gaps (mdl/executor 238, cmd/mxcli 95, mdl/backend 82) have no page mentioning them at all.

🤖 Generated with Claude Code