Finishing — What The Change Still Owes

Working is not finished. This page answers one question: what this change owes beside its code, and whether it has been paid. It states no rule of its own. Every row names an owner, the owner decides, and the row reports what the owner said.

Settled — do not re-propose

• Gating the audit on the change being finished. It is asked most often mid-change, as a reminder; a row the window cannot answer yet is not yet, and refusing the whole table over it would make the one moment it is most useful the one moment it does nothing.
• A third lane inside code-review. The review lanes settle code by reading it; most rows here have no diff to read at all — an unwritten test, an unswept ledger row, a diagram nobody updated — and folding them in would give the review skill a second topic, which is a split rather than a lane (skill-authoring, references/splitting-a-skill.md).
• Restating a row's rule here. Every row is a pointer by construction; the moment one explains its owner's rule there are two copies of it, and this is the page that goes stale first (skill-authoring, references/one-owner-per-topic.md).
• A row for the checks. pnpm format, typecheck, lint:fix and the tests are mechanical, batched once at the end, and owned by running-checks — an audit row for them would be a second place to forget them.

When it runs

The order of the ritual, and the commit and push that close it, are CLAUDE.md's "Finishing a change". This is the audit inside its review, test and docs steps — the ones before the checks.

It runs at any point, and it is a reminder before it is a gate. Asked mid-change it reports what is owed so far and marks the rows the window cannot answer yet as not yet — never a refusal to run, and never a demand that the change be finished first. Asked at the end it is the same table with every row answered. Run it unprompted at the end of a change; run it on request whenever.

Scope the audit to the same window a review would take (code-review, references/diff-window.md). A row is asked of that window, never of the repository.

The audit

The change may owe	Ask	Owner
A review, both lanes	quality and correctness over the window	code-review
A regression test — and the deletion of the ones it made redundant	what earns a test at all, and what no longer does	testing, test-values
A refactor the change exposed	the twin helper, the constant restated in two files, the special case that belonged in the mechanism	code-review quality lane, file-organization
Placement and cost	where the work runs and how it is shaped, not how fast the unit is	runtime-efficiency
Docs prose and its diagrams	the owning page, and every diagram whose edge or box the change made false	docs
A README	an inventory, a script table or a typedoc link the change made incomplete	readme-standards
A skill	a convention this change settled, or a skill claim it proved stale	skill-authoring
A bench report	whether a benched unit changed, so its committed report is stale until the bench reruns	bench
An open ledger row over the files touched	whether a sweep still lists this unit unswept, which is a commit ahead of the change	sweeps
An @TODO added, or one the change ended	whether a workaround waits on something external, and whether a bump or a closed issue ended one	todos
An enforcer instead of a repeated finding	whether the same finding has now been written twice	sweeps, oxlint
An enforcer the change made unnecessary	whether a lint rule or test now guards something deleted, or a shape that can no longer be written	invariants
An open proposal the change built, or moved the ground under	whether it shipped one, and whether it edited a file an open proposal's Key files table names	building-proposals, product-review
A decision the change made and left unwritten	a larger part it knowingly left out, or an idea it declined, that the next pass would find as a gap	docs, building-proposals
A finding the change did not fix	whether a review finding it declined or deferred has a home the next session will read	code-review

A behaviour change and a diagram have no name to grep and are missed for it: the stale sentence fails nothing and is found from the code instead (CLAUDE.md, "Finishing a change", the docs step), and a diagram's edge labels are read by no sweep (docs, references/diagrams.md). Both rows are only the reminder to run their owner's lookup.

Report every row, including the empty ones

A row is answered with a verdict, never with silence — "nothing owed" is a result, and an unmentioned row reads as an unasked one. Report the table, one line per row, then do the work the verdicts name.

The rows are independent, so an owner that finds nothing does not excuse the next: a change can be clean in both review lanes and still owe a diagram.