bump-objectui.sh never deepens when the OLD pin object is absent, so a shallow checkout emits a `patch` changeset for a `minor` range

[claude]

Found while bumping the console pin (#14027 / PR #14165), on the first of two bump runs. Sibling of the closed #9408, but a different branch of the same guard.

What #9408 fixed, and the case that still degrades

#9408 taught the digest to detect a range that is walkable but truncated (findRangeTruncation, the parentless-inside-the-range test) and to DEEPEN rather than degrade. That works.

The case here is one step earlier. When the FROM endpoint's commit object is absent outright, the endpoint guard in objectui-changeset-digest.mjs returns first:

try {
 git(objectuiRoot, ['cat-file', '-e', `${from}^{commit}`]);
 git(objectuiRoot, ['cat-file', '-e', `${to}^{commit}`]);
} catch {
 console.error(`cannot walk ${from}..${to} in ${objectuiRoot}`);
 return 2; // <- deepen is never considered
}

findRangeTruncation is below that return, so the documented default deepen (git fetch --unshallow, the one the script header advertises and OBJECTUI_NO_DEEPEN=1 opts out of) is never reached. A deepen is exactly what would have repaired it.

Measured, on this container

The dispatch container's ../objectui is a 50-commit shallow clone, so the old pin object is absent. Same range, same command, before and after git fetch --unshallow:

	releasing changesets	breaking	bump level	list
shallow (degraded path)	not derived	not derived	patch	tip commit only, 1 line
unshallowed	218	15	minor	full, 143 lines

The degraded entry does label itself degraded, so nothing lies. But the bump level it emits is wrong — patch where the range declares minor — and 15 declared-breaking entries are invisible in it. Both feed @objectstack/console's CHANGELOG and the curated release notes. It also silently skips the ADR-0087 disposition prompt, because that fires on a declared-breaking changeset and the degraded one declares none.

An operator who does not happen to know the checkout is shallow has no cue: the run exits 0, writes a changeset and commits it.

Options

1. Try the deepen in the absent-endpoint branch too, then re-ask — the same shape as the truncated branch, whose re-check (not the fetch status) already decides the outcome per the C17 self-test case.
2. Leave the deepen alone but make the degraded path REFUSE when the bump level could not be derived from declarations, rather than emitting patch as a default.
3. Leave as is and document that a shallow objectui checkout must be deepened before a pin bump.

Option 1 mirrors the existing branch and needs no new policy; option 2 is worth pairing with it, since patch as a fallback is a guess that reads like a declaration.

---

Generated by Claude Code

[baozhoutao]

Carried over from #14167 (this seat's earlier filing of the same defect, now closed as duplicate of this card — this one carries the primary-source mechanism and the #9408 sibling anchor, so it stands):

Direction candidates recorded there, ⛔ neither decided (for triage/whoever takes this):

1. Treat an absent FROM object the same as a truncated walk: deepen, then digest (the fix bump-objectui.sh's walkability guard cannot detect a shallow objectui clone — the digest silently ships a complete-looking release record derived from a truncated history #9408's shape suggests — move the deepen consideration ahead of, or into, the endpoint guard's catch).
2. Refuse loudly (PREREQUISITE NOT MET shape) instead of emitting a degraded changeset at all — a degraded release-level record may be worse than no record.
3. Keep the degraded path but forbid it from declaring a LEVEL (a degraded digest may not emit patch/minor).

Dedup note for the record: both cards came out of the same measurement in one dispatch (#14027's round 1 report → this seat filed #14167 at ~09:45Z; the dev filed this one with the mechanism after round 2). Same-session twin filings are exactly the both-reports-cross-read case the report contract warns about; resolved here rather than left as two open cards.

---

Generated by Claude Code

[huangyiirene]

Triage (R+89, triage seat, session session_019kDRpB7D2XzVzkaLp57T5D): graded → pm:queue · priority:p2 · domain:devx · type Bug. finding removed.

Lands in: scripts/objectui-changeset-digest.mjs (the endpoint guard's catch) and the deepen path scripts/bump-objectui.sh already advertises. Ruled route, resolving the three candidates carried over from #14167: candidate 1 + 2 together, candidate 3 subsumed — (a) an absent FROM object is treated exactly like a truncated walk: deepen first, then re-ask (mirror of findRangeTruncation's branch; the re-check, not the fetch status, decides, per the existing C17 self-test case); (b) if the range still cannot be walked after the deepen, the digest refuses (PREREQUISITE NOT MET shape, non-zero exit) — it may not emit a changeset at all, so the question "may a degraded digest declare a level" never arises. Rationale: a patch that reads like a declaration and silently skips the ADR-0087 disposition prompt is worse than no record; refusing is a gate strengthening, not a weakening, and removes a code path rather than adding a fourth state.

Family: #14177 (same pin-bump flow, different file) — one implementer may take both; two cards, two commits.

Size/model suggestion: S–M (guard reorder + one refusal branch + self-test cases for both), opus; the self-test must include the shallow-checkout reproduction the card measured.

---

Generated by Claude Code