Review Branch Changes and Fix Ref Spelling

Pi·GPT-5.6-terra·Soph·1mo ago·1146hr 59min·2 Checkpoints·15 file changes·+439/-16·316.6K tokens

can you review the changes in this branch

1mo ago·2.5s

Finding

  • P1 cmd/entire/cli/checkpoint/refs_store.go:480 — remote lookup still fetches only the canonical ref name. A checkpoint written on a case-insensitive machine may be pushed under the folded spelling; a fresh case-sensitive clone has neither local spelling, calls FetchCheckpointRef(canonical), receives “not found,” and never tries the folded remote ref. This also makes remote-list discovery show a stub that cannot hydrate. Retry refFetcher with FoldedRefName(cid) after canonical remote absence, then resolve locally.

Tests: checkpoint package passed as part of mise run test -- ...; the overall command exceeded the 120s tool timeout.

can you review the changes in this branch

1mo ago·5.7s

Re-reviewed current HEAD (3529c0cba). No remaining findings.

The remote fallback now probes both canonical and folded refs, with coverage for successful folded fetches and transport-failure short-circuiting.

Verified with targeted tests:

can you reduce the PR description

1mo ago·6.1s

Reduced the PR description: https://github.com/entireio/cli/pull/2484

there are pr comments, can you take a look

1mo ago·2.8s

I checked the review comments against current HEAD:

  • Remote fetch fallback: already fixed by 3529c0cba (resolveRefMaybeFetch now tries canonical then folded spelling). The Cursor/Copilot comments on that point are outdated.
  • Migration CAS / enqueue: valid. migrate.go still computes refName with RefName(cid), while refBase may resolve a packed folded ref. Migration can then CAS and enqueue the canonical name rather than the existing folded one, causing CAS retries/failure or a stale queue entry.
  • Existing queued canonical refs: valid. A queue entry created before this fix may name the canonical ref while the actual packed ref is folded. partitionLocalRefs does exact lookup, drops that entry as stale, and leaves the checkpoint unpushed until another write re-enqueues it.

I’d address the latter two together: make migration select the resolved write spelling, and make queued-ref validation recognize the folded spelling rather than pruning it.

do it

1mo ago·9.3s

Implemented both valid review fixes.

  • migrate.go now preserves an existing folded ref spelling for CAS and push-queue entries.
  • Push-queue processing resolves legacy canonical entries to a packed folded ref, normalizes the queue safely, and pushes the actual ref.
  • Added migration and queued-push regressions.

Validated:

9fcc485checkpoint: migrate through the resolving ref spelling refBase now returns the tip of a ref stored under its case-folded shard spelling, but migrate built its ref name with raw RefName. On a ref that only resolves folded — after git packed it — the CAS took `expected` from the folded tip and applied it to the canonical name, which holds nothing, so git refused the update and the migration failed its retries. The already-imported branch had the matching defect: it enqueued a name that cannot push. Routing migrate's ref name through writeRefName fixes all three sites at once, including the ErrObjectNotFound retry branch's Reference lookup. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRW9HR0VK1Y8R0JTMMZER+46/-1
89d6feestrategy: push the queued ref spelling that resolves locally A push queue written before folded-ref handling holds canonical ref names. Once git packs such a ref under its filesystem-selected shard directory the canonical name stops resolving, and partitionLocalRefs classified it as STALE — never pushable, so drop it. The queue entry was removed and the checkpoint was silently never pushed, permanently. partitionLocalRefs now resolves each queued name to the spelling that exists locally before deciding, which also covers the OPF pre-push rewrite: it reaches the same helper through Peek and never normalizes. flushCheckpointRefsQueue additionally rewrites the queue itself, so the stale spelling stops costing a lookup on every later push. Only a definite ErrReferenceNotFound triggers the rewrite — a transient lookup failure keeps the queued name rather than rewriting a queue entry on a flaky read. The rewrite enqueues before removing, so an interrupted normalization leaves the checkpoint queued twice (Drain dedups) rather than dropping it. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRY0CGP9H9NSFS7MPXES2+103/-4
f1792e4checkpoint: migrate through the resolving ref spelling refBase now returns the tip of a ref stored under its case-folded shard spelling, but migrate built its ref name with raw RefName. On a ref that only resolves folded — after git packed it — the CAS took `expected` from the folded tip and applied it to the canonical name, which holds nothing, so git refused the update and the migration failed its retries. The already-imported branch had the matching defect: it enqueued a name that cannot push. Routing migrate's ref name through writeRefName fixes all three sites at once, including the ErrObjectNotFound retry branch's Reference lookup. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRW9HR0VK1Y8R0JTMMZER+46/-1
4c92813strategy: push the queued ref spelling that resolves locally A push queue written before folded-ref handling holds canonical ref names. Once git packs such a ref under its filesystem-selected shard directory the canonical name stops resolving, and partitionLocalRefs classified it as STALE — never pushable, so drop it. The queue entry was removed and the checkpoint was silently never pushed, permanently. partitionLocalRefs now resolves each queued name to the spelling that exists locally before deciding, which also covers the OPF pre-push rewrite: it reaches the same helper through Peek and never normalizes. flushCheckpointRefsQueue additionally rewrites the queue itself, so the stale spelling stops costing a lookup on every later push. Only a definite ErrReferenceNotFound triggers the rewrite — a transient lookup failure keeps the queued name rather than rewriting a queue entry on a flaky read. The rewrite enqueues before removing, so an interrupted normalization leaves the checkpoint queued twice (Drain dedups) rather than dropping it. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRY0CGP9H9NSFS7MPXES2+103/-4
4476b20checkpoint: migrate through the resolving ref spelling refBase now returns the tip of a ref stored under its case-folded shard spelling, but migrate built its ref name with raw RefName. On a ref that only resolves folded — after git packed it — the CAS took `expected` from the folded tip and applied it to the canonical name, which holds nothing, so git refused the update and the migration failed its retries. The already-imported branch had the matching defect: it enqueued a name that cannot push. Routing migrate's ref name through writeRefName fixes all three sites at once, including the ErrObjectNotFound retry branch's Reference lookup. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRW9HR0VK1Y8R0JTMMZER+46/-1
f09d8e4strategy: push the queued ref spelling that resolves locally A push queue written before folded-ref handling holds canonical ref names. Once git packs such a ref under its filesystem-selected shard directory the canonical name stops resolving, and partitionLocalRefs classified it as STALE — never pushable, so drop it. The queue entry was removed and the checkpoint was silently never pushed, permanently. partitionLocalRefs now resolves each queued name to the spelling that exists locally before deciding, which also covers the OPF pre-push rewrite: it reaches the same helper through Peek and never normalizes. flushCheckpointRefsQueue additionally rewrites the queue itself, so the stale spelling stops costing a lookup on every later push. Only a definite ErrReferenceNotFound triggers the rewrite — a transient lookup failure keeps the queued name rather than rewriting a queue entry on a flaky read. The rewrite enqueues before removing, so an interrupted normalization leaves the checkpoint queued twice (Drain dedups) rather than dropping it. Refs #2402 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Entire-Checkpoint: 01M2JVRY0CGP9H9NSFS7MPXES2+95/-5