[ellie]

No description provided.

[greptile-apps]

Confidence Score: 4/5

[Medium risk] Fixes session history reconstruction for compacted AI conversations.

The PR should not merge until cycle reconnection cannot fast-forward a copy onto an unrelated rewind branch.

Findings

1.  Compaction joins the wrong branch ▶

Summary

The PR reconnects compacted Claude and Pi parent cycles during branch analysis and transcript restoration, and extends Claude boundary detection across attachments.

• Adds regression tests for compaction cycles and clock-skewed parent links.
• The branch-analysis fallback can connect a cycle to the wrong same-host rewind branch.

_{Reviews (1) · Last reviewed commit: "fix(ai-session): keep compacted Claude a..."}

[greptile-apps]

Compaction joins the wrong branch If a same-host rewind leaves a dead-end row timestamped just before a compaction cycle, this fallback can choose that row as the cycle’s parent even when the compaction belongs to the other branch. The dead end then appears on the restored head’s path, so sync can fast-forward the wrong local copy and refuse to fast-forward the copy on the actual continuation.

Knowledge Base Used: AI session persistence

[ellie]

supeerrrrrr unlikely

[greptile-apps]

Agreed that the exact arrangement is probably rare. The concern is less its frequency than that the fallback has no branch-affinity check: it selects the newest earlier row (same host first) that does not currently lead into the cycle. An unrelated earlier dead-end can therefore satisfy the predicate and become the cycle’s parent, silently changing path_to/fast_forward rather than failing loudly.

If this heuristic is intentional, I’d treat the comment as low priority; otherwise, a regression test covering two branches, a same-host dead-end, and a later cycle would make the intended invariant explicit.

[greptile-apps]

Fair — this is definitely an edge case requiring an already inconsistent or corrupted transcript. My concern is deterministic rather than probabilistic: once that cycle exists, before(true) has no branch-membership check, so any earlier same-host row that does not lead into the cycle can be selected, including a dead end from another branch.

If sync invariants guarantee that earlier same-host rows are always the intended continuation, then this fallback is safe; otherwise, the rare cycle is precisely where that assumption becomes unreliable.

[greptile-apps]

Fair — this requires a malformed or partial parent graph with an earlier same-host dead end that survives the cycle check, so it’s a very unlikely sync-state edge case rather than a practical concern for normal sessions. I wouldn’t block this change on it; at most, it’s worth a regression test if the cycle-repair fallback is intended to defend against corrupted graphs.

[ellie]

/merge

[atuin-bot]

Landed as e8a2c63. GitHub shows this merged once it reaches main.