fix(jcode): keep journal corrections turn-neutral so new turns still count

[Nanako0129]

When a session snapshot ends on a user message, a turn-start is pending as the append-only journal is merged. parse_jcode_messages runs the journal through the same mutable turn-state context as the snapshot, so a journal entry that merely corrects an already-snapshotted assistant id (a replace, whose is_turn_start is overwritten by the snapshot's flag during the merge) consumes that pending turn-start. A subsequent brand-new journal assistant is then never marked is_turn_start, under-counting the session's turn_count by one.

This makes a journal message that only replaces an already-emitted dedup_key turn-neutral: it no longer advances the turn-state machine (its turn flag is discarded anyway), so the pending turn-start survives for the next new message. The snapshot pass passes None and is unchanged.

journal_correction_of_snapshot_message_keeps_pending_turn_start builds a snapshot ending on a user message, a journal that first corrects the snapshot assistant and then opens a new turn, and asserts both turns are counted; it fails before the change (the new turn's is_turn_start is false) and passes after.

Bumps CACHE_SCHEMA_VERSION 24 -> 25 so caches holding the under-counted turn flags reparse. (If a sibling schema-touching fix lands first, this needs a trivial rebase to the next number.)

Fixes #818

---

Summary by cubic

Fixes under-counted turn tracking in Jcode sessions by making journal corrections turn-neutral, so a new assistant message after a user still starts a turn. Also bumps cache schema and includes a minor rustfmt cleanup.

• Bug Fixes

  • Use known snapshot dedup_keys to detect journal replacements and keep them turn-neutral.
  • Only non-replacement assistant messages consume a pending turn-start; snapshot parsing is unchanged.
  • Adds regression test journal_correction_of_snapshot_message_keeps_pending_turn_start.

• Migration

  • Bump CACHE_SCHEMA_VERSION to 26 to invalidate caches carrying under-counted is_turn_start.

^{Written for commit b46b13b. Summary will update on new commits.}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
tokscale	Ignored	Preview	Jul 5, 2026 6:50pm

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}

[junhoyeo]

Thanks @Nanako0129 — this is a really well-built fix. The issue write-up in #818, the turn-neutral replacement approach, the regression test that genuinely fails pre-fix (I verified: without the fix it panics at assertion failed: messages[1].is_turn_start, exactly the under-count you described), and the correctly-annotated cache schema bump all made this easy to verify. The only thing standing between this and green CI was rustfmt wanting the is_replacement binding on one line, so I pushed that one-line formatting commit to your branch rather than bounce it back to you. Merging once CI confirms. Nice catch on a subtle interaction between the snapshot/journal merge and the turn-state machine!

[junhoyeo]

Shipped in v4.0.11, included in the latest v4.1.0 🎉 @Nanako0129