fix(server): the turn after a stopped turn shows its diff again

[Ieruss]

After a stopped turn, the next completed turn can show “Failed to load turn diff” even though its Git checkpoints exist. The start lookup filters out both interrupted-run checkpoints and the runless baselines left when a run captured nothing.

For an adjacent turn range, use the ready checkpoint at the previous ordinal in the destination’s scope, matching checkpoint capture. Keep the completed-run lookup as the fallback. SQL and memory checkpoint contexts now expose the existing scope ordinal; no migration or public contract change is needed. Cumulative/full-thread selection and destination eligibility retain their current behavior.

Fixes #17133. This continues @Ieruss’s fix in #17150, preserving the author’s commits and merging main normally. The stopped turn’s own diff remains a separate issue.

Verification on Linux, Node 24.21 and Git 2.43, against main 2b5a879:

• Test takeover: 51 → 51 cases in two existing files (49 on main). Keep the two meaningful stopped-run variants; strengthen their exact scope/ref, cumulative, full-thread and interrupted-destination assertions. Reuse the existing projection-context test for real SQL and memory stores.
• Reverting production to main makes both stopped-turn cases and the projection-context case fail: 3 failed, 48 passed. Removing only the memory ordinal mapping independently fails that assertion.
• 59/59 focused tests pass across CheckpointDiffQuery, ProjectionStore, CheckpointService, CheckpointCaptureService and CheckpointScopeOwnership. Server typecheck and scoped lint/format checks pass.
• Real Git captures through production CheckpointStore, VCS registry and Git driver reproduce the missing-start error on main and return only the next completed turn’s patch on this head, for both captured-interrupted and runless-baseline histories. The query receives deterministic projected metadata; real SQL/memory metadata reads are covered separately above. Before/after output · reproduction script.

One fresh parallel whole-PR review round with GPT-6 Sol and Claude Opus 5.5 found no confirmed bugs or code blockers. The original implementation already matched the capture invariant; the takeover kept that small model and shortened its comment.

This is server-service proof; a complete provider turn, WebSocket/client rendering, macOS, Windows, iOS, Android, Electron desktop and remote/relay/tunnel sessions were not exercised here. No screenshot is presented as client verification. Current-head CI passes and Macroscope approved this exact head; there are no unresolved review threads. Superseded duplicate fingerprint runs are cancelled; the replacement fingerprint runs passed. Auto-merge is off; this PR remains unmerged.

Maintainer takeover: GPT-6 Astra tests, GPT-6.1 Sol implementation, GPT-6 Sol and Claude Opus 5.5 adversarial reviews, through T3 Code with Codex/Claude harnesses.

[macroscopeapp]

Approvability

Verdict: Approved at 5d9c71c

Macroscope's review found this PR approvable — This is a narrowly scoped server bug fix that selects the correct preceding checkpoint after stopped turns, with coverage for both stopped-run checkpoint variants and SQL/memory projection stores. Existing cumulative and full-thread behavior remains covered and unchanged in scope.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Path: .coderabbit.config.ts
• Review profile: CHILL
• Plan: Advanced
• Run ID: ad88df6b-bc2f-46da-89a6-87c3546966b6

📥 Commits

Reviewing files that changed from the base of the PR and between aa8bf8b and 5d9c71c.

📒 Files selected for processing: 4

• apps/server/src/checkpointing/CheckpointDiffQuery.test.ts
• apps/server/src/checkpointing/CheckpointDiffQuery.ts
• apps/server/src/orchestration-v2/ProjectionStore.test.ts
• apps/server/src/orchestration-v2/ProjectionStore.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 0 remain after this review.

📝 Walkthrough

Checkpoint context now includes each checkpoint’s scope ordinal. For consecutive turn diffs, the query can use the ready checkpoint immediately before the target checkpoint in the same scope, including when that checkpoint has no run ID.

Changes: Turn diff lookup

Layer / File(s)	Summary
Expose checkpoint scope ordinals apps/server/src/orchestration-v2/ProjectionStore.ts, apps/server/src/orchestration-v2/ProjectionStore.test.ts	Checkpoint context includes ordinalWithinScope in its decoded shape and SQL and memory store results. Tests check the expected context and missing-thread error in both stores.
Resolve consecutive-turn starting checkpoint apps/server/src/checkpointing/CheckpointDiffQuery.ts, apps/server/src/checkpointing/CheckpointDiffQuery.test.ts	For consecutive turn counts, getTurnDiff prefers the ready checkpoint one scope ordinal before the target, then falls back to the appRunOrdinal lookup. Tests cover interrupted runs with or without a captured checkpoint.

Priority: ➖ Normal

Estimated code review effort: 2 (Simple) | ~10 minutes

Change: Bug fix · Severity of issue fixed: Low

Suggested reviewers: juliusmarminge

Merge Risk: ⚪ Minimal · up to 5d9c7

This change fixes the turn diff after a stopped turn by looking up the previous checkpoint within the same scope, with the old lookup kept as a fallback. No actionable merge-blocking risk was found. The stopped turn's own diff and multi-turn ranges are intentionally unchanged.

Security architecture risk: 🔵 Low · up to aa8bf

The change restores comparisons after stopped turns while retaining thread and snapshot-scope restrictions. No new security issue was identified, although interruption and concurrent recovery behavior were not exhaustively verified.

Retained concerns
No architecture-level concerns identified.

Security review details

Security Blast Radius

• inferred — The changed selection is bounded to the requested thread's stored checkpoints and the selected target scope. It permits stopped-run or baseline content to participate in an existing comparison, but does not add caller-supplied filesystem paths or checkpoint refs. Thread filtering establishes data selection, not independently a tenant-authorization guarantee.

Trust Boundaries and Controls

• observed — The destination remains restricted to ready checkpoints belonging to completed runs. Only predecessor eligibility is broadened, with explicit same-scope and preceding-ordinal checks. The routed test helper declarations are local fixtures, not external entrypoints.

Resilience and Maintainability Implications

• observed — Existing capture controls use scope-derived checkpoint identity, a unique scope/ordinal database index, and non-ready statuses for failed captures. Settled captures are skipped on repetition, rolled-back runs are rejected, and checkpoint/run events are committed through an idempotent transactional command path. These controls support the normal-path predecessor invariant without proving every filesystem interruption outcome.

Pre-merge checks: 4

Check name	Status	Explanation
Linked Issues check		Issue #17133 requires the latest completed turn diff to load after a stopped turn. CheckpointDiffQuery now selects the ready checkpoint at the preceding ordinalWithinScope in the target scope for …
Out of Scope Changes check		The production changes support issue #17133. CheckpointDiffQuery.ts changes adjacent-turn start selection, and ProjectionStore.ts exposes the required ordinal in both store implementations. The te…
Title check		The title is concise, specific, and accurately describes the primary server fix for diffs after a stopped turn.
Description check		The description covers the problem, implementation, scope, linked issue, verification results, limitations, and review status. It does not use the template headings, but the required information is pr…

Full details: Linked Issues check

Explanation

Issue #17133 requires the latest completed turn diff to load after a stopped turn. CheckpointDiffQuery now selects the ready checkpoint at the preceding ordinalWithinScope in the target scope for an adjacent turn range. It retains the appRunOrdinal lookup as fallback. ProjectionStore supplies ordinalWithinScope from SQL and memory projections. Regression tests cover stopped-run checkpoints and runless baseline checkpoints, plus cumulative and full-thread selection and the stopped turn error.

Full details: Out of Scope Changes check

Explanation

The production changes support issue #17133. CheckpointDiffQuery.ts changes adjacent-turn start selection, and ProjectionStore.ts exposes the required ordinal in both store implementations. The test changes verify the regression and preserve existing selection and error behavior. No unrelated product behavior is demonstrated. The explicitly excluded multi-turn ranges, stopped-turn own diff, and provider-running state are not changed.

Full details: Description check

Explanation

The description covers the problem, implementation, scope, linked issue, verification results, limitations, and review status. It does not use the template headings, but the required information is present and detailed.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[chisewaguri]

The stopped turn's own diff still fails here, because the to side keeps the completed filter. Dropping only rolled-back runs from the ready set fixes that case too, as long as the stopped run captured a checkpoint of its own.

Dismissing prior approval to re-evaluate aa8bf8b

[mrGreazly]

Independent verification on a 103-turn thread (pi provider, Windows) — posted full details in
#17133. Short version for this PR: I simulated the parent-lookup against my projection rows and
all 16 broken single-turn diffs resolve (15 via the null-run/null-turn-count baselines left by
steer placeholders, 1 via an interrupted run's checkpoint), with 64 already-working cases
unchanged and none regressing. The trigger in my case is a queued message promoted to a steer —
no Stop involved — so this fix covers more than the stopped-turn repro in the issue.

[juliusmarminge]

Taking over this one for batch worktree-git-05. I’ll resolve the main conflict, keep the stopped-turn baseline regression focused, and verify SQL and memory projections before publishing here without merging.

Dismissing prior approval to re-evaluate 5d9c71c