You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Per-turn diff review (checkpoint-to-checkpoint), with unified and split views #148
Every turn is bracketed by hidden-git-ref checkpoints and the UI shows the diff of each turn, in unified and split views (internals).
Solenta today
The data already exists: per-turn auto-commits (coder-checkpoint: turn N) and per-checkpoint-pair shortstats (src/shared/ipc.ts:1146-1177). But the UI renders only shortstat review bars (src/components/ThreadView.tsx:201) plus one whole-worktree unified diff (src/components/ThreadView.tsx:1458-1466, tinted lines, truncated). No per-turn patch view, no split view.
Why it matters
Reviewing agent output turn-by-turn localizes mistakes to the turn that caused them; a 40-file cumulative diff hides that. Solenta already pays the git cost, so this is mostly UI work.
Deep-dive correctness recipe (Waku): per-turn diff attribution needs three refs per turn (start/end/diff-base), a turn-START snapshot distinct from the prior turn-end, and git merge-tree --write-tree for a virtual branch start so per-turn diffs contain only what the agent actually did — even across mid-turn branch switches. Plus isolated-index snapshotting (temp GIT_INDEX_FILE) so snapshots never disturb the user's staging area.
Research note (Zed): git.diff_base-style merge-base diffing — show worktree changes vs the DEFAULT BRANCH'S MERGE BASE, not HEAD. For long-lived threads where main has moved on, this is the diff the reviewer actually wants.
Round 27 intel: Orca's Annotate AI Diff is the tightest version of the review-to-agent loop seen so far: leave comments on any lines of the agent's diff, then send them all back to the agent in one batch to revise. No copying line numbers into the composer; the review comments are the follow-up prompt, with file/line context attached automatically. For Solenta's per-turn diff review: the comment UI should live on the diff itself, and "send annotations to agent" should be the primary action, not a copy-paste workflow. Pairs with #278 (inter-diff review) for the multi-round case.
Activity
currentbits commentedon Aug 17, 2026
Deep-dive correctness recipe (Waku): per-turn diff attribution needs three refs per turn (start/end/diff-base), a turn-START snapshot distinct from the prior turn-end, and git merge-tree --write-tree for a virtual branch start so per-turn diffs contain only what the agent actually did — even across mid-turn branch switches. Plus isolated-index snapshotting (temp GIT_INDEX_FILE) so snapshots never disturb the user's staging area.
currentbits commentedon Aug 18, 2026
Research note (Zed): git.diff_base-style merge-base diffing — show worktree changes vs the DEFAULT BRANCH'S MERGE BASE, not HEAD. For long-lived threads where main has moved on, this is the diff the reviewer actually wants.
currentbits commentedon Aug 18, 2026
Round 27 intel: Orca's Annotate AI Diff is the tightest version of the review-to-agent loop seen so far: leave comments on any lines of the agent's diff, then send them all back to the agent in one batch to revise. No copying line numbers into the composer; the review comments are the follow-up prompt, with file/line context attached automatically. For Solenta's per-turn diff review: the comment UI should live on the diff itself, and "send annotations to agent" should be the primary action, not a copy-paste workflow. Pairs with #278 (inter-diff review) for the multi-round case.
currentbits commentedon Sep 10, 2026
Per-turn checkpoint diffs: git.turnDiff plus unified/split panel on the review bar. Git pane remains the cumulative working-tree view.