aiti-git-review

❝ As a human, this is my last stubborn stand against AI. ❞

Git code review extension: by default reviews "all uncommitted changes" (git diff HEAD + untracked, with git concepts hidden from the user), and also supports arbitrary version baselines — comparing base (branch/tag/commitid, default HEAD) against ref2, the two versions against each other. Read-only git whitelist (status/diff/log/show/rev-parse/merge-base/ls-files/hash-object/for-each-ref), no write subcommands at all.

• Three context-aware entry points: top toolbar 👁️ (whole repo) / file toolbar 👁️ (that file) / folder toolbar 👁️ (that directory)
• Tasks are first-class citizens: a picker (create · check files in the file tree + AI pre-review option / open the existing list / delete / GC of empty tasks after 30 days), tasks persisted at <extRoot>/state/aiti-git-review/tasks/
• Version baselines: base (branch/tag/commitid, default HEAD) versus ref2, the two versions compared against each other; /refs lists branches/tags/remotes/last 100 commits grouped by day; the picker switches between "workspace / two versions"
• Monaco diff workspace: one-click switching across the four views side-by-side/merged/new-only/outline, F7/Shift+F7 hunk-level navigation across files, width-bucket layout memory, small-screen three-zone adaptation
• Line-level annotations (human + AI in one store): three entry points — right-click / ✍ mode / the pipeline form on the right; severity must/suggest/question; anchor flag-line snapshot + automatic re-mapping on content-hash drift (line numbers stay valid while editing mid-review; unfindable ones are marked stale with the original text kept); revision optimistic lock
• AI pre-review: aiti_review_* tools (status/diff/comment/comments/comment_update/summarize) are injected only into review task sessions (agentTools gated per session + boundSessionId persistence + ai-prebind lazy binding); day-to-day sessions get zero tools
• Summary final-review checklist: long-form findings as file#line [severity] comment + +N/-M stats; one click to send to the current session / open a new session / copy
• Rename/binary/untracked/no-HEAD all compatible: R-status dual-path diff keeps rename pairing, binary markers stay out of the diff, empty-tree baseline fallback before the first commit
• trace source + partial rollback: the review source is wired into the aiti-git-trace ledger; one-click partial rollback within a review (file/line level, write path consolidated into git-trace)
• Two-stage lazy-loading conversion (M3 batch ②, lazy-loading contract); the first case of cross-extension editor/folder bring-up (ensureMaterialized + fixed cross-extension registration ordering)
• services outlets: aiti-git-review.aiReview — groupReview (AI-refined commit grouping, consumed by aiti-git-ops M4) + commitMessage (AI-generated commit messages, consumed by the folder CommitDialog); spawns a prompt session + agent.onAgentSettled event-driven answer retrieval (single 180s deadline, no polling); when absent, consumer buttons are hidden

Agent tools (review task sessions only)

aiti_review_status           # survey the changeset (files + statuses)
aiti_review_diff             # structured diff (hunks/segments with line numbers)
aiti_review_comment          # AI annotates directly (severity + line + comment)
aiti_review_comments         # read existing annotations (avoid duplicates)
aiti_review_comment_update   # revise / mark resolved
aiti_review_summarize        # summary final-review checklist

ai-helper help point

skills/pre-review/SKILL.md — the AI pre-reviewer handbook (verify against the real code, annotate item by item, produce final-review recommendations; iron rule: never modify code).

Full design: internal plan document