prereview

prereview

Review your own changes — per line — before you push, and hand the comments to an LLM to act on.

A tiny local webapp for per-line review of your working tree — no commit, no PR, no GitHub round-trip. It ships as a Claude Code skill: /prereview launches a session on your current changes, you comment per line, hit "Hand off → Claude", and Claude applies them. Run it on a remote dev box and it auto-binds your Tailscale address — review and hand off straight from the Claude mobile app, over the tailnet, before anything is pushed.

Status: core flow (review → hand-off → LLM applies) works end-to-end and is in daily use; UI still being polished.

_{Per-line review & hand-off on desktop · the all-comments view · the same flow from your phone}

Features

• Per-line & range comments on your working tree — two-click range select; comment on any file (changed or not); a single file or a non-git directory works too.
• The hand-off loop — a "Hand off → Claude" button writes a marker; the Claude Code skill polls it, reads the CSV, and applies your comments. Review → hand-off → changes, without leaving the chat.
• Ships as a Claude Code skill — /prereview (or "review my changes") launches a session scoped to your current changes; prereview --install-skill installs it from the binary.
• Responsive, phone-friendly UI — review from an iPhone via the Claude mobile app + /remote-control, comment, hand off, push.
• Tailscale-aware binding — on a remote (SSH) box it auto-binds your tailnet address (reachable from your phone, never the public internet); locally it stays on 127.0.0.1. No --host juggling.
• Markdown renders, but comments by source line — .md shows formatted; tapping a rendered block selects its real line range, so it round-trips with the raw view and the CSV.
• Diff or full-file view — changed hunks with context, or the whole file; line numbers match so a comment resolves across both.
• Durable, boring storage — one RFC-4180 CSV, atomic writes (tmp+fsync+rename). Survives restarts; resume where you left off.
• Single Go binary — embeds every asset; no Node/npm, no JS runtime.

Install

Binary

go install github.com/livetemplate/prereview@latest

This puts prereview on your $PATH.

Claude Code skill (optional, for the LLM-driven flow)

The skill lets Claude Code launch and drive a review session for you — just say "review my changes" or /prereview instead of running the binary by hand.

Install it with one command (the binary embeds the skill):

prereview --install-skill
# → Installed prereview skill → ~/.claude/skills/prereview/SKILL.md

That writes ~/.claude/skills/prereview/SKILL.md (+ reference.md), available in every repo. Re-run after upgrading the binary to refresh the skill.

mkdir -p .claude/skills/prereview
cp skill/SKILL.md .claude/skills/prereview/SKILL.md

Requirements & notes:

• The skill shells out to the prereview binary, so install the binary (above) first and keep it on $PATH.
• Invoke it with /prereview, or natural-language triggers like "review my changes" / "per-line code review".
• New or renamed skills are normally picked up within the running session; if /prereview reports "unknown skill", run /reload or restart Claude Code (discovery is primarily a startup scan).
• skill/reference.md is supplementary (full CSV schema + filesystem contract). The skill works without it, but copy it alongside SKILL.md if you want Claude to have the deep reference handy.

Quick start

Standalone (manual review)

cd <your-repo>
prereview
# stdout: READY http://127.0.0.1:43029

Open the URL. Browse files in the left drawer, tap a line to select it, tap another to extend the selection, type a comment, save. Click Quit when done — the server shuts down. Your comments live in .prereview/comments.csv.

Skill mode (LLM-driven)

With the Claude Code skill installed, just ask Claude to "review my changes" (or run /prereview). Claude launches the session, hands you the URL, waits for your hand-off, then reads and acts on your comments — you don't run anything by hand.

Equivalent manual invocation (what the skill runs for you):

prereview --skill --repo "$(pwd)" --base HEAD &

The UI shows "Hand off → Claude" instead of "Quit". When you click it, .prereview/DONE is written with the path to the CSV; the skill polls for that file, reads the CSV, and acts on the comments.

See skill/SKILL.md for the LLM-side integration and skill/reference.md for the full schema + filesystem reference.

Flags

Usage tips

• Two-click range selection. Tap a line to anchor. Tap another to extend the selection. Tap again to reseat the anchor. Type and save.
• Edit/resolve/delete. Each saved comment has Edit, Resolve, Delete buttons in the inline card. Resolve keeps the comment in the CSV with resolved=true (audit trail); Delete removes it entirely (with an Undo toast).
• All-comments overview. The 💬 N chip in the top bar (desktop) or the overflow menu (mobile) jumps to a list view of every comment across all files; tap a comment to jump back to its source line.
• Diff vs File. Toggle "Diff" → "File" (overflow menu on mobile, chip on desktop). Diff shows only the changed hunks — changed lines plus 3 lines of context, with long unchanged runs collapsed to a "··· N unchanged lines ···" marker (an unchanged file has no diff, so it shows whole). File shows the entire current file with no diff at all — no add/del coloring, deletions omitted. Line numbers are identical in both modes, so a comment made in one resolves in the other.
• Markdown renders by default. .md/.markdown files show formatted (headings, lists, code, tables) instead of raw lines. You can still comment: tap a rendered block (heading, paragraph, list, code fence…) and it selects that block's source line range — the comment anchors to real line numbers, so it round-trips with the raw view and the CSV exactly like a line comment. A "Rendered" ⇄ "Raw" toggle switches to the source line view. Raw HTML embedded in the Markdown is not rendered (safe for untrusted repos).
• Base picker. The file drawer has a base dropdown listing HEAD~1/3/5/10, every local branch, and every remote-tracking branch (e.g. origin/main). For anything more exotic (a specific commit SHA, a tag), pass it once at launch with --base.
• File scope. The drawer defaults to changed files only (vs the base) — the right default on a large repo. A "Changed N · show all M" toggle switches to the full tracked-file list when you want to comment on something that didn't change. On a clean tree (nothing changed) it auto-shows everything and the toggle is hidden.

Output

<repo>/.prereview/comments.csv is the source of truth — RFC-4180 quoted, 8 columns, one row per comment:

id,file,from_line,to_line,side,body,created_at,resolved

See skill/reference.md for the full column documentation.

Architecture (at a glance)

• Single binary, embeds all assets (including the livetemplate client JS bundle) via //go:embed.
• State held server-side in livetemplate's session storage; WebSocket-driven UI patches. Pure Go on the server; no Node/npm.
• Atomic CSV writes via tmp+fsync+rename+parent-fsync. Read at any time without risk of a torn file.
• Syntax highlighting server-side via chroma — GitHub light theme. Cached per file path; falls back to plain text for unsupported languages.

Development

git clone https://github.com/livetemplate/prereview
cd prereview
make sync-client   # copies the latest livetemplate-client.js into internal/assets/client/
go build .
./prereview --repo "$(pwd)" --port 8765

E2E tests use chromedp + a headless browser:

go test -tags=browser ./...

License

MIT.