Claude Code Diff Review

Per-hunk review of every change Claude Code makes. Accept, reject, edit, or ask "why?" — with risk flags and a typecheck signal pointing you at what matters, without leaving VS Code and without touching git.

If you let Claude Code edit dozens of files in a single autonomous turn, you need a way to review each change before keeping it. This extension turns that review into a per-hunk Accept / Reject / Ask Claude flow — session-aware, snapshot-backed, no git required.

Highlights

• Per-hunk decisions — Accept, Reject, Edit-in-place, or Ask Claude on each hunk independently, even within one file.
• Decision support, not just mechanics — heuristic risk flags and a post-turn TypeScript build signal surface the hunks worth scrutinising so you don't review 50 changes blind.
• Keyboard-driven — fly through a review with j / k / a / r / e / ?; jump straight to flagged or build-breaking hunks.
• Snapshot-backed, no git — every reject/edit replays against a captured pre-edit snapshot; reversible at any time, independent of your VCS.
• Full history + Resume — every session is logged; reopen one mid-review with decisions intact, roll a turn back, or mine cross-session Insights.
• In-context AI chat — chat scoped to a hunk, augmented with the original prompt + surrounding tool calls that produced it.
• Pro / Max plan support — uses your claude /login OAuth automatically; API key also supported.

What it does

When Claude Code finishes a turn, the review panel opens with every file it touched. You decide on each change — keep it, drop it, fix it, or interrogate it — before it becomes part of your working tree.

Per-hunk review — four verbs

• ✓ Accept — keep the change, mark the hunk reviewed.
• ✗ Reject — revert just that hunk on disk (the rest of the file stays), mark reviewed.
• ✎ Edit — open the hunk's result inline, tweak it, and commit your version. The edited content is what lands on disk and what's recorded.
• 💬 Ask — open a chat scoped to that hunk.

Decisions are set-based and reversible: every reject/edit re-renders the file from the captured pre-edit snapshot plus the set of accepted (and edited) hunks, so any decision can be undone deterministically — no git required. Bulk actions (✓ File / ✗ File, Accept all / Reject all) and per-hunk / session Undo are built in.

Decision support

Reviewing dozens of hunks is only useful if you know which ones matter. Two signals triage them for you:

• Risk flags — heuristic, zero-cost triage at review time: sensitive paths (.env, secrets, auth, migrations), pure deletions, removed error-handling / null-checks, oversized hunks, lockfiles, test files. Flagged hunks get a chip; the header shows an N flagged count; Shift+J / Shift+K jump between them. Toggle with claudeReview.riskFlags.enabled.
• TypeScript build signal — after each turn the extension runs your workspace's tsc --noEmit in parallel with panel-open and marks every file + hunk that affects a failing typecheck. The header shows ⏳ tsc: running… → ✓ tsc: passed / 🚨 tsc: N errors; affected hunks get an inline badge with the error messages; Shift+N / Shift+P jump between them. Auto-detects tsconfig.json (or tsconfig.build.json), switches to tsc -b for project-references repos, force-kills on a configurable timeout, and never blocks the panel. Override the command with claudeReview.buildSignal.typecheckCommand (passed via argv — no shell injection) or disable with claudeReview.buildSignal.enabled.

Reject-with-feedback

Rejecting a hunk loses the most valuable signal — why. Add a reason on reject; reasons collect into a drafts queue in the chat overlay, then send as one consolidated message so Claude reworks with your feedback in hand.

Refactor grouping

When a rename touches the same identifier across many files, those hunks are grouped with a ↻ rename · N more chip — accept or reject the whole group in one action instead of clicking through twenty identical changes.

History, Resume & Insights

Every session is logged to a content-addressed event log (default 30-day retention; claudeReview.history.enabled, claudeReview.history.retentionDays). Open it via Claude Review: Open History Panel:

• Resume Review — reopen a session you closed mid-review, with all accept/reject/edit decisions reconstructed.
• Rollback this turn — restore every file in a session to its pre-edit content.
• Delete from history — remove the event log + its unreferenced blobs.
• Insights tab — cross-session analytics no other tool gives you: per-file accept rates ("README.md accepted 80%"), per-sub-agent acceptance, a 30-day rejection-rate trend, and recurring rejection-reason themes.
• Live — sessions started while the panel is open appear within ~300 ms; a ↶ N pending status-bar item tracks unfinished hunks across recoverable sessions.

Transcript-aware chat & sub-agent attribution

Clicking 💬 augments the prompt with the user's original message for the turn plus Claude's surrounding tool calls (read host-side from the session transcript — never forwarded to the webview; toggle with claudeReview.chat.transcriptContext). Files edited inside a Task invocation are labelled via Task: <description> in both the review and history panels, so you always know which sub-agent produced a change.

Resilient auth & editor integration

• Bearer token for the local hook server is stored in the OS keychain and reused across reloads. If a stale terminal returns 401, a burst-detector toast offers one-click recovery (Open New Terminal / Show Logs / Rotate Token); legacy hook entries are auto-stripped on activation.
• Files also appear in a dedicated Source Control group by review status, and CodeLens Accept / Reject buttons sit directly above each hunk in the editor.

Setup

1. Install

Install the extension from the VS Code Marketplace.

2. Open a workspace

Open the project folder you want Claude to edit. The extension writes its hooks to <workspace>/.claude/settings.json automatically.

3. Authenticate Claude

Three options, in resolution order:

Verify which auth path the extension picked: Claude Review: Use Claude Code Auth prints the active source as a toast (without revealing the token).

4. Run a Claude session

Open the integrated terminal in your workspace and run:

claude

Ask Claude to make some edits. When it finishes the turn, the review panel opens automatically.

Daily flow

1. Claude finishes a turn → review panel opens in a new editor tab; tsc starts in the background.
2. File list (left) shows every touched file, colour-coded by status, with risk chips and a build dot.
3. Header summarises progress, N flagged, and the build signal. Use 🏷 Flagged only to focus.
4. Diff pane (right, split or unified) — each hunk has ✓ Accept, ✗ Reject, ✎ Edit, 💬 Ask, plus a risk badge / 🚨 tsc errors badge / rename-group chip where relevant.
5. CodeLens above each hunk mirrors Accept / Reject; the status bar tracks pending hunks across sessions.

Bulk: per-file ✓ File / ✗ File, per-session Accept all / Reject all, and per rename-group. Reject/edit decisions are reversible via per-hunk and session Undo.

Keyboard shortcuts

Open the most recent panel with Ctrl/Cmd + Shift + R. Inside the review panel:

Configuration

Commands

Troubleshooting

Stop hook returns 401

Cause: the integrated terminal predates the extension's activation, so it doesn't have CLAUDE_REVIEW_TOKEN in its environment.

Fix:

1. Developer: Reload Window
2. Close every existing integrated terminal
3. Open a fresh terminal and re-run claude

Verify by running $env:CLAUDE_REVIEW_TOKEN.Length (PowerShell) or echo ${#CLAUDE_REVIEW_TOKEN} (bash). Expect 64.

Review panel is blank

Open the webview's DevTools: command palette → Developer: Open Webview Developer Tools → Console tab. Any unhandled error renders inside the panel via the built-in <ErrorBoundary>. If it's blank with no error, the React tree may have un-mounted; reload the window.

"Extension host did not start under 10 seconds"

On slower machines the activation may flirt with VS Code's 10 s timeout. The extension externalises the Anthropic SDK and defers the hook-config write to keep activation fast (~1–2 s typical). If you still hit this, file an issue with the Output → Claude Code Review panel contents.

Chat says "No Claude credential found"

Run Claude Review: Use Claude Code Auth. It prints which source the resolver picked, or a warning if none matched. Resolution order:

1. CLAUDE_CODE_OAUTH_TOKEN env var
2. CLAUDE_REVIEW_OAUTH_TOKEN env var
3. SecretStorage OAuth (set via the command)
4. ~/.claude/.credentials.json (the file claude /login creates)
5. SecretStorage API key

Reject says "Could not cleanly revert a hunk"

A formatter (or another tool) ran after Claude's edit and the post-edit content has drifted too far for jsdiff's fuzzFactor: 2. The banner offers a Revert file to original snapshot button — that writes the captured before content back in one disk write.

Port 53117 already in use

The server falls back to an OS-assigned dynamic port automatically and rewrites the hook URL accordingly. Check the Output → Claude Code Review panel for the actual port. To force a fixed port, set claudeReview.port to a number that's free.

Roadmap

• Multi-language build signal — jest / vitest / pytest / cargo test / go test via structured (--reporter=json) output, scoped to the languages a turn touched. Lands when a non-TypeScript user needs it.
• v1.0 — file-based hook-token resolution (so terminals spawned outside VS Code work without the env var), zero-config onboarding (credentials prompt only when you first use chat), OpenCode + multi-agent adapter support, multi-root workspace history, and Remote / SSH / Dev Container / WSL.

Feature requests, bug reports, and discussion are welcome at GitHub Issues.

Known limitations

• Single workspace folder. Multi-root workspaces use the first folder for .claude/settings.json and history. v1.0 may shard.
• Remote / SSH / Dev Container / WSL. v0.x is local-only. The hook server binds to 127.0.0.1; remote development needs the server to live in the remote host. Out of scope until v1.0.
• External terminals. Terminals spawned outside VS Code (Windows Terminal, tmux not inside an integrated terminal, etc.) don't inherit the CLAUDE_REVIEW_TOKEN env var. The burst-detector toast catches this and offers [Open New Terminal] in-VS-Code. v1.0 plans file-based token resolution to fix this fully.
• Claude Code version. Requires Claude Code ≥ 2.1.47 for last_assistant_message and stop_hook_active fields in the Stop hook payload.
• Localisation. English only.

Privacy & security

• API keys / OAuth tokens are stored exclusively in VS Code's SecretStorage (OS-keychain backed). They never cross into the webview, never appear in logs, and never get written to a file.
• Loopback server binds to 127.0.0.1 only. Bearer authentication uses a 256-bit token, stored in OS keychain (SecretStorage), reused across reloads, compared in constant time. Explicit rotation available via Claude Review: Rotate Bearer Token.
• Webview CSP sets connect-src 'none' — the webview cannot make network calls. All chat traffic flows through the extension host.
• Telemetry is opt-in (claudeReview.telemetry) and additionally gated by VS Code's global telemetry setting. Events are scrubbed for PII (file paths, message content, tokens) before emission.

Changelog

Full version history is in CHANGELOG.md. Latest: v0.6.1 — a performance + reliability pass (faster review rendering, lower IPC/I/O, fixed a chat-on-large-transcript stall and a mid-session sub-agent attribution bug).

License

MIT. See LICENSE.

Built to make Claude Code more reviewable. If something's off, run Claude Review: Show Log and file an issue with the output.