Code Review Gate

Distrust-driven code review: every line must be looked at. "Review = distrust" turned into a hard gate — read line-by-line + file-level AI analysis + dispose of every finding + cross-file global analysis. Only when all of that is done are you allowed to submit a conclusion.

VS Code extension · powered by the GitHub CLI and Copilot models

English | 简体中文

Core ideas

• Review is distrust: assume the author makes mistakes; no conclusion is allowed until the gate passes.
• Per-line coverage is real: based on a real VS Code editor + viewport tracking (visibleRanges) to judge that "the eyes actually scanned it", not a fake scroll; progress is bound to headSha — the moment the author pushes a new commit, old coverage is invalidated.
• AI is a copilot, not a judge: the model produces findings and fix proposals; whether to accept them is decided by a human, one finding at a time (fix / comment / ignore-with-reason).
• Native first: status bar, decorations, diagnostics use native VS Code APIs; the review workbench, document viewer, fix proposals, and global report use webviews for rich interaction.

Features

Review scope (workspace-first)

• The whole project folder (auto-skips node_modules / .git / dist / out / bin / obj / .vs)
• A self-built webview tree picker, locked to the project root, to pick any file subset
• The current branch's PR (via gh)
• Current branch vs base branch (pure git diff)
• Uncommitted working-tree changes (pure git)

Review workbench

• File tree + per-file / overall coverage progress + findings summary + action buttons, all in one webview
• Can be popped out into a separate window (like "Open in Agents Window")
• Auto-restores the review and workbench after a window reload — no need to re-pick the scope
• Token accounting (estimated): a HUD row shows estimated LLM tokens (↑input ↓output) with a per-operation breakdown on hover, accumulated over the review and persisted; estimated locally via countTokens, not the provider's billed counts

Per-file review

• Per-line coverage tracking and "jump to next unseen line"
• File-level AI analysis → findings with severity levels
• Every finding must be disposed: fixed / commented / ignored (reason required)
• Fix proposals: Copilot generates mutually-exclusive "multi-edit = one complete solution" candidates, one-click apply + one-click undo, with applied snapshots persisted
• Reviewer context before a fix: when the model's read of a finding is off, add a supplementary note in the fix panel; it overrides the finding when generating proposals and is persisted per finding
• Selection translation / code explanation annotations

Global analysis

• Cross-file logic analysis, optionally scoped to selected directories
• A dedicated global-report webview

Gate and conclusion

• Gate: every file is analyzed and all findings are disposed + the global conclusion is confirmed
• Conclusion: Approve / Request Changes / Comment
• When the scope is a PR, one click writes the conclusion back to the GitHub PR (gh); otherwise it is recorded locally

Gate model

gatePassed = all files ready && global analysis confirmed
file ready  = file-level analysis done && every finding in the file disposed

Per-line coverage is shown as a live progress signal (per-file / overall seen/total) to push "actually read it", displayed alongside the gate verdict.

Language (bilingual)

Code Review Gate is fully bilingual (English / 简体中文), English by default.

• codereview.language drives the whole experience — both the extension UI (status bar, prompts, webview panels) and all LLM output (findings, fix proposals, explanations, global conclusions).
  • en (default): everything in English
  • zh-CN: everything in Simplified Chinese
  • auto: follow the VS Code display language
• Switching the setting takes effect live — the status bar and any open webview panels re-render immediately, no reload needed.
• Command titles in the Command Palette follow the VS Code display language (a platform limitation for static contribution points), localized via package.nls.

Commands and settings

Commands

Settings

Architecture

src/
  extension.ts          activation entry, command registration, orchestration, gate
  i18n/                 language resolver + English/Chinese message catalogs
  gh/                   GitHub CLI wrapper (auth / pr view / pr diff / write back review)
  scope/                review scopes: PR / branch diff / working tree / file selection + tree-picker webview
  review/               review session model + persistence (ReviewStore abstraction)
  ai/                   model selection + file-level / global analysis + fix proposals + language directive
  ui/                   workbench / document / fix-proposal / global-report webviews + status bar / toast / progress

Persistence

• workspaceState (local), keyed by repo + scopeId + headSha.
• Applied-fix snapshots are persisted, so "undo fix" and "locate" still work after a reload.
• Progress is bound to headSha: old coverage is invalidated after the author pushes a new commit.
• Note: workspaceState does not follow your GitHub account and is not cross-device; cross-device capability is reserved via the ReviewStore interface for a future remote implementation.

Model and language

• Lists the models currently authorized for Copilot via the VS Code Language Model API, supports Auto, and remembers the choice per workspace.
• codereview.language controls the language of the whole experience (UI + all LLM output); see Language above.

Development and packaging

npm install
npm run compile            # or npm run watch; press F5 to launch the Extension Development Host
npm run package            # production build (esbuild --production)
npx @vscode/vsce package   # produce a .vsix

PR scope and writing back to a PR require the GitHub CLI installed and logged in: gh auth login.

Roadmap

• [ ] Cross-device / account-following review progress (a remote ReviewStore implementation).