Agent Reviewer

GitHub PR-style review comments in VS Code, delivered to your coding agent. Comment on any line or selection — in a diff view or a regular file — and the notes stay staged inside the editor, mirrored to a markdown file in your workspace. That file is the hand-off point: attach it to the Codex chat as a chip, and/or let a CLI hook — Kimi Code or Claude Code — pick it up on your next prompt. When a hook consumes the file, the staged comments clear themselves in VS Code.

中文介绍

How it works

1. Hover a line's gutter and click +, or select a range and press Ctrl+Alt+C (Cmd+Alt+C on macOS). Write as many notes as you like; the status bar keeps count.
2. Every staged change rewrites the pending reviews file (default .agent-reviewer/pending-reviews.md) within ~300ms.
3. Click the status bar item (or the send button in a diff editor's title bar) to flush. Depending on your targets:
   • Codex (agentReviewer.targets.codex): the file is attached to the Codex chat composer as a chip via the Codex extension's chatgpt.addFileToThread command.
   • Kimi (agentReviewer.targets.kimi) and Claude Code (agentReviewer.targets.claude): the file simply sits there — the next message you send in a CLI session in this workspace picks it up automatically (see below).
4. Staged reviews are cleared when a CLI hook consumes the file, or when you run Agent Reviewer: Clear Staged Reviews — e.g. after sending the reviews in Codex, which reads but never consumes the file.

The CLI hook

The extension and the CLI never talk to each other directly — they rendezvous through the pending reviews file:

• A UserPromptSubmit hook checks your session's workspace for .agent-reviewer/pending-reviews.md. If it contains review entries, the hook appends them to your prompt's context and renames the file to pending-reviews.md.consumed.
• The extension watches for that disappearance and clears the staged comments, threads, and status bar count automatically.
• If several sessions run in the same workspace, whichever submits a prompt first consumes the file; the others see nothing. Sessions in other workspaces are unaffected. That applies across agents too: with both hooks installed, whichever of Kimi or Claude Code you talk to first takes the round.

Kimi Code and Claude Code expose the same hook contract — a JSON payload carrying cwd on stdin, and hook stdout appended to the prompt's context — so hooks/inject-reviews.sh serves both without modification.

Kimi

This repository doubles as a Kimi Code plugin — kimi.plugin.json at the root declares the hook. In the Kimi TUI:

/plugins install https://github.com/NihilDigit/agent-reviewer

Confirm the third-party plugin prompt, then /reload or start a new session.

To install by hand instead, add to ~/.kimi-code/config.toml:

[[hooks]]
event = "UserPromptSubmit"
command = "bash /path/to/agent-reviewer/hooks/inject-reviews.sh"
timeout = 10

Claude Code

The repository is also a Claude Code plugin and its own single-plugin marketplace — .claude-plugin/ and hooks/hooks.json declare the hook. In Claude Code:

/plugin marketplace add NihilDigit/agent-reviewer
/plugin install agent-reviewer@agent-reviewer

To install by hand instead, add to ~/.claude/settings.json (or a project's .claude/settings.json):

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "bash /path/to/agent-reviewer/hooks/inject-reviews.sh",
            "timeout": 10
          }
        ]
      }
    ]
  }
}

Either route works if you customized agentReviewer.pendingFile — then also edit the path inside hooks/inject-reviews.sh to match.

What the agent receives

Each note becomes one objective citation: path, line range, side (the old side of a diff carries the commit SHA, the new side carries HEAD), the verbatim code, and your note text. No instructions are attached — whether the agent should explain the code or change it stays entirely up to your conversation.

## [ref #1] src/foo.ts:40-42

working tree (HEAD `1a2b3c4d5e`)

​```typescript
const a = 1;
const b = 2;
return a + b;
​```

> why is `b` added here?

Requirements

• VS Code ≥ 1.96
• Optional, per target: the official Codex extension (openai.chatgpt) for Codex chips; Kimi Code CLI or Claude Code with the hook installed for the CLI workflow

Install

npm install
npm run compile   # esbuild → out/extension.js

Then package with npx vsce package and install via code --install-extension, or copy package.json, out/, and README.md into ~/.vscode/extensions/NihilDigit.agent-reviewer-0.6.0/. For development, open this folder and press F5, or run code --extensionDevelopmentPath=<this repo> <a workspace>.

Usage

Action	Result
Hover the gutter, click +	Comment on that line (regular editors and both sides of a diff)
Ctrl+Alt+C (Cmd+Alt+C)	New review comment on the selection, or on the cursor line when nothing is selected
Click the status bar item	Flush reviews: attach the chip to Codex and/or leave the file for a CLI hook
Thread title buttons	Flush or delete a thread; single notes can be edited or deleted individually

Command palette: Agent Reviewer: Flush Reviews to File, Agent Reviewer: Clear Staged Reviews, and Agent Reviewer: Open Pending Reviews File.

Why the shortcut exists: the gutter + belongs to VS Code core. When several comment providers (e.g. GitHub Pull Requests) match the same line it shows a provider picker, and extensions cannot listen for modifier-key clicks on it. Ctrl+Alt+C creates the thread directly on this extension's own controller and skips the picker. Rebind it in Keyboard Shortcuts.

Settings

• agentReviewer.pendingFile (default .agent-reviewer/pending-reviews.md): workspace-relative path of the reviews file.
• agentReviewer.gitExclude (default true): adds the reviews file and its .consumed marker to the repository's .git/info/exclude so git status stays clean.
• agentReviewer.targets.codex (default true): attach the reviews file to the Codex chat composer when flushing.
• agentReviewer.targets.kimi (default true): watch the reviews file and clear staged reviews when the Kimi hook consumes it.
• agentReviewer.targets.claude (default true): the same, for the Claude Code hook. The watcher runs when either CLI target is on.

Known limitations

• Deleted lines can't be commented in inline diff mode. Use the diff editor's … menu → Toggle Inline View, and comment on the left (old) file in side-by-side mode.
• The comment widget is not styleable. Its size is fixed by VS Code's core comment component.
• Hook-injected reviews go straight into the model's context — they are not shown as an attachment in the conversation. Use Agent Reviewer: Open Pending Reviews File to inspect what will be sent.
• Codex consumes nothing: after sending the chip in Codex, clear the round manually with Agent Reviewer: Clear Staged Reviews — otherwise a later CLI prompt in the same workspace will pick the same reviews up again.
• Comment threads live in memory: reloading the window resets threads and the staged set (the reviews file keeps its last content and is rewritten on the next change).

License

MIT © NihilDigit