Break the
bot review
doom loop.

You fix a round of review comments, push, and new ones appear. agent-reviews handles the entire cycle: triage, fix, reply, watch, repeat.

Because nobody should spend their afternoon typing "Fixed in abc1234" forty-seven times.

Install GitHub

The CLI

A purpose-built interface for PR review comments. Think agent-browser, but for code reviews.

The gh CLI wasn't designed for review comments. Agents get the syntax wrong, miss pagination, can't detect reply status, and burn tokens parsing verbose output. agent-reviews fixes all of that with compact, structured output that works for both humans and LLMs.

Compact output
Minimal text, structured for agents. JSON mode for scripting.
Smart filtering
Bot vs. human, answered vs. unanswered, resolved vs. pending. Filters out meta-comments automatically.
Reply + resolve
Reply to any comment and resolve its thread in one command.
Watch mode
Poll for new comments after a push. Exits when bots go quiet.
gh api repos/.../pulls/414/comments
[
  {
    "id": 2870136380,
    "path": "scripts/provision-branch.js",
    "position": 42,
    "body": "### Hardcoded step label wrong\nin CI seed mode\n\n**Low Severity**\n\n`seedDemoData` has a hardcoded\n`logStep(\"8/10\", ...)` which is\ncorrect for local mode's 10-step\nprocess, but in CI mode the total\nsteps changed to 8...",
    "user": { "login": "cursor[bot]" },
    "created_at": "2025-12-19T14:23:01Z",
    "updated_at": "2025-12-19T14:23:01Z",
    "html_url": "https://github.com/..."
  },
  ... 21 more comments
  // no reply status, no filtering,
  // no pagination handling
]
agent-reviews -a --bots-only
Found 22 comments

[2870136380] CODE by cursor[bot] ○ no reply
  provision-branch.js:42
  Hardcoded step label wrong in CI seed...

[2870136382] CODE by cursor[bot] ○ no reply
  provision-branch.js:87
  Hardcoded "4/6" wrong when seed increases

[2870138016] CODE by codex[bot] ✓ replied
  provision-branch.js:112
  Fail CI when preview seeding incomplete
  └ 1 reply

The skills

The CLI is powerful on its own, but the skills are the killer feature. Install one, type a slash command, and walk away. Your agent uses the CLI under the hood to handle the full review cycle:

True positive
Fixes the bug, commits, replies with the commit hash.
False positive
Replies "Won't fix" with an explanation of why.
Uncertain
Asks you before touching anything.

Then it watches. When bots post new comments after your push (they always do), it processes those too. The loop continues until no new comments appear for 10 minutes.

/resolve-reviews All comments, human and bot.
/resolve-agent-reviews Bot comments only.
/resolve-human-reviews Human comments only.

Works with

Review bots: Copilot, CodeRabbit, Cursor Bugbot, Sourcery, Codacy, SonarCloud.

AI agents: Claude Code, Cursor, Codex, and any agent supporting Agent Skills.

Install

Agent Skills

recommended

Install once, then use /resolve-agent-reviews as a slash command.

npx skills add pbakaus/agent-reviews

CLI

For direct terminal use. List, filter, reply to, and watch comments.

npm install -g agent-reviews

Authenticates via gh CLI or GITHUB_TOKEN. Works locally, in Codespaces, and in remote agent environments.

What's new

v1.0.0 March 6, 2026
  • Three skills: /resolve-reviews, /resolve-agent-reviews, /resolve-human-reviews
  • Thread resolution via --resolve flag (GraphQL resolveReviewThread)
  • Expanded bot support: CodeRabbit, Sourcery, Codacy, SonarCloud, Copilot
  • Agent-harness universal: works with Claude Code, Cursor, Codex, and any agent supporting Agent Skills
  • Watch mode exits on new comments with 5s grace period for batch posts
  • --expanded flag for full detail (body, diff hunk, replies) in list mode
  • GH_TOKEN / GH_REPO env vars for cloud and proxy environments
  • Bot review body filtering and self-reply filtering to reduce noise