pbakaus/agent-reviews

agent-reviews

Manage GitHub PR review comments from the terminal and from AI coding agents.

PR review bots (Copilot, Cursor Bugbot, CodeRabbit, etc.) leave inline comments on your pull requests. agent-reviews gives you a CLI to list, filter, reply to, and watch those comments, plus agent skills that automate the entire triage-fix-reply loop.

Why

gh CLI is fragile for review comments. Agents frequently get the syntax wrong, fail to paginate, and can't reliably detect whether a comment has been replied to. agent-reviews provides a single, purpose-built interface that handles all of this correctly.

Bot reviews create a doom loop. You fix one round of findings, push, and new comments appear. Fix those, push again, more comments. This cycle can eat hours. The included skills solve this with an integrated watcher that keeps fixing and replying until the bots go quiet.

Works in cloud environments. Most solutions rely on local tooling that isn't available in cloud or remote agent environments. agent-reviews works everywhere, so you can kick off a session, let the agent resolve all findings autonomously, and come back to a clean PR.

Install

CLI (npm)

npm install -g agent-reviews

Agent Skills

Three skills are available, each as a slash command (no npm install required):

Works with any agent that supports Agent Skills (Claude Code, Cursor, Codex, etc.):

npx skills add pbakaus/agent-reviews@resolve-agent-reviews

Replace resolve-agent-reviews with whichever skill you want. Skills use npx agent-reviews at runtime, so the CLI is fetched automatically.

You can also use both: install the CLI globally for direct terminal use, and a skill for the agent workflow.

Authentication

The simplest method is the GitHub CLI. If you're logged in with gh auth login, agent-reviews picks up the token automatically. No configuration needed.

For cloud/remote environments or HTTPS proxy setups, set GITHUB_TOKEN or GH_TOKEN directly. agent-reviews includes undici ProxyAgent support and will route requests through HTTPS_PROXY automatically when set.

Resolution order (first match wins):

1. GITHUB_TOKEN environment variable
2. GH_TOKEN environment variable
3. .env.local in the repo root
4. gh auth token (GitHub CLI)

Fork pull requests

Branch discovery keeps the head repository separate from the PR's base repository. It uses the branch push remote, remote.pushDefault, tracking remote, or origin (in that order), and looks for the current branch in configured GitHub repositories and the fork's parent/source repositories. The selected remote's push URL determines the head repository, including when its fetch URL points upstream. An upstream remote is not required. A single GitHub remote with another name also works.

GH_REPO=organisation/project restricts the target repository without changing the head owner. Use GH_REPO=organisation/project agent-reviews --pr 123 to select a PR explicitly, including from detached HEAD. Malformed GH_REPO values fail explicitly. If multiple open PRs match, or a matching PR has a deleted/inaccessible head repository that cannot be verified, the CLI asks for an explicit target instead of choosing one. Discovery uses the GitHub API and works with token authentication without requiring gh.

Custom API host

Set GITHUB_API_URL to point agent-reviews at a GitHub Enterprise host or any API-compatible server (useful for testing, recording, or routing through a local mediator). Defaults to https://api.github.com.

# GitHub Enterprise Server
GITHUB_API_URL=https://github.example.com/api/v3 agent-reviews

# Local API-compatible server
GITHUB_API_URL=http://127.0.0.1:8080 agent-reviews

GraphQL endpoint resolution: agent-reviews uses ${GITHUB_API_URL}/graphql by default. For GitHub Enterprise Server, where REST lives under /api/v3 and GraphQL under /api/graphql on the same origin, the trailing /api/v3 is rewritten to /api/graphql automatically. Set GITHUB_GRAPHQL_URL directly if you need full control over the GraphQL endpoint.

CLI Usage

# List all review comments on the current branch's PR
agent-reviews

# Only unresolved comments
agent-reviews --unresolved

# Only unanswered bot comments
agent-reviews --unanswered --bots-only

# Full detail for a specific comment (diff hunk + replies)
agent-reviews --detail 12345678

# Reply to a comment
agent-reviews --reply 12345678 "Fixed in abc1234"
agent-reviews --reply 12345678 --body-file body.md

# JSON output for scripting / AI agents
agent-reviews --json

# Watch for new comments (polls every 30s, exits after 10 min idle)
agent-reviews --watch --bots-only

# After a push: wait for the head commit's checks, then print failures and unanswered bot comments
agent-reviews --wait-checks --bots-only --expect "Greptile Review"

# Target a specific PR (otherwise auto-detects from branch)
agent-reviews --pr 42

Unknown options, unexpected arguments, and invalid numeric values fail before authentication or network requests. To pass an inline reply starting with a dash, put options first and use agent-reviews --reply 12345 -- "- Fixed the issue". Everything after -- is literal text.

Options

To hide workflow status comments, use agent-reviews --ignore-author 'github-actions[bot]'. Repeat the flag for multiple authors, including with --watch. Matching is case-insensitive and uses the full GitHub login, including [bot] when present. Excluded authors' top-level findings are hidden, but their replies remain visible as context in other authors' threads. Explicit --detail lookups and replies are unaffected. No authors are excluded by default.

Agent Skills

The skills automate the full PR review resolution workflow:

1. Fetch unanswered comments (all, bot-only, or human-only depending on skill)
2. Evaluate each finding (true positive, false positive, actionable, etc.)
3. Fix real issues and run lint/type-check
4. Dismiss false positives with an explanation
5. Reply to every comment with the outcome
6. Wait for the head commit's checks (--wait-checks) and repeat until a pass is clean
7. Report a summary of all actions taken

Skill behavior

• True positives / actionable feedback get fixed and replied with Fixed in {commit}
• False positives get replied with Won't fix: {reason}
• Uncertain findings prompt the user for guidance
• All fixes are batched into a single commit before polling begins
• The bot and combined skills loop until every check on the head commit has passed and no comment waits; the human-review skill watches until reviewers go quiet

How It Works

Comment types

agent-reviews fetches three types of GitHub PR comments:

Meta-comment filtering

Bot review bodies (REVIEW type) are always filtered out since actionable findings come through as inline comments. Additionally, these bot issue comments are filtered:

Reply status

Each comment displays its reply status:

Waiting for checks

--wait-checks polls the PR's head commit until every check run and commit status has finished, then prints one line with the outcome, the failed checks with links, any expected check that never reported, and the unanswered comments (with the usual filters). It follows new pushes to the PR, stops at once when the PR conflicts with its base branch (no CI runs then), and waits 30 seconds after the last check finishes for reviewers that register late. Name a reviewer with --expect when it registers much later, for example only after a preview deploys; expected checks are not waited for on drafts.

Summary comments from review bots (Greptile, Copilot, CodeRabbit and others) are not counted as findings. PR-level comments (issue comments, review bodies) are listed but never change the exit code, since they have no thread to reply in and would otherwise hold every run at 1.

Watch mode

Polls the GitHub API at a configurable interval and reports new comments as they appear. Outputs both formatted text and JSON for AI agent consumption. Exits automatically after a configurable inactivity timeout (default: 10 minutes).

Reply files preserve Markdown and newlines exactly. For multiline or complex replies, write the text to a file and pass --body-file body.md to avoid shell quoting and command substitution. Both absolute and relative paths are supported; relative paths are resolved from the current directory. Prefer uniquely named temporary files outside the repository so later commits cannot accidentally include reply text. Use either a positional message or --body-file, not both; empty or whitespace-only replies are rejected. The flag works with --resolve and --json.

Changelog

1.3.0 (October 16, 2026)

• The bot and combined skills stop after a second round of fixes when new findings keep landing on lines the earlier fixes changed, and hand the rest to the user instead of patching on.
• --wait-checks rides out network drops of up to five minutes instead of five failed requests, and retries the PR lookup and the comment fetch through short drops.

1.2.0 (October 15, 2026)

• Wait for a PR's checks with --wait-checks: follows new pushes, stops on a conflict with the base branch, waits for late reviewers named with --expect, and prints only failed checks and unanswered comments, with an exit code for each outcome.
• The bot and combined skills wait with --wait-checks instead of a quiet window.
• Treat Greptile's PR summary as a summary, not a finding.

1.1.0 (September 10, 2026)

• Read exact UTF-8 reply text with --body-file, without shell command substitution.
• Reject unknown flags and invalid arguments before posting; use -- for replies beginning with a dash.
• Discover fork PRs across upstream repositories, separate push URLs, and GitHub Enterprise hosts.
• Exclude selected authors from listing and watching with repeatable --ignore-author.
• Reduce duplicate replies with fresh reply checks and watcher-startup reconciliation in all three skills.
• Update development dependencies to Vitest 5 and refresh website dependencies and audit fixes.

Development tests now require Node.js 22.12+, 24, or 26+ as supported by Vitest 5. CLI runtime support remains Node.js 18+.

1.0.2

• GitHub Enterprise Server support via GITHUB_API_URL env var (also works for local API-compatible servers used in testing). REST and GraphQL endpoints both honor it, with GHES /api/v3 automatically rewritten to /api/graphql. Optional GITHUB_GRAPHQL_URL for unusual setups.
• Gemini Code Assist meta-comment filtering: the bot's ## Summary of Changes issue comment is now dropped, while inline severity-badged findings are preserved.

1.0.0

Three skills instead of one. The single agent-reviews skill has been split into three, each tailored for different workflows:

• resolve-reviews resolves all comments (human + bot)
• resolve-agent-reviews resolves bot comments only
• resolve-human-reviews resolves human comments only

Thread resolution. The new --resolve flag marks GitHub review threads as resolved after replying. Uses the GraphQL resolveReviewThread mutation. Works with --reply in any argument order.

Expanded bot support. Added detection and meta-comment filtering for CodeRabbit, Sourcery, Codacy, SonarCloud/SonarQube Cloud, and Copilot PR reviewer, in addition to the existing Cursor Bugbot, Vercel, and Supabase filters.

Agent-harness universal. Skills now work with any agent that supports Agent Skills (Claude Code, Cursor, Codex, etc.), not just Claude Code.

Watch mode improvements. The watcher now exits immediately when new comments are found (with a 5s grace period for batch posts), designed for loop-based workflows where the agent processes comments and restarts the watcher.

New CLI options:

• --resolve resolves the review thread after replying (use with --reply)
• --expanded / -e shows full detail (body, diff hunk, replies) for each comment in list mode

Bug fixes:

• --json --resolve no longer emits plain-text status messages to stdout

Cloud and proxy support:

• GH_TOKEN environment variable support (in addition to GITHUB_TOKEN)
• GH_REPO environment variable for targeting repos in detached environments
• Curl-based HTTP fallback for environments without native fetch/undici
• Curl requests include timeouts (10s connect, 60s max)

Smarter filtering. Bot review bodies (summaries listing inline findings) are now automatically excluded, since actionable findings always come through as inline comments. Reply comments posted by agent-reviews itself (> Re: comment ...) are also filtered to avoid noise.

Simplified architecture. Skills now invoke npx agent-reviews at runtime instead of bundling their own scripts, reducing the package from ~4000 lines of duplicated code to a single CLI entry point. Skills no longer run redundant startup commands (version check, branch detection, PR lookup), relying on the CLI's own error handling instead.

License

MIT