pr-review

Request and manage pull request reviews using pr-agent (local CLI) with optional Claude second-pass. Use when needing to review a PR, check review comments, address feedback, or iterate on fixes. Trig

SKILL.md

GitHub

PR Review

Primary reviewer: pr-agent (Qodo Merge), installed locally via pipx. Optional Claude second-pass for code PRs: pr-review-toolkit:review-pr (model-invocable). code-review:code-review is user-invocable only, so Claude cannot call it; the user can type /code-review themselves.

Environment

pr-agent is configured entirely through environment variables — there is no user-global config file for the CLI (see Notes). Put this block in your shell rc (~/.zshrc) once, substituting your provider's endpoint, key var, and model:

export OPENAI__KEY="$YOUR_PROVIDER_KEY"
export OPENAI__API_BASE="https://your-endpoint/v1"   # your real endpoint URL — no angle brackets
export CONFIG__MODEL="<provider/model>"            # e.g. openai/<model-name>
export CONFIG__FALLBACK_MODELS='["<provider/fallback-model>"]'
export CONFIG__CUSTOM_MODEL_MAX_TOKENS=<tokens>    # required if model isn't in pr-agent's built-in list
export GITHUB__USER_TOKEN="$GITHUB_TOKEN"
  • SECTION__KEY double-underscore is dynaconf's env convention (pr-agent sets envvar_prefix=False, so no prefix). These vars are read only by pr-agent — they don't collide with other tools' OPENAI_API_KEY/OPENAI_API_BASE. OPENAI__* here means the OpenAI-compatible interface, not the OpenAI service specifically — it works with any provider that exposes an OpenAI-compatible endpoint.
  • CONFIG__MODEL overrides the default (gpt-*); prefix with the provider (openai/, anthropic/, …) to route a custom model to your OPENAI__API_BASE.
  • Setting OPENAI__API_BASE pins all openai-provider calls to your endpoint, so a stray real OPENAI_API_KEY in the environment can never reach api.openai.com via pr-agent — no unset needed.

Project-level config lives in .pr_agent.toml (if present).

Commands

All commands write to the PR under the user's GitHub identity. Confirm before running review, improve, or describe (which rewrites the PR title and body) on a PR the user didn't author.

In an interactive terminal the env block is already loaded, so just call pr-agent. Non-interactive callers (e.g. an agent's shell tool) that don't load your rc should load the env block first, for example by prepending source ~/.zshrc; to the command.

# Review — posts a top-level review comment (correctness, security, completeness)
pr-agent --pr_url <pr-url> review

# Improve — posts inline code suggestions
pr-agent --pr_url <pr-url> improve

# Describe — rewrites the PR title and description
pr-agent --pr_url <pr-url> describe

# Ask — ask a specific question about the PR
pr-agent --pr_url <pr-url> ask "<question>"

Workflow

  1. Identify the PR: gh pr view <number> to confirm scope, then grab the URL.
  2. Request pr-agent review:
    pr-agent --pr_url <pr-url> review
    
    For inline suggestions, also run improve. The Environment vars must be loaded (see the non-interactive note under Commands).
  3. Optional Claude second-pass (code PRs): invoke pr-review-toolkit:review-pr. Useful for a second-model perspective — pr-agent uses a different model by default, so Claude catches different patterns.
  4. Read the posted feedback with the commands under Checking Review Status.
  5. Address feedback — fix code, commit, push.
  6. Reply to each thread explaining what was fixed:
    gh api repos/OWNER/REPO/pulls/<PR_NUMBER>/comments/<COMMENT_ID>/replies \
      -X POST -f body="Fixed in <commit-sha>"
    
  7. Re-review — repeat from step 2. Stop when a round produces no new actionable comments after the previous round's fixes landed. Keep going while rounds surface real findings.

Round count guidance

  • 1 round: enough only when the first review finds nothing actionable.
  • 2-3 rounds: typical. A fix can introduce a new edge case the reviewer catches on re-pass, so the round after the last fix must come back clean.
  • 4+ rounds: expected for security issues, logic bugs, or when an earlier skip was later reversed.

Deciding What to Address

  • Address immediately: Security issues, bugs, missing error handling, secret leaks
  • Create issue for later: Style preferences, nice-to-have refactors, doc suggestions
  • Skip with explanation: Comments that don't apply to the project's context — reply on the thread so the rationale is on record. Be prepared for the reviewer to re-flag if the skip was wrong.

Checking Review Status

# List open PRs
gh pr list

# View PR details + diff
gh pr view <number>
gh pr diff <number>

# Check existing reviews
gh pr view <number> --json reviews --jq '.reviews[] | {author: .author.login, state, submittedAt}'

# Check inline review comments
gh api repos/OWNER/REPO/pulls/<number>/comments \
  --jq '.[] | {id, path, line, user: .user.login, body}'

Notes

  • pr-agent is a local CLI only — there is no GitHub Action workflow firing on push.
  • Install via pipx install pr-agent. Do not edit the bundled .secrets.toml inside the pipx venv — pipx upgrade/reinstall rebuilds the venv and wipes it. Configure via the env vars above instead (they live outside the venv and survive upgrades).
  • There is no user-global config file for the CLI (no ~/.pr_agent.toml). The org-level pr-agent-settings repo in the docs applies only to the hosted GitHub App. Environment variables are the only global mechanism.
  • Project-specific guidance for pr-agent goes in the project's .pr_agent.toml (for example extra_instructions).