Maintainer PR Triage and Review

This document describes the auto-triage workflow — a maintainer-driven tool that helps Apache Airflow maintainers triage and review incoming Pull Requests faster and in a more informed way. The tool is part of the Breeze development environment and is invoked via breeze pr auto-triage.

Overview

Apache Airflow receives a high volume of Pull Requests from contributors around the world. Maintainers need to assess each PR for basic quality criteria, CI status, merge conflicts, and code correctness before it can be merged. The auto-triage tool streamlines this process by combining deterministic checks with optional LLM-assisted analysis, while keeping the maintainer firmly in control of every decision.

Key principles:

• Maintainer-driven — The tool presents information and suggests actions, but every decision (drafting, closing, commenting, approving) is made by a human maintainer through interactive prompts. The tool never takes autonomous action on PRs.

• Two-stage process — The workflow is split into two distinct stages: a fast triage pass that checks basic quality criteria, and a deeper review pass that performs detailed code analysis with LLM assistance.

• Better-informed decisions — By aggregating CI status, merge conflicts, unresolved review threads, main-branch failure patterns, and LLM assessments into a single interactive session, maintainers can make faster and more consistent decisions.

Two-stage workflow

The auto-triage tool operates in two modes that correspond to the two stages of the PR lifecycle:

flowchart TD
    A[Contributor opens PR] --> B["Open PR (no label)"]
    B -->|"Maintainer runs
breeze pr auto-triage"| C["Stage 1: TRIAGE
Basic quality check"]
    C --> D{Outcome}
    D -->|Issues found| E[Convert to Draft
with comment]
    D -->|Looks good| F["Add 'ready for
maintainer review' label"]
    D -->|Suspicious| G[Close all PRs
by author]
    E --> H[Contributor fixes
and marks Ready]
    H --> I[Stage 2: REVIEW]
    F --> I
    I["Stage 2: REVIEW
Detailed code review with LLM
(breeze pr auto-triage --mode review)"] --> J{Outcome}
    J -->|Comments posted| K[Author addresses feedback]
    J -->|Approve| L[Merge]
    J -->|Request changes| K
    K --> I

Stage 1: Triage

The triage stage is the first pass over incoming PRs. It focuses on whether each PR meets the project's basic quality criteria and is ready for deeper review. It is invoked with:

breeze pr auto-triage

This is the default mode (--mode triage).

What the triage stage checks

The triage stage performs a series of deterministic checks on each PR:

1. CI status — Are the CI checks passing, failing, or still running? PRs with in-progress workflows are skipped until the next triage run.

2. Merge conflicts — Does the PR have merge conflicts with the base branch? If so, the author needs to rebase.

3. Unresolved review threads — Are there open review conversations that the author has not addressed?

4. Workflow approval — For PRs from first-time contributors, CI workflows need maintainer approval before they can run. The triage tool presents these PRs first so maintainers can review the diff for security concerns before approving.

After deterministic checks, PRs that pass are optionally sent to an LLM for quality assessment. The LLM evaluates the PR title, description, and metadata against the project's Pull Request guidelines and flags potential violations such as:

• Generic or unclear PR titles
• Missing or inadequate descriptions
• Missing Gen-AI disclosure (when AI-generated patterns are detected)
• Unrelated changes bundled together

How triage processes PRs

flowchart TD
    A[Fetch PRs via GraphQL] --> B[Filter: exclude already-triaged,
drafts with known issues]
    B --> C[Enrich: fetch CI checks,
merge status, review threads]
    C --> D[Deterministic checks
CI, conflicts, unresolved threads]
    C --> E[LLM assessment
title, description, quality criteria]
    D --> F
    E --> F
    F[Interactive maintainer session]

    subgraph F[Interactive maintainer session]
        direction TB
        G["For each PR, display:
- PR title, author, age, labels
- CI status with failure details
- Merge conflict status
- Unresolved thread summary
- LLM assessment (if available)
- Main-branch failure patterns"]
        G --> H["Maintainer chooses action:
[d]raft [c]omment [x]close [r]erun CI
[b]rebase [m]ark ready [s]kip [q]uit"]
    end

Available triage actions

When the tool presents a PR, the maintainer can choose from these actions:

CI failure analysis

The triage tool provides context to help maintainers distinguish between failures caused by the PR and systemic failures on the main branch:

• Main-branch failure patterns — The tool fetches recent merged PRs and identifies checks that are consistently failing across the repository. When a PR's failed check matches a known main-branch failure, this is highlighted so the maintainer knows not to penalize the contributor.

• Canary build status — The status of scheduled canary builds on the main branch is displayed, giving maintainers a quick view of overall CI health.

• Grace period — CI failures are given a grace window before the PR is flagged, so contributors have time to address them at their own pace:

  • 24 hours (default) — For new PRs without collaborator engagement. This gives contributors a full day to notice and fix CI issues.
  • 96 hours (extended) — When a collaborator, member, or owner has already left a review or comment on the PR. This extended window acknowledges that the contributor is actively working with a maintainer and should not be rushed.

  PRs within their grace window are skipped during triage. The tool displays which grace period applies and whether collaborator engagement was detected.

Stage 2: Review

The review stage is a deeper, LLM-assisted code review of PRs that have already passed triage and carry the ready for maintainer review label. It is invoked with:

breeze pr auto-triage --mode review

What the review stage does

flowchart TD
    A["Fetch PRs labeled
'ready for maintainer review'"] --> B[Deterministic checks
CI, conflicts, unresolved threads]
    A --> C[LLM code review
fetches full diff, analyzes code changes]
    B --> D
    C -->|runs in parallel| D

    subgraph D[Interactive review session]
        direction TB
        E["Phase 1: Deterministic failures
Present PRs with CI/conflict issues"]
        E --> F["Phase 2: LLM code review results
- Overall assessment
- Line-level review comments"]
        F --> G["For each comment, maintainer chooses:
[p]ost [e]dit [o]pen in browser
[s]kip [q]uit"]
    end

The LLM code review analyzes the full PR diff and produces:

• Overall assessment — A high-level verdict: approve, comment, or request changes.

• Summary — A concise summary of what the PR does and potential concerns.

• Line-level comments — Specific review comments attached to particular files and lines, covering areas such as:

  • Correctness and potential bugs
  • Security concerns
  • Performance implications
  • Architecture boundary violations
  • API design issues
  • Testing coverage
  • Style and best practices

Interactive comment submission

For each line-level comment proposed by the LLM, the maintainer reviews it and decides:

• [p]ost — Post the comment to the PR on GitHub.
• [e]dit — Edit the comment text before posting (opens $EDITOR).
• [o]pen — Open the relevant file in the browser for context.
• [s]kip — Skip this comment.
• [q]uit — Stop reviewing comments for this PR.

The tool detects duplicate comments (same file, line, and body as an existing review comment) and skips them automatically. Nearby existing comments are displayed for context to avoid redundant feedback.

After individual comments, the maintainer is prompted to post an overall review comment summarizing the review.

LLM integration and safety

The auto-triage tool optionally uses LLM providers (Claude or Codex) for PR assessment and code review. Several safety measures are in place:

• Read-only mode — The LLM runs in a restricted permission mode with access only to read-only tools (file reading, code search). It cannot modify files or execute code.

• Repository restriction — LLM-assisted triage is only enabled for trusted repositories (apache/airflow).

• No autonomous actions — The LLM produces assessments and review comments, but posting them to GitHub always requires explicit maintainer confirmation through the interactive prompts.

• Caching — LLM results are cached by PR number and commit SHA, so re-running triage on the same PR state does not incur redundant LLM calls. The cache can be cleared with --clear-llm-cache.

• Safety validation — The tool checks for dangerous environment variables that could bypass LLM sandboxing and blocks them. The maintainer must confirm before LLM processing begins.

Labels and PR states

The triage workflow uses specific labels and PR states to track progress:

Common usage examples

Triage all open PRs:

breeze pr auto-triage

Triage PRs with a specific label:

breeze pr auto-triage --label area:core

Triage a single PR:

breeze pr auto-triage --pr 12345

Review PRs that are ready for maintainer review:

breeze pr auto-triage --mode review

Review PRs assigned to you:

breeze pr auto-triage --mode review --reviews-for-me

Use only deterministic checks (no LLM):

breeze pr auto-triage --check-mode api

Dry run (no actions taken, just display what would be done):

breeze pr auto-triage --dry-run

Clear cached LLM assessments and start fresh:

breeze pr auto-triage --clear-llm-cache

For contributors

As a contributor, you don't need to use the auto-triage tool — it is a maintainer-facing workflow. However, understanding how it works helps you get your PRs reviewed faster.

There is no rush. We want contributors to work at their own pace. The triage tool gives you a generous grace period to fix issues — 24 hours by default, and 96 hours once a maintainer has started engaging with your PR. Comments posted by the tool will always remind you that you can take your time.

Here are some tips:

1. Start with Draft — Keep your PR in draft until CI passes and you're confident it meets the quality criteria.

2. Write a clear title and description — Both the deterministic checks and the LLM assessment evaluate your PR title and description. A descriptive title and meaningful description help your PR pass triage quickly.

3. Fix issues at your own pace — If your PR is converted to draft with a comment listing issues, address each one and mark the PR as "Ready for review" when you're done. The next triage run will re-evaluate it. There is no deadline.

4. Keep CI green — Resolve CI failures before requesting review. The triage tool flags PRs with failing checks, but gives you at least 24 hours before flagging (96 hours if a maintainer has already engaged).

5. Resolve review threads — Unresolved conversations block the triage process. Address reviewer feedback and resolve the threads.

6. Disclose Gen-AI usage — If you used Gen-AI tools, include a disclosure in the PR description per the Gen-AI guidelines. The LLM assessment flags PRs that appear AI-generated but lack disclosure.