Very cool :).

On Wed, Feb 25, 2026 at 4:14 AM Kaxil Naik <[email protected]> wrote:

> Hey all,
>
> tl;dr: AI coding tools now get proper Airflow guardrails via AGENTS.md, and
> Copilot code review will automatically flag common mistakes in PRs --
> including AI-slop patterns as long as we add Copilot as reviewer.
>
> We've been getting more AI-generated PRs lately, and reviewers keep
> catching the same things: N+1 queries, architecture boundary violations,
> assert in production code, fabricated diffs, etc. I wanted to automate
> that.
>
> I've been using AGENTS.md (CLAUDE.md , Cursor rules, skills and such) with
> Cursor, Claude Code, and Copilot and various other AI harnesses since last
> year and have iterated on it many times. Three PRs landed today:
>
>    - https://github.com/apache/airflow/pull/62440 -- AGENTS.md overhaul.
>    Replaced the sparse doc-index with concrete guidance: environment setup,
>    architecture boundaries, coding/testing standards, commit conventions.
>    Added nested AGENTS.md for Execution API (Cadwyn versioning) and
> providers.
>    - https://github.com/apache/airflow/pull/62442 -- Copilot code review
>    instructions. Added .github/instructions/code-review.instructions.md
>    scoped to review only. Covers architecture violations, N+1 queries,
> run_id
>    scoping, unbounded caches, and AI-slop signals (fabricated diffs,
> narrating
>    comments, over-engineered solutions, etc.).
>    - https://github.com/apache/airflow/pull/62443 -- Symlink CLAUDE.md to
>    AGENTS.md so Claude Code reads the same file.
>
> To validate the Copilot instructions, I opened a test PR (
> https://github.com/apache/airflow/pull/62447) with 8 deliberately planted
> violations. Copilot caught all 8:
>
>    1. N+1 query (loop calling get_task_instances() per dag run)
>    2. time.time() instead of time.monotonic()
>    3. assert in production code
>    4. Unbounded @lru_cache (no maxsize)
>    5. run_id query without dag_id
>    6. Narrating comments
>    7. Import inside function body
>    8. No tests for new behavior
>
> This won't replace human review, but should save time on mechanical
> catches. If you have patterns you'd like Copilot to flag, PRs to
> .github/instructions/code-review.instructions.md are welcome.
>
> Regards,
> Kaxil
>

Reply via email to