review-kernel-pr · git:20261117.82b2487 · 2026-11-17 · sha256 08f913e08ad9b868

review-kernel-pr git:20261117.82b2487A

Immutable. This exact content never changes and is served at /api/v1/blob/08f913e08ad9b868.

No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_ojxq2qp5cog27fxa/08f913e08ad9b868efaa346acc9406986eee65ee8d0316f97ef16ae0b153e877.json

Then /plugin install review-kernel-pr@markdownregistry-ojxq2qp5cog27fxa in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.

---
name: review-kernel-pr
description: Review AKLR kernel pull requests for integration quality, correctness, edge cases, sanitizers, performance, benchmark fairness, and reward hacking. Use when deciding whether a generated or ported kernel PR is ready to merge.
---

# Review a Kernel Pull Request

Review the kernel as part of AKLR's public execution path, not only as a fast
standalone implementation.

## Required Context

Read the harness [rules](../../rules.md), [gates](../../gates.md), current
[gate configuration](../../gate-config.md), the
[acceptance criteria](../../../docs/acceptance_criteria.md), and
[references/workflow.md](references/workflow.md). Check the change against
[bad-cases](../bad-cases/SKILL.md).

## Responsibilities

- Inspect provenance, naming, placement, public reachability, dispatch, and
  fallback behavior.
- Make the code preview understandable: separate host integration from kernel
  bodies and flag opaque numeric names, generated clutter, or hidden behavior.
- Run or verify unit, reference-correctness, edge-case, sanitizer, and relevant
  end-to-end tests.
- Compare performance with the actual AKLR incumbent in the same environment.
- Check the acceptance criteria: each kernel file has a workload at 1.05 or more over the
  baseline. A new plugin route, or one that now handles more cases, reports an end-to-end gain
  of 2% or more on a stated, representative workload, with its ops off and on, and commands to
  reproduce. Missing end-to-end evidence for such a route is `CHANGES REQUIRED`.
- Audit warm-up, synchronization, input selection, timeouts, cache state,
  fallback paths, failures, and aggregation for unfair benchmarking.
- Look for reward hacking, including benchmark-only reachability, shape
  overfitting presented as general support, skipped failures, and comparisons
  against a deliberately weak baseline.
- Distinguish verified facts from missing evidence and reviewer inference.

## Review Boundary

Review is read-only unless the user also asks for fixes. Do not approve based on
the PR description alone, and do not claim that an unavailable GPU or
downstream test passed.

## Completion

Return a verdict of `PASS`, `CHANGES REQUIRED`, or `BLOCKED`, with findings
ordered by severity and a concise evidence table. Use
[references/report-template.md](references/report-template.md).