[doc-only][agents] Update AGENTS.md with review-derived guidance#2106
[doc-only][agents] Update AGENTS.md with review-derived guidance#2106rparolin wants to merge 7 commits into
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
|
mdboom
left a comment
There was a problem hiding this comment.
The linting guidance is wrong and should be fixed. The rest I feel less strongly about.
…d-agents-guidance # Conflicts: # cuda_core/AGENTS.md
|
@mdboom I merged current Fix commit: 3c7d042 Could you re-review when you have a chance? |
Feedback addressed and user is away on vacation.
|
@mdboom @kkraus14 — this is doc-only and all your inline points are addressed:
mdboom, your review got dismissed when I pushed the fixes — could you re-review? CI is green and the branch is mergeable. One housekeeping item: the |
Summary
Adds review-derived guidance to the root and package-specific
AGENTS.mdfiles so future agent-generated changes better reflect recurring cuda-python code review feedback.The change updates:
AGENTS.mdcuda_bindings/AGENTS.mdcuda_core/AGENTS.mdcuda_pathfinder/AGENTS.mdcuda_python/AGENTS.mdHow this guidance was generated
This PR turns recurring review feedback from cuda-python maintainers and contributors into durable agent instructions.
I analyzed inline PR review comments created from August 5, 2024 at 13:50:18 UTC through May 16, 2026 at 07:25:14 UTC. The dataset covers PR #84 through PR #2087 and includes 6,687 inline review comments across 586 unique PRs.
Data source:
MEMBERcomments: 5,471 comments from 37 commentersThe content added to the
AGENTS.mdfiles was driven by themes that appeared repeatedly in those review comments. I used keyword-based matching over review comment bodies to quantify recurrence, then converted the highest-value themes into agent-facing guidance.It is a directional analysis to identify feedback patterns that are common enough to document up front.
Research findings that informed the updates
Across the full inline review comment corpus, 2,514 of 6,687 comments, or 37.59%, matched at least one theme reflected in this PR.
Looking only at comments from GitHub
MEMBERauthors, 1,990 of 5,471 comments, or 36.37%, matched at least one of these themes. That indicates the guidance is grounded in maintainer/team review feedback, not only in occasional external contributor discussion.cuda_pathfinderand repo-wide errors to avoid silent fallback paths, validateCUDA_HOME/CUDA_PATH, and surface user-actionable diagnostics.Mapping from research to file changes
The root
AGENTS.mdcaptures review themes that apply across the monorepo:The package-specific
AGENTS.mdfiles add narrower guidance where the review feedback was package-specific:cuda_bindings/AGENTS.mdadds generated-source, Cython-copy, CUDA-version compatibility, and external-contribution guidance because review comments repeatedly flagged one-off generated edits and compatibility risks in binding code.cuda_core/AGENTS.mdadds public API design, capability-query, CUDA compatibility, resource cleanup, and testing guidance because review feedback often focused on high-level API stability and user-observable CUDA behavior.cuda_pathfinder/AGENTS.mdadds discovery/fallback/environment-root guidance because review feedback around path discovery emphasized explicit diagnostics over silent success-shaped fallbacks.cuda_python/AGENTS.mdadds packaging and docs aggregation guidance because metapackage changes need to stay aligned with component versions, constraints, release tags, and generated docs behavior.