[PR #181] docs: add AGENTS.md and CLAUDE.md shim #200

New issue

Open

opened 2026-10-26 22:21:30 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-26 22:21:30 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/comfy-kitchen/pull/181
Author: @mattmillerai
Created: 9/16/2026
Status: 🔄 Open

Base: main ← Head: matt/be-3580-agents-md

📝 Commits (5)

7865da4 docs: add AGENTS.md and CLAUDE.md shim
e43c361 docs(AGENTS.md): qualify --no-cuda with --hip; ruff check is lint only
68e0b39 Merge origin/main into matt/be-3580-agents-md
e82b219 docs(AGENTS.md): document the Ascend backend and current dispatch order
48b30db docs(AGENTS.md): narrow HIP auto-selection; distinguish Linux/Windows CI wheels

📊 Changes

2 files changed (+75 additions, -0 deletions)

View changed files

➕ AGENTS.md (+73 -0)
➕ CLAUDE.md (+2 -0)

📄 Description

ELI-5

Different AI coding tools each look for their own instructions file. This repo had none, so every tool started cold and had to re-learn how to build, test and navigate the kernel library. This PR adds one short guidebook, AGENTS.md (the cross-tool standard), plus a tiny CLAUDE.md that just says "read AGENTS.md" — because Claude Code only reads CLAUDE.md and does not fall back to AGENTS.md on its own. No code changes, so nothing about the library's behavior moves.

What this adds

AGENTS.md (73 lines) — a tight, verified guide written from the repo's actual contents (README, pyproject.toml, pytest.ini, MANIFEST.in, the CI workflows, and the top-level layout): one-paragraph purpose, including the dispatch order (ascend → cuda → triton → eager by default, hip first under ROCm) and the optional torch-npu ascend backend; build/install commands and env-var knobs; the exact lint/test commands CI runs; a directory map; and repo-specific conventions/gotchas (ruff as the style source, DCO sign-off, the CLA workflow, mandatory submodules, MANIFEST.in/package-data for native sources, the abi3 Stable-ABI requirement, fail-closed HIP arch overrides, and the constraint-gated dispatch model).
CLAUDE.md — the canonical shim only (an HTML comment plus @AGENTS.md), so agent instructions have a single source of truth.

How this was verified

ruff check . (CI's lint step) passes clean on the branch — the change is Markdown-only at the repo root and does not touch Python, the build, or any CI-relevant path.
The build (pip install ., pip install -e ".[dev]", git submodule update --init --recursive) and test (ruff check ., python -m pytest tests/ -v --tb=short) commands documented in AGENTS.md are transcribed from README.md and the test_* jobs in .github/workflows/build-wheels.yml, which is the authoritative pipeline.
No .cursorrules exists anywhere in the tree, so none needed removing.
AGENTS.md is 73 lines (well under the 200-line cap); CLAUDE.md matches the required shim content byte-for-byte.

Provenance

Authored by: agent-work loop
Verified:
- ruff check . passed on the original authoring head. Every later commit changes only Markdown or merges main; ruff was not re-run on the current head.
- Dispatch order and backend list checked against comfy_kitchen/registry.py and comfy_kitchen/__init__.py on current main, which added the ascend backend.
- HIP auto-selection wording checked against get_extensions() in setup.py. CI wheel/test descriptions checked against the test_linux, test_windows and test_windows_arm jobs in .github/workflows/build-wheels.yml.
- Build/pytest commands checked against README.md and the CI workflow, not executed locally (see Residual).
Deviations:
- main was merged into the branch so AGENTS.md describes the tree it will land on (Ascend backend, new default priority).

Residual

The build and pytest commands in AGENTS.md were verified against CI config rather than executed locally: the authoring environment has no PyTorch, no CUDA/ROCm toolchain, and no compiler, so a source build and the test suite could not be run here. They are copied verbatim from README.md and the green test_* jobs in .github/workflows/build-wheels.yml. A follow-up could confirm them end-to-end by running pip install -e ".[dev]" + pytest on a machine with PyTorch (and, for the native paths, a CUDA/ROCm toolchain).

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/comfy-kitchen/pull/181 **Author:** [@mattmillerai](https://github.com/mattmillerai) **Created:** 9/16/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `matt/be-3580-agents-md` --- ### 📝 Commits (5) - [`7865da4`](https://github.com/Comfy-Org/comfy-kitchen/commit/7865da4a2a7dc0db7ed9bc258457ae8611373f73) docs: add AGENTS.md and CLAUDE.md shim - [`e43c361`](https://github.com/Comfy-Org/comfy-kitchen/commit/e43c3611849f3529d414732c0e6881a350529ec2) docs(AGENTS.md): qualify --no-cuda with --hip; ruff check is lint only - [`68e0b39`](https://github.com/Comfy-Org/comfy-kitchen/commit/68e0b397229517f1e94b62114bf4177ed65f5d94) Merge origin/main into matt/be-3580-agents-md - [`e82b219`](https://github.com/Comfy-Org/comfy-kitchen/commit/e82b2194fd283aefee8a4e3c6e6d6d985f94396f) docs(AGENTS.md): document the Ascend backend and current dispatch order - [`48b30db`](https://github.com/Comfy-Org/comfy-kitchen/commit/48b30db42b0e9831fd26b17f696f1a9a0c0c3ca9) docs(AGENTS.md): narrow HIP auto-selection; distinguish Linux/Windows CI wheels ### 📊 Changes **2 files changed** (+75 additions, -0 deletions) <details> <summary>View changed files</summary> ➕ `AGENTS.md` (+73 -0) ➕ `CLAUDE.md` (+2 -0) </details> ### 📄 Description ## ELI-5 Different AI coding tools each look for their own instructions file. This repo had none, so every tool started cold and had to re-learn how to build, test and navigate the kernel library. This PR adds one short guidebook, `AGENTS.md` (the cross-tool standard), plus a tiny `CLAUDE.md` that just says "read AGENTS.md" — because Claude Code only reads `CLAUDE.md` and does not fall back to `AGENTS.md` on its own. No code changes, so nothing about the library's behavior moves. ## What this adds - **`AGENTS.md`** (73 lines) — a tight, verified guide written from the repo's actual contents (README, `pyproject.toml`, `pytest.ini`, `MANIFEST.in`, the CI workflows, and the top-level layout): one-paragraph purpose, including the dispatch order (`ascend` → `cuda` → `triton` → `eager` by default, `hip` first under ROCm) and the optional torch-npu `ascend` backend; build/install commands and env-var knobs; the exact lint/test commands CI runs; a directory map; and repo-specific conventions/gotchas (ruff as the style source, DCO sign-off, the CLA workflow, mandatory submodules, `MANIFEST.in`/`package-data` for native sources, the abi3 Stable-ABI requirement, fail-closed HIP arch overrides, and the constraint-gated dispatch model). - **`CLAUDE.md`** — the canonical shim only (an HTML comment plus `@AGENTS.md`), so agent instructions have a single source of truth. ## How this was verified - `ruff check .` (CI's lint step) passes clean on the branch — the change is Markdown-only at the repo root and does not touch Python, the build, or any CI-relevant path. - The build (`pip install .`, `pip install -e ".[dev]"`, `git submodule update --init --recursive`) and test (`ruff check .`, `python -m pytest tests/ -v --tb=short`) commands documented in `AGENTS.md` are transcribed from `README.md` and the `test_*` jobs in `.github/workflows/build-wheels.yml`, which is the authoritative pipeline. - No `.cursorrules` exists anywhere in the tree, so none needed removing. - `AGENTS.md` is 73 lines (well under the 200-line cap); `CLAUDE.md` matches the required shim content byte-for-byte. ## Provenance - **Authored by:** agent-work loop - **Verified:** - `ruff check .` passed on the original authoring head. Every later commit changes only Markdown or merges `main`; ruff was not re-run on the current head. - Dispatch order and backend list checked against `comfy_kitchen/registry.py` and `comfy_kitchen/__init__.py` on current `main`, which added the `ascend` backend. - HIP auto-selection wording checked against `get_extensions()` in `setup.py`. CI wheel/test descriptions checked against the `test_linux`, `test_windows` and `test_windows_arm` jobs in `.github/workflows/build-wheels.yml`. - Build/pytest commands checked against `README.md` and the CI workflow, not executed locally (see Residual). - **Deviations:** - `main` was merged into the branch so `AGENTS.md` describes the tree it will land on (Ascend backend, new default priority). ## Residual - The build and `pytest` commands in `AGENTS.md` were **verified against CI config rather than executed locally**: the authoring environment has no PyTorch, no CUDA/ROCm toolchain, and no compiler, so a source build and the test suite could not be run here. They are copied verbatim from `README.md` and the green `test_*` jobs in `.github/workflows/build-wheels.yml`. A follow-up could confirm them end-to-end by running `pip install -e ".[dev]"` + `pytest` on a machine with PyTorch (and, for the native paths, a CUDA/ROCm toolchain). --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi added the

pull-request

label

2026-10-26 22:21:30 +00:00

bsakaguchi referenced this issue

2026-10-26 22:21:42 +00:00

[PR #200] [CLOSED] [HIP] Add software tiled GEMM support to older HIP architectures (RDNA1 and Vega) #220