# Slangpy Github

> GitHub operations for SlangPy. Clone, branch, PR, issues, CI.

- Skill: `shader-slang/slangpy-github` (Agent Skill)
- Install (CLI): `npx skillmds@latest add shader-slang/slangpy-github`
- Raw SKILL.md: https://api.skillmd.com/api/skills/shader-slang/slangpy-github/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: MIT
- Author: shader-slang (https://skillmd.com/u/shader-slang)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/shader-slang/slangpy-github

---


## Repository

- **URL:** https://github.com/shader-slang/slangpy
- **Clone:** `git clone --recursive --tags https://github.com/shader-slang/slangpy.git`
- **Docs:** https://slangpy.shader-slang.org/en/latest/
- **Discussions:** https://github.com/shader-slang/slang/discussions
- **Discord:** https://khr.io/slangdiscord

## PR process

1. Fork -> clone -> branch (`feature/description`)
2. Implement, test, format (`pre-commit run --all-files`)
3. Push to fork, create PR against `shader-slang/slangpy:main`
4. PR requires review approval + CI pass
5. Squash merge -- rewrite final commit message to be descriptive

**Autonomous PR creation:** Create PRs without asking for confirmation. If tests pass and changes are minimal, open the PR immediately via `gh pr create`. Include: what changed, why, and test results in the PR body. Only stop and notify the user if CI fails after 2 rerun attempts.

> **`gh pr create` vs REST — decided by the PR's _base_ repo.** Normal PRs (base = `shader-slang/slangpy`, including a head branch on the `slang-coworkers/slangpy` fork) → use `gh pr create`; it routes via GraphQL/App token, which has write on shader-slang. **Only a _cross-fork_ PR whose _base_ is a contributor's fork** must use REST — `gh api -X POST repos/<owner>/slangpy/pulls -f head="slang-coworkers:<branch>" -f base="<their-branch>" …` — because `gh pr create` (GraphQL) gets the App token there and 403s; REST `/repos/*` gets the `nv-slang-bot` user PAT that can open it.

During review, push follow-up commits (don't rebase after PR creation). Sync with upstream via merge, not rebase:

```bash
git fetch upstream main
git merge upstream/main
git submodule update --recursive
```

## CI

CI workflows in `.github/workflows/`:

| Workflow | Purpose |
|----------|---------|
| `ci.yml` | Main CI: configure, build, test on multiple platforms |
| `ci-latest-slang.yml` | Test against latest Slang compiler build |
| `ci-benchmark.yml` | Performance benchmarks |
| `wheels.yml` / `wheels-dev.yml` | Python wheel builds |
| `slangpy_torch.yml` | PyTorch integration tests |
| `claude.yml` | Claude-driven CI tasks |

Key CI commands:

```bash
gh run list --repo shader-slang/slangpy --workflow=ci.yml --limit 5
gh run view <id> --log-failed
gh run rerun <id> --failed
```

## Issues

```bash
gh issue list --repo shader-slang/slangpy
gh issue view <number> --repo shader-slang/slangpy
gh issue create --repo shader-slang/slangpy --title "..." --body "..."
```

## Conflict handling

On merge conflict:
1. Identify conflicting files: `git diff --name-only --diff-filter=U`
2. For conflicts in generated files: take `ours` where the file is regenerated by a build step
3. For conflicts in source files: apply the change that is additive (preserves both sides) where possible
4. If the correct resolution is ambiguous, stop and notify the user with the specific conflict

## Related repositories

- **Slang compiler:** https://github.com/shader-slang/slang -- the Slang shading language compiler that SlangPy depends on
- **slang-rhi:** GPU abstraction layer wrapped by the SGL core layer

## From project

- `CONTRIBUTING.md` -- fork/clone/branch process, PR requirements, squash merge policy, sync instructions
- `AGENTS.md` -- repository structure, CI system description
- `.github/workflows/` -- CI workflow definitions