docs(contributing): align PR, code, and CI guidance by yaoyu-33 · Pull Request #3194 · NVIDIA-NeMo/Gym

yaoyu-33 · 2026-09-09T20:26:44Z

Summary

This PR establishes repo-wide contributor and AI-agent guidance for Gym:

documents PR titles, focused changes, simplicity-first code guidance, typing and public-API expectations, logging/error-handling conventions, DCO, and validation handoff in AGENTS.md and CONTRIBUTING.md
adds a nemo-gym-pr-checks-and-labels skill for affected-file history review, CI/check selection, missing-check diagnosis, and Gym's live label taxonomy
updates the PR template and contributor docs to request exact validation, compatibility impact, and rollout evidence when applicable
adds native Claude/Codex skill discovery links and regression coverage

The simplicity guidance distills the core idea from NVIDIA NeMo labs-molt's simplicity-first skill directly into Gym's always-loaded code-style rules: prefer the smallest clear solution for current requirements, avoid speculative abstractions, and preserve required behavior.

The guidance adapts only conventions that Gym currently supports. In particular, Gym uses Ruff rather than Black, area:* labels rather than Megatron Bridge title prefixes, and does not claim a repository-wide strict mypy or Ruff-docstring CI gate.

Part of #3122.

Relationship to #3123

This PR is independent of #3123 and can merge first.

docs(skills): align repository skills with current contracts #3123 updates the technical contracts and skills for adding environments, benchmarks, and Fern documentation.
docs(contributing): align PR, code, and CI guidance #3194 defines the more general repository foundation: contribution, PR, code-style, validation, CI, and label guidance.

The PRs overlap in a few documentation and skill-discovery files, so whichever lands second will need a mechanical rebase/conflict resolution. There is no semantic dependency, and reviewing #3123 first is not required.

What counts as a smoke rollout?

For behavior-changing environment or agent code, a representative real smoke rollout is the smallest model-backed end-to-end run that exercises the changed runtime path--normally one or a few representative tasks. The contributor should inspect the agent output or tool interactions and the verifier result, then record the command/configuration and observed outcome in the PR.

Unit tests, mocked requests, and health checks do not replace this evidence. A full evaluation, reward profile, or training run is not required. Docs-only and metadata-only catalog or manifest changes may record rollouts as N/A.

Compatibility and benchmark impact

None. This PR changes contributor policy, documentation, repository skills, discovery links, and their tests; it does not change runtime behavior or benchmark results.

Validation

Agent Skills structural validation
pytest tests/unit_tests/test_skills.py tests/unit_tests/test_blade_toolkit.py -q -- 26 passed
python tests/unit_tests/test_fern_docs_links.py -- 14/14 passed
pre-commit run --all-files
make docs-check -- 0 errors; redirect validation skipped because Fern authentication was unavailable locally
git diff --check
exercised the documented PR inspection, check, label, and REST fallback commands against live Gym PRs
forward-tested the affected-file history workflow against test: [1/3] add L1 environment verifier acceptance coverage #3127, including the constraint recorded in feat: unified judge failure handling #2113
three independent review rounds found no remaining correctness blockers