---

name: repo-development description: "Use for NeMo Speech repository maintenance: installs, code style, focused tests, docs builds, bug reproduction, optional backends, safe verification, and PR/CI guidance." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Repo Development

Use this sub-skill when changing, validating, or reviewing the NeMo Speech repository itself: installing a developer environment, choosing optional dependency groups, adding/fixing tests, running formatting checks, building docs, triaging import/backend failures, planning safe native verification, and preparing PR/CI guidance.

Route end-user model work elsewhere. Classic ASR transcription/fine-tuning/decoding belongs in ../asr/SKILL.md; audio enhancement/separation belongs in ../audio/SKILL.md; SpeechLM2/SALM/voice-agent work belongs in ../speechlm2/SKILL.md; speaker diarization/VAD/forced alignment belongs in ../speaker-diarization/SKILL.md. TTS synthesis and generic data/tokenizer tools should use their sibling sub-skills when present.

Start Here

1. Read references/maintainer-workflows.md for repository architecture, edit policy, bug-fix flow, docs, PR/CI, and safe source-script handling.
2. Read references/testing-and-formatting.md before choosing pytest, isort, black, docs, or native verification commands.
3. Read references/dependency-backends.md before installing extras, CUDA groups, docs/test groups, or optional compiled backends.
4. Read references/troubleshooting.md when imports, optional dependencies, CUDA/backends, manifests/configs, Hydra overrides, docs, tests, or CI fail.
5. Run python scripts/select_tests.py --help to plan focused commands from changed paths or capability keywords. The helper prints commands only; it never runs tests, downloads data, trains, formats, or mutates files.

Core Rules

• Reproduce bugs first with a minimal focused test, add the reproduction as a unit test, then fix the source and verify the targeted test passes.
• Keep changes focused; do not modify CI workflow files, delete tests, push branches, or change unrelated formatting unless explicitly asked.
• Use NeMo formatting settings: 119-character lines, Black required version 24 style, skip_string_normalization = true, and isort profile = black with nemo as first-party.
• Prefer focused validation: run relevant tests with -m "not pleasefixme", then run isort --check and black --check on touched Python paths.
• Treat original repo examples, utility scripts, and workflow files as evidence unless a safe bundled helper exists here. Long-running, network-bound, GPU/training-heavy, checkpoint-mutating, or checkout-dependent scripts are reference-only in this sub-skill.

Safe Bundled Tool

• scripts/select_tests.py maps changed paths and capability keywords to focused command suggestions such as pytest tests/collections/asr -m "not pleasefixme" -v, isort --check <paths>, black --check <paths>, and docs build commands.
• Example: python scripts/select_tests.py --changed nemo/collections/asr/models/foo.py tests/collections/asr/test_foo.py --keyword docs --keyword asr.

Evidence Base

This sub-skill distills repository evidence from AGENTS.md, CLAUDE.md, README.md, CONTRIBUTING.md, pyproject.toml, docs/source/starthere/install.rst, docs/README.md, tests/**, nemo/core/**, nemo/utils/**, and .github/workflows/**. Runtime guidance is self-contained; future agents should not need to reopen those source files or run original repo scripts to complete maintainer tasks.