yang2sdk

Generate a Pydantic v2, IDE-friendly SDK for your network device directly from the YANG modules it actually runs.

flowchart LR
    A[yang-downloader<br/>get-schema off the box] --> B[pyang -f tree<br/>you pick root modules]
    B --> C[yang2restconf / yang2netconf<br/>AST → IR → Jinja]
    C --> D[versioned Python package]
    D --> E[HAL + adapters<br/>multi-vendor automation]
    D -. lab-only .-> F[sdk-verify<br/>read + RPC + gated CRUD]

• For network engineers: get an SDK in 4 commands · write safely · validate on lab
• For contributors: gates and tests · CONTRIBUTING.md · docs/dev.md · AGENTS.md · SECURITY.md

Both targets share the same navigator intent — retrieve / update / replace / create / delete, RPC __call__ — but the per-class surface is asymmetrical by protocol (e.g. RESTCONF plural lists expose create, not update; NETCONF plural lists expose update, not delete); only transport/encoding differs (JSON vs XML). See docs/usage.md and AGENTS.md parity table.

Example

from g30_1_4_0 import RestconfClient

# Auth comes from args or DEVICE_USER / DEVICE_PASS at runtime.
# Nothing is baked into generated code. verify=True is the default.
client = RestconfClient(management_ip="192.0.2.10")

# Top-level data is "<module>_<node>"; lists take key(s) in `key` order.
iface = client.data.ietf_interfaces_interface("ethernet-1/1")

port = iface.retrieve(content="config", depth=2)
port.description = "uplink-to-core"
iface.update(port)  # PATCH merge — the verb you normally want

Pydantic validates on assignment, so the IDE catches bad values before they reach the device:

port.admin_status = "dowm"
# ValidationError: Input should be 'up' or 'down'

replace() is PUT / nc:operation="replace" — the body is the whole resource (RFC 8040 §4.5, RFC 6241 §8.2.1). It refuses a depth-truncated model instead of silently deleting the rest; read deeper, use update(), or pass allow_partial=True when deletion is intended. See docs/usage.md.

For network engineers

Requires Python >= 3.12 and uv.

git clone https://github.com/ConsortiumGARR/yang2sdk.git
cd yang2sdk
uv sync --locked --extra lab
cp .env.example .env   # never commit .env; DEVICE_USER / DEVICE_PASS live here

uv run yang-downloader
mkdir -p temp/yang_tree
uv run pyang -p temp/yang_modules/<device>/ -f tree temp/yang_modules/<device>/*.yang > temp/yang_tree/<device>.txt
uv run yang2restconf temp/yang_modules/<device>/file1.yang [file2.yang ...] --device <device>
uv run yang2netconf  temp/yang_modules/<device>/file1.yang [file2.yang ...] --device <device>

temp/ is an untracked workspace — every command creates its own output dir on demand, so a fresh clone has no temp/ until the first run (and rm -rf temp/ only loses re-downloadable/recompilable output).

Each compile emits a self-contained package at temp/<protocol>_clients/<device>_<os-version>/ (client, models, navigators, pyproject.toml, README, MANIFEST.yang-revisions.json, py.typed):

uv add path/to/temp/restconf_clients/<device>_<os-version>

from <device>_<os_version> import RestconfClient  # or NetconfClient

One package per device and OS version — different YANG revisions produce different models. Next: compile flags and HAL sketch · lab validation with sdk-verify.

For contributors

uv sync --locked --group dev --extra lab
uv run ruff check . && uv run ruff format --check . \
  && uv run ty check && uv run pyrefly check \
  && uv run pytest tests/test_matrix.py   # offline gate: no docker, no device

ruff is the only linter/formatter. ty and pyrefly must both pass on src/ (temp/ output is excluded). The four gates are pinned exactly and run from the venv, never via uvx — see docs/dev.md for why, and for the bump procedure. Live-matrix tests need docker / lab gear — see docs/dev.md for the test/CI matrix and AGENTS.md for the normative generator contract.

Docs

Status

Public prototype. Both protocols generate; wire shapes are pinned offline in tests/test_matrix.py and exercised live against notconf simulators, SR Linux (full vendor tree, non-NMDA), and the Groove G30. Alternatives (pydantify, pyangbind) cover data modelling but not network operations — this project generates models and the operation code on top of pyang.