Network automation standards — the network as code
Criteria verified as of August 2026. Re-verify on the web before committing to anything (§8).
1. Scope and triggers
Applies when automating the configuration and verification of network devices: source of truth
and its governance, intent data model, configuration generation, device interfaces (CLI,
NETCONF/RESTCONF, gNMI), testing and operational-state validation, virtual lab, network CI/CD with
batched rollout and tested rollback, drift detection, streaming telemetry, and the custody of the
credentials that grant access to the entire fleet.
Triggers: *.clab.yml, containerlab, netlab, vrnetlab, nornir_config.yaml, hosts.yaml,
groups.yaml, napalm, netmiko, scrapli, pynetbox, ansible.netcommon,
cisco.ios/arista.eos/junipernetworks.junos/nokia.srlinux, gnmic, pygnmi, ncclient,
pyang, batfish, pyats/genie, network configuration *.j2, "config context",
"config template", "source of truth", "drift", "pre-check"/"post-check", "commit-confirm",
"streaming telemetry", "dial-in"/"dial-out".
Not applicable: each skill decides a different thing:
networking-standards (the trunk, the parent: decides addressing and IPAM, VLANs, routing
fundamentals, MTU/MSS and that NetBox is the intent SoT; none of that is repeated here, here what
gets decided is how that intent is modelled, how the configuration is generated and how it reaches
the device without breaking anything); routing-switching-standards (decides what the campus and
edge configuration must say: STP, LACP, MLAG, IGP, BGP policy, RPKI, CoPP, AAA);
datacenter-fabric-standards (decides what the fabric configuration must say: Clos, EVPN, VRF,
MTU, lossless networking). Those two design; this one automates what they design.
network-troubleshooting-standards (decides the reactive method when the automated change broke
something); iac-standards (decides Terraform and Ansible as tools and how they are written —
here only what is network-specific); secrets-management-standards (decides the secrets manager,
rotation and ephemeral credentials); observability-standards (decides the platform, the
thresholds and the alerts — here only where the device data comes from and over which protocol);
cicd-standards (decides the pipeline engine); firewall-policy-standards (decides the
filtering policy and its governance, even when it is applied with these tools);
opensource-licensing-standards (decides the acceptable licence of every tool adopted). Also
bordering: sre-practice-standards, dns-standards, vpn-standards, kubernetes-standards,
onprem-standards (umbrella), identity-access-management-standards, linux-hardening-standards,
vulnerability-management-standards, finops-standards, offensive-security-standards (this
skill is defensive), and network-vendors-standards, wan-legacy-standards,
telco-5g-standards, high-speed-interconnect-standards and datacenter-facilities-standards.
Governing principle: automating a badly designed network breaks it faster and in more places at
once. Automation does not fix the design: it multiplies it. That is why the order is first read
and check, then generate, and only at the end apply, and why the hard rule is that a network
change either rolls itself back or it is not applied.
2. Default decisions
Versions and licences verified Aug 2026 against the PyPI JSON API, the Atom release feeds
and the raw LICENSE file. Re-verify before committing to anything (§8).
| Area |
Default |
Justifiable alternative / vetoed |
| Source of truth |
NetBox 4.6.7 (30 Jul 2026; Apache-2.0, verified in the raw LICENSE.txt) as the intent SoT |
❌ Spreadsheet; ❌ the device configuration as SoT; ❌ populating it by discovery and calling that intent |
| Data model |
Versioned declarative intent (YAML or NetBox objects) with a schema validated in CI |
❌ Intent embedded in the templates; ❌ schema-less data |
| General orchestrator |
Ansible (ansible-core 2.21.2, GPL-3.0-or-later) with network collections, if it is already the house standard |
Slow execution model on large fleets and crude per-host error handling |
| Python framework |
Nornir 3.6.0 (2 Aug 2026, Apache-2.0) when real logic, concurrency and tests are needed |
Nornir is a framework, not a toolbox: it brings inventory and parallelism, you supply the drivers (nornir-napalm 0.6.0, nornir-netmiko) |
| Multi-vendor abstraction |
NAPALM 5.2.0 (27 Jul 2026, Apache-2.0) for normalised get_*, compare_config and confirmed commit |
Limited and uneven platform coverage: verify the driver for your NOS before designing on top of it |
| CLI transport |
netmiko 4.7.0 (12 May 2026, MIT) as the safe option; scrapli (MIT) for performance, async or NETCONF |
scrapli is in transition (§8): the latest non-prerelease on PyPI is 2026.2.20 and the 2.0 rewrite is at release candidate. Do not pin it as the default yet |
| Templates |
Jinja 3.1.6 (BSD-3-Clause), with dumb templates and rich data |
❌ Business logic in the template: code without tests |
| Device interface |
gNMI for telemetry and, where support is solid, for configuration; NETCONF (RFC 6241) for transactional configuration |
CLI only when there is no data model; RESTCONF (RFC 8040) where it is the only thing available |
| gNMI client |
gnmic 0.46.0 (14 May 2026, Apache-2.0, under the openconfig organisation); pygnmi (BSD-3-Clause) from Python |
Writing your own gRPC client without needing to |
| YANG models |
OpenConfig first in a multi-vendor environment; IETF for the basics (interfaces, IP, routing); native only for what neither covers |
❌ Designing the whole automation on native models: that is CLI with a different syntax |
| Virtual lab |
containerlab v0.77.0 (28 Jun 2026; BSD-3-Clause, copyright Nokia, verified raw) |
netlab (package networklab 26.7, MIT) on top; GNS3/EVE-NG and vrnetlab if only VM images exist. NOS images have their own licence and redistribution restrictions: that is a legal problem |
| Validation |
Operational-state pre-checks and post-checks, plus static configuration analysis and intent validation |
❌ Validating only that the configuration got written |
| Telemetry |
Streaming telemetry (gNMI Subscribe) where the NOS supports it; SNMP as a fallback |
❌ Mass SNMP polling as the only telemetry on a large fleet |
3. The right order and its structure
The order, and no step gets skipped
- Read and check — inventory, current state, drift against the SoT. Everything read-only is
automated first: it delivers value without risk and builds confidence and knowledge of the real
estate before touching anything.
- Generate — configuration rendered from the intent, into a reviewable artifact. The diff is
the object under review, not the template.
- Apply — only at the end, in batches, with rollback armed.
The source of truth, decided before the tool
- An intent SoT and a discovery SoT are different things and do not get mixed. NetBox holds
what must be; discovery produces what there is; the difference is drift, and drift is a
finding with an owner. Auto-populating the SoT from the network erases exactly that signal.
- Where each thing lives (what
networking-standards does not decide): data and relations in
NetBox (sites, devices, interfaces, prefixes, VLANs, circuits); parameters that modulate the
template in hierarchical config contexts; rendering in config templates or in the Git
repository. Split rule: if two systems can answer the same question with different answers, one
of them is redundant.
- Every piece of intent data has a schema and is validated in CI: a schema-less YAML is a
production failure waiting for the first typo. Fail in the pipeline, not on the device.
- The SoT closes in both directions: after every change, the SoT reflects the intended reality or
the change is not finished.
Device interfaces: the data model is what makes automation portable
- CLI by screen-scraping works on everything and breaks with anything: a format change in a
minor version, a banner, a warning, paginated output. There is no contract, no transaction and no
validation; and the parsing (TextFSM, TTP, home-made templates) is code to maintain per platform
and per version.
- NETCONF (RFC 6241) with YANG (RFC 7950) brings what the CLI does not have: candidate
datastore, transaction, prior validation and confirmed commit. With NMDA (RFC 8342) the
separation between intended configuration, applied configuration and operational state stops being
ambiguous — which is exactly the distinction that lets you verify intent against reality.
- gNMI/gRPC is the best option for telemetry (subscription, high frequency, efficient) and
increasingly valid for configuration; its declared weak point is the transaction, where it is
more limited than NETCONF.
- The data model is the portability: writing against OpenConfig or IETF models means the same
code serves another vendor; against native models or CLI, it means rewriting it at the next
purchase. The realistic answer is hybrid: OpenConfig is operationally complete, not exhaustive,
and real interoperability between vendors is still imperfect. Design with a standard model and
isolate in an adapter whatever demands a native model or CLI.
Templates: rich data, dumb templates — every decision that can be taken in the data is taken in
the data. One template per role, composable in blocks, with deterministic rendering (same data →
same text, byte for byte) so that the diff means something.
4. Testing, validation and network CI/CD
- Virtual lab with the same NOS versions as production, and with the topology generated from
the same SoT: if the lab is described by hand, it tests a different network.
- Network pipeline, in increasing cost order: (1) lint and schema validation of the intent;
(2) rendering and configuration
diff as a reviewable artifact in the PR; (3) static analysis
of the rendered configuration (reachability, policy, design errors) where the tooling allows it;
(4) deployment to the lab and functional tests; (5) canary batch in production with
post-checks; (6) the rest of the fleet in batches, with an abort criterion between batches.
- Pre-checks and post-checks on operational state, not on configuration: what matters is not that
the line got written, but that adjacencies are still up, the expected prefixes are still there,
interfaces are not accumulating errors and application traffic passes. They are captured before,
compared after, and the comparison is automatic.
- Intent validation as a continuous gate: periodic checking that the network behaves as the SoT
says (routes, neighbours, VLANs, zero drift). A failure here is a finding with an owner.
- The hard rule: a network change either rolls itself back or it is not applied. Timed rollback
(
commit-confirm or equivalent) armed before the change, with a window sized to the
verification time, and the rollback plan tested in the lab as part of the change: an unrehearsed
rollback is not a plan, it is a hope. Always with an OOB console available.
- Verified idempotence: the second run changes nothing. If it changes something, the playbook is
lying about state and is useless for detecting drift.
5. Security: the automation holds the credentials to the whole network
This is the dominant risk of this skill: whoever controls the automation system controls every
device at once, with changes that also look legitimate.
- No credential in the repository or in the inventory: they are fetched at run time from the
secrets manager (
secrets-management-standards); in CI, OIDC/workload identity rather than
long-lived static keys.
- Named service accounts, with least privilege: a read-only one for inventory, drift and
telemetry — which is most of the work — and a different, write-capable one only for the apply step.
TACACS+ allows per-command authorisation: use it to bound what the automation account can do
(
routing-switching-standards, identity-access-management-standards).
- Ephemeral credentials and rotation; never a password shared between humans and automation,
because it destroys the traceability of who changed what.
- A record of every change, correlatable end to end: who requested it, which PR approved it, which
run applied it, to which devices and with which diff. If the device logs and the automation logs
cannot be cross-referenced, a malicious change is indistinguishable from a routine one.
- The network configuration repository is a high-value target: it contains topology, addressing,
filtering policy and — if someone was careless — credentials. Restricted access, mandatory review,
commit signing, branch protection and secret scanning as a gate that breaks the build. A
secret that reached the history is compromised: it gets rotated.
- Backed-up configurations: encrypted and sanitised — they contain password hashes, pre-shared
keys, SNMP communities and certificates.
- The automation system is critical infrastructure: management network, minimal surface, patching
up to date, MFA for whoever operates it, and its own recovery plan (if it goes down, you can
neither change nor roll back).
- Supply chain: collections, PyPI packages and NOS images pinned by version or digest, with
SCA in CI. A compromised collection runs against the entire fleet.
6. Telemetry and operability
- Streaming versus polling: subscription (gNMI
Subscribe, dial-in or dial-out) gives higher
frequency, lower CPU cost on the device and data already structured per the YANG model. Mass SNMP
polling scales badly and misses short events. Criterion: streaming where the NOS supports it,
SNMP as a fallback for the older estate and for what the model does not expose — it is not
switched off out of dogma. The platform that receives, stores and alerts belongs to
observability-standards.
- The automation's own signals: success rate per run and per device, drift detected (count and
age), time from commit to applied change, number of rollbacks, and unmanaged devices — those
last ones are what break rollouts.
- Declared coverage: what percentage of the estate is automated and what is left out, with a
reason. A half-automated estate is more dangerous than a manual one if nobody knows which is which.
- Toil: a manual task repeated three times is a candidate for automation; and automation that
requires routine manual intervention is badly built (
sre-practice-standards).
7. Sustainability and prohibitions
- Start with reading: inventory, drift and telemetry first. Immediate value, zero risk, and it
builds what makes the write phase viable.
- Cadence: review version, maintenance status and licence every quarter; pin versions and
upgrade deliberately. A tool that changes licence or goes into maintenance mode is an architecture
decision, not a footnote (
opensource-licensing-standards).
- Deprecation: unused templates, playbooks and scripts get deleted — dead code gets executed by
accident one day.
FORBIDDEN
- ❌ Applying to the whole fleet with no test batch. Canary first, batches after, with a written
abort criterion between batches.
- ❌ Automating without a reliable inventory: without a complete and correct SoT, errors propagate
at scale.
- ❌ Storing configurations with cleartext credentials in the repository (or secrets in
inventories, group variables, templates or run logs).
- ❌ Relying on the CLI when there is a data model available on the platform.
- ❌ Applying a change without an armed and tested rollback, or without an OOB console available.
- ❌ Changing configuration by hand in production "and I will put it in the repo later".
- ❌ Populating the SoT by automatic discovery and calling it intent.
- ❌ Intent data with no schema and no validation in CI.
- ❌ Business logic inside the Jinja templates.
- ❌ Reviewing the template in the PR instead of the rendered configuration diff.
- ❌ Post-checks that only verify that the configuration got written.
- ❌ Running with a shared administrator account, or with the same credential for reading and writing.
- ❌ Long-lived credentials in CI when OIDC/workload identity is available.
- ❌ Collections, packages or images without a pinned version or digest.
- ❌ Automating a network whose design is not settled: fix the design first
(
routing-switching-standards, datacenter-fabric-standards).
- ❌ A non-idempotent playbook used as a drift detector.
- ❌ Redistributing virtual-lab NOS images without checking their licence.
8. Mandatory web verification
Methodology: versions from the PyPI JSON API and the GitHub Atom release feeds
(api.github.com returns 403 unauthenticated); licences read from the raw LICENSE file, not
from the label GitHub's interface shows nor from a summary.
Verified Aug 2026 (version | date | licence checked raw). NetBox 4.6.7 |
30 Jul 2026 | Apache-2.0 (LICENSE.txt on main and master; plain LICENSE returns 404 —
the path matters when verifying), with v4.6.8-rc2 in flight (4 Aug 2026). Nornir 3.6.0 |
2 Aug 2026 | Apache-2.0 (slow but alive cadence: 3.5.0 was from Jan 2025). NAPALM 5.2.0 |
27 Jul 2026 | Apache-2.0. netmiko 4.7.0 | 12 May 2026 | MIT. Jinja2 3.1.6 |
BSD-3-Clause. ansible-core 2.21.2 | GPL-3.0-or-later. pygnmi 0.8.15 |
BSD-3-Clause. gnmic 0.46.0 | 14 May 2026 | Apache-2.0, and it lives under the
openconfig organisation, not under the original personal repository: the project moved, and
searching for the old repo gives a false impression of abandonment. containerlab v0.77.0 |
28 Jun 2026 | BSD-3-Clause with Nokia copyright (read in the raw LICENSE): it is not
Apache-2.0, which is the usual assumption. netlab = PyPI package networklab 26.7 |
MIT (ipspace and contributors), CalVer, and the package name does not match the project name.
Declared discrepancy — scrapli: GitHub and PyPI give different versions for the same thing.
The Atom feed shows v2.0.0-rc.16 (20 Jul 2026); PyPI publishes that artifact as 2026.7.20rc16 and
declares 2026.2.20 (Feb 2026) as the current version, because under PEP 440 the CalVer sorts
above 2.0.0rc. Operational conclusion: the 2.0 rewrite is at release candidate, a
pip install scrapli without --pre installs the Feb 2026 CalVer line, and pinning scrapli 2.0 as
the default today is premature.
State of OpenConfig/gNMI (qualitative): active project (gnmic, ygot, ygnmi, gNOI and
featureprofiles with activity in 2026). The models are written in YANG 1.0 and are
"operationally complete", not exhaustive; real interoperability between vendors is still
imperfect, and gNMI is weaker than NETCONF for transactional configuration (although the
specification requires that a SetRequest spanning several origins be treated as a transaction
with rollback). No authoritative adoption-share figure was found: treat any number about
OpenConfig or gNMI adoption as a Declared gap.
RFCs verified one by one against rfc-editor.org: NETCONF RFC 6241 (Jun 2011, updated by
7803 and 8526); RESTCONF RFC 8040 (Feb 2017, updated by 8527); YANG 1.1 RFC 7950 (Aug 2016,
updated by 8342 and 8526); NMDA RFC 8342 (Mar 2018); YANG Library RFC 8525 (Mar 2019,
obsoletes RFC 7895).
Declared gaps — do NOT fill from memory:
- Versions and status of the Ansible network collections (
ansible.netcommon, cisco.ios,
arista.eos, junipernetworks.junos, nokia.srlinux): not verified; their cadence is
independent of ansible-core.
- NAPALM's real coverage per platform and NOS version: not verified, and it is what decides
whether NAPALM is usable in your estate.
- Status, version and licence of Batfish and of pyATS/Genie: not verified; they are mentioned
as a category, not as a pinned default.
- Status and licence of GNS3, EVE-NG and
vrnetlab: not verified. EVE-NG has commercial
editions: check the specific edition.
- Licence and redistribution of lab NOS images: vendor-specific and not verified. It is a
legal question (
opensource-licensing-standards).
- gNMI and NETCONF support per platform and NOS version, and the concrete models exposed: not
verified. Check it with
Capabilities on the real device.
- The relationship between NetBox Community (Apache-2.0) and NetBox Labs' commercial offerings
and what functionality falls outside the open edition: not verified.
ncclient, pyang, TextFSM/ntc-templates and TTP: version, maintenance and licence not
verified.
If the web contradicts this document, the web wins — flag the discrepancy.
1---2name: network-automation-standards3description: Network as code — source of truth, generation, validation and safe rollout of device configuration. Use when driving devices from Ansible network collections (ansible.netcommon, cisco.ios, arista.eos, junipernetworks.junos, nokia.srlinux), Nornir with nornir-napalm or nornir-netmiko, NAPALM get_facts/compare_config/commit_config, netmiko send_config_set, scrapli or scrapli-netconf, Jinja templates rendering device config, NetBox as source of truth with pynetbox, custom fields, config contexts, config templates or the NetBox Ansible/Nornir inventory plugin, YAML intent data and schema validation, NETCONF (RFC 6241), RESTCONF (RFC 8040), YANG 1.1 (RFC 7950), NMDA (RFC 8342), candidate datastores and confirmed-commit, gNMI Get/Set/Subscribe with gnmic or pygnmi, OpenConfig versus IETF versus native YANG models, streaming telemetry replacing SNMP polling, containerlab .clab.yml topologies, netlab, vrnetlab or GNS3/EVE-NG virtual labs, pre-checks and post-checks, config diff and dry-run, drift detection against intent, batched canary rollout with tested rollback, commit-confirm timers, Batfish or pyATS/Genie operational-state validation, or storing device configs and credentials in a git repository.4metadata:5 source: https://skillmd.com/skills/serialexperimentslainnnn/network-automation-standards6---78# Network automation standards — the network as code910Criteria verified as of **August 2026**. Re-verify on the web before committing to anything (§8).1112## 1. Scope and triggers1314Applies when **automating the configuration and verification of network devices**: source of truth15and its governance, intent data model, configuration generation, device interfaces (CLI,16NETCONF/RESTCONF, gNMI), testing and operational-state validation, virtual lab, network CI/CD with17batched rollout and tested rollback, drift detection, streaming telemetry, and the **custody of the18credentials that grant access to the entire fleet**.1920Triggers: `*.clab.yml`, `containerlab`, `netlab`, `vrnetlab`, `nornir_config.yaml`, `hosts.yaml`,21`groups.yaml`, `napalm`, `netmiko`, `scrapli`, `pynetbox`, `ansible.netcommon`,22`cisco.ios`/`arista.eos`/`junipernetworks.junos`/`nokia.srlinux`, `gnmic`, `pygnmi`, `ncclient`,23`pyang`, `batfish`, `pyats`/`genie`, network configuration `*.j2`, "config context",24"config template", "source of truth", "drift", "pre-check"/"post-check", "commit-confirm",25"streaming telemetry", "dial-in"/"dial-out".2627**Not applicable**: each skill **decides** a different thing:28`networking-standards` (**the trunk, the parent**: decides **addressing and IPAM, VLANs, routing29fundamentals, MTU/MSS and that NetBox is the intent SoT**; none of that is repeated here, here what30gets decided is **how that intent is modelled, how the configuration is generated and how it reaches31the device without breaking anything**); `routing-switching-standards` (**decides what the campus and32edge configuration must say**: STP, LACP, MLAG, IGP, BGP policy, RPKI, CoPP, AAA);33`datacenter-fabric-standards` (**decides what the fabric configuration must say**: Clos, EVPN, VRF,34MTU, lossless networking). **Those two design; this one automates what they design.**35`network-troubleshooting-standards` (**decides the reactive method** when the automated change broke36something); `iac-standards` (**decides Terraform and Ansible as tools and how they are written** —37here only what is network-specific); `secrets-management-standards` (**decides the secrets manager**,38rotation and ephemeral credentials); `observability-standards` (**decides the platform, the39thresholds and the alerts** — here only where the device data comes from and over which protocol);40`cicd-standards` (**decides the pipeline engine**); `firewall-policy-standards` (**decides the41filtering policy and its governance**, even when it is applied with these tools);42`opensource-licensing-standards` (**decides the acceptable licence** of every tool adopted). Also43bordering: `sre-practice-standards`, `dns-standards`, `vpn-standards`, `kubernetes-standards`,44`onprem-standards` (umbrella), `identity-access-management-standards`, `linux-hardening-standards`,45`vulnerability-management-standards`, `finops-standards`, `offensive-security-standards` (**this46skill is defensive**), and `network-vendors-standards`, `wan-legacy-standards`,47`telco-5g-standards`, `high-speed-interconnect-standards` and `datacenter-facilities-standards`.4849**Governing principle**: **automating a badly designed network breaks it faster and in more places at50once.** Automation does not fix the design: it multiplies it. That is why the order is **first read51and check, then generate, and only at the end apply**, and why the hard rule is that **a network52change either rolls itself back or it is not applied**.5354## 2. Default decisions5556> Versions and licences verified Aug 2026 against the **PyPI JSON API**, the **Atom release feeds**57> and the **raw `LICENSE` file**. Re-verify before committing to anything (§8).5859| Area | Default | Justifiable alternative / vetoed |60|---|---|---|61| Source of truth | **NetBox 4.6.7** (30 Jul 2026; **Apache-2.0**, verified in the raw `LICENSE.txt`) as the **intent** SoT | ❌ Spreadsheet; ❌ the device configuration as SoT; ❌ populating it by discovery and calling that intent |62| Data model | **Versioned declarative intent** (YAML or NetBox objects) **with a schema validated in CI** | ❌ Intent embedded in the templates; ❌ schema-less data |63| General orchestrator | **Ansible** (`ansible-core` 2.21.2, **GPL-3.0-or-later**) with network collections, if it is already the house standard | Slow execution model on large fleets and crude per-host error handling |64| Python framework | **Nornir 3.6.0** (2 Aug 2026, **Apache-2.0**) when real logic, concurrency and tests are needed | Nornir is a framework, not a toolbox: it brings inventory and parallelism, **you supply the drivers** (`nornir-napalm` 0.6.0, `nornir-netmiko`) |65| Multi-vendor abstraction | **NAPALM 5.2.0** (27 Jul 2026, **Apache-2.0**) for normalised `get_*`, `compare_config` and confirmed commit | Limited and uneven platform coverage: **verify the driver for your NOS before designing on top of it** |66| CLI transport | **netmiko 4.7.0** (12 May 2026, **MIT**) as the safe option; **scrapli** (**MIT**) for performance, async or NETCONF | **scrapli is in transition** (§8): the latest non-prerelease on PyPI is `2026.2.20` and the 2.0 rewrite is at *release candidate*. **Do not pin it as the default yet** |67| Templates | **Jinja 3.1.6** (**BSD-3-Clause**), with **dumb templates and rich data** | ❌ Business logic in the template: code without tests |68| Device interface | **gNMI** for telemetry and, where support is solid, for configuration; **NETCONF (RFC 6241)** for transactional configuration | **CLI** only when there is no data model; **RESTCONF (RFC 8040)** where it is the only thing available |69| gNMI client | **`gnmic` 0.46.0** (14 May 2026, **Apache-2.0**, under the **openconfig** organisation); **`pygnmi`** (BSD-3-Clause) from Python | Writing your own gRPC client without needing to |70| YANG models | **OpenConfig** first in a multi-vendor environment; **IETF** for the basics (interfaces, IP, routing); **native** only for what neither covers | ❌ Designing the whole automation on native models: that is CLI with a different syntax |71| Virtual lab | **containerlab v0.77.0** (28 Jun 2026; **BSD-3-Clause**, copyright **Nokia**, verified raw) | **netlab** (package `networklab` 26.7, **MIT**) on top; GNS3/EVE-NG and `vrnetlab` if only VM images exist. **NOS images have their own licence and redistribution restrictions: that is a legal problem** |72| Validation | **Operational-state pre-checks and post-checks**, plus static configuration analysis and intent validation | ❌ Validating only that the configuration got written |73| Telemetry | **Streaming telemetry (gNMI `Subscribe`)** where the NOS supports it; **SNMP as a fallback** | ❌ Mass SNMP polling as the only telemetry on a large fleet |7475## 3. The right order and its structure7677**The order, and no step gets skipped**781. **Read and check** — inventory, current state, drift against the SoT. Everything read-only is79 automated first: it delivers value without risk and builds confidence and knowledge of the real80 estate before touching anything.812. **Generate** — configuration rendered from the intent, into a reviewable artifact. **The diff is82 the object under review**, not the template.833. **Apply** — only at the end, in batches, with rollback armed.8485**The source of truth, decided before the tool**86- **An intent SoT and a discovery SoT are different things and do not get mixed.** NetBox holds87 **what must be**; discovery produces **what there is**; the difference is **drift**, and drift is a88 finding with an owner. Auto-populating the SoT from the network erases exactly that signal.89- **Where each thing lives** (what `networking-standards` does not decide): data and relations in90 **NetBox** (sites, devices, interfaces, prefixes, VLANs, circuits); parameters that modulate the91 template in hierarchical **config contexts**; rendering in **config templates** or in the Git92 repository. Split rule: **if two systems can answer the same question with different answers, one93 of them is redundant.**94- **Every piece of intent data has a schema and is validated in CI**: a schema-less YAML is a95 production failure waiting for the first typo. Fail in the pipeline, not on the device.96- **The SoT closes in both directions**: after every change, the SoT reflects the intended reality or97 the change is not finished.9899**Device interfaces: the data model is what makes automation portable**100- **CLI by *screen-scraping*** works on everything and **breaks with anything**: a format change in a101 minor version, a banner, a warning, paginated output. There is no contract, no transaction and no102 validation; and the parsing (TextFSM, TTP, home-made templates) is code to maintain per platform103 and per version.104- **NETCONF (RFC 6241) with YANG (RFC 7950)** brings what the CLI does not have: **candidate105 datastore, transaction, prior validation and confirmed commit**. With **NMDA (RFC 8342)** the106 separation between intended configuration, applied configuration and operational state stops being107 ambiguous — which is exactly the distinction that lets you verify intent against reality.108- **gNMI/gRPC** is the best option for **telemetry** (subscription, high frequency, efficient) and109 increasingly valid for configuration; **its declared weak point is the transaction**, where it is110 more limited than NETCONF.111- **The data model is the portability**: writing against **OpenConfig** or IETF models means the same112 code serves another vendor; against native models or CLI, it means rewriting it at the next113 purchase. **The realistic answer is hybrid**: OpenConfig is operationally complete, not exhaustive,114 and real interoperability between vendors is still imperfect. Design with a standard model and115 **isolate in an adapter** whatever demands a native model or CLI.116117**Templates**: rich data, dumb templates — every decision that can be taken in the data is taken in118the data. One template per role, composable in blocks, with **deterministic rendering** (same data →119same text, byte for byte) so that the `diff` means something.120121## 4. Testing, validation and network CI/CD122123- **Virtual lab with the same NOS versions as production**, and with the topology **generated from124 the same SoT**: if the lab is described by hand, it tests a different network.125- **Network pipeline, in increasing cost order**: (1) lint and **schema validation** of the intent;126 (2) rendering and **configuration `diff`** as a reviewable artifact in the PR; (3) static analysis127 of the rendered configuration (reachability, policy, design errors) where the tooling allows it;128 (4) deployment to the **lab** and functional tests; (5) **canary batch** in production with129 post-checks; (6) the rest of the fleet **in batches**, with an abort criterion between batches.130- **Pre-checks and post-checks on operational state, not on configuration**: what matters is not that131 the line got written, but that **adjacencies are still up, the expected prefixes are still there,132 interfaces are not accumulating errors and application traffic passes**. They are captured before,133 compared after, and the comparison is automatic.134- **Intent validation as a continuous gate**: periodic checking that the network behaves as the SoT135 says (routes, neighbours, VLANs, zero drift). A failure here is a finding with an owner.136- **The hard rule: a network change either rolls itself back or it is not applied.** Timed rollback137 (`commit-confirm` or equivalent) armed **before** the change, with a window sized to the138 verification time, and the rollback plan **tested in the lab** as part of the change: an unrehearsed139 rollback is not a plan, it is a hope. Always with an OOB console available.140- **Verified idempotence**: the second run changes nothing. If it changes something, the playbook is141 lying about state and is useless for detecting drift.142143## 5. Security: the automation holds the credentials to the whole network144145This is the dominant risk of this skill: **whoever controls the automation system controls every146device at once, with changes that also look legitimate.**147148- **No credential in the repository or in the inventory**: they are fetched at run time from the149 secrets manager (`secrets-management-standards`); in CI, **OIDC/workload identity** rather than150 long-lived static keys.151- **Named service accounts, with least privilege**: a **read-only** one for inventory, drift and152 telemetry — which is most of the work — and a different, write-capable one only for the apply step.153 **TACACS+ allows per-command authorisation**: use it to bound what the automation account can do154 (`routing-switching-standards`, `identity-access-management-standards`).155- **Ephemeral credentials and rotation**; never a password shared between humans and automation,156 because it destroys the traceability of who changed what.157- **A record of every change, correlatable end to end**: who requested it, which PR approved it, which158 run applied it, to which devices and with which diff. If the device logs and the automation logs159 cannot be cross-referenced, a malicious change is indistinguishable from a routine one.160- **The network configuration repository is a high-value target**: it contains topology, addressing,161 filtering policy and — if someone was careless — credentials. Restricted access, mandatory review,162 **commit signing**, branch protection and **secret scanning as a gate that breaks the build**. A163 secret that reached the history is compromised: it gets **rotated**.164- **Backed-up configurations: encrypted and sanitised** — they contain password hashes, pre-shared165 keys, SNMP communities and certificates.166- **The automation system is critical infrastructure**: management network, minimal surface, patching167 up to date, MFA for whoever operates it, and its own recovery plan (if it goes down, you can168 neither change nor roll back).169- **Supply chain**: collections, PyPI packages and NOS images **pinned by version or digest**, with170 SCA in CI. A compromised collection runs against the entire fleet.171172## 6. Telemetry and operability173174- **Streaming versus polling**: subscription (gNMI `Subscribe`, dial-in or dial-out) gives higher175 frequency, lower CPU cost on the device and data already structured per the YANG model. Mass SNMP176 polling scales badly and misses short events. **Criterion**: streaming where the NOS supports it,177 **SNMP as a fallback** for the older estate and for what the model does not expose — it is not178 switched off out of dogma. The platform that receives, stores and alerts belongs to179 `observability-standards`.180- **The automation's own signals**: success rate per run and per device, **drift detected** (count and181 age), time from commit to applied change, number of rollbacks, and **unmanaged devices** — those182 last ones are what break rollouts.183- **Declared coverage**: what percentage of the estate is automated and what is left out, with a184 reason. A half-automated estate is more dangerous than a manual one if nobody knows which is which.185- **Toil**: a manual task repeated three times is a candidate for automation; and automation that186 requires routine manual intervention is badly built (`sre-practice-standards`).187188## 7. Sustainability and prohibitions189190- **Start with reading**: inventory, drift and telemetry first. Immediate value, zero risk, and it191 builds what makes the write phase viable.192- **Cadence**: review version, **maintenance status and licence** every quarter; pin versions and193 upgrade deliberately. **A tool that changes licence or goes into maintenance mode is an architecture194 decision**, not a footnote (`opensource-licensing-standards`).195- **Deprecation**: unused templates, playbooks and scripts get deleted — dead code gets executed by196 accident one day.197198**FORBIDDEN**199- ❌ **Applying to the whole fleet with no test batch.** Canary first, batches after, with a written200 abort criterion between batches.201- ❌ **Automating without a reliable inventory**: without a complete and correct SoT, errors propagate202 at scale.203- ❌ **Storing configurations with cleartext credentials in the repository** (or secrets in204 inventories, group variables, templates or run logs).205- ❌ **Relying on the CLI when there is a data model** available on the platform.206- ❌ Applying a change without an armed and **tested** rollback, or without an OOB console available.207- ❌ Changing configuration by hand in production "and I will put it in the repo later".208- ❌ Populating the SoT by automatic discovery and calling it intent.209- ❌ Intent data with no schema and no validation in CI.210- ❌ Business logic inside the Jinja templates.211- ❌ Reviewing the template in the PR instead of the **rendered configuration diff**.212- ❌ Post-checks that only verify that the configuration got written.213- ❌ Running with a shared administrator account, or with the same credential for reading and writing.214- ❌ Long-lived credentials in CI when OIDC/workload identity is available.215- ❌ Collections, packages or images without a pinned version or digest.216- ❌ Automating a network whose design is not settled: fix the design first217 (`routing-switching-standards`, `datacenter-fabric-standards`).218- ❌ A non-idempotent playbook used as a drift detector.219- ❌ Redistributing virtual-lab NOS images without checking their licence.220221## 8. Mandatory web verification222223**Methodology**: versions from the **PyPI JSON API** and the **GitHub Atom release feeds**224(`api.github.com` returns 403 unauthenticated); **licences read from the raw `LICENSE` file**, not225from the label GitHub's interface shows nor from a summary.226227**Verified Aug 2026 (version | date | licence checked raw)**. **NetBox 4.6.7** |22830 Jul 2026 | **Apache-2.0** (`LICENSE.txt` on `main` and `master`; plain `LICENSE` **returns 404** —229the path matters when verifying), with `v4.6.8-rc2` in flight (4 Aug 2026). **Nornir 3.6.0** |2302 Aug 2026 | **Apache-2.0** (slow but alive cadence: 3.5.0 was from Jan 2025). **NAPALM 5.2.0** |23127 Jul 2026 | **Apache-2.0**. **netmiko 4.7.0** | 12 May 2026 | **MIT**. **Jinja2 3.1.6** |232**BSD-3-Clause**. **ansible-core 2.21.2** | **GPL-3.0-or-later**. **pygnmi 0.8.15** |233**BSD-3-Clause**. **gnmic 0.46.0** | 14 May 2026 | **Apache-2.0**, and it **lives under the234`openconfig` organisation**, not under the original personal repository: the project **moved**, and235searching for the old repo gives a false impression of abandonment. **containerlab v0.77.0** |23628 Jun 2026 | **BSD-3-Clause with Nokia copyright** (read in the raw `LICENSE`): **it is not237Apache-2.0**, which is the usual assumption. **netlab** = PyPI package **`networklab` 26.7** |238**MIT** (ipspace and contributors), CalVer, and **the package name does not match the project name**.239240**Declared discrepancy — scrapli**: **GitHub and PyPI give different versions for the same thing.**241The Atom feed shows `v2.0.0-rc.16` (20 Jul 2026); PyPI publishes that artifact as `2026.7.20rc16` and242declares **`2026.2.20`** (Feb 2026) as the current version, because under PEP 440 the CalVer sorts243above `2.0.0rc`. **Operational conclusion**: the 2.0 rewrite is at *release candidate*, a244`pip install scrapli` without `--pre` installs the Feb 2026 CalVer line, and **pinning scrapli 2.0 as245the default today is premature**.246247**State of OpenConfig/gNMI (qualitative)**: active project (gnmic, ygot, ygnmi, gNOI and248`featureprofiles` with activity in 2026). The models are written in **YANG 1.0** and are249"operationally complete", not exhaustive; **real interoperability between vendors is still250imperfect**, and gNMI is **weaker than NETCONF for transactional configuration** (although the251specification requires that a `SetRequest` spanning several *origins* be treated as a transaction252with rollback). **No authoritative adoption-share figure was found**: treat any number about253OpenConfig or gNMI adoption as a **Declared gap**.254255**RFCs verified one by one against `rfc-editor.org`**: NETCONF **RFC 6241** (Jun 2011, updated by2567803 and 8526); RESTCONF **RFC 8040** (Feb 2017, updated by 8527); YANG 1.1 **RFC 7950** (Aug 2016,257updated by 8342 and 8526); NMDA **RFC 8342** (Mar 2018); YANG Library **RFC 8525** (Mar 2019,258obsoletes RFC 7895).259260**Declared gaps — do NOT fill from memory**:2611. **Versions and status of the Ansible network collections** (`ansible.netcommon`, `cisco.ios`,262 `arista.eos`, `junipernetworks.junos`, `nokia.srlinux`): **not verified**; their cadence is263 independent of `ansible-core`.2642. **NAPALM's real coverage per platform and NOS version**: **not verified**, and it is what decides265 whether NAPALM is usable in your estate.2663. **Status, version and licence of Batfish and of pyATS/Genie**: **not verified**; they are mentioned267 as a category, **not as a pinned default**.2684. **Status and licence of GNS3, EVE-NG and `vrnetlab`**: **not verified**. EVE-NG has commercial269 editions: check the specific edition.2705. **Licence and redistribution of lab NOS images**: **vendor-specific and not verified**. It is a271 legal question (`opensource-licensing-standards`).2726. **gNMI and NETCONF support per platform and NOS version**, and the concrete models exposed: **not273 verified**. Check it with `Capabilities` on the real device.2747. **The relationship between NetBox Community (Apache-2.0) and NetBox Labs' commercial offerings**275 and what functionality falls outside the open edition: **not verified**.2768. **`ncclient`, `pyang`, TextFSM/`ntc-templates` and TTP**: version, maintenance and licence **not277 verified**.278279If the web contradicts this document, **the web wins** — flag the discrepancy.