change-validation

name	change-validation
description	Snapshot a node's state before and after a change, then report exactly what changed and whether the fabric is still healthy
state	draft

Change validation

Lab 1b solution and the second composable skill. Use this skill to wrap any network change in a pre/post check so you can prove the change did what you intended and broke nothing else. It composes device-state-query: same snapshot shape before and after, then a diff. This is the skill form of the smoke-test.sh discipline.

Scope

Validates the effect of one change on one or more named nodes. It does NOT make the change; the operator (or another skill) applies the change between the two snapshots. Read-only on both ends.

Steps

1. Pre-snapshot. For each node in scope, capture state using the device-state-query procedure. Store as before.
2. Hand back. Tell the operator the pre-snapshot is captured and to apply their change now. Wait for confirmation that the change is applied.
3. Post-snapshot. Capture the same nodes again. Store as after.
4. Diff. Compare before and after field by field:
   • interfaces that changed oper state,
   • BGP sessions that changed state (established -> not, or the reverse),
   • new or removed neighbors.
5. Judge. The change is valid only when the intended difference appears AND no unintended regressions appear (no session that was established is now down, no interface that was up is now down, unless that was the stated intent).

Output

{
  "nodes": ["leaf1"],
  "intent": "<one line the operator gave>",
  "changed": [
    {"node": "leaf1", "field": "bgp[10.1.1.1].state", "before": "established", "after": "established"}
  ],
  "regressions": [],
  "valid": true,
  "summary": "Change applied. No regressions. 2/2 BGP sessions still established."
}

Failure modes

• Operator never confirms the change: do not fabricate a post-snapshot; stop and report valid: false, reason: "change not confirmed".
• Pre-snapshot itself unhealthy: warn loudly before the change; you cannot validate a change from a broken baseline. Suggest ./scripts/reset.sh.
• A regression appears: set valid: false, list every regression, and recommend rollback. Never report valid: true with a non-empty regressions.