Runs a device software upgrade on whichever vendor/platform the caller is targeting, following that
platform's own documented procedure, safety mechanisms, and pass/fail criteria exactly as published —
never an improvised or by-analogy sequence.
When to use this skill
- A network device or stack needs an OS/firmware version upgrade and a documented, deterministic
procedure for the target platform exists under
references/. - Not for initial device bring-up, or any vendor/platform-family/mode not yet listed in
references/INDEX.md — see that file before assuming coverage.
Universal concepts (the shape every platform's procedure follows, not a wire standard)
There is no single IETF/wire spec grounding this the way RFC 1035 grounds DNS — but every platform
procedure under this skill follows the same underlying shape, worth stating once instead of per vendor:
- Stage the new image before touching the running state. Where a platform separates image
transfer from activation, the transfer step does not itself change what's running. Some platforms
(e.g. Prisma SD-WAN ION) can chain download and install in one scheduled request; there, a
download-only request is the staging step, and the request that includes install is the gated one.
- Activation is the risky step, and some platforms give it a reversible window — not all do. Where a
platform provides a safety net (an auto-abort timer, a manual rollback command), never treat that
window as equivalent to explicit confirmation — they solve different problems (the timer protects
against an unreachable device, not against proceeding without sign-off). Where a platform has no
documented safety net (e.g. Arista's standard
reload/install ... reload), treat every activation
step as needing the same operator confirmation discipline a timer would otherwise partially provide —
don't assume one exists just because other platforms in this skill have one. - Where a platform has a separate, later "commit" (or equivalent) step, that step is what makes the
change durable. Until then, reverting to the prior version is the platform's own built-in path, not a
manual reconstruction. Treat it as a required confirmation gate, not just on platforms that call it
"commit" by name. Some platforms have no such step (e.g. Prisma SD-WAN ION upgrades run to completion
once scheduled): there, the confirmation gate sits before the step that schedules or starts the install.
- Where a platform documents an idempotency/already-on-this-version check, run it first rather than
discovering the no-op by re-running the whole procedure. Not every platform documents one explicitly
(e.g. none of Cumulus Linux's three methods state an idempotency check) — where it's absent, confirm
the current version before starting rather than assuming the platform will short-circuit a redundant
upgrade on its own.
See references/INDEX.md for which platforms have a procedure, and each platform file for how it
implements this shape concretely (exact commands/calls, exact state names, exact error conditions).
Procedure
- Identify vendor + platform family + mode — from the caller's request or arguments. Check
references/INDEX.md; if no matching row exists, stop: improvising a generic upgrade sequence for an
undocumented platform is out of scope for this skill (get the vendor's own procedure documented as a
new reference file first). - Open the matching
references/<vendor>/<platform-family>.md and follow it exactly — numbered
steps, safety mechanisms, error handling, patterns, known limitations, tools, and verification
checklist all live there. - Report the outcome using that platform file's own status vocabulary (e.g.
SUCCESS,
IDEMPOTENT, PARTIAL) — don't normalize it away.
Known limitations
- Platform coverage is exactly what's listed in
references/INDEX.md — currently Cisco IOS-XE INSTALL
mode, Cisco IOS XR classic RPM install mode, Arista EOS Standard upgrade/downgrade, Juniper Junos OS
EX Series Standard Installation, Extreme EXOS Standard upgrade, NVIDIA Cumulus Linux, and Prisma SD-WAN ION devices. A
platform/mode not listed needs a new reference file before this skill can be used for it. - Does not cover initial device bring-up or downgrade procedures beyond what a platform's own built-in
rollback/abort mechanism provides.
Contributing to this skill
SKILL.md only changes for two reasons: a bug in the universal-shape/router logic above, or adding a
pointer to a newly-documented vendor/platform-family. This category is expected to accumulate the most
vendor/platform entries of any skill in this repo — resist the temptation to inline any platform's
procedure here as it grows. Every platform-specific fact (exact commands, exact state names, a new error
condition, a version-specific quirk) belongs in that platform's own references/<vendor>/*.md.
Tools
Each platform procedure assumes its own execution mechanism (CLI/console session, REST API session) —
see references/INDEX.md for which, and that platform's file for the specific mechanisms.
Verification checklist
- [ ] Vendor + platform family + mode confirmed against
references/INDEX.md before proceeding - [ ] That platform's own procedure file was followed exactly, including its own confirmation gates and
error handling — not a generic/improvised sequence
- [ ] Outcome reported using the platform procedure's own status vocabulary