[Plopix]

Every skill was checked against the Crystallize source, and what was wrong is fixed. One commit per skill, plus a few follow-ups where one skill's fix changed what another should say, so it reads best commit by commit.

Why

The skills ship to everyone who installs the plugin, and agents act on them without checking. They had drifted: many examples could not run at all, and some behaviour was described backwards.

What changed

Skill	Main corrections
agent-tenants	Provisioning has its own rate limit and can answer 503. Invites work once. Promoting a human claims the tenant
bookable-resources	A lost booking race is refused, not silently orphaned. Availability and hydrate pairing. New /booking/admin reference
content-json	Text is valid on formatting and block nodes. Image src in metadata renders nowhere
content-model	SKILL.md from 1,046 to under 500 lines. Component defaults, limits and choice config. Upsert replaces whole definitions
data-creation	No longer claims bulk import. Maps its payload onto the API input
information-architecture	The folder example could not run. How paths, shortcuts, ordering and redirects behave
js-api-client	Bearer tokens work on Core and PIM only. Pipelines take ids. shopBookingApi cannot reach the booking admin API
mass-operations	Runs act as the task creator, not as tenant admin. Examples could not run. Fields the runner silently drops
mutation	Removed cart mutations that do not exist. What deleting, unpublishing and publishing do
payments	SKILL.md from 641 to under 500 lines. Helpers built invalid mutations. Crystallize does not retry webhooks. Shop writes can undo stage moves
permissions	Replaced invented role GraphQL with the real API. Tokens belong to users
plugins	Placements that never rendered, install flags, payload claims, where the Backend Token works
pricing	The lowest price wins, not the most specific list. Cart, Discovery and mass operation examples
query	Most Core examples failed validation. Discovery auth and paging, Shop order paging
responsive-images	Variant formats depend on the source. Discovery's focalPoint. The Image component's real behaviour
subscriptions	Crystallize renews contracts on schedule and never bills, the reverse of the old guidance. Shop and Core sync
taxonomy	Removed a component and a filter that do not exist. How to ship a topic tree
vector-ranking	The rerank window is the first N by sorting, not the top matches. Taste changes need a publish, not an ignition

Also adds five new references (crystallize-payments-lib, booking-admin, shape-consolidation, mass-operation-example, role-api) with eval cases for them. No version bump: releases are a separate workflow.

Review these first

These reverse what the skills used to say, so they matter most if they are wrong:

1. subscriptions: a contract renews by itself near renewAt, and calling renewSubscriptionContract on top moves it a second period. The old text told integrators to renew it themselves. It was verified live, but a scheduled renewal would not show up in a short test.
2. pricing: the lowest price wins across price lists.
3. mutation and payments: a stage moved in Core or the App does not reach the Shop order, and the next Shop write puts it back.
4. mass-operations: the runner acts as the user who created the task.
5. vector-ranking: how the rerank window is chosen.

How it was checked

• Each claim was checked by reading the code, not by running it. Nothing ran against a live tenant. Where a skill's own live measurements could not be settled from the code, they stay as written.
• crystallize-payments-lib.md type-checks against @crystallize/js-api-client 7.6.0.
• The mass-operation validator confirmed that it refuses price tiers and bare variant price and stock.
• All 300 internal links resolve, every SKILL.md is under 500 lines, and every description is within 1,024 characters.
• tsc, oxlint and the 211 tests pass. A /code-review pass found ten issues, all resolved in the last three commits; the one that cannot be fixed in code (syncStages writes pipelines one at a time) is documented.

Not settled

• The new eval cases have not been scored: the live eval needs AI Gateway access. All expected references resolve.
• data-creation: the json:itemData grid format itself could not be checked, so those claims stand as written.
• pricing: "customer groups are a dead end" stays as written. The code shows no cause, so it needs a live test.
• Provider references in payments: the Crystallize side was checked, the providers' own APIs were not.

The rest of the unsettled claims were reported in the internal review notes.