Rp Discovery

rp-discovery

Discover and document the source-platform schema for the active migration project.

Purpose

Use this skill to inspect the source system, identify entities, relationships, fields, identifiers, media, rich content, and platform-specific constraints. Examples include Shopify, WordPress, WooCommerce, and custom CMS platforms.

This skill owns the platform-agnostic discovery process and its output contract (source-profile.md + source-schema.json). Platform-specific details — how to capture a given source, its auth model, REST quirks — live in a dedicated source adapter skill, not here. For WordPress / WooCommerce, that adapter is rp-source-wordpress. To support a new platform, add a sibling adapter (e.g. rp-source-shopify) and leave this skill unchanged.

Inputs

Expected inputs may include:

• migrations/<project>/orchestration/run.json

• migrations/<project>/orchestration/decisions.json

• source site/app URL

• source platform name when it cannot be inferred

• source acquisition mode when the platform offers multiple read paths

• source API docs

• credentials, tokens, or local dump files when available

• export files when the run is file-based rather than URL-based

• current project under migrations/<project>/

• project-local config under migrations/<project>/config/

Config gate before capture

Before running source capture, verify the project-local config files created by wix-replatform:

• config/wix.env should always exist with WIX_SITE_STRATEGY, WIX_SITE_ID, and WIX_AUTH_TOKEN keys, even though discovery itself may not use Wix credentials yet.

• config/source.<platform>.env should exist once the source platform is known. A WordPress-only source uses config/source.wordpress.env; a source with confirmed WooCommerce evidence uses config/source.woocommerce.env and records sourcePlatform=woocommerce. Both use the rp-source-wordpress discovery adapter.

If a required non-sensitive key is missing or blank, ask the user for that value and fill the config file before continuing. If a required sensitive key is missing or blank, use the secure Secrets Manager flow: check for an existing non-placeholder local env value, then an existing non-placeholder Wix secret under the resolved site, and ask the user to update the dashboard secret only for keys missing from both places. Never ask the user to paste sensitive values into chat. Never print secret values back to the user; report only present/missing.

Use skills/wix-replatform/scripts/source-secrets.js resolve for sensitive source keys. Its output is limited to key names and statuses. It always returns postHydrationReport ({ exists, keys }: present/blank/placeholder/missing for each requested key in the local env file). Read that field to confirm a hydration succeeded; do not open the env file to check.

Treat migrations/<project>/config/*.env as secret-bearing once they may contain real values. Do not inspect them with whole-file reads that echo contents into tool output. Use secret-safe checks only: existence, required key names, and present / blank / missing status.

Site creation precedence

If this skill encounters conflicting Wix guidance about how to create a new site + headless destination, wix-replatform's migration contract wins.

• Route headless destination creation back to resources/rp-destination/, which scaffolds via npm create @wix/new@latest headless. This is the verified way to get a genuine headless site; the account-level Projects API is deprecated for this workflow (it produced non-headless sites).

• Discovery is source-side and does not create the site itself — just defer to that section.

Workflow

1. Confirm the active project under migrations/<project>/.

2. Start from the source URL when available and try to identify the source platform yourself before asking the user. Use lightweight signals such as a REST index, headers, HTML/application markers, or platform-specific route patterns. Only ask the user to name the platform if detection remains inconclusive.

3. Once the platform is inferred, resolve the acquisition mode before requesting source credentials when the platform has materially different read paths.

   • For Shopify URL-based migrations, ask whether to use the Admin API or only publicly available storefront data. Admin API records sourceMode=private_data and uses the rp-source-shopify adapter, read through the operator's Shopify CLI session; it collects no secrets. storefront keeps today's public route (sourceMode=public_content).

   • For WordPress / WooCommerce URL-based migrations, ask whether to import public content only or also include private/authenticated data.

   • Only the WordPress / WooCommerce also include private/authenticated data branch should trigger the secure Secrets Manager collection flow for sensitive credentials. The public content only branch proceeds without credentials and should be described as limited to public data. For WooCommerce, this branch should still probe public Store API catalog routes such as /wc/store/v1/products and /wc/store/v1/products/categories before declaring commerce out of scope. Treat file/export ingestion as a separate flow that starts from user-provided files instead of a site URL probe; do not offer exports as a third option in the URL-based acquisition-mode question.

4. Then select the matching source adapter skill (e.g. rp-source-wordpress for WordPress / WooCommerce, rp-source-csv when the run is file-based). If no adapter exists for the platform, capture entities manually following the same output contract.

   • Shopify private data (sourceMode=private_data, sourcePlatform=shopify) uses rp-source-shopify.

   • File-based runs (sourceMode=files_only, sourcePlatform=csv) use rp-source-csv regardless of which system produced the files; that adapter identifies the originating vendor from the header row. There is no acquisition-mode question and no credentials request for this path.

5. Run the adapter's capture step to produce a raw, machine-captured dump under <migrations-root>/<project>/data/<source>-discovery/. For WordPress, the capture script lives in rp-source-wordpress/scripts/ — run it from that skill directory (see rp-source-wordpress Capture section and CONVENTIONS.md). When using project-local secret-bearing config, prefer the script's deterministic --env-file path over shell sourcing. The adapter owns the capture mechanics, auth model, and platform quirks; this skill consumes its output. For long runs, pass --progress-log <path> and poll it per CONVENTIONS.md#progress-log-polling.

   • Distinguish supported entities (advertised by the source) from used entities (those with recordCount > 0). Entities advertised but empty should be flagged, not mapped as if they hold data.

   • A capture made without credentials is usually incomplete (gated entities, private fields, PII return 401/403). Do not treat an unauthenticated capture as authoritative — the adapter documents what auth a complete run requires.

   • For WordPress / WooCommerce captures, read data/wp-discovery/skipped-routes.json when present. Treat it as the canonical route-scope audit trail: skipped routes are evidence, not source entities, unless they were explicitly force-included by an audited override.

   • For WordPress / WooCommerce captures, also read data/wp-discovery/plugin-coverage.json and data/wp-discovery/plugin-inventory.json. These are the plugin-tier evidence: which plugins were detected, which are installed but unrecognized, which entities were derived generically, and what each capability's coverage status is. Do not re-derive plugin knowledge by reasoning about route names.

   • For Shopify private-data captures, run rp-source-shopify/scripts/shopify-discovery.js (see rp-source-shopify → Discovery). Exit 3 means a merchant question is pending: stop, and re-run after it is answered; recorded steps are not read again. Synthesize from data/shopify-discovery/entities/*.json only after it exits 0.

   • For CSV captures, the capture script lives in rp-source-csv/scripts/csv-discovery.js and takes the whole file set in one run (--file is repeatable) so roles and split files resolve together. Read data/csv-discovery/fileset.json — it is the canonical machine capture, and source-schema.json is synthesized from it:

     • carry sourceFiles[] (with role, vendor, partOf), vendor, dialect, drift, mappingHints, and csvInputRoot into sourceMeta, keeping file paths relative to csvInputRoot so the project stays movable;

     • give every entity an origin (file-rows | row-group | column-values) with the parameters that origin needs, and set hierarchical: true on nested derived entities so the mapper's faithfulness-ledger rule fires;

     • surface drift.unmappedColumns as unknowns so the mapper handles them explicitly;

     • honor halt: true. An ambiguous layout, an unknown file role, conflicting split-file headers, or a near-miss vendor detection is a question for the user, not something to resolve by picking the highest-scoring candidate. The warning text names the decision to put to them.

6. Capture field-level schema details, including type, cardinality, requiredness, and example values.

7. When bundled Wix domain knowledge recognizes a source route or source entity, annotate the discovered entity with sourceMeta.candidateTargetRefs[] such as ["stores/product"]. Discovery must still record source facts only; these refs are mapper hints, not target decisions.

8. Note operational constraints such as pagination, rate limits, auth model, and incremental sync options. If the source base URL or discovered media/file URLs use localhost, 127.0.0.1, or another private-only host, record a media reachability note in source-profile.md. Localhost is fine for discovery and local source reads, but Wix Media import is URL-based and Wix servers cannot fetch the user's localhost. This is an optional preparation step and, as far as we know today, only affects media import. State the two acceptable choices:

   • expose the source with a public HTTPS tunnel such as ngrok before live media import

   • skip/defer media import while continuing non-media entities

   Include concise ngrok setup instructions when relevant:

   Copy

   brew install ngrok
   ngrok config add-authtoken "<YOUR_AUTHTOKEN>"
   ngrok http 8090
   export WP_BASE_URL=https://<id>.ngrok-free.app

9. Synthesize the raw capture into the normalized artifacts below.

10. Promote the adapter's source-coverage record (see Source coverage below) as the last discovery write. Discovery is not promoted until this succeeds.

Artifacts to create or update

• migrations/<project>/discovery/run.json

• migrations/<project>/discovery/entities/

• migrations/<project>/discovery/warnings.json

• migrations/<project>/discovery/llm-handoff.json

• migrations/<project>/discovery/plugin-coverage.json

• migrations/<project>/discovery/review/plugin-coverage.md

• migrations/<project>/discovery/source-coverage.json

• migrations/<project>/orchestration/checkpoints.json

• migrations/<project>/data/<source>-discovery/: raw machine-captured output from the source adapter. Treated as evidence, not a hand-off artifact — downstream skills reference it for traceability but do not read it wholesale.

• migrations/<project>/source-profile.md: source platform, access method, limits, auth, and operational notes. Synthesized from the raw capture. Capture the operational facts the adapter documents (auth model, pagination, rate limits) so rp-import-codegen has them without re-deriving.

• migrations/<project>/source-schema.json: machine-readable schema for entities and fields. Synthesized from the raw capture — this and source-profile.md are the canonical hand-off to rp-mapper. Include traceability pointers so the mapper can drill into a specific entity's raw file when needed:

  • top-level rawDiscovery: relative path to the raw capture dir, e.g. data/wp-discovery/.

  • per-entity rawFile: file name within that dir, e.g. wp-v2--posts.md.

  • per-entity recordCount and inUse so consumers can distinguish supported vs. actually-used entities.

  • per-entity relations derived from the source-declared relationships in the raw capture, so relationships are evidence-backed rather than guessed. Each relation should carry an evidence pointer back to the source signal it came from.

  • For WordPress / WooCommerce, synthesize entities only from sampled backend_data and accepted backend_metadata route artifacts. Do not synthesize entities from routes listed in skipped-routes.json unless the skipped-route record has includedByOverride: true; in that case, include originalDiscoveryCategory, includedByOverride: true, and overrideReason when present in the entity sourceMeta.

  • Follow the adapter's source-schema.example.json for the shape (e.g. rp-source-wordpress/source-schema.example.json). It is a template to follow, not a strict schema to validate against — keep the platform-agnostic core stable and push platform quirks into each entity's open sourceMeta blob.

• Optional supporting notes under migrations/<project>/research/ if needed.

Plugin coverage (source adapters that support it)

When the source adapter produces plugin evidence — WordPress / WooCommerce does — promote it into the canonical artifacts rather than leaving it in the raw capture:

• Copy the adapter's coverage rows into migrations/<project>/discovery/plugin-coverage.json (one artifact: capability rows plus the per-plugin projection), and render a short user-facing discovery/review/plugin-coverage.md grouped by the four statuses — Migration planned · No need to migrate · Pending · Requires development. The statuses are already customer-readable; phrase the rest in user terms ("your events will come across as a CMS collection"), never internal jargon.

• For selected WordPress plugin storage without a complete supported REST reader, follow rp-source-wordpress/PLUGIN-READ-PLANS.md. A missing permanent profile does not end discovery. Reuse actual schema/storage evidence to prepare and validate a project-local plan. Preserve its canonical sourceEntityRef, bridgeDesignated: true, and exact sourceMeta.extractionPlan reference in source-schema.json. Unknown storage remains a typed entity deferral with producer remediation; do not invent a query or record.

• Emit plugin-derived entities into source-schema.json like any other entity, with sourceMeta.plugin, sourceMeta.capability, sourceMeta.channel, sourceMeta.recognized (a profile matched) and sourceMeta.candidateTargetRefs[].

• Derived (unrecognized) entities carry sourceMeta.origin: "generic-post-type" | "generic-taxonomy" and, when the source taxonomy is nested, "hierarchical": true so the mapper's mandatory hierarchy-lossiness rule fires.

• Entities read from a property embedded in a parent record carry sourceMeta.origin: "embedded" plus the parent route, so codegen extracts them from the parent fetch instead of issuing a second request.

• Record in source-profile.md: whether the plugin list was available, how many plugins were detected/unrecognized, and the counts by status. Downstream stages must not have to parse the raw capture to learn whether plugin coverage was complete.

Classifying an unrecognized plugin is a judgement, not a lookup. For every installed plugin with no profile, you — not the [js] layer — answer the only question that matters: is there anything here to migrate at all? The no-migration-needed list (plugins/no-migration-needed.json) is an input to that answer: a slug hit is strong evidence and not by itself the answer, because the list says nothing about what this particular installation is holding. Add your own reading of the registered types, taxonomies and data-shaped routes, and record the reasoning either way. Three exits, and only three:

• Nothing to move → No need to migrate, with basis: list (a list entry you confirmed) or basis: proposed (your own reading) and the rationale filled in — a list hit with no recorded confirmation is a failure.

• It holds data → propose a Wix capability; the row is Migration planned as proposed, decided at the mapping review. When the capability has no REST route, check whether wix-wp-plugin-v2 can reach it as a structure-bridge-plugin fulfillment before assuming a user-file/CSV request is the only path — see rp-source-wordpress/plugins/README.md's "Resolving a db-only capability" section. The bridge uses the same WordPress Application Password as other private reads.

• Cannot tell → leave the row Pending with reason: cannot-tell. An honest "cannot tell" is the point of Pending — it is what stops this step from becoming a silent way to drop data. Never resolve it yourself; the mapping review is the only exit.

"Cannot tell" means tried and still uncertain, not merely unexamined. Before landing a plugin on Pending, make the same best-effort reachability pass rp-source-wordpress/plugins/README.md's "Resolving a db-only capability" section describes for a confirmed capability — REST route, built-in export, public docs/source for its storage mechanism, wix-wp-plugin-v2 reachability — so the question put to the human at the mapping review is "here is what we found, decide the target" rather than "we didn't look, please figure this out." A live-verified negative result (e.g. every candidate table 404s, or the plugin's own source shows the data never leaves its vendor's cloud) is itself a valid, useful finding to record on the row — it is not the same as never having checked, and it changes what the human is actually being asked to decide. This does not license guessing a candidate target on your own judgement (see the "It holds data" exit above and rp-mapper's pending-row-verdict section) — it means the row that reaches the human is backed by an actual attempt, not a name and a shrug.

You may never conclude that Wix cannot do something — Requires development is reachable only from the human-signed register or from the human at the gate.

A recognized (profiled) plugin is not yours to re-decide: its rows come from the profile and the registers. That includes a No need to migrate row with basis: decision — a human-signed per-capability entry in no-migration-needed.json (capabilities[]), settled and carrying its decidedBy / decidedOn. Report it; do not re-open it as Pending.

• Collect every blocked[] entry across rows (a missing credential, a file only the user has, a changed surface — Blocked — recoverable, never a status) and ask the user once, as a single batched request with each item individually skippable. Record each answer in orchestration/decisions.json under pluginBlocker:<capability>:<kind> with value provided or declined — the discovery script reads these back so a declined ask renders as declined, not as unanswered. Never block the run on a skipped capability.

• Beside it, pluginCmsFallback:<capability> holds the merchant's answer to a CMS-fallback ask (recognized capabilities that a human signed as CMS-eligible). Discovery reads it but never writes it: it is asked at the mapping review and recorded only by orchestration-state.js decide-cms-fallback (see rp-mapper). Copy each coverage row's cmsFallback block into discovery/plugin-coverage.json unchanged, and list pending CMS-fallback rows under Pending in plugin-coverage.md with their featureLoss sentence.

Router reason cms_fallback_inputs_changed — reclassify, do not refetch

The router sends a run here with this reason when a CMS-fallback answer, the eligibility register, or anything that could resolve an in-scope capability natively (target knowledge, a detected plugin's current profile, the other signed registers) has changed since plugin-coverage.json was written. Rebuild coverage without any source request:

node skills/wix-replatform/resources/rp-source-wordpress/scripts/wp-discovery.js \
  --reclassify-coverage --out-dir migrations/<project>/data/wp-discovery

It reads the stored plugin-inventory.json, the prior coverage's entityStats and review evidence, and the current profiles, registers, target knowledge and decisions, then rewrites both data/wp-discovery/plugin-coverage.json and discovery/plugin-coverage.json (record counts, genericEntities, childSamples, notes and statusInvariant carried forward unchanged). Refresh discovery/review/plugin-coverage.md for the changed rows. If it exits 3 with outcome: needs-full-discovery (coverage written before entityStats existed, or a current profile route it never sampled), it wrote nothing: run normal discovery instead (one source fetch). Never hand-edit coverage rows to satisfy the router.

Source coverage (every adapter)

Anything the adapter advertised or attempted but did not emit as a source-schema.json entity must stay visible: a 401/403 entity, a route that failed capture, a data route skipped by rule, an unsupported route surface, a CSV file with an unknown role or a halted layout. Each adapter writes these as rows in data/<source>-discovery/source-coverage.json — always, with rows: [] when there is nothing to report. Frontend, theme, admin and transient route skips stay route-scope evidence only and get no row.

After source-schema.json is synthesized, promote the record (run from this skill directory):

node scripts/promote-source-coverage.js <migrations-root>/<project> --adapter wordpress   # or csv, shopify

It validates the adapter file, writes discovery/source-coverage.json, and stamps sourceCoverageContract: 1 on the project-root source-schema.json. It exits 2 and refuses when the adapter file is missing or malformed, when source-schema.json does not exist yet, or for an adapter it does not know; there is no unstamped path for a current adapter. Never hand-write or edit the coverage file or the stamp. Quick modes write their own record at plan.

The same command also promotes the adapter's source-locale.json (the store's payment currency, time zone, language and country; a value the source hides is an explicit unknown with its reason, never a guess) to discovery/source-locale.json, and refuses (exit 2) when that file is missing or invalid. Quick modes and the SFCC technique write discovery/source-locale.json directly. It feeds setup (the Wix site/locale entity), the execution plan and the import gate. To supply a missing value, put it in the adapter file as state: "supplied", re-run promote, and the regional section is shown again before anything is written.

Do not synthesize an entity for a coverage row. A row is resolved downstream: rp-mapper gives it exactly one disposition. Recoverable rows (credentials, a missing route, a force-include) join the single batched blocker ask under Plugin coverage, individually skippable; record the answer through rp-mapper/scripts/coverage-reconcile.js decide (never by hand), and re-run discovery for a provided answer. A skipped capability never blocks the run.

Output quality rules

• Separate confirmed facts from assumptions.

• Record per-entity volume (record counts) so downstream skills know what the site actually uses, not just what it supports.

• Preserve source-specific identifiers exactly.

• Include enough detail for downstream mapping and code generation.

• Flag unknowns explicitly instead of inventing structure.

Final discovered-set census

As the last standard-mode discovery step, after source-schema and plugin/source-coverage outputs are complete, run node scripts/discovered-census.js build <projectDir> from the toolkit root. The deterministic builder writes discovery/census.json from metadata only, including populated registered sub-parts, distinct plugin capabilities and unlinked plugins. Preserve unknown counts as unknown; do not invent entities or decisions for uncaptured source-coverage rows. This adds no discovery pause. Mapping reconciles each decision row to its plan or explicit gap before review. Quick mode keeps its existing path.

Public multilingual discovery

Before object discovery, every public-capable adapter determines whether the source topology is single, multi, unknown, or not-applicable. Write the adapter result with scripts/public-multilingual.js; it validates and produces discovery/public-source-languages.json and discovery/public-multilingual-content.ndjson. Only proven multi topology may produce content rows. Every additional locale must carry a value, explicit absence, or public-unavailable state for each row. Unknown or unreachable topology blocks only the multilingual sidecar; default discovery continues unchanged. Never choose the default locale here—discovery/source-locale.json remains authoritative.

Prefer the source platform's public locale/configuration endpoint when it explicitly enumerates locales. Treat locale URL prefixes, hreflang/alternate links, feeds, and repeated storefront responses as corroboration and reachability evidence; they cannot choose the default locale by themselves. Preserve the authoritative response reference and every corroborating route in the topology evidence. If those signals conflict, or an enumerated locale cannot be read through its public route, record unknown/unreachable rather than guessing from page text.