DSC Doc Scraping

Inputs

• url (required) -- a DSC URL: either the top-level /docs/apis catalog, a product-area landing /docs/<product>/<area>/references, a specific reference root /docs/<product>/<area>/references/<name>, or a single slug .../references/<name>?meta=<slug>.
• out (required) -- an absolute path to the output root directory.

If the user's request is vague ("pick one from Commerce Cloud", "scrape any endpoint from Data Cloud"), bootstrap from the top-level catalog: scrape https://developer.salesforce.com/docs/apis first, read the resulting <out>/_catalog.json, pick the product whose title matches the user's request, and follow its referenceUrl into the area-landing scrape. Each product in the catalog has title, body (short description), referenceUrl, referenceShape, and searchKeys (an array of acronyms and synonyms drawn from the product's landing page plus a hand-curated map for community jargon like "SCAPI"). Match the user's hint case-insensitive substring against title, body, and searchKeys together – the third surface is what makes acronyms like "OCI" or "FSC" resolve without depending on the model's training data. The referenceShape tells you whether the reference target is scrapeable (area-landing, reference-root) or out-of-scope (atlas, static-html).

Catalog product names drift from what a user might say. Salesforce has rebranded "Data Cloud" to Data 360, for example, so matching the literal string "Data Cloud" against catalog titles finds nothing; match on topic keywords. A handful of products (notably Marketing Cloud Growth / Marketing Cloud Next, and Agentforce) don't surface in /docs/apis at all but do have /references/ pages reachable by direct URL. After a catalog miss, lowercase the user's hint and substring-match it against the keys in lib/scrape/aliases.js (the CATALOG_MISSING_ALIASES map) before asking the user for a URL. Don't substitute a related-looking product when the catalog and alias map both miss – e.g. for "Agentforce" the catalog returns no match and the alias map is the canonical next step, not Einstein Gen AI or any other product whose name overlaps semantically. If both miss, ask the user for a URL.

Watch out for Summary-page redirects. On reference roots, DSC auto-redirects bare /references/<name> URLs to ?meta=Summary -- a valid slug, but it's the reference's overview prose, not an operation. If the user asked for "an endpoint" (or said something that implies one, like "a POST call" or "a specific operation"), scrape the reference root first; the output includes a per-reference _index.json with the full slug list, and you can pick a verb-shaped slug (names starting with get, create, update, delete, search, list, etc.) for the user's follow-on question.

URL shapes

Flow

Single call -- run this skill's scripts/scrape.js with Node. Both <url> and <out> are positional arguments; there is no --cache-root / --cache / --out flag. The only flags accepted are --all and --force. The out-root must be a concrete shell-expanded path (e.g. /Users/me/.cache/dsc-scrape), not a literal ~/... string – tilde expansion is the caller's job:

node ~/.claude/skills/dsc-scrape/scripts/scrape.js \
  'https://developer.salesforce.com/docs/commerce/commerce-api/references/orders' \
  ~/.cache/dsc-scrape

The script is a thin wrapper around the shared scrape library at lib/scrape/ (reached via the lib -> ../_shared symlink); js-yaml ships there. If invoked with a misparsed flag (--cache-root, --out, etc.) or a literal ~ as the second positional, the script exits with code 2 and a usage hint instead of silently creating a stray directory.

The script classifies the URL, fetches the reference page HTML to extract the refList (from reference-set-config or ReDoc's reference-config), fetches the static spec file (OAS 3 YAML for rest-oa3, AMF JSON sidecar for rest-raml, Swagger 2 JSON or YAML for rest-oa2), parses it, and writes one JSON file per slug. No browser, no external tools required.

For successful runs, the script prints a JSON summary to stdout listing slugsWritten, format, specUrl, files[], and refreshed (boolean). refreshed: false means the cached reference was still fresh and no fetch happened (fast path); refreshed: true means new bytes were fetched and written. Relay the file count and (for reference-root scrapes) the path to _index.json back to the user. If refreshed: false, mention that the cache was already fresh.

Freshness

The script uses a 1-hour TTL that matches DSC's own cache-control: max-age=3600 on spec files. On each scrape:

• Read _index.json.scrapedAt from the output directory.
• If it's less than 1 hour old, skip the fetch entirely -- no network, no parse, no disk writes. Result: refreshed: false.
• Otherwise fetch, parse, and overwrite _index.json + every slug file. Result: refreshed: true.

Overrides:

• --force CLI flag ignores the TTL and always re-fetches.
• DSC_CACHE_TTL_MS=<ms> env var replaces the default (set to 0 to always refresh).

Single-slug scrapes (?meta=<slug>) bypass the TTL -- the caller explicitly wants that file on disk, possibly rewritten.

Vague-URL recipe

When the user asks for "an endpoint" or a product without naming a specific reference ("scrape any endpoint from Commerce Cloud", "pick a SCAPI call"):

1. Scrape /docs/apis -- writes _catalog.json with every top-level product and its referenceUrl.
2. Pick the product whose title matches the user's request. Skip any whose referenceShape is atlas or static-html -- those aren't scrapeable.
3. Scrape the product's referenceUrl (an area-landing). This writes _landing/<area>.json listing every reference in that area with its id, title, and referenceType (rest-oa3 / rest-raml are scrapeable; markdown isn't).
4. Scrape a specific reference root (/references/<name>) from that list. This writes every slug -- Summary, endpoints, types, and _index.json.
5. Read _index.json, pick a verb-shaped slug from its slugs list, and point the user at the corresponding file on disk.

The catalog scrape and area-landing scrape are list-only by design -- they don't eat spec bandwidth for refs the user doesn't want. TTL is the same 1-hour freshness window used for references, so once you've bootstrapped a product area, follow-on scrapes are free cache hits.

Only re-scrape a single slug to a different output path if the user was specific about where one endpoint JSON should land.

Output layout

<out>/
├── _catalog.json                    top-level /docs/apis product index (only if scraped)
├── _landing/
│   ├── <product>_<area>.json        product-area landing (list of refs in the area)
│   └── <slug-of-html-path>.json     ReDoc / non-slug landing URLs
└── <area>/                          e.g. commerce_commerce-api, revenue_subscription-management
    └── <reference>/
        ├── _index.json              reference-wide metadata: title, source, full slug list, siblings
        ├── Summary.json             reference overview
        ├── <operationId>.json       one file per endpoint
        └── types/
            └── <TypeName>.json      one file per named type

Each per-slug JSON has a unified envelope -- kind (endpoint/type/summary), reference, slug, url, scrapedAt, source.{format, specUrl} -- followed by an endpoint / type / summary payload. OAS, AMF, and Swagger 2 sources produce identical envelope shape; consumers don't branch on format. Swagger 2 specs have their $ref paths normalized from #/definitions/X to #/components/schemas/X so type lookups by ref work the same way OAS 3 does.

_catalog.json has products[] with title, body, overviewUrl, guidesUrl, referenceUrl, and referenceShape (area-landing / reference-root / atlas / static-html / unknown). Only area-landing and reference-root shapes are scrapeable.

_landing/<product>_<area>.json has references[] with id, title, referenceType (rest-oa3 / rest-raml / rest-oa2 are scrapeable; markdown is not), source, and href.

Reading the output

Downstream questions like "which scopes do I need for getProducts?" or "what query params does searchOrders take?" are answered directly from the scraped JSON -- no need to re-scrape or visit the site. The endpoint payload has the same shape for both OAS and AMF sources. Fields worth knowing when fielding such questions:

• endpoint.method / endpoint.path / endpoint.url -- HTTP verb, templated path, full URL with the spec's base server prepended.
• endpoint.operationId -- may be null for specs that don't set it. In those cases the slug is synthesized as {method}-{path-with-slashes-and-braces-stripped}, so the file on disk still has a predictable name.
• endpoint.parameters[] -- path, query, and header params. Each has name, in (path / query / header), required, schema, description.
• endpoint.body -- request body shape, if any. Either schemaRef (#/components/schemas/...) or inline schema, with optional examples.
• endpoint.responses[] -- one entry per status code with code, description, and schemaRef or schema.
• endpoint.security[] -- auth requirements. Each entry is {scheme, scopes[]}. This is where OAuth scope questions live.

OCAPI specifics: endpoint.operationId may be human prose like "Get multiple products". The slug used as the on-disk filename comes from a fallback <method>-<path> derivation (e.g. get-products-ids) for that reason -- search by method/path or endpoint.operationId, not by slug, when an OCAPI question references a verb-shaped name. OCAPI URLs in endpoint.url carry {host} and {siteId} placeholders rather than concrete hostnames; that mirrors the spec, where the runtime host is the customer's instance.

Type references resolve by path: endpoint.body.schemaRef = "#/components/schemas/Product" -> read <area>/<reference>/types/Product.json for the full type shape. Type files carry type.schema with the same structure OAS/RAML produces.

Citing in the answer

Every customer-facing answer this skill composes from the scraped JSON cites public developer.salesforce.com URLs. Never cite the local cache path (~/.cache/dsc-scrape/...) and never cite the skill's own files (~/.claude/skills/...) – those are skill internals; engineers forward these answers to customers. The URLs are already in the JSON: _landing/<product>_<area>.json carries the area url, each per-slug file carries a url field, and _index.json per reference carries source.specUrl. When the answer is a list of references rather than a single endpoint – e.g. "which references does Marketing Cloud Growth expose?" – cite each reference's URL inline alongside its name. Listing reference names without URLs is not a citation. A discovery-style summary still has to be forwardable.

Even when the answer is a negative finding – "the queried product doesn't expose an endpoint for X" – cite the queried product's references-area URL so the user can verify and explore. Pivoting to related products that do expose X is fine and often helpful; pivoting away from the queried product's URL is not. An answer that concludes "MCG has no event-ingestion API, here's where to look in Data Cloud and Pub/Sub" must still cite the MCG references-area URL alongside the alternatives. The citation contract is "cite the queried product, then optionally cite alternatives," not "cite some developer.salesforce.com URL."

Scope

• In scope: DSC /references/ pages that expose either reference-set-config (OAS 3 YAML, RAML, or Swagger 2 / OCAPI) or a <doc-redoc-reference> element (ReDoc-rendered OAS 3), plus the top-level /docs/apis catalog. All render via static HTML the skill fetches directly.
• Atlas books (URLs contain atlas. and end in .htm, with no /references/ segment). Decline -- different viewer. Example: the core Platform REST API guide is an atlas book. The catalog parser flags these as referenceShape: "atlas" so callers can filter them before attempting to scrape.
• Static HTML references (legacy, e.g. Pardot/Account Engagement's guide/version3.html). Decline for the same reason. Flagged as referenceShape: "static-html" in the catalog.
• MuleSoft docs at docs.mulesoft.com are a separate platform. Decline.
• Guides, concept pages, release notes: Decline. This skill is for reference pages.

The script's classifier catches these and exits with a decline message. If the user hands you a clearly-out-of-scope URL, call the script anyway and relay its decline -- don't second-guess the classifier.

Error handling

• Decline (classifier) -> exit 3 with a message. No file written. Relay the reason.
• Unreachable spec (catalog lists a reference but the static file is 404 -- e.g. conversation-service-api is a known stale catalog entry) -> exit 1 with the URL and status. In --all mode, the error is recorded per-reference and the loop continues.
• Malformed spec -> exit 1 with the parse error.

Never retry. Surface the error to the user.

Key invariants

• One slug -> one file. The per-reference _index.json is the only file that carries the full slug list and sibling list. Don't duplicate that data into individual slug files.
• Type slugs (type:<Name>) write to <area>/<reference>/types/<Name>.json. Other slugs write to <area>/<reference>/<slug>.json. The slug field in the JSON keeps the type: prefix; the filesystem layout is purely a disk concern.
• Endpoint slug = the spec's operationId. When a spec has no operationId for an operation (e.g. Data 360 Connect), a fallback <method>-<path-with-slashes-and-braces-stripped> is synthesized -- get-ssot-activation-targets, post-ssot-activations-activationId-actions-publish. In both cases the slug is filesystem-safe, so slug and filename (minus .json) are the same string.
• rest-oa3, rest-raml, and rest-oa2 dispatch to different parsers but produce identical-shape output. Consumers don't need to branch on source.format.

See also

• README.md (in this skill dir) -- design, internals, test instructions.