Index of every page on this site: llms.txt
All skills
apollographql avatar

/graphos-factory

@9dda139 official
by Apollo GraphQLapollographql/skills134 stars
14
Run it yourself

Build and iterate on an Apollo Connectors subgraph for a GraphOS supergraph from a REST API, with or without an OpenAPI or Swagger spec, in a dedicated git workspace that records what the API offers, the user's operation and field selection, every design decision, and the evidence each verification layer produced. Use whenever the user wants to wrap a REST API as a GraphQL subgraph for their supergraph, add or remove operations or fields from an existing connector subgraph, refresh one against a new spec, or record live API traffic as test fixtures, even if they do not say "connector", "subgraph" or "Apollo".

Use this Skill: https://skilld.dev/gh/apollographql/skills/graphos-factory

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

graphos-factory-corereferencesschema-authoring.md

≈19k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Schema authoring

Each policy here is stated once, with its reason, so that thirty operations authored across three sessions come out looking like one service. When you depart from one, the departure goes in decisions.json as a tracked decision (graphos-factory-core decisions) — that is what makes it reviewable instead of drift. Applying a policy as written is not a decision: it is recorded only as a finding (graphos-factory-core findings add --cites …), and only when an instrument reads its omits or affects (workspace-contract.md rule 3).

Envelopes

Most list endpoints wrap their payload: { "widgets": [...], "page_info": {...} }. Two honest choices:

  1. Flatten when the envelope carries nothing but the list: the field returns [Widget_Co_Widget] and the selection descends into the key.
  2. Keep the envelope when it carries pagination or totals the caller needs: the field returns Widget_Co_WidgetList with widgets and nextCursor.

Pick one per connector and apply it everywhere. Mixed envelopes are the most common way a hand-authored connector starts feeling arbitrary.

Which key the payload sits under is a judgement, and it is written down. It lives in selection.yaml, per operation, never in inventory.json:

"get:/lists":
  include: true
  response:
    envelope: lists      # or null: the field returns the whole body
    confirmed: true

graphos-factory-core selection draft proposes one per included operation and marks it confirmed: false; the same run drafts one links: entry per candidate_entity_link fact, confirmed: false too (§ Cycles and depth). Confirm every block with the user before you apply. Until you do, reconcile names it as still the tool's draft and lint warns (response-envelope-unconfirmed); an operation with no block at all falls back to the suggestion and reconcile says which one it used (no-response-envelope).

The proposal comes from facts the inventory records about the response shape — root_property_count, array_root_properties, link_root_properties, sole_root_property, total_items_property, cursor_root_properties, root_is_array — and one rule: a root object earns an envelope when it has exactly one non-link property, or when exactly one non-link property is an array and a sibling names a total or carries a next cursor, or when that array has exactly one other non-link property (a status flag). Anything else is a resource that happens to embed a list, and flattening to that list would drop the rest of the payload. Read the facts, not the proposal, when you disagree with it — they are checkable against the document in seconds. Then write the answer you want into selection.yaml; never "correct" the inventory, which is built, never edited.

envelope must name a root property of the operation's response shape: reconcile and lint both refuse one that names nothing (unknown-envelope), because a typo would otherwise flatten the field to a type with no fields.

envelope names the wrapper key for the instruments; it does not decide flatten or keep. The instruments read it to find where the payload sits: reconcile's coverage delta (envelope_for), scaffold's stub bodies and the sparse-fieldsets check. Whether the field flattens to the list or keeps a wrapper type is the authoring choice above, made in the schema. pagerduty's get:/incidents records envelope: "incidents", and pagerduty_listIncidents still returns Pagerduty_IncidentList, which keeps limit, offset, more and total beside incidents. gitea's searchRepos does the same with envelope: "data" and Gitea_RepositorySearchResult { ok, data }. The same envelope value serves both choices.

A flattened list needs $.. When the field really returns [Type], write $.webhook_endpoints { ... }, not webhook_endpoints { ... }. A single key with a sub-selection keeps the key (mapping-language.md § The single-key rule), so the bare form selects { webhook_endpoints: [...] } into a list type and compose fails with SELECTED_FIELD_NOT_FOUND.

Pagination

Connectors have no pagination primitive, and synthesising a Relay Connection buys nothing at the Federation level. Mirror the REST shape:

widget_co_listWidgets(limit: Int, cursor: String): Widget_Co_WidgetList

with nextCursor: page_info.next_cursor in the selection. An inventory operation that declares a paging parameter carries a pagination block naming the request parameter, the size parameter and the response path detected for that operation; an operation with no block is not paginated; api.pagination summarises the whole API — the style most operations carry, with counts per style — and can disagree with an operation (PagerDuty is offset-paged at the API level, {cursor: 2, offset: 16, unknown: 2}, cursor-paged on its two audit endpoints). Read the operation's block; selection.yaml's per-operation pagination block is what actually gets exposed. Expose the cursor and the page size and nothing else — offset and cursor on the same field is a bug the caller will find at runtime. Graph-API paging (Meta: { data[], paging: { cursors: { before, after }, next } }) records response: paging.cursors.after and keeps the URL as next_url: paging.next: map nextCursor to the cursor, not to the URL.

A list that takes no paging parameter but returns a link to the next page is style: next_link. Salesforce SOQL's nextRecordsUrl and OData's @odata.nextLink are examples. The block records the link's path as next_url. When the document declares exactly one GET that fetches the linked page (the list's path plus one parameter ending in locator, cursor, token, page or next, and returning the same link), it records that GET as next_operation, e.g. Salesforce's get:/services/data/v67.0/query/{queryLocator}. Expose the link as nextUrl and the follow-up as its own field taking the locator. The caller pages by calling the follow-up, not by passing a cursor back to the list. Salesforce's link is a relative path whose last segment is the locator: from /services/data/v67.0/query/01g…-2000 the caller passes queryLocator: "01g…-2000", and the nextUrl doc comment says so. An absolute link with no next_operation, such as OData's @odata.nextLink, has nothing in the inventory to follow it: expose it as nextUrl only and record the missing follow-up with graphos-factory-core decisions.

When the inventory's size parameter carries default or maximum, the pagination argument's doc comment must state both, e.g. "Page size (provider default: 20, maximum: 100).". When the source declares no bound, the doc comment must say so explicitly, e.g. "(default unknown; no documented maximum).". Numeric page limits map to Int, never String. This is the page-size case of the general rule in § Argument constraints; the page-size wording above is what the pagination-bounds-* lint rules read, so keep it. When that default or maximum lies outside GraphQL Int (int64's 9223372036854775807 is how some specs write "no limit"), no Int argument can reach it: state no practical maximum (or no practical default) in place of the number, e.g. "Page size (provider default: 5; no practical maximum).". pagination-bounds-undocumented accepts the phrase for a bound outside Int only.

For page-indexed pagination (pagination.style: "page" in the inventory), the page-index argument's doc comment must state the offset formula: "0-based page index (records skipped = page_index * page_limit)." or the 1-based variant "(records skipped = (page_number - 1) * page_limit).". The doc comment must say: when the page size changes mid-iteration, restart at page index 0 (or 1 for 1-based)—continuing at the old index skips records. When the inventory declares a maximum inside Int on the size parameter, the description should note values above it are rejected, so an agent hitting a rejection knows the supported bound without probing.

Paginated list operation descriptions

The operation description must state the result is one page, not the complete collection, e.g. "Retrieves one page of widgets. Iterate through all pages to get the complete collection." The stop condition: iterate until an empty or short page; OR when the inventory's response facts name a completion signal (total_items_property, cursor_root_properties, or pagination.response cursor path), tell the caller to use it—e.g. "Use the total field to know when all records have been fetched, or iterate until a short or empty page." The restart-at-index-0 rule belongs on the page-index argument's doc comment, not repeated in the operation description.

The Returns: line that § Descriptions requires is appended after the one-page/stop-condition sentence above, never in place of it.

Sparse fieldsets

Some APIs return only the fields a GET names in a query parameter (Graph's fields=id,name,business{name}). When a selected GET's inventory operation has a string query parameter with the workspace's sparse name (workspace.yaml sparse_fieldsets: {param: fields}; fields when absent), the root field declares it and forwards it:

meta_ads_campaign(
  campaignId: ID!
  """The Graph API fields to return, comma-joined. Defaults to every field this type exposes; pass fewer to narrow the response."""
  fields: String = "daily_budget,id,name,objective,status"
): Meta_Ads_Campaign
  @connect(source: "meta_ads", http: { GET: "/{$args.campaignId}", queryParams: "fields: $args.fields" }, selection: "…")

The default is derived, not chosen. It is the sorted, comma-joined wire names (never the GraphQL renames) that the connector's own selection: reads from the entity the operation returns:

  • a node GET (the response is the entity object): its selected top-level fields.
  • a paged edge (a pagination fact, and data the only array root): the selected fields of the data[] item. The page wrapper (the data key itself, and every key the pagination fact reads, such as paging) is left out. Meta's own SDKs send exactly the item fields on an edge.
  • an expansion boundary (a property carrying x-expansion, a relationship to another node): the selected children render as a group, business{name}, recursively (business{primary_page{name}}), whenever they go beyond the boundary's verified default projection. Inside that projection (business { name } under a verified ["id","name"]) no group is needed. An unverified default always groups. An embedded value with no annotation is requested whole.

When the default cannot be derived, do not guess. That covers upstream aliases (name.as(x)), paging modifiers (campaigns.limit(10){name}), edges requested as expansions, a page selecting a key beside its items, a list response with no pagination fact ({data: [...]}, {results: [...], count}, {data: [...], meta: {...}}: any response whose suggested envelope names an array root; neither a node nor a recognised page), and a selected path with no single wire name. Scaffold's stubs for a list keep the wrapper and narrow its items. Record the literal instead, in a resolved decision that names the operation and carries the literal verbatim. Choosing the literal is a call — how many, which fields, inline or not — so it carries the alternatives you weighed:

graphos-factory-core decisions add . --title "Ad account reads ten campaign names inline" \
  --question "What fields default does get:/act_{ad_account_id} send?" \
  --choice 'inline:ten campaign names inline' --choice 'ids:campaign ids only' \
  --resolved --chosen inline --note "callers list an account's campaigns by name" \
  --decision 'Send `campaigns.limit(10){name},id` as the fields default for get:/act_{ad_account_id}.'

A literal a reference settles instead (the vendor documents the one projection that works) may be a current finding (findings add --cites …) naming the operation the same way; the rule reads both.

Both halves match exactly. The operation key must appear whole in the decision's text or as an affects entry: a decision about get:/act_{ad_account_id}/campaigns does not name get:/act_{ad_account_id}. The literal must be a whole backtick span in the resolution (`id`, not the bare word; a fenced block counts too), and an empty default is never excused.

A decision adds; it never subtracts. When the default is derivable, a recorded literal passes only if it requests every field of the derived list, compared as a tree. Order, whitespace and .modifier(…) suffixes do not count: id,campaigns.limit(5){name} keeps campaigns{name},id, and business{id,name,extra} keeps a derived bare business. A literal may add a paging modifier, a wider group or an edge the connector does not map. A literal that drops a mapped field (business for a derived business{name}) fails the rule, because the source would stop sending that field and the schema would resolve it null.

A fields parameter that means something else (the fields a search looks in, Google's items(id) partial-response syntax) is not a sparse fieldset. Turn the rule off with sparse_fieldsets: {enabled: false} in workspace.yaml and record why in decisions.json. Lint then skips the argument, and scaffold treats it as an ordinary argument.

lint's sparse-fieldsets rule (an error) checks four things: the argument exists, it is String, its default equals the derived list or a recorded literal that keeps it, and queryParams forwards $args.<param>. A literal fields: $("…") in queryParams fails the rule, because the caller could never narrow it. A non-string parameter gets an info finding and the rule does not apply. Lint validates the declared default and the forwarding. It cannot prove anything about a value a caller passes at run time; that narrowing is the caller's, and the argument's description says so.

Argument constraints

Any argument whose inventory entry carries default, minimum, maximum or enum states them in its doc comment. The inventory records those four keys for every parameter and request-body property, not only page-size ones, and an argument that drops them leaves the caller — usually another agent — to guess or to probe:

  "Sort order (default asc, one of asc|desc)."
  sort: String
  "Search radius in metres (default 25, min 1, max 100)."
  radius: Int

An array-typed request-body property also carries minItems/maxItems when the spec declares them; state them the same way: "Custodians to add (1 to 100)." A query parameter's array bounds are not recorded yet, so read those from the spec.

How to spell it, so two services read alike:

  • Cap an enum list at 8 values, then (+N more).
  • Keep an empty-string default rather than omitting it: q (default "").
  • Spell an enum-typed default bare, like the listed values: default asc, not default "asc".
  • An argument typed as a GraphQL enum states its values too. The consumer reads the argument before it follows the type reference, and one of closed|open|all on the argument saves that hop. The enum type stays the machine-readable list — wire-enum-drift checks it against the spec and checks nothing about the prose, so when the spec's value set moves, fix both.
  • An argument whose inventory entry carries none of the four gets no such sentence. Never invent a bound the source does not state; for page size specifically, say the gap out loud (§ Pagination).
  • Spell an exclusive bound exclusive min N / exclusive max N. OpenAPI 3.1 writes it as the number (exclusiveMaximum: 100); OpenAPI 3.0 and Swagger 2.0 write a boolean that makes minimum/maximum exclusive (maximum: 100, exclusiveMaximum: true). The inventory keeps the form the source wrote, so read the boolean with its maximum: both spell exclusive max 100, never max 100.
  • A default outside Int on an Int argument is spelled exactly (default: unbounded: the source's 9223372036854775807 does not fit Int; omit the argument to accept it), with the source's number. The router never sends a default, so the value reaches coercion only if a caller copies it from the prose. This spelling is the one doc comment that clears int-overflow, and it clears the default only. A minimum/maximum outside Int takes the same shape (max: unbounded: the source's … does not fit Int), but on any argument except the page size it is still an int-overflow error: retype the argument (§ Scalar choice).
  • A workspace written before crate 0.5.26 states the number, as this section then prescribed: (default 9223372036854775807, min 0). That is still an int-overflow error, because a caller can copy the number. The finding quotes the clause (default 9223372036854775807) and gives the replacement with the number filled in; replace only that clause and keep the rest (, min 0). Adding default: unbounded beside the old clause does not clear the finding: the number is still there to copy. The same holds for default: no upper bound, the wording some 0.5.1 workspaces carry: the finding quotes it as the clause to replace, and while it sits beside default: unbounded the finding says to delete it. When a doc comment holds more than one legacy clause, the finding names each one.
  • For a page-size argument use § Pagination's spelling ("Page size (provider default: 20, maximum: 100).", or the explicit-gap wording): that is the wording the pagination-bounds-* lint rules read, and this section does not restate it.

The omission sentence, a fifth trigger. An optional argument (no !) whose inventory description has a sentence saying what happens when the argument is omitted keeps that sentence verbatim in its doc comment. It is a source fact like any other: covered by the schema, or a recorded decision says why not (source coverage). A sentence is one when it contains, in any case:

  • a default: by default, defaults to, default to (will default to), default is, default value is;
  • a condition on absence: if or when, up to four words, then not passed|provided|specified|given|set|supplied|sent|present|included, or omitted, absent, unset, left empty|blank|out (if not passed, If this parameter is not provided, when unset);
  • if no … (up to three words) is|are passed|provided|specified|given|set|supplied|sent (If no card is given);
  • unless (otherwise) specified|provided|set|given, or omit (it|this) to.

missing is not a cue: "if the user is missing a role" is about something else. Venmo's payment_card_id ("ID of the payment card to use for the transaction. If not passed, Venmo balance will be used.") is the case: the first-sentence trim drops the second sentence, and nothing else in the schema can say that no card means paying from the balance.

  "If not passed, Venmo balance will be used."
  paymentCardId: Int
  "Sort order (default asc, one of asc|desc). Ascending when omitted."
  sort: String
  • It must be about this argument. A sentence that names another of the operation's parameters or body properties does not count, because it describes how two arguments interact. A name counts as named when it appears case-sensitively, [] dropped, with no letter, digit, _ or - directly before or after it. PagerDuty's since has "Defaults to 2 weeks before until if an until is given.", which names until, so it does not count. A sentence that names only the argument itself still counts. Gitea's order (ignored if "sort" is not specified) names sort, so it does not count. For a key nested in the request body, the keys beside it count as its siblings too.
  • Keep the first sentence that qualifies. A sentence ends at ., ! or ? followed by a space and something other than a lowercase letter (but not after e.g., i.e. or etc., and not inside parentheses or brackets: Defaults to 10 (max. 100 per page). is one sentence), at a blank line, and at a list item.
  • Where it goes: after the clause, as a sentence of its own, separated by one space from the clause's period. Keep the source's final punctuation, and add a period when there is none: (default asc, one of asc|desc). Ascending when omitted. With no clause the sentence is the whole doc comment, or it follows the opt-in source sentence in the same way. Never cap it at 150 characters and never paraphrase it.
  • Write it once, never twice. When the omission sentence is the source's first sentence and the workspace opted in to the source sentence, part (1) of § Argument descriptions already carries it ("include private repositories this user has access to (defaults to true)"), and nothing is added after the clause. When the workspace has not opted in, write the sentence as described above.
  • A defaults to sentence can repeat the clause's default. Keep both: the clause is the uniform spelling, and the sentence often says what no default key can (Defaults to current time.).
  • Page-size and page-index arguments keep § Pagination's wording and get nothing else. A required argument gets no omission sentence, because the caller cannot omit it.

The clause is the default-on half of an argument's doc comment; whether a source description goes in front of it — opt-in, one sentence — and what never goes in one at all (the argument's type or requiredness) is § Argument descriptions.

Never write = literal on an argument. An SDL default is executable — the router sends it when the caller omits the argument — and a consuming agent reads a published default as "the value to use": a default of 5 against a maximum of 20 produced four times more pagination calls than the API needed. The provider's default belongs in the doc comment, which the rule above guarantees is there. The one exception is a sparse-fieldsets argument (§ Sparse fieldsets): its default is the fields list the connector maps, not a provider default, and it must be declared.

What is checked mechanically: the page-size half of the rule above, by pagination-bounds-undocumented and pagination-bounds-unknown; and, for every other constrained argument, that it has a doc comment at all, by argument-constraints-undocumented (a warning), whose message spells the clause. Whether an existing doc comment states the clause, and the = literal ban everywhere, are applied by hand. The same rule also reads what an optional argument's doc comment says: when the omission sentence above is missing, whether or not the argument has a doc comment, the rule warns and quotes the sentence. The sentence counts as present when, without its final period, it appears in the doc comment after both are whitespace-normalised and compared case-insensitively. source-coverage --check also fails on the same sentence, and a resolved behaviour waiver silences both the check and this warning. Both halves trace an argument to its source entry through a queryParams key, an aligned path segment, a header whose value is the argument ({ name: "From", value: "{$args.from}" }), and a body key, flat or nested in object literals, $({ … }) included, through methods such as ->map(…)->first. An argument sub-selection ($args.input { … }) or a deeper path ($args.input.x) sends an input type's fields: their doc comments are not read, so write their omission sentences by hand. The only lint rule that reads an SDL default is sparse-fieldsets, on its own argument. int-overflow reads one clause, default: unbounded, on an Int argument whose source default lies outside Int, and quotes a default N clause stating that number, or a default: no upper bound clause, as the text to replace.

Nullability

Be conservative. A field is non-null (!) only when the API is documented non-null and you have seen it populated in a recorded sample or a live probe. A mapped field the payload lacks yields null with no error — a wrong ! turns that silent null into a runtime error for the whole response, and a spec that says non-null is not evidence: vendors ship summary: null against a required declaration routinely.

Path parameters are the exception: their arguments are always !, because the URI cannot be built without them.

Scalar choice

GraphQL Int is a signed 32-bit integer. A source that declares a numeric property or parameter format: int64, uint64 or uint32 is saying the value may not fit — uint32 is unsigned, so its upper half is past 2^31 too. Past 2^31 the router nulls the field and reports a coercion error rather than returning the number, and every offline layer stays green because the recorded fixtures carry small values. The int-overflow lint (testing.md) is an error on exactly this. Two fixes, and there is no third — no doc comment silences the rule, because a sentence beside a field does not change what the router coerces:

  • ID for an identifier — a row id, a foreign key, or a resource's own number used to address it. ID accepts an integer literal as input, so gitea_issue(index: 1) in a test document keeps working when index: Int! becomes index: ID!, and the value is interpolated into the URL verbatim. This is the right answer for an argument. An output field is a different judgement: ID is for something a consumer passes back, not for a number it does arithmetic on, and this router hands an ID-typed numeric property back as a bare JSON number rather than a string — so an output that happens to be the same value as an ID argument still follows the magnitude rule below. The gitea pilot is the worked example: gitea_issue(index: ID!) and Gitea_Issue.number: String are the same number in the two positions.

  • String mapped with path->match([null, null], [@, @->jsonStringify]) for everything else — a magnitude (bytes, kilobytes, seconds, money in minor units) and an unbounded counter alike.

    The null guard is not decoration. A bare path->jsonStringify turns a JSON null into the four-character string "null", which a nullable String field happily carries: the same silent wrong answer int-overflow exists to prevent. ->match([null, null], …) hands back null; $(path?->jsonStringify ?? null) does too, but reconcile does not read a $(…) expression block as a mapping of its path and reports every such leaf as unmapped, so prefer the ->match form. .github/scripts/null-mapping-probe.sh pins both halves against rover.

    The mapping must be aliased: size: size->match(…) composes, a bare size->match(…) does not (connectors-language.md).

    A converted field's doc comment says so, e.g. "Size in kilobytes. int64 in the API; String keeps the value exact and a null stays null." — otherwise the next reader sees a String holding digits and takes it for an oversight.

A declared minimum/maximum is the source's own word and beats the format hint: an int64 property bounded inside Int is not a finding, and a bound outside Int is one whatever the format says. This is why Numeric page limits map to Int, never String (§ Pagination) does not contradict this section — a pagination size parameter is a small number by construction and is exempt from int-overflow outright. An exclusiveMinimum/exclusiveMaximum counts as the bound it is, and a default outside Int is a finding too (§ Argument constraints has the one spelling that clears a default on an argument).

An ID argument mapped into a JSON body is not re-typed to a string. The router sends the literal the caller wrote: milestone: 1 reaches the body as the number 1, and serviceId: "PSVC001" as a string. This was measured at e2e on Router 2.17.0 (the gitea pilot's create_issue, the pagerduty pilot's create_incident). rover connector test sends every $args scalar as a string (the gitea pilot's .factory/memory.md), so a numeric id in a body is proven at e2e only. Check what type the source wants, and record the choice.

Every numeric predicate over a converted field has to be restated. In jq, .count >= 1 on a JSON string compares as a string and is true for every value the API will ever send — a green assertion proving nothing. Pipe it through tonumber, which fails loudly on anything that is not a number (the gitea pilot's tests/live.yaml).

Enums

Map a closed enum to a GraphQL enum when the values are valid GraphQL names and the set is genuinely closed. Otherwise use String:

  • x-extensible-enum (Zalando's convention) means an open enum. String is the correct mapping; a closed GraphQL enum silently rejects values the API will send.
  • Values that are not valid GraphQL names (leading digits, hyphens, dots) need a ->match translation in both directions. That is real complexity; take it only when the enum is load-bearing, and note it as a decision in decisions.json.
  • Keeping the vendor's own casing avoids the translation entirely when the values serialise straight back into request bodies.

graphos-factory-core lint checks the first sentence (closed-enum-as-string): a selected argument or response leaf typed String while the source's enum is all strings and every value is a valid GraphQL name is a warning until it is an enum in wire casing or the reason it is not is on record. Keeping String over a valid enum is a call (the enum is the alternative), so the reason is a resolved decision whose affects names the slot — Type.field for a leaf (Pagerduty_Reference.type), the prefixed root field with the argument in parentheses for an argument (gitea_listIssues(state), Query.gitea_listIssues(state), or several at once, gitea_listIssues(state, type)):

graphos-factory-core decisions add . --title "…" --question "…" \
  --choice 'string:keep String' --choice 'enum:an enum in wire casing' \
  --resolved --chosen string --decision "…" --affects "Pagerduty_Reference.type"

When the source itself settles it (the vendor documents the vocabulary as open), a current finding with the same --affects counts too (findings add --cites …). An open decision is a pending question and does not count. Two conditions keep the rule quiet where an enum would be wrong or empty, and neither needs a decision. A one-value enum is a constant discriminator (Pagerduty_Service.type is always service), and a one-member GraphQL enum documents nothing the field does not already say, so the rule wants two or more values. A field that several selected operations reach — Pagerduty_Reference.type, one type for ten reference slots, where the spec enumerates the vocabulary at only one of them — is judged on every spec property that reaches it and fires only when all of them declare such an enum, then proposes the union of their values; an enum built from one slot would fail coercion on the others. A reach the spec does not document — an operation with no response shape, or a parent object absent from it — counts as declaring none. The second sentence is not a finding either: a vocabulary with a leading digit or a hyphen (2–10, in-progress) is String by this rule, and a slot the connector maps with ->match is the mapping's to spell. The gitea pilot's Gitea_ObjectFormatName (sha1 | sha256, reached by two operations that both declare it) is the worked conversion; the pagerduty pilot's five type fields are the worked silence — one shared reference type and two one-value constants, no finding and no decision.

Cycles and depth

REST payloads embed each other; GraphQL types can reference each other freely, so a cycle in the shape is not itself a problem. Depth is: a max_depth of 6 in selection.yaml is the default because deeper nesting almost always means the API embedded an unrelated resource. At the limit, stop and expose the nested resource's id, with its own operation if one exists. Say so in the type's doc comment.

The inventory flags exactly this pattern for you: a property whose name and type family match the trailing path parameter of a canonical GET-by-id operation whose response is one record carrying that key carries a candidate_entity_link fact, wherever it sits in the response shape — a top-level property, a list item ([]>album_id), or a nested object or list (songs[]>album_id, owner>account_id). graphos-factory-core inventory links . prints every fact flat, with the operations that return the host shape and whether the by-id operation is selected, then the GET-by-id operations it refused as targets and why — a read keyed by an email that returns a balance (AppWorld splitwise) is one: it is not the record the email names, so no property gets a hint for it; inventory list shows the same as links: lines under each operation. It is a hint, never a finding. Route it through the selection, not the schema: graphos-factory-core selection draft . writes one links: entry per fact with confirmed: false; agree each with the user; a confirmed one is applied as a relationship field on the host type — graphos-factory-core links apply . --dry-run prints the field-level @connect keyed by $this for you to paste (connectors-language.md § Relationship fields); no @key and no decision are involved. Decline one with include: false and a reason on its entry; a service-wide "no relationship fields" is one decision in decisions.json, with wiring them as its alternative (graphos-factory-core decisions add . --question … --choice … --resolved …), the way gitea's D-0013 did it. The entity form (@key, a type-level or $batch connector) is the exception for cross-subgraph references and batching, and that one does need a decision.

The opaque JSON policy

Untyped beats unreachable, but typed beats untyped.

Emit Widget_Co_JSON only when the shape cannot be typed honestly:

  • polymorphism the connector cannot type. At connect/v0.3 that is every oneOf/anyOf. At v0.4 a oneOf of objects is a union, or an interface when its members share fields (mapping-language.md § Abstract types). It stays JSON only when:
    • no wire value tells the members apart;
    • a member is not an object;
    • the value sits under another type's field, and no key can re-fetch it through a connector of its own;
  • a genuinely free-form object — type: object with a description and no properties is a real and common shape meaning "arbitrary JSON here", not a defective spec;
  • additionalProperties-only maps.

Every JSON-scalar field carries a doc comment saying why: a reader of the schema, and any policy your supergraph attaches to @tag, learns nothing about a field typed JSON except from that comment, and a field nobody can describe is a field nobody can use with confidence. graphos-factory-core lint fails on an undocumented one under the default opaque_json_policy: forbid; allow_with_reason relaxes it, and typed fields are still the rule.

The doc comment is for a human; graphos-factory-core spans json-accounting . [--check] is the machine check. It enumerates every field the rendered SDL still types as the scalar — including a field nested inside another object type and one wrapped in a list at any depth ([Widget_Co_JSON], [Widget_Co_JSON!]!, …) — and matches each, keyed "<TypeName>.<fieldName>", against a json_reasons entry on a resolved .factory/decisions.json record (graphos-factory-core decisions add --json-reason 'Type.field|reason'; never a selection.yaml field or a schema doc comment — the decisions-only rule, because reasons recorded in the schema confused the model reading it), for a reason from a closed vocabulary: free-form-object, recursive, vendor-undocumented, polymorphic-without-discriminator, depth-cap. defaults.fields: all decides which fields are selected; it never counts as a reason for one that is still typed JSON.

Reasons are re-checked every run, not recorded once and trusted forever. Each is a predicate over the current .factory/inventory.json shape at the field's own path (the type's bare shape name, then the property matching the field name or its snake_case wire form) — recursive means the $ref chain loops back to itself; free-form-object means type: object with no declared properties; vendor-undocumented means no shape is recorded there at all; polymorphic-without- discriminator means a oneOf/anyOf with no discriminator; depth-cap means the workspace cut the field at its stated depth: its type sits at defaults.max_depth on some path from a root field (6 when unset), or the shape is a recursion, and the inventory has a structured shape beyond the cut. A holding depth-cap outranks recoverable, because re-typing the field would undo the cut the reason states. A field with no matching entry, or an entry whose reason is outside the closed list, is unaccounted; a recorded reason whose predicate no longer holds is stale — the field moved, the recorded claim did not, usually because the vendor now documents something different. A field the inventory has since made fully typed is recoverable regardless of what reason (if any) is on file — that overrides the other three, because a reason that used to be true is moot once the source has caught up, and the fix is to re-type the field, not to re-justify it. A recorded reason on a type whose shape cannot be located at all (not by name, not through its root field's operation, not through a parent's property) is unresolved: an open finding, never accounted or stale. --check fails closed on anything but accounted.

What is never acceptable: a typed guess that drops payload data. If you type four of a payload's nine keys, the other five are unreachable through this schema and no test will ever tell you.

Errors

Map errors once, on the @source, with a ?? fallback in the message expression — several statuses on every API have no JSON body at all. When you write a test that asserts an error mapping, tests/router.yaml needs include_subgraph_errors: { all: true } or the snapshot shows only Subgraph errors redacted and the assertion proves nothing.

Take the message path from the inventory, never guess it. Read the selected operations' errors[] in .factory/inventory.json; each entry names a status and either a shape_ref or null. Follow every non-null shape_ref into shapes and write the path the documented body actually carries — $.message for {message: string}, $.error.message for {error: {message: string}}, and so on for whatever shape is there. When different statuses on the same @source document different bodies, chain their paths with ?? in declared order; a status whose shape_ref is null documents no body at all, which is why the chain still ends in a string fallback. In older inventories, an error body written inline in the spec — not a $ref — was silently dropped from the inventory (shape_ref: null) even though the body was fully documented; a path copied from another API's convention (.detail is FastAPI's default, not necessarily this API's) is the failure that produced. graphos-factory-core lint's error-path-unresolved checks this mechanically: every $. path in an errors block — each side of a ??, and each ->first/->last step into an array item, checked separately — must resolve in at least one documented error shape for the operations the block covers, or it warns. It is silent when no shape is documented at all, and a free-form body (opaque JSON, no properties) counts as resolving.

Descriptions

Every root field and every type gets a doc comment. Prefer the vendor's own summary over a paraphrase; it is what the API's users already know, and it is what an MCP client shows to a model choosing between tools. A spec with no summaries (every Google discovery-derived one) gives the operation's description instead. lint's undocumented-root-field (an error) fails a selected root field with no doc comment while that text exists, and quotes it. Where the connector deliberately differs from the API (an excluded field, a flattened envelope, a renamed operation), say so in the doc comment as well as in decisions.json — the README's "Limitations & exclusions" section is generated from those decisions (the resolved records). An argument carrying a default, minimum, maximum or enum in the inventory states it in its own doc comment — see § Argument constraints for the general rule and § Pagination for the page-size case.

A service that mints its own credential states the rule on the login root field. App-level facts a consumer needs go on the credential-minting root field's doc comment, before its Returns line, and on the token type's description. Never put them in a schema description (a description on the schema definition): the composed supergraph keeps only one, so a consumer of any other service never sees it. Never put them in a # comment either: that is not SDL and never reaches a consumer. The facts are: which operation mints the credential; that every secured operation takes it as an argument, named (access_token); and which account property is the username, when the account-creation shape in the inventory determines it. Take the username fact from the inventory shape; never invent it. A service whose credential lives on @source and in the README (all three pilots) is unchanged. § Argument descriptions' credential bullet names both homes — the source directive and README for a forwarded credential, the login root field for a self-minted one — and either way the rule is stated once, not on every argument. credential-source-undocumented warns when two or more selected root fields take a credential-named argument, a selected root returns a type declaring it, and that root's doc comment does not name it before the Returns line (testing.md has the name list).

Object field descriptions

The rule above covers root fields and types; it says nothing about object-type fields (properties), and the gap shows: across the three pilots only 54 of 395 object fields carry a doc comment (14%), even though the inventory already holds far more source material than that — gitea's inventory.json alone carries a description on 732 of 1334 properties (55%). COPY_KEYS (crate/src/openapi.rs) copies description verbatim from the OpenAPI/Swagger document onto every property already; the gap is guidance, not missing data.

Opt-in per service. Unlike Pagination and the rule above, this is not on by default: a workspace adopts it explicitly, recorded as a resolved decision in decisions.json (graphos-factory-core decisions, never hand-edited — workspace-contract.md). The trade is unmeasured — bigger schema text, a larger context budget for every consumer, against filling a real documentation gap — and unlike the Returns-line rule, which shipped only once the apollo-conn-gen benchmark had measured the turn it closes, there is no bundled evidence yet that this helps. Its effect is to be measured per service, not asserted.

Once a workspace opts in:

  • A field whose inventory property carries a non-empty description gets a doc comment using that text, trimmed to one sentence — cut at the first sentence boundary. A source with no sentence boundary (one long run-on) is capped at roughly 150 characters and closed with … rather than cut mid-word. created_at sourced from "The time the issue was created. Read-only; set by the server on POST and never revised." becomes "The time the issue was created."; html_url sourced from a run-on such as "The fully qualified URL that resolves to this issue in the web UI, suitable for use in a notification or chat message so the reader lands on the exact issue without reconstructing the path from the id and repo name" becomes "The fully qualified URL that resolves to this issue in the web UI, suitable for use in a notification or chat message so the reader lands on the exact…". Never paraphrase, and never invent a description the source does not contain.
  • A field with no source description gets none — this is opt-in enrichment of what the inventory already states, not a mandate to describe every field.
  • A field that already carries a doc comment is never touched. This rule fills gaps; it does not relitigate existing curation.

Argument descriptions

Root-field arguments have the same gap and the same source material as object fields — the inventory copies every parameter's and request-body property's description verbatim — with one difference: this surface already carries a default-on rule, the constraint clause of § Argument constraints, and a lint-read wording for page-size arguments (§ Pagination). What it did not have was a rule for the rest, and the gap was measured: an authoring run with no rule for this surface produced 1,565 argument doc comments where the previous run had produced 0 — 691 of them verbatim OpenAPI parameter text, 367 a per-argument "requires access_token" sentence — and the schemas' description text grew from 28 KB to 229 KB. A documentation surface with no rule fills with vendor text.

Opt-in per service for the source sentence, the same mechanism and the same reasoning as Object field descriptions above. The constraint clause stays default-on because a benchmark measured what its absence costs. The source sentence has no such measurement behind it, its cost side is the usual one — bigger schema text, a larger context budget for every consumer, vendor prose of uneven quality — and in the run above it was 691 lines nobody asked for. A workspace adopts it as a resolved decision in decisions.json (graphos-factory-core decisions, never hand-edited); promotion to default-on waits on a measured comparison, not on this rule.

An argument's doc comment is composed of at most three parts, in this order, and nothing else:

  1. The source sentence — opt-in. The inventory parameter's (or request-body property's) description, trimmed exactly as Object field descriptions trims: cut at the first sentence boundary, a run-on with none capped at roughly 150 characters and closed with …. Verbatim — the source's own casing, wording and punctuation, never paraphrased, never invented. No description in the inventory, no sentence. Written as a single-line "…" string, never a """ block (why, under part 3).
  2. The constraint clause — default-on. § Argument constraints, unchanged: (default asc, one of asc|desc). The omission sentence of § Argument constraints is also default-on, and comes verbatim after the clause: (default asc, one of asc|desc). Ascending when omitted. When it is the source's first sentence and part (1) is opted in, part (1) carries it and it is not repeated.
  3. Page-size and page-index wording. A page-size or page-index argument keeps § Pagination's wording and gets nothing else — the source sentence is not added to it, opted in or not. pagination-bounds-undocumented accepts any standalone number in the argument's doc comment that equals the inventory bound, and pagination-bounds-unknown accepts no documented, unknown, not specified or no bound anywhere in it (lint_pagination, crate/src/lint.rs); a source sentence that happens to carry a number or one of those phrases satisfies the rule without the schema saying what the rule wants said. The mechanism is measured, not a gitea case: Gitea's whole two-sentence order description — the second sentence ends ignored if "sort" is not specified. — placed on a page-size argument silences the gap warning; the one-sentence trim this rule would actually produce carries no such phrase and leaves lint where it was. The exclusion guards against a source whose first sentence carries a bound number or a gap phrase. The single-line form under part (1) guards the same rule from the other side: arg_doc_comment reads the doc comment leftmost-first, so a """ block on the page-size argument makes it read from the first earlier """ block in the argument list onward, and a source sentence on any earlier argument then satisfies the rule (also measured).

When (1) and (2) both apply, spell them as the source sentence without its final period, then the clause in parentheses, one period — the constraint-clause spelling with the source sentence in front:

  "Sort order (default asc, one of asc|desc)."
  sort: String
  "owner of the repo"
  owner: String!

No doc comment when none of the three applies.

  • Never restate type or requiredness. The SDL already says both: String! is the whole of "a required string". Two tests, applied to the source sentence as much as to anything hand-written. Requiredness is never written in any form — Required., Required argument, (required), Optional. — and is cut from a source sentence wherever it appears (a cut, not a rewording; a sentence that is only this is dropped). Type is judged by what the text adds to the declaration and the argument's own name, not by which words it contains: a sentence that only names the type — a bare string or integer, the ID of …, list of … ids — is dropped, not trimmed, so "milestone id" on milestone: ID and "list of label ids" on labels: [ID!] get no doc comment; "search string" on q: String stays verbatim, because q alone does not say the value is a search term.
  • A credential the connector forwards (access_token, an API-key argument) gets parts (1)–(3) like any other argument and nothing more. No per-argument "Requires access_token" sentence: where the credential comes from is stated once — on the source directive and in the README, or, for a credential the service mints itself, on the login root field (§ Descriptions) — not on every argument that carries it.
  • An argument that already carries a doc comment keeps it. As with object fields, this rule fills gaps; it does not relitigate existing curation. The two rules above that are not opt-in still apply to it: a restatement of type or requiredness is removed, and a constraint the inventory carries is still owed (§ Argument constraints).

Nothing new is checked mechanically. pagination-bounds-* read a page-size argument's doc comment (§ Pagination says what) — and, when that doc comment is a """ block and an earlier argument carries one too, everything from the earlier block onward, which is why part (1) is a single-line string; no rule counts argument doc comments, measures their length or reads them for type words. Review by reading the schema. argument-constraints-undocumented also checks that part (2)'s omission sentence appears in the doc comment (§ Argument constraints).

Returned field names

A root field's doc comment ends with the names its response type exposes, so an agent that has chosen the operation can write the selection set without spending a turn introspecting the return type. Spell the names as the schema spells them (GraphQL-visible, post-alias) in declaration order, comma separated:

  • Object response — Returns: id, name, createdAt. — the order the type declares, not alphabetical.
  • List response — Returns a list of items with: id, name, createdAt.
  • Envelope response — an object whose one object-list field carries the payload and whose other fields are paging plumbing or a status flag (the shapes § Envelopes recognises): write both lines. Returns: incidents, limit, offset, more, total. then Each item in `incidents` has: id, incidentNumber, title, … The envelope's own names do not answer the question the line exists to answer, so naming them alone is not enough. The payload field stays a bare name in the first line: the second line is its expansion, and the brace form below would name the item fields twice.
  • Scalar, union or opaque-JSON response — no line at all. gitea_version: String gets none. Never guess a shape the schema does not state.

An object-typed entry renders one level deep, and a small leaf-only object inside it one more. A field whose type is an object, or a list of objects, carries that type's field names in braces — teams { id, type, summary, self, htmlUrl } — in the nested type's declaration order, comma separated, capped at 6 and closed with (+N more) inside the braces: owner { id, login, loginName, sourceId, fullName, email (+16 more) }. Inside the braces, an object field whose type has at most 6 fields, every one a leaf (scalar, enum or opaque-JSON scalar), gets braces of its own: cards { code, value, suit, image, images { svg, png } }, and on a splitwise-shaped type shares { debtor { name, email }, debtAmount }. A second-level type with any object-typed field, or more than 6 fields, stays a bare name — piles { name, remaining, cards }, because a card carries images — and nesting stops there. A second-level brace group is one entry toward the nested 6; its own list needs no cap, because it is at most 6 by definition. Scalars, enums and opaque-JSON fields are names only, at every level. A bare name for an object field reads as a leaf, and a consuming agent acts on that reading: in one AppWorld run a text message's flat sender, receiver produced selection errors in 13 tasks, and flat reference types invite guesses such as first_name (lessons.md).

Cap the outer list at 14 entries and close it with (+N more). The 14 are the first 14 in declaration order, never a hand-picked "interesting" subset — a reviewer checks the line by reading the type top to bottom, and a subset nobody can re-derive is a judgement with no audit. A brace group is one entry toward the 14, whatever it holds; the nested cap of 6 is counted inside the braces and separately. Wrap between entries and keep a brace group on one line where it fits: the breaks are cosmetic inside a """ block, but a group split across lines is harder to check against the type.

One shape still leaves the introspection turn in place, and it is not a reason to widen the rule. A very wide type: the cap leaves most of it unnamed — Gitea's Gitea_Repository has 46 fields, so gitea_repo ends (+32 more), and the nested cap does the same one level down (owner { … (+16 more) } on the 22-field Gitea_User). The resource that merely embeds a list — not an envelope, because § Envelopes wants exactly one other non-link property — was the rule's second limit until the brace form closed it: piles { name, remaining, cards } names the item type's fields where a bare piles did not. What stays unnamed is a second-level type that is not small and leaf-only (the Card fields under cards inside piles), and everything below the second level, by design.

The line is the last thing in the doc comment. On a paginated operation it goes after the one-page/stop-condition sentence (§ Paginated list operation descriptions), never instead of it: list-completion-missing looks for a pagination keyword anywhere in the description, so a Returns: line that names an envelope's total field satisfies the rule by itself and would hide a missing stop condition.

Root fields only. Nested types keep their own doc comments — the turn this removes is on the operation the agent has already picked, and a Returns: line on every nested type would collide with the entity-type descriptions for no measured gain. Whether the line pays for itself is an eval question, not something this rule asserts.

returns-line-nesting (warn) reads each selected root field's Returns line against the SDL: a name the type does not declare, a bare object-typed entry at the first level (the envelope's payload field in the first line excepted), a bare small leaf-only object at the second level, and braces on a second-level type that is too wide or nests. It does not check the order, the caps or the (+N more) arithmetic, and a doc comment with no Returns line is not its business.

Which operation a root field serves

reconcile, lock and source-coverage find a root field's operation from its @connect method and path: the literal path, else the one operation whose template matches once parameter names are erased. Where several do — a Graph-style API serves every node at a bare /{id}, so GET /{$args.campaignId} matches get:/{campaign_id} and get:/{ad_id} alike — the field is attributed to the operation whose selection entry names it: <field_prefix>_<graphql.name> under graphql.root. Keep the field name exactly that. An included entry's claim wins; an excluded entry's graphql block still names its field until apply removes it, which is how reconcile knows what to remove. A type-level @connect (an entity) on such a path is attributed to the entry marked graphql.entity: true whose root field returns that type. A field or type nothing declares, or one two entries of the same rank both claim, is a reconcile selection error listing the candidates. For a root field, lock also refuses to record it, codify refuses it, and lock --check and lint (unattributed-span) report it as a tie to settle in the selection, not a hand edit. The tool never picks one.

Coverage: what the schema leaves on the wire

Source coverage: every fact the source offers is covered by the schema, or a recorded decision says why not. A schema is finished when that holds, not when every selected operation compiles. graphos-factory-core source-coverage . OP-KEY reads one operation's request body and response out of inventory.json and classifies each leaf. It was once spans obligations, and that spelling still runs. It does not classify path, query or header parameters, with one exception (the behaviour section, below). Check the rest by hand against the operation's parameters in the inventory: every one should be an argument, a fixed value in the connector's URL, queryParams or headers, or a recorded decision.

Class Meaning
mapped a schema field reads or writes it, including as a later operand of a ??/?! chain in the selection (title: name ?? label)
consumed no field, but a connector expression reads it (an error path, a ?? fallback in isSuccess/errors) — response side
omitted-decided an omits entry on a resolved decisions.json record or a current findings.json finding covers it, with a reason; one on an open or superseded record does not count, and the row stays unaccounted with a note naming the decision
unaccounted nothing does — the schema silently drops it
unresolved the tool could not tell: an inventory shape it cannot resolve, a dropped spec construct, or a connector expression outside the grammars it parses (a request-side ->match, a request-body expression)
unverified-default an expansion boundary whose default projection is not verified: one row for the relationship, its children not enumerated — response side
transport-expansion-missing the schema maps a boundary child the request never asks for, so the source never sends it — response side, not counted as offered

Run it per operation while authoring, and again with --check (non-zero when anything is unaccounted, unresolved, unverified-default or transport-expansion-missing, naming which on stderr) before lock. Zero of all four in both directions is the completion bar: --check fails closed, because an unresolved path is not known to be handled. An omits record does not clear an unresolved or an unverified-default row.

Behaviour facts. A third section, behaviour, appears when an optional argument's source says what omitting it does (the omission sentence of § Argument constraints, "If not passed, Venmo balance will be used."). One row per such argument:

Class Meaning
documented the argument's doc comment carries the sentence
waived a behaviour omit on a resolved decision or a current finding records why it does not
unaccounted neither — --check fails on it (behaviour unaccounted N)

The row's path is the source the argument reaches — query:NAME, path:NAME, header:NAME or body:KEY.PATH — and the row names the carrying argument (Query.shop_pay(paymentCardId)) and quotes the sentence. Fix an unaccounted row by writing the sentence into the doc comment, or record why not. When the sentence is not about omitting this argument (a "by default" about a server setting), the source settles it: a finding.

graphos-factory-core findings add . --title T --body TEXT --cites REF \
  --omit 'get:/x|behaviour|query:sort|not-applicable'

When you chose not to state it, that is a call with stating it as the alternative: a decision.

graphos-factory-core decisions add . --title T --question Q \
  --choice 'omit:leave the sentence out' --choice 'state:state it' \
  --resolved --chosen omit --decision TEXT \
  --omit 'get:/x|behaviour|query:sort|editorial'

not-applicable and editorial are the two reasons for a behaviour omit; consumed is a wire reason and is refused, and findings add refuses editorial. The waiver clears the sentence, not the constraint clause. It covers exactly the one source it names, and only on a resolved decision or a current finding. An argument sent through an input type ($args.input { … }, $args.input.x) has no row yet: the trace stops at the argument, so write those sentences by hand.

Do not make an optional source parameter required to get past this. ! on an argument whose source parameter or body key is optional is the lint error argument-required-optional-in-source: the caller could no longer leave it out, so the source's omission behaviour would be unreachable. Make it optional, or, when it must be required, record why with the decision form of the behaviour waiver above (--omit 'get:/x|behaviour|query:x|editorial').

A bare $ selection returns the whole response, and the classifier reads it that way: selection: "$" on a [String] field over a root array of strings maps [], and over an object or an array of objects maps every row beneath it as mapped (json). $ { … } walks its sub-selection from the root, and a method on the bare $ ($->first) leaves every row unresolved. Record no omits entry for a root row a bare $ returns; one recorded before 0.5.67 still resolves, but the row now reads mapped first.

Every selected operation at once. With no OP-KEY, graphos-factory-core source-coverage . [--json] [--check] runs the same classifier on every operation selection.yaml includes (include: true), in inventory order. Do not loop over inventory list: it pages, and it lists operations the selection leaves out, which need no coverage.

  • Text: one line per operation with both directions' counts (and unverified-default / transport-expansion-missing when nonzero), FAIL on each that misses the bar, a totals line, and N of M selected operations fail the bar. It prints no rows; run the single-key form on a failing operation to see them.
  • --json: {"operations": [...], "totals": {...}, "failing": [...]}. Each entry of operations is byte for byte the object source-coverage . OP-KEY --json prints. totals holds selected, passing, failing, errors and the summed response and request counts; failing lists the operation keys.
  • --check: exits 1 when any selected operation fails, with one source-coverage --check: OP fails — … line per failing operation and a count line on stderr.
  • An operation the classifier cannot build a report for (a selected key missing from inventory.json) is listed with its error ({"op_key", "error"} in --json), counted failing, named with its error on stderr, and makes the run exit 1 with or without --check. A selection that includes no operation, or a workspace with no selection.yaml, exits 1: nothing checked is not a pass.

Stale omits. An omits entry on a resolved decision or a current finding can outlive the reason for it: the schema now maps the path, a classifier upgrade now reads a ->match it once reported unresolved, the envelope now reads the path (for an editorial entry), or a refreshed source no longer offers the path or the operation. Such an entry is stale: source-coverage counts it per direction as stale-omit N (JSON stale_omit, only when nonzero), lists it after that direction's rows as stale omit: D-id omits path (direction) on OP, but … (JSON stale_omits: decision, operation, direction, path, why — mapped, consumed, not-offered or no-operation — message, fix), and with no OP-KEY lists every stale entry in the workspace, on any operation, selected or not. It is not part of the bar: a stale entry leaves no offered path unaccounted, so --check does not fail on it. lint warns stale-omit on each, naming the decision and the row. Fix it through the log, never by hand: graphos-factory-core decisions supersede . --id D-id, then record any of its omits that still apply on a new resolved decision (for a finding, graphos-factory-core findings supersede . --id F-id and a new finding). An entry is only reported when staleness is known: a row it covers that is unresolved or unverified-default counts as needing it.

Expansion boundaries. A response property whose inventory node carries x-expansion is a relationship to another node, and the walk stops there instead of unfolding the target's graph:

  • unexpanded, verified default: it offers exactly the default leaves (business.id), mapped or omitted like any leaf;
  • unexpanded, unverified: it offers one unverified-default row. Nothing in the workspace clears it; verify the default (a raw probe of the unexpanded field, or version-specific documentation), record it with its evidence where the source document is generated, and refresh;
  • expanded on the wire (the fields default or literal carries business{…}): it offers what the expression requests, so a requested child you do not map is unaccounted.

To expose a child beyond the default, put the group in the fields default (§ Sparse fieldsets). Mapping the child without it is transport-expansion-missing. An operation with no fields parameter (any POST) cannot ask at all: expose such a child only when a resolved decision names the operation and records the path, as a backtick span (`owner.name`), as what the source is verified to send. A GET that takes and sends the fields parameter gets no such exemption; its fix is the group in the default.

The classifier reads the forms this skill recommends:

  • map->entries over a dictionary maps the dictionary's row;
  • field->match([from, to], …) on a scalar maps field, as value translation (over an object or array, without a sub-selection, it stays unresolved: an arm can send any part of the value);
  • field->map(@->match(…)) over an array of scalars maps field, the same translation applied to each element; anything else inside the map, or a method after it, stays unresolved. In a request body it is still unresolved, as every request-side ->match is;
  • a nested object literal in a request body is read key by key;
  • an argument sub-selection (key: $args.x { wire: field }) is read wire key by wire key;
  • a request body that is one $args.<name> expression, with or without a sub-selection (body: "$args.input", $args.input { Wire: field }), is classified at the request root; a body that follows that expression with key: expr pairs is not, so write one form or the other;
  • a deeper argument path ($args.a.b) forwards the field it names;
  • a ... spread of disc->match(…) arms, the abstract-type form (mapping-language.md § Abstract types), maps the discriminator and each arm's fields. Any other spread leaves every path under its object that the selection does not otherwise map unresolved.

An unresolved row names the construct the classifier does not read, for example request expression not parsed: ->map->first or selection method not parsed: ->entries->first. Such a row still comes from the classifier's grammar, not from the inventory. Never hand-edit inventory.json to make one go away: it is facts, and lint reports an edit to it (unacknowledged-inventory-edit). Instead:

  • keep the idiom;
  • record a finding naming the operation, the unresolved paths, and the grammar that leaves them unresolved (findings add --cites …: the classifier's grammar settles it, nobody chose it);
  • report the operation as not verified by this layer (e2e and conformance still prove it), never as done.

Only a dropped spec construct or a depth-limit marker points at the source document. That is fixed in the pinned working copy (codify --source), never in inventory.json.

An unaccounted leaf is not a lint nit: it is a field the source returns that no caller can ever see, and nobody decided that.

Clearing one has exactly two honest endings — map the field, or record why you did not. Leaving it off on your own judgement is a decision, with exposing it as the alternative:

graphos-factory-core decisions add . --title "Drop the internal audit block from candidate reads" \
  --question "Expose results.auditTrail on post:/candidate.info?" \
  --choice 'omit:leave it off' --choice 'expose:map it as a field' \
  --resolved --chosen omit --note "Internal-only; no caller has asked for it." \
  --omit 'post:/candidate.info|response|results.auditTrail|editorial'

A path an expression reads is a fact, not a call — a finding:

graphos-factory-core findings add . --title "success is read by the @source error mapping" \
  --body "The errors block reads success on every operation; it is not a field." \
  --cites "references/schema-authoring.md § Errors" \
  --omit 'post:/candidate.info|response|success|consumed'

--omit operation|direction|path|reason is repeatable on both; direction is response or request, reason is editorial (deliberately not exposed; a decision only — findings add refuses it) or consumed (used by an expression, not surfaced as a field). An omit on a parent path covers its children, and never matches by type name — it is a path on one operation. . is the root path, the whole direction: it is the only way to record an EmptyResponse 204, whose one offered row sits at the empty root (--omit 'delete:/ecommerce/stores/{store_id}|response|.|editorial'). The classifier reads these records back, so the same record that justifies the gap is what closes it. Never widen an omit to silence a leaf you have not actually considered; that converts a finding into a lie the next --check will not catch.

Writes

A write operation gets the same care as a read plus two things: an explicit graphql.root: mutation in selection.yaml, and a hand-authored WireMock case, because the unit layer cannot assert an object-valued argument at all (testing.md). A tag an operation needs (tags: [beta]) is set in selection.yaml at selection time, not later.

Copy and update workflows

A copy or clone operation reads one object and creates another. The caller must explicitly carry over any field affecting behavior or presentation — provider defaults for omitted optional fields are not "the source object's values". Identifying behavior-affecting fields: booleans and enums on the read shape that are also writable on the create or update request shape (e.g. pinned, archived, starred, visibility, status, locked, draft); readOnly: true marks fields the API never accepts on write. When the source metadata's description names the field's effect, surface it in the doc comment. Copy and clone descriptions must: (1) state they create a new object from a source; (2) list behavior-affecting fields to carry over explicitly, warning that omitted fields take provider defaults; (3) when the operation takes a full input object, warn unset keys default and do not mirror. Example: a synthetic notes API — GET /notes/{id} returns { id, title, content, pinned, archived, color, created_at }, POST /notes accepts { title, content, pinned?, archived?, color? }. The copy operation's description: "Creates a note. To copy an existing note, read it first and pass its title, content, pinned, archived, and color values. Fields omitted from the request take provider defaults (e.g. pinned: false), not the source note's values." The e2e case for a copy flow must demand the preserved field in the outbound body — see testing.md.

Source: SKILL.md on GitHub

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No alerts20d3 checks · Risk SAFE
  • Gen Agent Trust Hub20d

    The skill is a professional tool for building Apollo Connector subgraphs. It installs and uses official vendor tools (Apollo Rover, Apollo Router, WireMock) from official sources. It follows security best practices for credential management, including redaction and environment variable usage. It requires explicit user consent for tool installation and licensing. While it processes untrusted OpenAPI specifications, it includes instructions for auditing results to mitigate indirect prompt injection risks.

  • Socket20d

    No alerts

  • Snyk20d

    Risk: LOW · No issues

Signed by skilld at 9dda139. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 3 weeks ago.

Activeupdated last month
metadata
{
  "version": "0.12.3"
}
Other metadata
compatibility
macOS or Linux (WSL on Windows) with bash, curl, git and jq; Java 17+ for the end-to-end and live layers. Any Agent Skills host (Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI and others) that can run shell commands.

README badge

README badge for apollographql/skills/graphos-factory