fraiseql_cli/schema/converter/mod.rs
1//! Schema Converter
2//!
3//! Converts `IntermediateSchema` (language-agnostic) to `CompiledSchema` (Rust-specific)
4
5mod cascade_types;
6mod directives;
7pub mod functions;
8mod identity;
9mod interface_conformance;
10mod mutation_error_union;
11mod mutations;
12mod queries;
13mod relay;
14mod subscriptions;
15pub(crate) mod tenancy;
16mod types;
17
18#[cfg(test)]
19mod tests;
20
21use std::collections::HashSet;
22
23use anyhow::{Context, Result};
24use fraiseql_core::{
25 compiler::fact_table::{
26 DimensionColumn, DimensionPath, FactTableMetadata, FilterColumn, MeasureColumn, SqlType,
27 },
28 schema::{CompiledSchema, FieldType},
29 validation::CustomTypeRegistry,
30};
31use fraiseql_functions::FunctionsConfig;
32use tracing::{info, warn};
33
34use super::{
35 artifact::CompiledArtifact,
36 intermediate::{IntermediateFactTable, IntermediateInjectDefaults, IntermediateSchema},
37};
38
39/// The directory a project's function modules live in when nothing says otherwise.
40///
41/// The default lives here, in the compiler, rather than on the wire type: every
42/// compiled artifact carries an explicit `module_dir`, so a *hand-written*
43/// `schema.compiled.json` that omits it still fails loudly instead of quietly
44/// resolving somewhere its author did not choose.
45const DEFAULT_FUNCTION_MODULE_DIR: &str = "functions";
46
47/// Converts intermediate format to compiled format
48pub struct SchemaConverter;
49
50/// The author-declared enum, interface, union and custom-scalar names of one schema
51/// (`#923`, extended by `#1018`).
52///
53/// `parse_field_type` matched the twelve builtin scalar names and routed everything else
54/// to `FieldType::Object`, because it is a free function with no view of the document it
55/// is parsing a name from. So `status: OrderStatus`, where the same file declares
56/// `enum OrderStatus`, compiled to `{"Object": "OrderStatus"}` — and
57/// `FieldType::Enum`/`Interface`/`Union` had **no producer at all** for authored schemas;
58/// their only writers were the compiler's own synthesized cascade types.
59///
60/// Everything downstream branches on the variant:
61///
62/// * introspection (`field_resolver`) reports `TypeKind::Object` for an enum-typed field, so an
63/// introspection-driven client (Apollo, Relay, graphql-codegen) is told a scalar enum is an
64/// object with no fields;
65/// * the TypeScript client generator treats `Object` as requiring a nested selection and emits
66/// `status { … }` for a scalar value;
67/// * `--emit-ddl` maps `Object` to `JSONB` rather than to the enum's own Postgres type.
68///
69/// `#1018` is the same defect one variant over: a name declared in `custom_scalars` also
70/// fell through to `Object`, so `FieldType::Scalar` had no producer either.
71///
72/// Held by name rather than by definition because that is all resolution needs, and
73/// because it must be built *before* the type arrays are consumed by conversion.
74#[derive(Debug, Default)]
75pub(crate) struct DeclaredTypeNames {
76 objects: HashSet<String>,
77 enums: HashSet<String>,
78 interfaces: HashSet<String>,
79 unions: HashSet<String>,
80 scalars: HashSet<String>,
81}
82
83impl DeclaredTypeNames {
84 /// Collect the names an intermediate schema declares.
85 fn collect(schema: &IntermediateSchema) -> Self {
86 Self {
87 objects: schema.types.iter().map(|t| t.name.clone()).collect(),
88 enums: schema.enums.iter().map(|e| e.name.clone()).collect(),
89 interfaces: schema.interfaces.iter().map(|i| i.name.clone()).collect(),
90 unions: schema.unions.iter().map(|u| u.name.clone()).collect(),
91 scalars: schema.custom_scalars.iter().flatten().map(|s| s.name.clone()).collect(),
92 }
93 }
94
95 /// Resolve a non-builtin type name to the variant the rest of the codebase branches on.
96 ///
97 /// Falls back to `FieldType::Object` for a name declared nowhere. That never reaches an
98 /// artifact: `SchemaValidator` refuses an undeclared field type by name (#1530), and the
99 /// server refuses an artifact whose field names an undeclared object at load.
100 ///
101 /// A name declared in `custom_scalars` resolves to `FieldType::Scalar` (#1018). Without
102 /// it that variant had no producer for an authored schema — the #923 defect, one variant
103 /// over — and three consumers branched on the wrong answer: `--emit-ddl` emitted `JSONB`
104 /// where `TEXT` is correct, introspection reported `OBJECT` for a leaf, and the
105 /// TypeScript generator emitted a sub-selection for a scalar.
106 fn resolve(&self, name: &str) -> FieldType {
107 let owned = || name.to_string();
108 if self.objects.contains(name) {
109 // A declared type keeps its name, even one the engine also knows as a rich
110 // scalar (`Money`, `Image`, `Color`, …): the author declared it.
111 FieldType::Object(owned())
112 } else if self.enums.contains(name) {
113 FieldType::Enum(owned())
114 } else if self.interfaces.contains(name) {
115 FieldType::Interface(owned())
116 } else if self.unions.contains(name) {
117 FieldType::Union(owned())
118 } else if self.scalars.contains(name) || fraiseql_core::schema::is_known_scalar(name) {
119 // A declared custom scalar (#1018), or one of the engine's own rich scalars
120 // (#1530): `Email`, `Hostname`, … are scalars without anyone declaring them, and
121 // fell through to an object reference that was served without its value.
122 FieldType::Scalar(owned())
123 } else {
124 FieldType::Object(owned())
125 }
126 }
127}
128
129/// Optional, default-off behaviours for [`SchemaConverter::convert_with_options`].
130///
131/// Threaded here rather than on [`IntermediateSchema`] so adding an option never
132/// breaks the many full-literal `IntermediateSchema { .. }` constructions, and
133/// the plain [`SchemaConverter::convert`] keeps its historical signature.
134#[derive(Debug, Clone, Default)]
135pub struct ConvertOptions {
136 /// Auto-synthesize a shared `MutationError` type + per-mutation result unions
137 /// and rewrite object-returning mutations to those unions (`[fraiseql.mutations]
138 /// auto_error_union`).
139 pub auto_error_union: bool,
140}
141
142impl SchemaConverter {
143 /// Convert `IntermediateSchema` to `CompiledSchema` with default options.
144 ///
145 /// # Errors
146 ///
147 /// See [`SchemaConverter::convert_with_options`].
148 pub fn convert(intermediate: IntermediateSchema) -> Result<CompiledSchema> {
149 Self::convert_with_options(intermediate, &ConvertOptions::default())
150 }
151
152 /// Convert `IntermediateSchema` into the whole compiled **artifact** — the
153 /// schema plus the platform sections written beside it (#1325).
154 ///
155 /// [`convert_with_options`](Self::convert_with_options) produces only the
156 /// `CompiledSchema`, which is what almost every caller wants. The sections that
157 /// are *siblings* of it in `schema.compiled.json` — currently `functions` — are
158 /// assembled here, so there is exactly one place where an authored section
159 /// becomes a compiled one and exactly one place a probe has to look.
160 ///
161 /// The `[functions]` TOML settings (`module_dir`, `dlq_store`) are paired with
162 /// the authored definitions here; the two halves have one owner each.
163 ///
164 /// # Errors
165 ///
166 /// See [`convert_with_options`](Self::convert_with_options).
167 pub fn convert_artifact(
168 mut intermediate: IntermediateSchema,
169 options: &ConvertOptions,
170 ) -> Result<CompiledArtifact> {
171 let functions = Self::take_functions_section(&mut intermediate)?;
172 let schema = Self::convert_with_options(intermediate, options)?;
173 // Unconditional, including when the project declares no function at all:
174 // a query whose `function = "<name>"` names nothing is exactly the case an
175 // `if let Some(functions)` here would wave through, and it is the one an
176 // author hits by deleting a function and forgetting the query (#1329).
177 functions::validate_against_schema(functions.as_ref(), &schema)?;
178 Ok(CompiledArtifact { schema, functions })
179 }
180
181 /// Pair the authored function declarations with their `[functions]` settings.
182 ///
183 /// Returns `None` when the project declares no functions and no enabled source, so
184 /// a schema that has never heard of them serialises exactly as it did before.
185 ///
186 /// # Errors
187 ///
188 /// Refuses a `[functions]` table when neither a function nor an enabled source would
189 /// use it.
190 /// There is nothing to apply the settings to, so carrying them would write a
191 /// section the server reads and cannot act on, and dropping them would be the
192 /// silent-default failure this whole seam exists to remove (#1008 refuses a
193 /// `storage` section on the same grounds).
194 fn take_functions_section(
195 intermediate: &mut IntermediateSchema,
196 ) -> Result<Option<FunctionsConfig>> {
197 let settings = intermediate.functions_config.take().unwrap_or_default();
198 // An empty list is absence, not an empty section, and two producers reach this.
199 // Some SDKs always emit the key and some omit it, so without this rule the same
200 // project compiled through two of them yields different artifacts. More
201 // importantly `seam::empty_accumulator` seeds EVERY array section with `[]`, so
202 // every TOML-workflow compile arrives here with an empty list whether or not the
203 // project has ever heard of functions — and an emitted section is one #1326
204 // refuses to boot on a build that cannot serve it.
205 let definitions = intermediate.functions.take().unwrap_or_default();
206 // An enabled source's connector is loaded from `module_dir` too (#1399), so a
207 // project whose only code is a connector still needs the section — with no
208 // definitions in it. Refusing the table here made the documented connector
209 // layout impossible to compile.
210 let runs_a_connector = intermediate.sources.iter().flatten().any(|source| source.enabled);
211 if definitions.is_empty() && !runs_a_connector {
212 anyhow::ensure!(
213 settings.module_dir.is_none() && settings.dlq_store.is_none(),
214 "the `[functions]` table configures the functions subsystem but the schema \
215 declares no function and no enabled source, so nothing would use it. Declare \
216 a function or a source, or remove the table."
217 );
218 return Ok(None);
219 }
220
221 Ok(Some(FunctionsConfig {
222 module_dir: settings
223 .module_dir
224 .unwrap_or_else(|| std::path::PathBuf::from(DEFAULT_FUNCTION_MODULE_DIR)),
225 definitions,
226 dlq_store: settings.dlq_store,
227 }))
228 }
229
230 /// Convert `IntermediateSchema` to `CompiledSchema`.
231 ///
232 /// This performs:
233 /// 1. Type conversion (intermediate types → compiled types)
234 /// 2. Field name normalization (type → `field_type`)
235 /// 3. Optional mutation-error-union synthesis (`options.auto_error_union`)
236 /// 4. Validation (type references, circular refs, etc.)
237 ///
238 /// # Errors
239 ///
240 /// Returns an error if any type, query, mutation, interface, subscription,
241 /// or directive conversion fails, if federation/security/observer config JSON
242 /// cannot be deserialized, or if compiled schema validation detects unknown
243 /// type references.
244 pub fn convert_with_options(
245 intermediate: IntermediateSchema,
246 options: &ConvertOptions,
247 ) -> Result<CompiledSchema> {
248 info!("Converting intermediate schema to compiled format");
249
250 // Aggregate the per-type `@subscribable(tables=[...])` annotations (#366)
251 // into the top-level `CompiledSchema.subscribable` list BEFORE the types
252 // are consumed below. Empty `tables` is treated as "not subscribable".
253 let subscribable: Vec<fraiseql_core::schema::SubscribableEntity> = intermediate
254 .types
255 .iter()
256 .filter_map(|t| {
257 t.subscribable_tables
258 .as_ref()
259 .filter(|tables| !tables.is_empty())
260 .map(|tables| fraiseql_core::schema::SubscribableEntity {
261 entity_type: t.name.clone(),
262 tables: tables.clone(),
263 pre_image: t.subscribable_pre_image,
264 })
265 })
266 .collect();
267
268 // Refuse a schema that declares observers, rather than validating them and throwing
269 // them away (#779).
270 //
271 // The `observers` block is emitted by the Python SDK's `@fraiseql.observer` (its
272 // shipped example declares five), it binds to `IntermediateSchema.observers`, and
273 // `SchemaValidator` spends ~220 lines checking names, entities, events, actions and
274 // retry policy — a typo in any of them fails the compile, which tells the author
275 // emphatically that the block is honoured. It was not: this function set
276 // `observers: Vec::new()` under a comment claiming they were "populated from
277 // IntermediateSchema", nothing else ever wrote the field, and no webhook, Slack
278 // message or email ever fired for any declared event.
279 //
280 // The runtime loads observers exclusively from `tb_observer` / the admin API and
281 // reads nothing from the compiled schema. So *carrying* them here would only move
282 // the silent drop one layer down and produce a compiled artifact whose `observers`
283 // array is decoration. Until a runtime consumer exists, the honest outcome is to
284 // fail and name the mechanism that does work.
285 if let Some(observers) = intermediate.observers.as_ref().filter(|o| !o.is_empty()) {
286 let names: Vec<&str> = observers.iter().map(|o| o.name.as_str()).collect();
287 anyhow::bail!(
288 "This schema declares {} observer(s) ({}), but declared observers are not \
289 loaded by the runtime — it reads them from the `tb_observer` table and the \
290 admin API only, so compiling them would produce a schema whose observers \
291 never fire.\n\nDefine them in `tb_observer`, or register them at runtime \
292 with `POST /api/observers`, and remove the `observers` block from the \
293 authored schema.",
294 observers.len(),
295 names.join(", ")
296 );
297 }
298
299 // Refuse a schema that declares `aggregate_queries`, rather than carrying it through
300 // the seam and dropping it (#956).
301 //
302 // `aggregate_queries` is listed in `AUTHORABLE_ARRAY_SECTIONS`, so it is valid input
303 // on every compile path and every loader and merger carries it faithfully — and then
304 // this converter maps `fact_tables` and never reads it. `CompiledSchema` has no
305 // corresponding field, so the definitions reach the end of the seam and evaporate
306 // under a `✓ Schema compiled successfully`. That is the #755 shape surviving inside
307 // the very seam built to kill it.
308 //
309 // #624 gave `[[analytics.queries]]` real semantics by lowering each entry onto an
310 // ordinary `QueryDefinition`, so the capability exists and has a supported spelling.
311 // Refusing here — the `#779` disposition — names it, instead of accepting a second
312 // spelling that does nothing.
313 if let Some(aggregates) = intermediate.aggregate_queries.as_ref().filter(|a| !a.is_empty())
314 {
315 let names: Vec<&str> = aggregates.iter().map(|a| a.name.as_str()).collect();
316 anyhow::bail!(
317 "This schema declares {} aggregate_queries entry/entries ({}), which no \
318 compile path lowers into the compiled schema — `CompiledSchema` has no \
319 field for them, so they would vanish silently.\n\nDeclare them as \
320 `[[analytics.queries]]` in fraiseql.toml instead (#624): each entry \
321 becomes a real list-returning, view-backed query. Then remove the \
322 `aggregate_queries` block from the authored schema.",
323 aggregates.len(),
324 names.join(", ")
325 );
326 }
327
328 // #923: collect the declared enum / interface / union names BEFORE the arrays are
329 // consumed below, so a field or argument naming one resolves to that kind instead of
330 // to a phantom `FieldType::Object`.
331 let declared = DeclaredTypeNames::collect(&intermediate);
332
333 // Split the authored `types` array on `is_input` before converting anything (#848).
334 //
335 // A type marked `is_input` is an input object that four SDKs happen to declare in
336 // the `types` array — Elixir's exporter emits no `input_types` key at all, so this
337 // is its only route. Routing it here rather than converting it as an object type is
338 // what keeps the compiled schema GraphQL-legal: an argument may only be typed with
339 // an input type (§3.10).
340 let (input_marked_types, object_types): (Vec<_>, Vec<_>) =
341 intermediate.types.into_iter().partition(|t| t.is_input);
342
343 let types = object_types
344 .into_iter()
345 .map(|t| Self::convert_type(t, &declared))
346 .collect::<Result<Vec<_>>>()
347 .context("Failed to convert types")?;
348
349 // Extract query_defaults before consuming intermediate.queries.
350 // unwrap_or_default() → all-true, matching historical behaviour when no
351 // [query_defaults] section is present in fraiseql.toml.
352 let defaults = intermediate.query_defaults.unwrap_or_default();
353
354 // Apply `[inject_defaults]` (#847) to any operation that has not named the
355 // parameter itself.
356 //
357 // This runs here, in the converter, rather than in `commands::compile` — the
358 // converter is the single path every caller of the public API goes through, so a
359 // default cannot be lost by reaching the compiled schema some other way.
360 //
361 // It is also deliberately *after* `[fraiseql.tenancy]` validation (which
362 // `commands::compile` performs on the intermediate schema before calling this).
363 // Tenancy auto-injects when `inject_params` is empty and errors when it is
364 // non-empty but lacks the annotated field, so applying defaults first would make a
365 // single unrelated default break every tenancy-annotated operation.
366 let inject_defaults = intermediate.inject_defaults.unwrap_or_default();
367 let query_defaults_inject = inject_defaults.for_queries();
368 let mutation_defaults_inject = inject_defaults.for_mutations();
369
370 // Convert queries. A query that opted into `count` emits a second,
371 // derived `<name>Count` definition (#938) — built by `count_sibling` so
372 // its tenancy, role and filter inheritance has exactly one definition.
373 let mut queries: Vec<fraiseql_core::schema::QueryDefinition> =
374 Vec::with_capacity(intermediate.queries.len());
375 for mut q in intermediate.queries {
376 IntermediateInjectDefaults::apply_to(
377 &query_defaults_inject,
378 &mut q.inject,
379 &q.exclude_inject_defaults,
380 &q.name,
381 )?;
382 let wants_count = q.count;
383 let converted = Self::convert_query(q, &defaults, &declared)
384 .context("Failed to convert queries")?;
385 let sibling = wants_count.then(|| converted.count_sibling());
386 queries.push(converted);
387 queries.extend(sibling);
388 }
389
390 // A generated sibling must not silently displace an author's own query of
391 // the same name — whichever won, the other would vanish from the schema
392 // with no diagnostic.
393 let mut seen: HashSet<&str> = HashSet::new();
394 for q in &queries {
395 if !seen.insert(q.name.as_str()) {
396 anyhow::bail!(
397 "Query name '{}' is declared twice. A `count = true` query generates \
398 '<name>Count'; rename the conflicting query or drop count=true.",
399 q.name
400 );
401 }
402 }
403
404 // Convert mutations
405 let mutations = intermediate
406 .mutations
407 .into_iter()
408 .map(|mut m| {
409 IntermediateInjectDefaults::apply_to(
410 &mutation_defaults_inject,
411 &mut m.inject,
412 &m.exclude_inject_defaults,
413 &m.name,
414 )?;
415 Self::convert_mutation(m, &declared)
416 })
417 .collect::<Result<Vec<_>>>()
418 .context("Failed to convert mutations")?;
419
420 // Convert enums
421 let enums = intermediate.enums.into_iter().map(Self::convert_enum).collect::<Vec<_>>();
422
423 // Convert input types — the `input_types` array plus the `is_input`-marked entries
424 // lifted out of `types` above. Both routes land in one registry, so a duplicate
425 // between them is an authoring conflict and is refused rather than shadowed.
426 let mut input_objects = intermediate.input_types;
427 for marked in input_marked_types {
428 let lifted = Self::input_object_from_marked_type(marked)?;
429 if input_objects.iter().any(|existing| existing.name == lifted.name) {
430 anyhow::bail!(
431 "'{}' is declared both in `input_types` and as a type marked \
432 `is_input: true`. Keep exactly one declaration — two definitions of one \
433 input object cannot be reconciled, and whichever won would depend on \
434 compile order.",
435 lifted.name
436 );
437 }
438 input_objects.push(lifted);
439 }
440 let input_types =
441 input_objects.into_iter().map(Self::convert_input_object).collect::<Vec<_>>();
442
443 // Convert interfaces
444 let interfaces = intermediate
445 .interfaces
446 .into_iter()
447 .map(|i| Self::convert_interface(i, &declared))
448 .collect::<Result<Vec<_>>>()
449 .context("Failed to convert interfaces")?;
450
451 // Convert unions
452 let unions = intermediate.unions.into_iter().map(Self::convert_union).collect::<Vec<_>>();
453
454 // Convert subscriptions
455 let subscriptions = intermediate
456 .subscriptions
457 .into_iter()
458 .map(|s| Self::convert_subscription(s, &declared))
459 .collect::<Result<Vec<_>>>()
460 .context("Failed to convert subscriptions")?;
461
462 // Convert custom directives
463 let directives = intermediate
464 .directives
465 .unwrap_or_default()
466 .into_iter()
467 .map(|d| Self::convert_directive(d, &declared))
468 .collect::<Result<Vec<_>>>()
469 .context("Failed to convert directives")?;
470
471 // Convert fact tables from Vec<IntermediateFactTable> to HashMap<String, FactTableMetadata>
472 let fact_tables = intermediate
473 .fact_tables
474 .unwrap_or_default()
475 .into_iter()
476 .map(|ft| {
477 let name = ft.table_name.clone();
478 let metadata = Self::convert_fact_table(ft);
479 (name, metadata)
480 })
481 .collect();
482
483 let mut compiled = CompiledSchema {
484 types,
485 enums,
486 input_types,
487 interfaces,
488 unions,
489 queries,
490 mutations,
491 subscriptions,
492 directives,
493 fact_tables, // Analytics metadata
494 // Refused above when non-empty (#779); an empty vector is the only reachable
495 // value, so this is not a drop.
496 observers: Vec::new(),
497 sources: intermediate.sources.clone().unwrap_or_default(), /* #573 scheduled ingress
498 * sources */
499 subscribable, // @subscribable capture-trigger declarations (#366)
500 federation: intermediate
501 .federation_config
502 .map(serde_json::from_value)
503 .transpose()
504 .context("federation_config: invalid JSON structure")?,
505 security: intermediate
506 .security
507 .map(serde_json::from_value)
508 .transpose()
509 .context("security: invalid JSON structure")?,
510 auth: intermediate
511 .auth
512 .map(serde_json::from_value)
513 .transpose()
514 .context("auth: invalid PKCE OAuth-client structure (#621)")?,
515 observers_config: intermediate
516 .observers_config
517 .map(|mut oc| {
518 // #631: `handlers` reaching this seam (SDK-authored
519 // `observers_config.handlers`, bypassing the TOML
520 // validator's #612 bail) must fail the compile, not land in
521 // the compiled schema as decoration the runtime never
522 // reads. tb_observer / the admin API is the single observer
523 // concept. An EMPTY array is inert boilerplate some
524 // exporters emit — strip it so `deny_unknown_fields` on the
525 // compiled type does not reject it.
526 match oc.get("handlers").and_then(serde_json::Value::as_array) {
527 Some(handlers) if !handlers.is_empty() => {
528 anyhow::bail!(
529 "observers_config declares {} handler(s), but compiled \
530 handlers are not a runtime concept (#631): the runtime \
531 loads observers only from the `tb_observer` table and the \
532 admin API. Define them in `tb_observer` or via \
533 `POST /api/observers`, and remove `handlers` from the \
534 authored schema.",
535 handlers.len()
536 );
537 },
538 Some(_) => {
539 if let Some(map) = oc.as_object_mut() {
540 map.remove("handlers");
541 }
542 },
543 None => {},
544 }
545 serde_json::from_value(oc).context("observers_config: invalid JSON structure")
546 })
547 .transpose()?,
548 subscriptions_config: intermediate.subscriptions_config, /* Subscriptions config from
549 * TOML */
550 validation_config: intermediate.validation_config, // Validation limits from TOML
551 debug_config: intermediate.debug_config, // Debug config from TOML
552 mcp_config: intermediate.mcp_config, // MCP config from TOML
553 rest_config: intermediate.rest_config, // REST config from TOML
554 grpc_config: intermediate.grpc_config, // gRPC config from TOML (#780)
555 naming_convention: intermediate.naming_convention, // Naming convention from TOML
556 session_variables: intermediate.session_variables.unwrap_or_default(),
557 locale: intermediate.locale,
558 hierarchies_config: intermediate.hierarchies_config,
559 changelog: intermediate.changelog_config, // Changelog exposure config from TOML
560 schema_sdl: None, // Raw GraphQL SDL
561 custom_scalars: CustomTypeRegistry::default(), // Custom scalar registry
562 ..Default::default()
563 };
564
565 // Populate custom scalars from intermediate schema
566 if let Some(custom_scalars_vec) = intermediate.custom_scalars {
567 for scalar_def in custom_scalars_vec {
568 let custom_type = Self::convert_custom_scalar(scalar_def)?;
569 compiled
570 .custom_scalars
571 .register(custom_type.name.clone(), custom_type)
572 .context("Failed to register custom scalar")?;
573 }
574 }
575
576 // Changelog exposure requires the observer system: the views it exposes read
577 // from tables (`tb_entity_change_log`, `tb_transport_checkpoint`) that the
578 // observer install convention supplies. Emitting GraphQL types over absent
579 // tables would only fail later, at runtime — reject it here instead.
580 if let Some(ref cl) = compiled.changelog {
581 if cl.expose && !compiled.observers_config.as_ref().is_some_and(|o| o.enabled) {
582 anyhow::bail!(
583 "[changelog] expose = true requires [observers] to be enabled: the tables \
584 it exposes (tb_entity_change_log, tb_transport_checkpoint) are installed by \
585 the observer system."
586 );
587 }
588 // #497: the change-log surface is a per-database stream, not a federated
589 // value type — the injected `EntityChangeLog` type and `entity_change_logs`
590 // root query are not `@shareable`. Two subgraphs in one supergraph that both
591 // `expose` it inject the identical type/root field, which `rover supergraph
592 // compose` rejects with INVALID_FIELD_SHARING. We can't detect the collision
593 // here (each subgraph compiles alone), so warn on the federated+exposed combo
594 // and point at the single-owner pattern. This is a reminder, not an error —
595 // the owning subgraph legitimately sets both.
596 if cl.expose && compiled.federation.as_ref().is_some_and(|f| f.enabled) {
597 warn!(
598 "[changelog] expose = true on a federation subgraph: EntityChangeLog / \
599 entity_change_logs are not @shareable, so expose the change-log in exactly \
600 one subgraph per supergraph (others keep capturing via [changelog] \
601 write_enabled). Two exposing subgraphs fail `rover supergraph compose` with \
602 INVALID_FIELD_SHARING. See the changelog-graphql guide for the single-owner \
603 pattern."
604 );
605 }
606 }
607
608 // Canonicalize the Trinity external identity (`id: UUID` → `id: ID`, ADR-0017)
609 // before the interface-forcing passes, so Relay `Node` / `CascadeNode` see a
610 // conformant `id: ID` on every entity. Wire-transparent (a UUID is an `ID`).
611 identity::normalize_entity_identity(&mut compiled);
612
613 // Inject synthetic Relay types (PageInfo, Node interface, XxxConnection, XxxEdge).
614 relay::inject_relay_types(&mut compiled)?;
615
616 // Inject the changelog GraphQL surface (EntityChangeLog / TransportCheckpoint
617 // types + cursor query + point lookup + upsert mutation) when opted in.
618 fraiseql_core::schema::inject_changelog(&mut compiled);
619
620 // Synthesize the typed cascade surface (CascadeNode interface + envelope
621 // types + per-mutation `<Name>Payload` wrappers) for mutations that opt in
622 // via `cascade = true`. Runs BEFORE error-union synthesis so a cascade
623 // payload becomes the success member of the result union
624 // (`<Name>Result = <Name>Payload | MutationError`). Inert with no cascade
625 // mutations.
626 cascade_types::synthesize_cascade_types(&mut compiled)?;
627
628 // Auto-synthesize a shared MutationError type + per-mutation result unions
629 // when opted in (`[fraiseql.mutations] auto_error_union`), so the runtime's
630 // success/error discrimination has a union to resolve against.
631 if options.auto_error_union {
632 mutation_error_union::synthesize_mutation_error_unions(&mut compiled);
633 }
634
635 // IR honesty (#456): `parse_field_type` has no `Input` variant, so an
636 // argument referencing a declared input type was resolved to
637 // `FieldType::Object`. Rewrite those argument types to `FieldType::Input`
638 // now that every input type is registered (after relay / rich-filter /
639 // changelog / error-union injection) so the IR honestly represents the
640 // GraphQL input position — which also makes introspection report the correct
641 // `INPUT_OBJECT` kind for those args. The runtime already accepts both
642 // shapes, so this is purely representational.
643 Self::promote_input_type_args(&mut compiled);
644
645 // Validate the compiled schema
646 Self::validate(&compiled)?;
647
648 info!("Schema conversion successful");
649 Ok(compiled)
650 }
651
652 /// Rewrite mutation/query argument types of the form `Object(name)` (and lists
653 /// thereof) to `Input(name)` when `name` is a registered input type, so the IR
654 /// honestly represents input-position references (#456). Input type names are
655 /// disjoint from output type names per the GraphQL spec, so the lookup is exact
656 /// — an output-object argument is never misclassified.
657 fn promote_input_type_args(schema: &mut CompiledSchema) {
658 fn promote(ft: &mut FieldType, input_names: &HashSet<String>) {
659 match ft {
660 FieldType::Object(name) if input_names.contains(name) => {
661 *ft = FieldType::Input(name.clone());
662 },
663 FieldType::List { item: inner, .. } => promote(inner, input_names),
664 _ => {},
665 }
666 }
667
668 let input_names: HashSet<String> =
669 schema.input_types.iter().map(|i| i.name.clone()).collect();
670 if input_names.is_empty() {
671 return;
672 }
673
674 for mutation in &mut schema.mutations {
675 for arg in &mut mutation.arguments {
676 promote(&mut arg.arg_type, &input_names);
677 }
678 }
679 for query in &mut schema.queries {
680 for arg in &mut query.arguments {
681 promote(&mut arg.arg_type, &input_names);
682 }
683 }
684 }
685
686 #[allow(clippy::cognitive_complexity)] // Reason: comprehensive schema validation with many field-level checks
687 fn validate(schema: &CompiledSchema) -> Result<()> {
688 info!("Validating compiled schema");
689
690 // Build type registry
691 let mut type_names: HashSet<String> = HashSet::new();
692 for type_def in &schema.types {
693 type_names.insert(type_def.name.to_string());
694 }
695
696 // Build interface registry. Interfaces are *also* valid query/mutation
697 // return types (a field may return an interface, narrowed via inline
698 // fragments), so register them in `type_names` too — previously omitted,
699 // which made a query/mutation returning an interface silently fail the
700 // reference check below.
701 let mut interface_names = HashSet::new();
702 for interface_def in &schema.interfaces {
703 interface_names.insert(interface_def.name.clone());
704 type_names.insert(interface_def.name.clone());
705 }
706
707 // Add input types — valid as mutation argument types (fraiseql/fraiseql#190)
708 for input_type in &schema.input_types {
709 type_names.insert(input_type.name.clone());
710 }
711
712 // Add union type names — valid as mutation/query return types
713 for union_def in &schema.unions {
714 type_names.insert(union_def.name.clone());
715 }
716
717 // Add enum type names — valid as field and argument types
718 for enum_def in &schema.enums {
719 type_names.insert(enum_def.name.clone());
720 }
721
722 // Add built-in scalars
723 for scalar in crate::schema::builtin_scalar_names() {
724 type_names.insert(scalar.to_string());
725 }
726
727 // Add **declared** custom scalars, from the registry the compiler actually carries.
728 //
729 // This used to register every type name appearing in any object field, which made a
730 // field-type typo legalize itself as a return type too (#724 item 2) — the same
731 // blindness as `SchemaValidator`, duplicated here.
732 for (name, _) in schema.custom_scalars.list_all() {
733 type_names.insert(name);
734 }
735
736 // Collect every problem, then report them together.
737 //
738 // This tier used to `bail!` on the first error, with no suggestion, after redundantly
739 // `warn!`ing the same text (#724 item 4). Errors that only surface after synthesis —
740 // relay, cascade, changelog-injected types — are validated *here*, so the same class
741 // of mistake got a materially worse experience depending on which tier caught it, and
742 // a user with three typos needed three compile cycles for information the compiler
743 // had in hand on the first.
744 let mut problems: Vec<String> = Vec::new();
745 let mut known: Vec<&str> = type_names.iter().map(String::as_str).collect();
746 known.sort_unstable();
747
748 let describe = |name: &str| -> String {
749 let similar = fraiseql_core::runtime::suggest_similar(name, &known);
750 if similar.is_empty() {
751 String::new()
752 } else {
753 format!(" (did you mean: {}?)", similar.join(", "))
754 }
755 };
756
757 // Validate queries
758 for query in &schema.queries {
759 if !type_names.contains(&query.return_type) {
760 problems.push(format!(
761 "Query '{}' references unknown type '{}'{}",
762 query.name,
763 query.return_type,
764 describe(&query.return_type)
765 ));
766 }
767
768 for arg in &query.arguments {
769 let type_name = Self::extract_type_name(&arg.arg_type);
770 if !type_names.contains(&type_name) {
771 problems.push(format!(
772 "Query '{}' argument '{}' references unknown type '{}'{}",
773 query.name,
774 arg.name,
775 type_name,
776 describe(&type_name)
777 ));
778 }
779 }
780 }
781
782 // Validate mutations
783 for mutation in &schema.mutations {
784 if !type_names.contains(&mutation.return_type) {
785 problems.push(format!(
786 "Mutation '{}' references unknown type '{}'{}",
787 mutation.name,
788 mutation.return_type,
789 describe(&mutation.return_type)
790 ));
791 }
792
793 for arg in &mutation.arguments {
794 let type_name = Self::extract_type_name(&arg.arg_type);
795 if !type_names.contains(&type_name) {
796 problems.push(format!(
797 "Mutation '{}' argument '{}' references unknown type '{}'{}",
798 mutation.name,
799 arg.name,
800 type_name,
801 describe(&type_name)
802 ));
803 }
804 }
805 }
806
807 // Validate interface implementations
808 for type_def in &schema.types {
809 for interface_name in &type_def.implements {
810 if !interface_names.contains(interface_name) {
811 let mut candidates: Vec<&str> =
812 interface_names.iter().map(String::as_str).collect();
813 candidates.sort_unstable();
814 let similar =
815 fraiseql_core::runtime::suggest_similar(interface_name, &candidates);
816 let hint = if similar.is_empty() {
817 String::new()
818 } else {
819 format!(" (did you mean: {}?)", similar.join(", "))
820 };
821 problems.push(format!(
822 "Type '{}' implements unknown interface '{interface_name}'{hint}",
823 type_def.name
824 ));
825 continue;
826 }
827
828 // Validate that the type has all fields required by the interface
829 if let Some(interface) = schema.find_interface(interface_name) {
830 for interface_field in &interface.fields {
831 let type_has_field = type_def.fields.iter().any(|f| {
832 f.name == interface_field.name
833 && f.field_type == interface_field.field_type
834 });
835 if !type_has_field {
836 problems.push(format!(
837 "Type '{}' implements interface '{interface_name}' but is \
838 missing field '{}'",
839 type_def.name, interface_field.name
840 ));
841 }
842 }
843 }
844 }
845 }
846
847 // Relationships the REST embed executor could not follow (#1266) — an undeclared
848 // target, a join column no field publishes, a target no list query returns.
849 //
850 // Borrowed wholesale from the load-time check rather than restated here: the two
851 // must agree, and the way they stop agreeing is one of them being taught a rule
852 // the other is not. `CompiledSchema::finish_load` runs the same function, which
853 // is what covers the hand-edited artifact this compile-time refusal cannot reach.
854 problems.extend(schema.relationship_violations());
855
856 // Federation `@key`s the router would refuse (#1395) — the same function the load
857 // path runs, for the same reason as the relationship check above.
858 problems.extend(schema.federation_key_violations());
859
860 if !problems.is_empty() {
861 let count = problems.len();
862 let noun = if count == 1 { "problem" } else { "problems" };
863 anyhow::bail!(
864 "Schema validation found {count} {noun}:\n{}",
865 problems.iter().map(|p| format!(" - {p}")).collect::<Vec<_>>().join("\n")
866 );
867 }
868
869 info!("Schema validation passed");
870 Ok(())
871 }
872
873 /// Extract type name from `FieldType` for validation
874 ///
875 /// Built-in types return their scalar name, Object types return the object name
876 fn extract_type_name(field_type: &FieldType) -> String {
877 // A built-in scalar's name is the one the author wrote, from the same
878 // table `parse_field_type` reads — this is that function's inverse.
879 if let Some(name) = field_type.authoring_name() {
880 return name.to_string();
881 }
882 match field_type {
883 FieldType::Scalar(name) => name.clone(),
884 FieldType::Object(name) => name.clone(),
885 FieldType::Enum(name) => name.clone(),
886 FieldType::Input(name) => name.clone(),
887 FieldType::Interface(name) => name.clone(),
888 FieldType::Union(name) => name.clone(),
889 FieldType::List { item: inner, .. } => Self::extract_type_name(inner),
890 // Reason: non_exhaustive requires catch-all for cross-crate matches
891 _ => "Unknown".to_string(),
892 }
893 }
894
895 /// Convert `IntermediateFactTable` to `FactTableMetadata`.
896 fn convert_fact_table(ft: IntermediateFactTable) -> FactTableMetadata {
897 FactTableMetadata {
898 table_name: ft.table_name,
899 type_name: ft.type_name,
900 measures: ft
901 .measures
902 .into_iter()
903 .map(|m| MeasureColumn {
904 name: m.name,
905 sql_type: Self::parse_sql_type(&m.sql_type),
906 nullable: m.nullable,
907 additivity: m.additivity,
908 })
909 .collect(),
910 dimensions: DimensionColumn {
911 name: ft.dimensions.name,
912 paths: ft
913 .dimensions
914 .paths
915 .into_iter()
916 .map(|p| DimensionPath {
917 name: p.name,
918 json_path: p.json_path,
919 data_type: p.data_type,
920 })
921 .collect(),
922 },
923 denormalized_filters: ft
924 .denormalized_filters
925 .into_iter()
926 .map(|f| FilterColumn {
927 name: f.name,
928 sql_type: Self::parse_sql_type(&f.sql_type),
929 indexed: f.indexed,
930 hierarchy: f.hierarchy,
931 })
932 .collect(),
933 calendar_dimensions: vec![],
934 native_measures: ft.native_measures,
935 native_dimension_mapping: ft.native_dimension_mapping,
936 }
937 }
938
939 /// Parse a SQL type string into a `SqlType` enum variant.
940 fn parse_sql_type(s: &str) -> SqlType {
941 match s.to_uppercase().as_str() {
942 "INT" | "INTEGER" | "SMALLINT" | "INT4" | "INT2" => SqlType::Int,
943 "BIGINT" | "INT8" => SqlType::BigInt,
944 "DECIMAL" | "NUMERIC" | "MONEY" => SqlType::Decimal,
945 "REAL" | "FLOAT" | "DOUBLE" | "FLOAT8" | "FLOAT4" | "DOUBLE PRECISION" => {
946 SqlType::Float
947 },
948 "JSONB" => SqlType::Jsonb,
949 "JSON" => SqlType::Json,
950 "TEXT" | "VARCHAR" | "STRING" | "CHAR" | "CHARACTER VARYING" => SqlType::Text,
951 "UUID" => SqlType::Uuid,
952 "TIMESTAMP" | "TIMESTAMPTZ" | "TIMESTAMP WITH TIME ZONE" | "DATETIME" => {
953 SqlType::Timestamp
954 },
955 "DATE" => SqlType::Date,
956 "BOOLEAN" | "BOOL" => SqlType::Boolean,
957 "LTREE" => SqlType::Ltree,
958 _ => SqlType::Other(s.to_string()),
959 }
960 }
961}
962
963/// Parse an authored `requires_actor` list into [`ActorType`]s (#966).
964///
965/// One parser for queries and mutations, so the two cannot disagree about which
966/// tokens exist.
967///
968/// # Why an unknown token is an error
969///
970/// `requires_actor` is an allow-list, so a dropped entry does not open the
971/// operation — it *closes* it further, and a typo'd `"human"` would refuse every
972/// human user while the schema reads as though it admits them. And a whole list
973/// that failed to parse would leave the field empty, which means *unrestricted*:
974/// the failure mode of silence is a security gate that is not there. Naming the
975/// bad token at compile time is the only reading where the author finds out.
976///
977/// # Errors
978///
979/// Returns an error naming the offending token and listing the valid ones.
980pub(crate) fn parse_requires_actor(
981 operation_kind: &str,
982 operation_name: &str,
983 tokens: &[String],
984) -> Result<Vec<fraiseql_core::security::ActorType>> {
985 use fraiseql_core::security::ActorType;
986
987 tokens
988 .iter()
989 .map(|token| {
990 ActorType::from_token(token).ok_or_else(|| {
991 anyhow::anyhow!(
992 "{operation_kind} '{operation_name}': requires_actor entry '{token}' is not a \
993 known actor type. Valid values: {}",
994 ActorType::ALL.map(ActorType::as_str).join(", ")
995 )
996 })
997 })
998 .collect()
999}