Skip to content

Update Versions - #276

Open
github-actions[bot] wants to merge 2 commits into
mainfrom
changepacks/main
Open

github-actions[bot] wants to merge 2 commits into
mainfrom
changepacks/main

Conversation

@github-actions

@github-actions github-actions Bot commented Nov 14, 2026 •

Copy link
Copy Markdown
Contributor

Changepacks

vespera@0.8.0 → 0.8.1 - crates/vespera/Cargo.toml

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera_core@0.8.0 → 0.8.1 - crates/vespera_core/Cargo.toml

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera_inprocess@0.8.0 → 0.8.1 - crates/vespera_inprocess/Cargo.toml

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera_jni@0.8.0 → 0.8.1 - crates/vespera_jni/Cargo.toml

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera_macro@0.8.0 → 0.8.1 - crates/vespera_macro/Cargo.toml

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera-bridge@0.8.0 → 0.8.1 - libs/vespera-bridge/build.gradle.kts

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

vespera-bridge-gradle-plugin@0.8.0 → 0.8.1 - libs/vespera-bridge-gradle-plugin/build.gradle.kts

Patch

  • vespera-bridge audit fixes: a non-web Spring Boot app without spring-web starts again (the core autoconfiguration no longer loads PathPatternParser); on virtual threads SmartDispatchModeResolver streams the response of safe requests (like DIRECT's overflow stash on platform threads) instead of buffering it, so large downloads and Accept-less event streams work (opt back into parking SYNC with new SmartDispatchModeResolver(direct, sync, true)); 1xx / 204 / 304 responses no longer carry Content-Length: 0; unpaired UTF-16 surrogates in wire header strings are encoded as U+FFFD instead of a \uD800 escape the Rust parser rejects; a prefix-mapped DispatcherServlet path is stripped from the forwarded path by decoded segments (%61pi, //, dot segments, %2F); an ASYNC dispatch that hits Spring's async timeout answers the proxy's plain-text 503 with the request id instead of the /error page; the springdoc group-mode document endpoint is mapped from the normalized docs-path (trailing slash, blank) and a root or pattern docs-path fails the startup clearly; dispatchDirectPooled / BufferTooSmallException / DispatchMode / VesperaClient docs describe the overflow stash and poll-first parking.
  • vespera-bridge: multipart uploads through the servlet proxy now reach Rust byte-for-byte. The multipartResolver bean is wrapped in VesperaProxyMultipartResolver so DispatcherServlet no longer parses (drains) proxied multipart requests and Spring's multipart size limits no longer apply to them; Spring controllers keep normal multipart handling (opt out: vespera.bridge.multipart-passthrough=false). A multipart body consumed before the proxy is answered 500 with a clear message instead of being forwarded empty.
  • vespera-bridge: a WebFlux explicit stream that fails before its first flush hint now answers the proxy 500 instead of committing Rust's status, headers and a partial body. A servlet ASYNC client disconnect is reported as 499 (no WARN) like WebFlux. A WebFlux multipart body consumed by a WebFilter answers 500 instead of reaching Rust empty. An unparsable Rust Content-Type passes through ASYNC like the other modes. The proxy restores the caller's MDC request id, a buffered read with Content-Length 0 and Transfer-Encoding honours the cap, and a cyclic cause chain no longer hangs the async completion. A DIRECT stash entry cleared by a runtime shutdown answers the retryable 503. The native loader checks the ancestors of the extraction directory and reports an invalid vespera.native.dir as UnsatisfiedLinkError without leaking the resource stream. VesperaClient rejects a '#' in rawQuery, encodes lone surrogates as U+FFFD in query and body, and rejects decode-only charsets clearly. Gradle plugin: bundleNativeLib, verifyNativeLibs and the native jar tasks read the vespera block lazily, and arch aliases the bridge loader never looks up (arm64, amd64, armv7) are rejected. Doc and test fixes.
  • vespera-bridge: the servlet 413 for a buffered body without Content-Length no longer reports the read limit as the body's actual size. The WebFlux not-initialized startup warning names the proxy only when one exists (VesperaClient-only apps are told their client requests fail). Gradle plugin and README: autoBuildCargo builds with cargoProfile (default release), and the vesperaNative publication needs a dedicated project or a distinct artifactId. Test fixes for shutdown-on-close, async cancellation linking and the demo app.
  • vespera-bridge round-5 fixes: the async response executor bean no longer suppresses Boot's applicationTaskExecutor; POMs and JARs declare/ship the Apache-2.0 license; WebFlux streams bodyless GETs with a wildcard Accept (event streams no longer buffer forever); ASYNC bodies are written byte for byte (no Content-Disposition, no Spring 304/412); malformed Content-Type reaches Rust; loader, platform detection, config parsing, streaming executor, validation errors, DIRECT stash, client stream/header hardening; workflow template supports a Gradle build in a subdirectory
  • vespera-bridge: a VesperaRequest InputStream body whose read fails is a VesperaClientException; the proxy honours a SmartDispatchModeResolver subclass; oversize-policy=throw streams the request once instead of a 500 or a false overflow; ASYNC HEAD has no body object; async CORS dedupe and reset keep the filter headers; explicitly configured native knobs count after init; shutdown(Duration) no longer overflows; springdoc spec-paths keys with underscores bind, merged components rename consistently, public Rust operations do not inherit the host's security, and native images register the docs mapping hint; docs and weak tests fixed
  • vespera-bridge: a VesperaRequest InputStream read that fails because the native side closed the source after the response drained no longer fails the send; DispatchEvent.mode reports SYNC / STREAMING when a DIRECT request is downgraded or re-routed, and an unknown-app 404 reports its path; isInitialized() is published only after pending configure* values were applied; the public dispatchDirectPooled heap fallback parks on a virtual thread instead of pinning the carrier; Gradle plugin: extraNativeLibs without cargoRoot / targetDir bundles only the prebuilt libraries; docs on DIRECT overflow, version check, configureStreaming ranges and JaCoCo merge of isolatedJvmTest fixed
  • vespera-bridge fixes: the forwarded principal is also capped by bytes, so a user with many long authorities keeps the wire header within the Rust 1 MiB limit instead of getting 400 on every request (dropped entries are logged once); a HeaderSource callback that encodes another request on the same thread no longer corrupts the outer header; BufferTooSmallException.dispatched() tells the oversize-policy=throw rejection (nothing dispatched, requiredSize is the request size) from a response overflow; the unused byte[] ASYNC response builder is removed; the 413 for an unknown-length body over the built-in 64 MiB bound names that bound instead of a configured cap. Docs: a system property present at VesperaBridge.init fixes that native knob before Spring can override it; named-app CORS preflights reach the default app; client header values use the Latin-1 mapping; the native max-buffered-response-bytes default under Spring is 64 MiB; the Smart resolver splits on safe vs unsafe methods. Gradle plugin: native libraries are wired only into the main processResources, autoBuildCargo without cargoRoot fails with a clear message, nativeJarPlatforms without a source directory reports the intended error, and the cargo package lookup reads only package names. Demo tests: the shutdown probe timeout applies, mid-body panic cases follow the configured chunk size, and several assertions now check what they claim.
  • vespera-bridge: the native loader no longer decides who the current user is from a probe file inside the directory it is still verifying (procfs, a java.io.tmpdir probe or the user.name lookup come first; the in-directory probe is a guarded, uncached last resort), closing a local-user library swap race. The servlet proxy reports a request whose error response could not be written, and tags a failing DispatchModeResolver with the selected app. A closed child context no longer disables the DIRECT buffer pool. A spec-conforming case-insensitive getHeaders no longer forwards duplicated header values. A sub-millisecond native dispatch-timeout rounds up to 1 ms and a negative one fails startup. The WebFlux 413 with an unlimited cap names the built-in bound. Doc fixes in the README and the Gradle plugin targets KDoc.
  • vespera-bridge: the WebFlux proxy strips ;path parameters like the servlet proxy; WebFlux cors=spring applies WebFluxConfigurer#addCorsMappings to proxied paths (preflights answered by Spring); vespera.native.dir works with a missing or read-only java.io.tmpdir; ASYNC responses without a Rust Content-Type no longer get one from Accept, and ASYNC proxy errors echo the request id
  • OpenApi::merge pushes a merged document's top-level security / servers down onto its own operations instead of adopting them for every operation; a nullable schema whose oneOf / allOf / not / anyOf rejects null is wrapped in anyOf [schema, null] even when it has a type; JNI async dispatch symbols moved to their own module (no behaviour change)
  • vespera_core: OpenApi::merge no longer lets the receiver's top-level security govern a merged operation that was unauthenticated (it gets an explicit empty security list), keeps each side's path-level parameters and servers on its own operations when two apps share a path, and pushes only security (not servers) onto merged webhooks; a nullable $ref with a sibling allOf/oneOf/not is wrapped in anyOf [schema, null] so it really accepts null. vespera_jni: a null InputStream/OutputStream answers 400 on the header-first streaming symbols too, like dispatchStreaming/dispatchFullStreaming, instead of 500 JNI streaming setup failed. Weak tests tightened
  • vespera_inprocess: vespera_dispatch_into answers a NULL request with a non-zero length with the 400 wire, like vespera_dispatch, instead of VESPERA_ERR_INVALID_ARGUMENT when out happened to lie in [0, request_len). vespera_core: OpenApi::merge no longer lets an operation-less receiver path item's path-level parameters/servers govern the incoming operations (or hide the incoming path-level values). Test and README fixes: the schema_type! rename_all camelCase default is documented, a cfg-dead test fixture moved to a source-only file
  • TypedMultipartWithLimits<T, ..> documents a multipart/form-data request body, can be wrapped in Validated / ValidatedWith and derefs to T like TypedMultipart. Unknown and nameless multipart parts ignored in non-strict mode are drained through the meter, so their bytes count against max_total_bytes. MeteredField::headers exposes the part headers to custom field parsers. vespera_core: deserializing type: [T, "null"] keeps the null next to a oneOf with a {type: null} branch, a not of a non-null schema or an allOf of {type: null}; OpenApi::merge keeps the path-level parameters / servers written next to an inlined #/components/pathItems reference, and an unresolvable receiver reference no longer adopts incoming path-level parameters / servers. jni_apps! docs name the "_default" entry as the way to register the default app.
  • Docs: the JNI demo build steps publish the bridge and the Gradle plugin to mavenLocal (as CI does); extractor tables document Form on GET/HEAD as query parameters and TypedHeader names derived from T; naive date/time and unit return mappings corrected; the landing site gains 0.7 / 0.8 migration notes, WebFlux / forwarded principal / reactive-streaming / max-request-producers / springdoc UI exclude docs and a C ABI page; SKILL.md covers serve_listener. No API change.
  • C ABI: vespera_dispatch_async releases its async in-flight slot before invoking the callback (as JNI dispatchAsync does), so a callback that dispatches again is no longer answered 503 at the cap
  • schema_type!/schema! audit: field types, serde function paths and relation paths resolve in the source module; sources found in crate roots, inline and #[path] modules; storage hits must match the requested module; relation loads select by the target FK column and fail loudly when unresolvable; adapter wrappers implement Clone/Deserialize and convert in from_model; field schema/garde/deprecated/cfg attributes are carried; type-correct NOW() defaults; primary_key with auto_increment = false is not a default; raw identifiers and non-identifier names no longer panic; split rename_all is kept; tuple/unit sources are rejected; generic sources take the macro's type arguments; schema! validates pick/omit names and filters flattened structs; relation_adapters, rename and partial entries are validated with diagnostics at the offending literal
  • schema_type!: module-aware SeaORM relations, cycles and lookups (crate roots, tests/examples/bin targets, call-site imports), generic sources resolve call-site arguments and defaults correctly, shared relation adapters with serde container conversions, cfg_attr serde / deprecated / raw names carried, naming a dropped field is a compile error, typed SeaORM defaults
  • Gradle plugin: a native-jar consumer (crateName + bundleNative = false, no cargoRoot / targetDir) no longer fails tasks --all, IDE sync or tasks.withType<Sync>().configureEach {} with "Cannot query the value": bundleNativeLib / verifyNativeLibs resolve the cdylib location lazily and verifyNativeLibs names the missing setting only when it runs. cargoProfile "test" bundles from target/debug and "bench" from target/release, where Cargo writes them.
  • vespera-bridge-gradle-plugin: a build with no cdylib (and no extraNativeLibs) now fails in the new verifyNativeLibs task instead of silently producing a jar without a native library; a host platform key mapped to a musl triple keeps the host build in native//
  • vespera_inprocess: a bidirectional request body that outlives its dispatch (moved into a detached task) no longer reads the host source after the dispatch returned; a stopped response sink is traced as a client disconnect (status 499, DEBUG) inside the dispatch span instead of an ERROR 500; app names are trimmed with ASCII-only whitespace everywhere (matching the compile-time checks); vespera.h and API docs corrected
  • vespera_inprocess: a fully materialised DIRECT overflow (a large hoisted 422, a synthesized error wire) now streams its body through the body callback instead of inside the header callback, so a stashed overflow streamed by the bridge no longer answers an empty body. vespera: cron schedulers no longer probe the caller's Tokio runtime for a time driver by catching a panic (noisy stderr, and an abort under panic = "abort"); they sleep on a vespera-owned timer thread and run on the current runtime whether or not it enables timers. Test and bench integrity fixes: the dispatch_path bench clones outside the timed region, the shutdown drain test proves it waited for the lease, header-first disconnect tests stop on a failed data write, allocation budgets check the measured status, and the multipart strict / validation envelope checks pin the shipped behaviour.
  • vespera_inprocess: the 422 validation_errors hoist no longer reads a JSON array (an array error element or a top-level array) as an error object. With the tracing feature the span's app field is the resolved, trimmed app name and is never recorded for an invalid or unknown name. A shared-runtime hook registered while another thread builds the runtime is no longer left waiting. vespera_jni: a header consumer that throws on a pre-dispatch error header (null stream 400, 413, setup failure, runtime unavailable, expired stash, panic fallback) now aborts the call with VesperaStreamAbortedException instead of returning normally. vespera_core: OpenApi.paths and Operation.responses are optional on deserialize (OpenAPI 3.1); a null in a type array no longer widens a schema whose enum or composition rejects null; OpenApi::merge inlines colliding pathItems references and never places a $ref next to operations. vespera: an invalid multipart enum value answers 422 like Validated and WrongFieldType (was 400). Doc and test fixes.
  • C ABI: vespera_dispatch_into answers an out buffer below VESPERA_MIN_OUT_BYTES (1024, NULL/0 size probe included) with VESPERA_OVERFLOW and *written = 1024 without running the handler, so a panic/timeout/shutdown error never forces a re-run; an oversized request_len is a 400 wire regardless of buffer addresses; bench header ladder respects MAX_REQUEST_HEADERS; stronger streaming/inline-pull tests and refreshed docs
  • Bidirectional streaming: a started request producer whose body outlived the dispatch without being polled no longer blocks forever on the full channel (and no longer makes the dispatch wait the 2 s join grace); it gives up the send once the dispatch finished, drops its sender and frees its blocking thread, producer slot and pull closure. C ABI: an async dispatch dropped by vespera_shutdown releases its in-flight slot before its 503 callback, like a normal completion; vespera.h documents that vespera_shutdown called on a runtime worker returns without waiting. Stronger tests and corrected bench docs
  • vespera_inprocess: a request whose path already carries a query and that also sends a query field joins the two with & instead of a second ? (the old URI decoded page=2?limit=10 as one value). vespera_core: OpenApi::merge inlines a path item or webhook that is a $ref to the merged document's components.pathItems, so its operations receive the pushed-down top-level security and servers.
  • JNI: panic fallbacks of the void symbols actually run (panic wire / completed future instead of a swallowed panic); dispatchFullStreaming no longer flushes the caller's stream on flush hints; a panic after body bytes in the non-header streaming symbols throws VesperaStreamAbortedException(panicked); a DirectStash fetch panic answers a 500 panic wire instead of looking expired; dispatchAsync releases its in-flight permit before completing a future on the calling thread
  • vespera! / export_app! build files: the library, a binary, the cargo test harness and rust-analyzer get separate cache and sidecar files, so one build no longer prunes the include_str! file another still reads; the target directory is read off rustc's --out-dir (dependencies and a relative CARGO_TARGET_DIR agree with the parent; VESPERA_TARGET_DIR overrides it); openapi.json and exported specs never describe #[cfg(test)]-only routes or schemas, so every build writes the same file; one path spelled with different parameter names (/{id} vs /{user_id}) and docs_url / redoc_url / spec_url colliding with an app or merged route are compile errors; undeclared security schemes are checked without a generated document; an ambiguous merge path lists the exporting crates; stale rust-analyzer route / cron entries of renamed handlers are ignored; sibling symlinked route directories are scanned; cron jobs outside src/ get real module paths or a clear error; two #[derive(Multipart)] fields reading one multipart field are a compile error.
  • Schema validation: portable usize bounds (32-bit targets), minimum/maximum checked against the integer field type (compile error instead of a wrapping cast; fractional bounds round inward), and 422 error paths use the serde deserialize name. Multipart derive: every form_data / try_from_multipart key is read in any order, an invalid or overflowing limit is a compile error, serde(skip) / skip_deserializing fields default instead of being required, and generic structs are supported.
  • OpenAPI request/response accuracy: OptionalQuery, WithRejection and Cached extractors are documented; Option<Path> types path params; Query/Path structs follow serde from/try_from and Option fields with a custom deserializer are required; Path maps/sequences type each segment; empty query structs no longer emit a fake parameter; response parts (CookieJar, HeaderMap, header arrays, AppendHeaders, TypedHeader, Extension) carry no body; Json<(A, B)> is a JSON array; error tuples document their payload; Box/Cow are text/plain, Cow<[u8]>/[u8; N] octet-stream; NoContent is 204 and Redirect 3XX; unresolved types inside Result are description-only; the Schema check no longer flags foreign paths like serde_json::Value and now checks Path; alias defaults may reference earlier parameters.
  • Fix round-10 macro parser findings: an adjacently tagged Option newtype variant now gets its {Name}Input component (content key optional on input), externally tagged data-variant branches allow exactly one key (maxProperties: 1, dropped where the enum is flattened or an internally tagged newtype payload), an adjacently tagged unit variant always documents its optional content key as null, and defaulted tuple-struct / tuple-variant positions (or a tuple struct container default) are optional trailing items in the request view
  • vespera_macro: an internally tagged newtype over an externally tagged enum documents its unit variants as {"A": null} keys next to the tag (a unit-only inner enum no longer fails to compile), and a non-object inline payload (Uuid, NaiveDate) is a compile error instead of vanishing behind a bare discriminator. A map payload's length bounds count the tag. The request view of a mixed externally tagged enum accepts a unit variant's {"Name": null} map form. A tuple position with skip_serializing_if is optional in the response view and later positions may shift into its slot. status = N re-keys only the inferred success, never an explicit responses = [(200, T)], and 1xx / 3xx declared statuses are not labelled Error response. Every garde length rule (one per mode) applies, inner(..) reaches Box<[T]> and &[T], and a pattern using Rust-only regex syntax is skipped. Doc comments drop their common indentation like rustdoc. A SeaORM relation $ref names a known component ({Module} / {Module}Schema) or documents any value. Recursive generic instance names escape the base name. Path types each segment positionally. A Result nested in a response is axum's Ok body plus another error response. crate:: / super:: return aliases resolve in the crate root / parent module, and use anyhow; makes anyhow::Result external.
  • Fix round-5 macro parser findings: serde default/skip/flatten detection after other keys and in literals, schema(default) lookup, transparent structs per serde side, Result and HasMany schemas, enum deny_unknown_fields, flattened maps and Option in request views, Cached<Validated<..>> 422, generic Query/Path structs, Path pairs/alias/flatten, raw identifiers, module-qualified renamed types, serde_repr enum matching by position, codec/value_type internally tagged newtypes, Option<Path> checks and the route dir mod.rs module path
  • Fix round-6 macro parser findings: request-view parameters, transparent and from/try_from query/path extractors, value_type internally tagged newtypes, serde_repr and serde(other) request schemas, annotation-only garde rules, recursive generic definitions and colliding instance names, deny_unknown_fields with map-bearing flatten, flattened unit structs and unit variants, Result error status keys and Infallible errors, nested response tuples, scoped Result alias resolution, fast-path extractor checks, declared 4xx/2xx statuses and 205 bodies, wide integer examples, decorated block docs, schema(ref) with nullable or value_type, BelongsTo relations, user schemas named Json, and generic parameters inside value_type/into/from
  • Fix round-7 macro parser findings: serde(transparent) Query/Path structs follow serde's transparent-field rule (PhantomData and serde(default) fields excluded) and tuple structs, generic value_type substitution keeps commas inside angle brackets (HashMap<String, T>), request-view recursive generic instances use a .Input suffix that never collides with a type named *Input, two flattened maps write anyOf (not allOf) in the response view, a NoContent error side is labelled as a successful 204, a bare argument of unknown type keeps the string path placeholder, and doc and test assertion fixes
  • Fix round-8 macro parser findings: flatten and internally tagged payload analysis of a module-qualified renamed type (b::Foo) uses the definition its module declares, request-view propagation resolves module-qualified references through their scoped key, a qualified return-type alias (v2::ApiResult) expands the alias of the module it names and an external glob import (use axum::response::*) keeps a crate-wide alias from applying, error_status = [422] next to a Validated extractor keeps the validation-error envelope, and fallback-path tests really reach the raw-token fallback
  • Fix round-9 macro parser findings: a flattened skip_serializing(_if) member requires none of its keys in the response view, newtype / transparent / tuple / newtype-variant payloads keep their #[schema] constraints and #[garde] rules, an adjacently tagged Option newtype accepts a missing content key in the request view, an internally tagged unit payload is a closed tag object and an Option payload is its inner type, an enum without emitted variants is {not: {}}, malformed #[schema] keys on containers, tuple fields and enum variant fields are compile errors, inner(..) garde rules on Option keep explicit pattern/format, lifetime-only generics are referenced, (T) is substituted, and a same-named struct in the handler scope stops the crate-wide return alias lookup
  • OpenAPI request extraction fixes: Result<Extractor, Rejection> arguments are documented like the bare extractor (required body / query / path / header params; Result<Validated<..>> still asserts garde::Validate); Option<Form> and Result<Form, _> on GET become query parameters; a Path<(..)> tuple longer than the route's own params (nest prefix) is aligned from the end; a parameter after a static prefix (/report-{year}) is documented, text after a parameter (/img/{name}.png, which axum cannot route) is a compile error, and escaped literal braces {{x}} are no longer a parameter (documented as %7Bx%7D).
  • vespera_macro / vespera: the route cache key covers the cfg(test) verdicts of schemas and route files, so gating a module in src/lib.rs no longer serves a stale openapi.json; Set / Duration fields get Input request views without any serde attribute and unreached materialised .Input generic instances are pruned; fractional range bounds compile through nested Option layers, Option aliases and macro type groups; missing Schema derives inside tuples, arrays, slices and references are compile errors. Security schemes reject fields their type does not define and non-absolute OAuth2 / OpenID Connect URLs; server URLs with {var} templates and route folders with whitespace are rejected; repeated responses statuses are errors; a rejected #[route] keeps its handler; cron discovery scans modules of a call site outside src/ and stops at module cycles; nested inline modules inside cfg-gated ones are not rejected. schema_type!: ignore + name registers an explicitly named schema aliased by its Rust ident, schema! resolves renamed types like the OpenAPI document, partial naming a relation is an error, keyword key columns become raw fields. Fuzz, script, workflow and doc fixes.
  • vespera_macro round-11 fixes: a rejected #[vespera::cron] (invalid expression or time zone) keeps the function next to its error, so vespera! no longer adds an unresolved-function error for the job the module scan still finds; the schema_type! relation_adapters duplicate-field error points at the repeated field name. The bench gate skips version-only PRs again (Cargo.lock member versions are ignored like the Cargo.toml ones); the bench scripts remove the temp work dir they created. The axum example's raw multipart upload answers 400 on a malformed body instead of panicking; the structural OpenAPI test covers TRACE. README, rust-jni-demo README, CI comment and macro doc fixes.
  • Multipart missing fields follow serde: a container #[serde(default)] / default = "fn" fills them from the container value (the struct now needs Default, as with serde), a field default = "fn" also applies to Option / Vec fields; schema_type!(..., multipart) keeps the source's serde default / skip keys, container default and #[try_from_multipart(strict)]; serde default functions resolve from the struct's inline module and through crate:: paths; route param vs catch-all conflicts and segments after a catch-all are compile errors; generic #[route] responses types are rejected; export namespace renames no longer swap schemas; garde integer bounds are exact; schema!/schema_type! find a source by its Rust name and inside inline modules; collection / wide-integer / Date serde defaults are documented; DateTimeLocal NOW() no longer documents a fixed default
  • vespera_macro fixes: schema_type! of a SeaORM Local column with a SQL-function default (NOW()) compiles again; same-named types in sibling inline modules of one file resolve to the right module for schema! / schema_type! / default functions; schema_type! carries serde default values into the derived component; Multipart derives on generic types add the Default bounds their defaults need; a stale route or cron registration is dropped only inside rust-analyzer; two export_app! of one crate with the same name are a compile error; derived garde validation skips serde-skipped fields, does not nest flattened field paths, peels nested Option, keeps bounds beyond 2^53 exact (and loosened to the documented f64 so runtime and OpenAPI agree) and rejects non-finite bounds; a default on a $ref property is documented as a sibling; more literal defaults are resolved (negative numbers, to_owned / into / to_string, String::new / Vec::new, JSON array / object strings); a renamed self-referencing type resolves to its own component; invalid route file name suggestions are valid identifiers; OAuth2 flows reject URL fields they do not take; raw-identifier modules, Path<()> and inline modules in rust-analyzer resolve; use x::{self} aliases resolve for merge; merge route conflicts are checked before any spec file is written; a sidecar spec without a fingerprint falls back to its content hash; escaped braces and percent escapes are one route shape for overlap checks; the cron module-tree scan runs only with the cron feature and ignores a cron attribute imported from another crate.
  • vespera_macro / vespera: cron jobs start once per crate even when several vespera! apps exist, raw-identifier cron fns are not scheduled twice, custom [lib] path crate roots are scanned for cron jobs, and uncallable cron fns (generic, unsafe, methods, nested) are compile errors. #[route] bumps the file-cache epoch; module-qualified responses types and cfg-gated route conflicts are resolved; docs URLs reject route-parameter syntax; relative server URLs are accepted from the environment; fractional bounds on aliased numeric types round toward the valid range; serde(transparent) keeps the parent error path. schema_type!: a trusted storage hit uses its defining module, pub use re-exports are followed, cfg-inactive structs are skipped, partial fields rewrite a carried cfg_attr skip_serializing_if, the to diagnostic quotes the raw value, and a relation target Schema is called the way its own schema_type! generates it. Cache tags and build dirs separate profiles, feature sets and literal inputs; openapi.json is written atomically; the compile bench clears the route cache between runs.
  • vespera_macro / vespera: integer bounds on float type aliases compile and fractional bounds no aliased integer reaches reject every value; cron jobs move to a dedicated thread on a runtime without timers, start from export_app! crates, must return () and be Send, are found in #[path] modules and in nested const / closure scopes, and do not leak between targets in rust-analyzer; raw-identifier routes keep their inline-module scope. Merged child catch-alls with another name, cfg-gated routes against merged children and ambiguous bare responses types are compile errors; an unknown qualified responses type refs its last segment. schema_type!: pick = [] is rejected, text defaults containing a parenthesis stay literals, module segments starting with Schema survive the Model lookup, and a bare source name uses the call-site definition. Example and bench fixes.
  • Router codegen and build cache: duplicate (method, path) routes and routes overlapping a merged child app are compile errors (with or without a spec); export_app specs are crate-scoped, merge resolves use aliases and checks the merged app at compile time; include_str sidecars are content-addressed and written atomically; route(path = "/") maps to the file path without a trailing slash; unknown route keys list the valid ones; duplicate tags / security_schemes are errors and declared tags keep their order; dir must be a module folder under src/ (nested "api/v1" supported); symlinked route files are followed; src/.rs edits and inline-module handlers are tracked; cron jobs in src/bin crates get their real module path.
  • Schema fidelity: a struct #[serde(tag)] is documented as a required tag property in the response view (left out of the {Name}Input request view, since serde never reads it); Option of a type-less inline schema (oneOf / allOf) keeps null via anyOf, and Option of an inline string enum lists null; multipart field names use form_data(field_name) before serde rename in the request view, and the Multipart derive reads the deserialize-side serde rename / rename_all
  • Runtime/core fixes: shutdown_runtime(Duration::MAX) no longer panics after closing the gate (the runtime is shut down and cleanup runs); a whole-number constraint of exactly 2^63 is emitted as a float instead of saturating to i64::MAX; Schema deserializes JSON Schema boolean schemas (true -> {}, false -> {"not":{}}) in components.schemas and $defs, and the OpenAPI 3.0 boolean exclusiveMinimum / exclusiveMaximum (true moves minimum / maximum into the 3.1 exclusive bound); OpenAPI 3.1 fields no longer dropped on deserialize: Components pathItems / links / callbacks, Response links, Header required / deprecated / style / explode / example / examples / content, Parameter deprecated / allowEmptyValue / allowReserved / examples / content, Example externalValue, and x-* extensions on OpenApi / Info / Operation / PathItem (new public fields; Parameter / Response / Header / Example / OpenApi now implement Default, use ..Default::default() in struct literals); JNI dispatchDirect0 answers an out buffer below MIN_DIRECT_OUT_BYTES (1 KiB) with -(1024) without dispatching, so a panic / timeout wire always fits and a retry never re-runs a failing handler; the cancel registry hasher spreads sequential tokens (Fibonacci multiply); doc fixes (streaming dispatch_timeout scope, Validated 422 builder, envelope API limits).
  • Fix round-11 schema_type!, Schema and Multipart macro findings: Multipart accepts serde alias part names (and rejects alias collisions), form_data(default) fields are optional in the request view, absent partial multipart fields stay None under a container default, a field with an existing serde default keeps it over a sea_orm default_value, cfg(true/false) literals and cfg_attr(.., schema(..)) are honoured, module lookup skips cfg-dropped modules and items, follows chained and glob imports, respects private glob names and resolves bare names inside inline modules, paths through another crate report an error, default functions resolve through the module tree, skipped fields are not default-validated, the relation adapter follows clone = false, duplicate generated field names are rejected, and file-cache fingerprints track out-of-tree path modules
  • Generated schemas now match what serde writes and reads: flattened maps, closed, unit and Option structs merge into one object; alias and other are accepted in request schemas; tuple fields honour codecs and one-sided skips; time crate types, sets, Duration, schema(any), generic defaults and instance names are fixed.
  • CI/release hardening: release-artifacts.sh isolates every per-package step (a failed SBOM no longer drops all attestations) and verifies each published crate was packaged from the released commit; Gradle wrapper validation + distributionSha256Sum; tarpaulin image pinned by digest; bench gates fail when nothing was measured and trigger on toolchain/lockfile changes.
  • Cron jobs sleep in slices of at most 60 s and re-read the wall clock, so a suspend or forward clock step no longer delays a run by the remaining monotonic time; cron::shutdown_and_wait needs no Tokio time driver (returns at once when idle); a cron expression with no occurrence left now logs a warning. vespera_core: nullable enums list null, a nullable schema without type or $ref whose oneOf/allOf/not/anyOf rejects null serializes as anyOf [schema, null], Operation.callbacks is callback name -> runtime expression -> PathItem (new Callback alias), SchemaRef deserializes boolean schemas (true as {}, false as {"not":{}}), OpenApiVersion adds 3.0.4, 3.1.1 and 3.1.2. Multipart errors echo part names bounded to 64 chars with control characters escaped; the DefaultBodyLimit docs now match axum (the extractor applies it on every transport, 2 MB by default).

@github-actions
github-actions Bot force-pushed the changepacks/main branch 30 times, most recently from d70ec6f to 14daa33 Compare November 15, 2026 10:19
@github-actions
github-actions Bot force-pushed the changepacks/main branch 27 times, most recently from 37c0c02 to 50b952b Compare November 17, 2026 06:36

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants