fix(server): preserve response and streaming contracts

.github/workflows/ci.yml

@@ -29,12 +29,17 @@ jobs:
 
   openai-sdk-compat:
     runs-on: ubuntu-latest
+    timeout-minutes: 15
     steps:
       - uses: actions/checkout@v4
         with:
           submodules: true
       - uses: dtolnay/rust-toolchain@stable
       - uses: Swatinem/rust-cache@v2
+        with:
+          workspaces: |
+            . -> target
+            examples/server-openai-responses -> target
       - uses: actions/setup-python@v5
         with:
           python-version: "3.12"
@@ -50,6 +55,34 @@ jobs:
           official_openai_python_sdk_matches_generated_axum_server --
           --ignored --exact --nocapture
 
+  anthropic-sdk-compat:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+        with:
+          workspaces: |
+            . -> target
+            examples/server-anthropic-messages -> target
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: pip
+          cache-dependency-path: tests/python/anthropic_sdk_compat/requirements.txt
+      - name: Install pinned official Anthropic Python SDK
+        run: python -m pip install --requirement tests/python/anthropic_sdk_compat/requirements.txt
+      - name: Test official SDK against generated Axum server
+        env:
+          ANTHROPIC_COMPAT_PYTHON: python
+        run: >-
+          cargo test --test anthropic_sdk_compat_test
+          official_anthropic_python_sdk_matches_generated_axum_server --
+          --ignored --exact --nocapture
+
   clippy:
     runs-on: ubuntu-latest
     steps:
@@ -86,6 +119,10 @@ jobs:
       - uses: dtolnay/rust-toolchain@stable
       - uses: Swatinem/rust-cache@v2
       - run: scripts/install-smoke.sh
+        env:
+          # Keep cargo-install artifacts inside the target tree restored by
+          # rust-cache; the script still isolates and removes its package root.
+          CARGO_TARGET_DIR: ${{ github.workspace }}/target/install-smoke
 
   msrv:
     runs-on: ubuntu-latest
@@ -103,6 +140,10 @@ jobs:
   spec-compile:
     if: github.event_name != 'schedule' && github.event_name != 'workflow_dispatch'
     runs-on: ubuntu-latest
+    env:
+      # Share scratch-crate dependencies across runs through rust-cache's
+      # root-workspace target directory.
+      SPEC_COMPILE_TARGET_DIR: ${{ github.workspace }}/target/spec-compile
     steps:
       - uses: actions/checkout@v4
       - uses: dtolnay/rust-toolchain@stable
@@ -114,6 +155,8 @@ jobs:
     if: github.event_name == 'schedule' || github.event_name == 'workflow_dispatch'
     runs-on: ubuntu-latest
     timeout-minutes: 240
+    env:
+      SPEC_COMPILE_TARGET_DIR: ${{ github.workspace }}/target/spec-compile
     steps:
       - uses: actions/checkout@v4
       - uses: dtolnay/rust-toolchain@stable

CHANGELOG.md

@@ -4,6 +4,36 @@ All notable changes are recorded here. This project follows semantic versioning,
 with one pre-1.0 qualification: a minor release may change generated Rust APIs
 when correcting output that was wrong or incomplete on the wire.
 
+## [0.10.0] - 2026-07-26
+
+### Added
+
+- Every-PR compatibility coverage for the generated Anthropic Messages server
+  through the pinned official Python SDK, including unary and SSE responses.
+
+### Changed
+
+- Regenerated server response enums now use the declared status in bodyless and
+  SSE variant names and require a runtime status for wildcard/default variants.
+  This is a source-breaking correction for existing server trait implementations.
+
+### Fixed
+
+- Config-driven `server list` and `server add` now apply
+  `generator.schema_extensions`, so overlay-provided operations and SSE media
+  types match generation.
+- Schema extensions accept the documented JSON, YAML, and YML formats with
+  path-rich parse errors.
+- Generated server response enums retain reusable Response Object references,
+  including structurally compatible local refs stored outside
+  `components.responses`, plus bodyless status codes, vendor/problem JSON media
+  types, SSE status codes, and runtime status values for wildcard/default
+  responses.
+- Server generation rejects response sets that contain only unsupported media
+  types and reports normalized Rust identifier collisions between distinct tags.
+- Server example tests use Cargo's current integration-test binary instead of
+  a potentially stale hard-coded `target/debug` executable.
+
 ## [0.9.1] - 2026-07-26
 
 ### Fixed
@@ -152,7 +182,8 @@ when correcting output that was wrong or incomplete on the wire.
   signed enum values, recursive unions, parameter collisions, optional request
   bodies, range response codes, and path-segment encoding.
 
-[Unreleased]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.9.1...HEAD
+[Unreleased]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.10.0...HEAD
+[0.10.0]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.9.1...v0.10.0
 [0.9.1]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.9.0...v0.9.1
 [0.9.0]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.8.0...v0.9.0
 [0.8.0]: https://github.com/gpu-cli/openapi-to-rust/compare/v0.7.0...v0.8.0

Cargo.toml

@@ -6,7 +6,7 @@ exclude = ["tmp"]
 
 [package]
 name = "openapi-to-rust"
-version = "0.9.1"
+version = "0.10.0"
 edition = "2024"
 rust-version = "1.88.0"
 autobins = false

README.md

@@ -245,7 +245,7 @@ Two complete examples are in the repo:
 | `streaming.rs` | SSE streaming **client** with event parsing (when configured) |
 | `server/mod.rs` | Module re-exports for the server (when `[server]` is set) |
 | `server/api.rs` | `trait <Tag>Api { async fn <op>(&self, …) -> <Op>Response; }` per tag |
-| `server/errors.rs` | `enum <Op>Response { Ok(T), BadRequest(E), …, OkStream(Sse<…>) }` with `IntoResponse` |
+| `server/errors.rs` | Status-typed response enums with JSON, bodyless, wildcard/default, and status-correct SSE variants plus `IntoResponse` |
 | `server/router.rs` | Per-tag `Router` factory; combined `build_router<…>(…)` for multi-tag selections |
 | `server/validation.rs` | Offline JSON Schema validators and sanitized Problem Details rejections |
 | `mod.rs` | Module declarations + re-exports |
@@ -665,10 +665,11 @@ The compile tiers are intentionally different:
   OpenAPI specs. The bundled Gitea Swagger 2.0 document is reported as skipped.
 - Local `scripts/spec-compile.sh` runs the same full tier; pass spec names as
   arguments for a smaller targeted run.
-- Every pull request also starts the generated OpenAI Responses Axum example
-  on loopback and exercises it through the pinned official OpenAI Python SDK.
-  The compatibility gate covers unary and streaming Responses, input-item
-  path/query parameters, and the organization costs query.
+- Every pull request also starts the generated OpenAI Responses and Anthropic
+  Messages Axum examples on loopback and exercises them through pinned official
+  Python SDKs. The compatibility gates cover unary and streaming responses,
+  with the OpenAI row also covering input-item path/query parameters and the
+  organization costs query.
 
 ## Examples
 

examples/server-anthropic-messages/Cargo.toml

@@ -4,6 +4,8 @@ version = "0.0.0"
 edition = "2021"
 publish = false
 
+[workspace]
+
 [dependencies]
 async-trait = "0.1"
 axum = "0.8"

examples/server-anthropic-messages/README.md

@@ -14,14 +14,18 @@ cargo run --manifest-path examples/server-anthropic-messages/Cargo.toml
 curl -s http://127.0.0.1:3001/v1/messages \
   -H 'content-type: application/json' \
   -d '{"model":"claude-x","max_tokens":50,"messages":[{"role":"user","content":"hi"}]}'
+
+# Stream the response as server-sent events:
+curl -N -s http://127.0.0.1:3001/v1/messages \
+  -H 'content-type: application/json' \
+  -d '{"model":"claude-x","max_tokens":50,"stream":true,"messages":[{"role":"user","content":"hi"}]}'
 ```
 
-**Streaming caveat.** Anthropic's published OpenAPI spec doesn't
-declare `text/event-stream` on the 200 response, so the generator
-can't emit an `OkStream` variant. Tracked as `openapi-generator-in6`
-— the planned fix is a schema-extension overlay that adds the
-streaming content type, after which this example will gain a
-streaming branch with no manual changes.
+Anthropic's published OpenAPI spec does not declare `text/event-stream` on the
+200 response. The example config applies `sse-overlay.json` through
+`generator.schema_extensions`, adding that response media type before analysis
+so generation emits both `Ok(Message)` and `OkStream` response variants.
 
-The integration test in `tests/server_examples_test.rs` guarantees
-the example keeps building.
+The integration test in `tests/server_examples_test.rs` regenerates and tests
+the example. PR CI additionally calls the generated server with the pinned
+official Anthropic Python SDK, covering both unary and streaming responses.

examples/server-anthropic-messages/src/main.rs

@@ -32,6 +32,7 @@ use futures_util::stream;
 use gen::CreateMessageParams;
 use gen::server::{MessagesPostResponse, ServerApi, server_api_router, sse_response};
 use std::convert::Infallible;
+use std::io::Write;
 
 #[derive(Clone)]
 struct AppState;
@@ -62,7 +63,7 @@ fn messages_unary() -> MessagesPostResponse {
         model: gen::Model::Custom("claude-demo".into()),
         role: gen::MessageRole::Assistant,
         stop_details: None,
-        stop_reason: None,
+        stop_reason: Some(gen::StopReason::EndTurn),
         stop_sequence: None,
         r#type: gen::MessageType::Message,
         usage: gen::Usage {
@@ -71,7 +72,7 @@ fn messages_unary() -> MessagesPostResponse {
             cache_read_input_tokens: None,
             inference_geo: None,
             input_tokens: 0,
-            output_tokens: 0,
+            output_tokens: 2,
             server_tool_use: None,
             service_tier: None,
         },
@@ -86,7 +87,10 @@ fn messages_streaming() -> MessagesPostResponse {
     // subset; production code mirrors the full sequence from the
     // upstream model.
     let events = stream::iter(vec![
-        sse_event("message_start", r#"{"type":"message_start","message":{"id":"msg_demo"}}"#),
+        sse_event(
+            "message_start",
+            r#"{"type":"message_start","message":{"id":"msg_demo","type":"message","role":"assistant","model":"claude-demo","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":0,"output_tokens":0}}}"#,
+        ),
         sse_event(
             "content_block_start",
             r#"{"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}}"#,
@@ -100,6 +104,10 @@ fn messages_streaming() -> MessagesPostResponse {
             r#"{"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"world"}}"#,
         ),
         sse_event("content_block_stop", r#"{"type":"content_block_stop","index":0}"#),
+        sse_event(
+            "message_delta",
+            r#"{"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"output_tokens":2}}"#,
+        ),
         sse_event("message_stop", r#"{"type":"message_stop"}"#),
     ]);
     MessagesPostResponse::OkStream(sse_response(events))
@@ -112,8 +120,14 @@ fn sse_event(name: &str, data: &str) -> Result<Event, Infallible> {
 #[tokio::main]
 async fn main() {
     let app = server_api_router(AppState);
-    let listener = tokio::net::TcpListener::bind("127.0.0.1:3001").await.unwrap();
-    println!("listening on http://{}", listener.local_addr().unwrap());
+    let bind_address =
+        std::env::var("OPENAPI_SERVER_ADDR").unwrap_or_else(|_| "127.0.0.1:3001".to_string());
+    let listener = tokio::net::TcpListener::bind(&bind_address).await.unwrap();
+    println!(
+        "OPENAPI_SERVER_URL=http://{}",
+        listener.local_addr().unwrap()
+    );
+    std::io::stdout().flush().unwrap();
     axum::serve(listener, app).await.unwrap();
 }