Repository navigation
Merge main into spiceai - #41
Merged
Merged
Conversation
* utils.rs for wasm * image api group for wasm * download.rs not availbale in wasm * config.rs for wasm * completions.rs for wasm * client.rs for wasm * chat.rs for wasm * types/input_source.rs for wasm * async-openai/src/types/impls.rs for wasm * async-openai/src/assistants/runs.rs for wasm * async-openai/src/assistants/threads.rs for wasm * async-openai/src/audio/speech.rs for wasm * async-openai/src/audio/transcriptions.rs for wasm * async-openai/src/responses/responses_.rs for wasm * async-openai/src/types/assistants/stream.rs for wasm * async-openai/src/types/audio/sdk.rs for wasm * async-openai/src/types/audio/stream.rs for wasm * async-openai/src/types/chat/chat_.rs for wasm * async-openai/src/types/completions/completion.rs for wasm * async-openai/src/types/images/stream.rs for wasm * async-openai/src/types/images/sdk.rs for wasm * async-openai/src/types/images/impls.rs for wasm * async-openai/src/types/realtime/client_event.rs for wasm * async-openai/src/types/responses/stream.rs for wasm * async-openai/src/error.rs for wasm * README.md for wasm * add examples/wasm-responses * async-openai/Cargo.toml for wasm * add getrandom in example cargo toml * async-openai/src/types/images/impls.rs warning fix for wasm * wasm-pr-checks.yml * cargo fmt * fix: targets -> target * fix clippy * update docs * remove clippy from wasm pr checks * fix wasm-pr-checks.yml * add .cargo/config.toml https://docs.rs/getrandom/latest/getrandom/\#opt-in-backends * updated Cargo.toml * fix: wasm pr check * fix example
…it#512) Add serde attribute to omit None values for include_usage and include_obfuscation. Add tests to verify correct serialization and deserialization behavior.
* updated spec - and corresponding code * cargo clippy fixes * cargo clippy fixes for assistants feature
* fix: make open_page url optional * fix: replace `find` with `find_in_page` * Update async-openai/src/types/responses/response.rs * Update async-openai/src/types/responses/response.rs * Update async-openai/src/types/responses/response.rs --------- Co-authored-by: Himanshu Neema <himanshun.iitkgp@gmail.com>
Allowed values are in_memory and 24h https://platform.openai.com/docs/guides/prompt-caching#configure-per-request
* feat: add Skills API support Add types and API wrappers for the Skills API endpoints including skill CRUD, skill versions, and content downloads. * add name in RealtimeServerEventResponseFunctionCallArgumentsDone * skill types * updates to response type * image generation action in ImageGenTool * new types in image.rs * updated spec * enviroment * handle keepalive in stream_mapped_raw_events * cargo fmt * add new feature flags * add Skills in Readme
…ion (64bit#527) default EasyInputMessage.type to message when missing in JSON input default WebSearchApproximateLocation.type to approximate when missing in JSON input add focused response-types integration tests for both deserialize-default + canonical serialization register responses_input_item_serde integration test in Cargo.toml with required-features Refs: - openapi.documented.yml: EasyInputMessage (required role+content; type const not required) - openapi.documented.yml: WebSearchApproximateLocation (type has default approximate) - https://github.com/openai/openai-node/blob/03ee9c5f42c6ce9636fbbb365a97746feb17e8cd/src/resources/responses/responses.ts#L411 - https://github.com/openai/openai-python/blob/656e3cab4a18262a49b961d41293367e45ee71b9/src/openai/types/responses/easy_input_message_param.py#L13 - https://github.com/openai/openai-go/blob/aeccc5f54d88662b744cee4eb3d256142d6ec76b/responses/response.go#L10755
* updates to response types * update MCPTool type * updated ImageDetail type * updated Voice type * video character and edit apis + types * fix examples compilation from updated types * updated spec * cargo clippy fix * cargo fmt
While this is a required field in the OpenAPI spec, the spec is wrong; OpenAI itself accepts a missing `id` and popular tools like OpenCode send requests without an `id`. This aligns the behavior with the real world usage to allow deserializing OpenCode (and other) requests
* add agents.md * agents.md * update adgents * updated to async-openai/src/types/vectorstores/vector_store.rs: * updates to async-openai/src/types/responses/conversation.rs * updated types in async-openai/src/types/responses/response.rs * update example for updated types * updated spec
) - Bump reqwest from 0.12 to 0.13 - Add "query" feature (now opt-in in reqwest 0.13) - Replace abandoned reqwest-eventsource 0.6 with aha-reqwest-eventsource 0.1 (fork that supports reqwest 0.13), aliased via package rename so no source-level changes are needed - Update rustls/rustls-webpki-roots feature mappings: reqwest 0.13 unified TLS roots behind the 'rustls' feature (uses rustls-platform-verifier); separate native-roots/webpki-roots features were removed upstream
64bit#538) * Add From<OutputItem> for Item / InputItem (multi-turn responses) The Responses API requires Reasoning items to be echoed back into the next request's input alongside the FunctionCall items they preceded. Without that pairing the server rejects the request with: Item 'fc_...' of type 'function_call' was provided without its required 'reasoning' item: 'rs_...' In the Python and Node SDKs the canonical pattern is `input_list += response.output` because their input/output item types are the same. In Rust the schemas split into Item vs OutputItem, with several variants that have *Resource (output, required id/status) and *ItemParam (input, optional id/status) flavors, so callers had to hand-write a per-variant conversion to do what's a one-liner in the official SDKs. Add From<OutputItem> for Item and From<OutputItem> for InputItem, plus the supporting per-variant From impls (status enum mappings, shell environment/outcome mappings, apply_patch operation mappings, resource-to-param mappings). Required id/status fields fold into Some(...) on the way in; the output-only created_by annotation is dropped because the input schema doesn't accept it. The canonical SDK pattern now works: input_items.extend(response.output.into_iter().map(InputItem::from)); Includes round-trip tests covering the headline reasoning + function_call pairing case plus the resource-to-param folding for ApplyPatch, Compaction, and FunctionCallOutput. New responses-multi-turn-reasoning example demonstrates the pattern with a reasoning model. Closes 64bit#492. * chore: fmt
* initial implementation of middleware * allow error to be Into<OpenAIError> * documentation and default reqwest service * namspaced in middleware * park tower error handling for now TODO * working compilation of all crates * support for middleware in wasm * impl Default for ReqwestService * add tower example * add tower-wasm example * introduce Boxed variant * add both display and debug * update comment * updated doc * updated doc * refactor: rename http_executor to executor; create a middleware dir for doc.rs * move retry stuff in its own module * update examples * document and log retry headers * add openai retry layer * fix logical error in openai layer * optimize openai retry layer * return early on success * update doc * updated doc * add middleware to full * add middleware section * nest retry stuff inside ::middleware::retry * change of module path * updated doc * updated doc * update doc * updated doc * fix the warning * update tower example * cleanup * middleware.md * only compile it for wasm * cleanup client * cleanup client * cleanup * renmae midddlware * update readme * cleanup streaming and form creation * fix test * add larger delay for retrieval to work * make image-edit example work * allow tower user input to be random * update readme * update workflows to build middleware feature * fix warnings * fix warnings * only keep wasm32-unknown-unknown for workflows checks * bump async-openai-macros version * update doc comment * remove duplicate From tower BoxError for wasm/non-wasm * remove redundant doc comment * cleanup comment * in readme use link to github instead of relative path - so that it works on crates.io * remove comment * update lib.rs WASM section * add docs on middleware error handling
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
build_request_builder applies the Config's default headers, then the JSON
factory appended Content-Type: application/json. A Config that forwards a
caller's headers (the Codex gateway) already supplies Content-Type, so the
outgoing request carried two Content-Type headers. The ChatGPT Codex backend
rejects that with {"detail":"Unsupported content type"}.
Set Content-Type by inserting (replacing) on the built request instead of
appending on the builder, so there is always exactly one.
When a Responses request is deserialized and reserialized by a gateway (Spice in front of Codex), replayed input items broke two ways: hard deserialize failures and silent field drops. Real Codex request bodies exposed: - OutputMessage.status was required, but a replayed assistant message omits it -> the whole input array failed to deserialize (HTTP 422). Now optional. - OutputTextContent.annotations was required, but a replayed assistant message omits it -> same 422. Now defaults to empty and is skipped when empty. - InputMessage had no id field, so message ids were dropped on reserialize. - Reasoning dropped its context field (Codex sends context: all_turns). - CustomToolCall dropped its status field on reserialize. - OutputTextContent.logprobs reserialized as null; now skipped when None. Adds a sweep test that deserializes real captured Codex bodies into CreateResponse and diffs a reserialize against the original to catch both hard failures and silent field loss.
Replace local paths, installation/session ids, and other machine-specific values in the captured fixtures with neutral placeholders. The sweep only checks item structure and self-consistent roundtrip, so placeholder values keep it passing while keeping the corpus free of local details.
Codex tool-call turns rely on x-codex-turn-state: the backend returns it as a response header and replays it on the next request in the same turn. Client::post/post_stream discarded response headers entirely, so a caller had no way to read it back. Add create_with_headers and create_stream_with_headers alongside the existing create/create_stream.
The merge from main kept the spiceai [features] block, which dropped the rustls-no-provider feature. The TLS feature CI matrix still builds it.
krinart
approved these changes
Oct 9, 2026
…iceai # Conflicts: # async-openai/tests/ser_de.rs
krinart
approved these changes
Oct 9, 2026
…iceai # Conflicts: # async-openai/tests/ser_de.rs
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
20 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
mainbranch into thespiceaibranchupgrade-async-openaiagent skill documenting the required two-stage update process and focused validationValidation
cargo check -p async-openai --features responsescargo test -p async-openai --features response-types --test responses_input_item_serde --test responses_output_to_inputcargo test -p async-openai --features responses --lib## Codex gateway fixesThese commits let Spice act as a gateway in front of the Codex CLI. A gateway deserializes each Responses request into
CreateResponseand serializes it again before it forwards the request. Real Codex traffic showed two failure types: a hard deserialize failure (HTTP 422) and a silent loss of fields. This change corrects both.Request transport:
Content-Typeheader on each JSON request. The client applied the config headers and then appended a secondContent-Type. A forwarding config already suppliesContent-Type, so the request carried two. The Codex backend rejects twoContent-Typeheaders withUnsupported content type.Responses input items (found with a sweep of captured Codex bodies):
OutputMessage.statusis now optional. A replayed assistant message omits it, so a required field failed the whole input array.OutputTextContent.annotationsnow defaults to empty and is skipped when empty. A replayed assistant message omits it.InputMessagekeeps a new optionalid. The gateway forwards it instead of dropping it.Reasoningkeeps a new optionalcontext(for exampleall_turns).CustomToolCallkeeps a new optionalstatus.OutputTextContent.logprobsis skipped when it isNone, so the body does not gain anullfield.Validation
cargo test -p async-openai --no-default-features --features "rustls full" --test codex_gateway_sweepCreateResponseand compares a re-serialize against the original to find hard failures and lost fields. All fixtures pass.