fix(parser): recognize raw json_schema response format in auto-parsing checks

[hsusul]

Problem

When providing a raw json_schema response format object ({ type: 'json_schema', json_schema: {...} }) in response_format (Chat Completions) or text.format (Responses API) without wrapping it in an SDK helper like zodResponseFormat, hasAutoParseableInput returned false.

Root Cause

hasAutoParseableInput in both src/lib/parser.ts and src/lib/ResponsesParser.ts checked isAutoParsableResponseFormat(format), which only returns true when the `` property is present on the format object. Consequently, maybeParseChatCompletion and `maybeParseResponse` bypassed auto-parsing, leaving the fallback `JSON.parse(content)` logic unreachable and returning `parsed: null` / `output_parsed: null`.

Solution

Updated hasAutoParseableInput in src/lib/parser.ts and src/lib/ResponsesParser.ts to recognize objects with type === 'json_schema'.

Validation

• Added unit tests in tests/lib/parser.test.ts and tests/lib/ResponsesParser.test.ts verifying maybeParseChatCompletion and maybeParseResponse with raw json_schema inputs.
• Verified with pnpm build, pnpm test, pnpm lint, pnpm format, and git diff --check.

Generated Code Impact

None. Changes are strictly confined to manually maintained code in src/lib/ and tests/lib/.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0c57565d03

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Return a usable type for raw schema results

When a raw schema reaches this new branch through responses.stream(...).finalResponse(), ExtractParsedContentFromParams still resolves its parsed type to null because it recognizes only branded AutoParseableTextFormat values; the analogous Chat Completions extractor has the same problem. The runtime now returns the JSON.parse result while output_parsed, content parsed, and message.parsed are declared to be null, preventing TypeScript consumers from using the newly supported result without a cast. Update both extractors to represent raw JSON-schema results with a usable type such as unknown or a JSON value type.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Parse raw schemas in stream completion events

When a raw json_schema format is used with chat.completions.stream(), this branch makes finalChatCompletion() return parsed JSON, but the stream's content.delta snapshots and content.done event still report parsed: null: #getAutoParseableResponseFormat() only accepts branded formats, and both event paths are gated through it. Consumers listening to the documented parsing events therefore receive a result inconsistent with the final completion; the raw-schema case also needs to be recognized in those streaming paths, using generic JSON parsing where no $parseRaw callback exists.

Useful? React with 👍 / 👎.

[jbeckwith-oai]

Raw-schema recognition is currently added only to the coarse hasAutoParseableInput() gate, while the type extractors and streaming parser still use the old brand-only definition. That split leaves two blocking inconsistencies:

1. Both ExtractParsedContentFromParams aliases still resolve an unbranded { type: "json_schema", ... } format to null. The methods now return parsed JSON at runtime while message.parsed, content[].parsed, and output_parsed are declared as null. Please represent raw-schema output with a usable type (for example unknown or a JSON value type) in both Chat Completions and Responses, with compile-time coverage.

2. ChatCompletionStream still gates partial/final event parsing through #getAutoParseableResponseFormat(), which accepts only branded helper formats. I reproduced a raw-schema stream where both content.delta.parsed values and content.done.parsed were null, but finalChatCompletion().choices[0].message.parsed was { city: "SF" }. Please unify the parsing decision so raw schemas are handled consistently in stream events too, and add regression coverage.

This is also a maintainability issue: adding the same special-case recognition at only two entry gates creates competing definitions of “auto-parseable.” A canonical predicate/parser path should own the branded-callback versus generic-JSON behavior.

Validation at 0c57565d03bd75c029b3f658b42e6e9adde60588: the three focused parser/stream suites passed (28 tests, 26 snapshots); changed-file Prettier and ESLint passed; TypeScript passed; package build passed; and git diff --check passed. The streaming inconsistency above was reproduced against the built package.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1b161ba320

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve raw-schema parsed types for typed params

When callers pass a params variable typed as the public ChatCompletionStreamParams/ResponseStreamParams rather than an inline literal, Params['response_format'] is the whole optional response-format union, so this conditional does not extend ResponseFormatJSONSchema and ParsedT still resolves to null. The fresh case is typed stream params: the new hasAutoParseableInput() gate now parses raw json_schema streams via JSON.parse, so finalChatCompletion().choices[].message.parsed (and the analogous Responses output_parsed) are declared null while containing parsed JSON; detect the json_schema member of the union here and mirror the fix in ResponsesParser.

Useful? React with 👍 / 👎.