fix(streaming): move error event handling before thread.* check#3452
fix(streaming): move error event handling before thread.* check#3452C1-BA-B1-F3 wants to merge 5 commits into
Conversation
Guard against empty type args when using bare dict or list annotations (no type parameters, e.g. instead of ). Previously: - construct_type() raised ValueError on bare dict - construct_type() raised IndexError on bare list - _transform_recursive() raised IndexError on bare dict Now these cases gracefully return the data as-is, matching the behavior of parameterized types when the data already matches. Fixes openai#3338, openai#3341
…I responses Fixes openai#3179 The API can return null for the action field in web search calls (e.g., when the search is still in progress or the action hasn't been determined yet). This change makes the action field Optional[Action] with a default of None, matching the actual API behavior. This allows users to safely check action.type without getting an AttributeError when action is None.
When NO_PROXY or other proxy environment variables contain newline characters (common in Docker, .env files, or shell scripts), httpx's get_environment_proxies() only splits by comma and fails with InvalidURL. This fix adds sanitize_proxy_env_vars() that removes newlines from proxy environment variables before httpx reads them. Fixes openai#3303
The ActionSearchSource type only supported type='url' with a required url field, but the API returns specialized data sources (weather, sports, finance) with type='api', a name field, and url=None. Changes: - Add 'api' to the type Literal in both response and param types - Make url optional (absent when type='api') - Add optional name field for the data source identifier Fixes openai#2736
The error event check inside the `if sse.event.startswith("thread.")` block
was unreachable because "error" never starts with "thread.". This meant
error events from the SSE stream were silently ignored in the thread event
path, instead of raising an APIError.
Fix: Move error event handling to the top of the event loop, before the
thread.* check, so error events are always handled regardless of context.
Fixes openai#2796
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: b6554a06b8
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| raise APIError( | ||
| message="An error occurred during streaming", | ||
| request=self.response.request, | ||
| body=data, | ||
| ) |
There was a problem hiding this comment.
Preserve assistant error events
When Assistants streams emit event: error, the payload is the top-level ErrorObject used by AssistantStreamEvent.ErrorEvent, and the assistant stream handlers still have an event.event == "error" branch. This unconditional fallback now raises APIError for that valid event shape because it lacks a nested error key, so consumers of Stream[AssistantStreamEvent]/the assistant helpers can no longer receive the typed error event; the async block has the same behavior.
Useful? React with 👍 / 👎.
Problem
The error event check inside the block in is unreachable because never starts with . This means error events from the SSE stream are silently ignored in the thread event path, instead of raising an .
This is a regression from commit ("fix(streaming): correct indentation").
Fix
Move error event handling to the top of the event loop, before the check, so error events are always handled regardless of context.
Before (unreachable code):
After (error events handled first):
Changes
Fixes #2796