oaistream

func ConvertMessages ¶

func ConvertMessages(ctx context.Context, messages []chat.Message, id modelsdev.ID, store *modelsdev.Store, override *modelinfo.CapsOverride) []openai.ChatCompletionMessageParamUnion

ConvertMessages converts chat.Message slices to OpenAI message params, resolving attachment capabilities from the provided modelsdev.Store. A non-nil override declares the model's attachment capabilities explicitly, bypassing the models.dev lookup (issue #2741).

func ConvertMessagesWithCaps ¶ added in v1.87.0

func ConvertMessagesWithCaps(ctx context.Context, messages []chat.Message, mc modelinfo.ModelCapabilities) []openai.ChatCompletionMessageParamUnion

ConvertMessagesWithCaps is the caps-injectable variant of ConvertMessages. It is used by providers whose models are not in the models.dev catalog (e.g. Docker Model Runner), which must supply attachment capabilities explicitly rather than resolving them from a store.

func ConvertMessagesWithTransform ¶ added in v1.152.0

func ConvertMessagesWithTransform(ctx context.Context, messages []chat.Message, mc modelinfo.ModelCapabilities, transform func(int, []openai.ChatCompletionMessageParamUnion)) []openai.ChatCompletionMessageParamUnion

ConvertMessagesWithTransform decorates each source message's converted items before deferring tool attachments. The transform receives the source index, keeping provider-specific metadata attached to its original content.

func ConvertMultiContent ¶

func ConvertMultiContent(ctx context.Context, multiContent []chat.MessagePart, id modelsdev.ID, store *modelsdev.Store, override *modelinfo.CapsOverride) []openai.ChatCompletionContentPartUnionParam

ConvertMultiContent converts chat.MessagePart slices to OpenAI content parts, resolving attachment capabilities from the provided modelsdev.Store. A non-nil override declares the model's attachment capabilities explicitly, bypassing the models.dev lookup (issue #2741).

func ErrorBodyMiddleware ¶

func ErrorBodyMiddleware() option.Middleware

ErrorBodyMiddleware returns an OpenAI SDK middleware that preserves full error details in HTTP error responses.

The OpenAI SDK extracts only the "error" field from error response bodies (via gjson). When a provider returns a body without an "error" object (e.g. a string "error" field, plain text, or a different JSON structure), the details are silently lost. This middleware rewrites such responses into {"error": <original body>} so the SDK preserves the full content.

func MergeConsecutiveMessages ¶

func MergeConsecutiveMessages(openaiMessages []openai.ChatCompletionMessageParamUnion) []openai.ChatCompletionMessageParamUnion

MergeConsecutiveMessages merges consecutive system or user messages into single messages. This is needed by some local models (like those run by DMR) that don't handle consecutive same-role messages well.

func ReplayReasoning ¶ added in v1.153.0

func ReplayReasoning(providerOpts map[string]any) bool

ReplayReasoning reports whether assistant reasoning stored in the session may be replayed to a model's endpoint. Replay defaults on. An explicit provider_opts.replay_reasoning overrides the default. A value that is present but not a boolean turns replay off instead of falling back to the default: the option is how an operator keeps reasoning away from an endpoint, so a typo must not leave replay enabled.

Replaying chat.Message.ReasoningContent lets a reasoning model see its own previous reasoning after a tool call (issue #4363). Because reasoning_content is a non-standard extension, the explicit false value is the compatibility escape hatch for endpoints that reject or must not receive it.

func WithoutReplayedReasoning ¶ added in v1.153.0

func WithoutReplayedReasoning(messages []chat.Message) []chat.Message

WithoutReplayedReasoning returns messages with assistant reasoning cleared, for a model that opted out with ProviderOptReplayReasoning. Only the reasoning field is affected: content, tool calls, and media are left alone, so an assistant turn that had both reasoning and content still converts to a valid message.

The input is never mutated -- the slice is copied only when something has to change -- because the messages belong to the session, not to the provider.

func WrapOpenAIError ¶ added in v1.32.3

func WrapOpenAIError(err error) error

WrapOpenAIError wraps an OpenAI SDK error in a *modelerrors.StatusError to carry HTTP status code and Retry-After metadata for the retry loop. Non-OpenAI errors (e.g. io.EOF, network errors) pass through unchanged. Exported so openai/response_stream.go can reuse it without duplication.