provider/openai

The OpenAI adapter reaches the GPT family. It ships inside the core module, so it costs you no dependencies — it is written against net/http and encoding/json.

Reference#

func New(apiKey string, opts ...Option) *Provider

Parameters

apiKey — sent as Authorization: Bearer.
opts — functional options, applied in order.

Option	Signature	Default	Description
`WithBaseURL`	`func WithBaseURL(url string) Option`	OpenAI's own host	Overrides the API host — an Azure deployment, a proxy, or the sandbox.
`WithHTTPClient`	`func WithHTTPClient(hc *http.Client) Option`	http.DefaultClient	Supplies the *http.Client used for every request.
`WithOrganization`	`func WithOrganization(org string) Option`	—	Sets the OpenAI-Organization header, for accounts that bill per org.
`WithProject`	`func WithProject(project string) Option`	—	Sets the OpenAI-Project header.
`WithMaxTokensField`	`func WithMaxTokensField(field string) Option`	max_completion_tokens	Chooses which wire field carries Request.MaxTokens. OpenAI renamed max_tokens to max_completion_tokens; a host that has not followed needs the old name.

Caveats

It shares one implementation with openaicompat (internal/oai), so request mapping, response parsing, streaming and error classification are byte-identical. See the comparison.
Thinking is ignored unless both Enabled and Effort are set. So &Thinking{Enabled: false} does nothing — an explicit "off" has no wire representation here.
Effort: max is sent literally, and OpenAI does not define that value. Expect a 400.
ProviderOptions is a shallow top-level merge. Setting a nested object replaces it wholesale.
ToolResult.IsError is lossy: it prefixes "error: " to the content and drops the flag entirely when the content is empty.
Model listing reports almost nothing — no display name, context window or output cap. The gap is upstream.
StopStopSequence is never produced. A stop-sequence hit arrives as plain stop, so it becomes StopEndTurn.

Usage#

Constructing

goCompiles

client := skyl.New(openai.New(os.Getenv("OPENAI_API_KEY")))

client := skyl.New(openai.New(os.Getenv("OPENAI_API_KEY")))

An organisation and project

goCompiles

p := openai.New(key,
	openai.WithOrganization(os.Getenv("OPENAI_ORG")),
	openai.WithProject(os.Getenv("OPENAI_PROJECT")),
)

p := openai.New(key,
	openai.WithOrganization(os.Getenv("OPENAI_ORG")),
	openai.WithProject(os.Getenv("OPENAI_PROJECT")),
)

Azure or a host that has not renamed max_tokens

goCompiles

// OpenAI renamed max_tokens to max_completion_tokens; a host that has not
// followed needs the old name.
p := openai.New(key,
	openai.WithBaseURL("https://my-resource.openai.azure.com/openai/v1"),
	openai.WithMaxTokensField("max_tokens"),
)

// OpenAI renamed max_tokens to max_completion_tokens; a host that has not
// followed needs the old name.
p := openai.New(key,
	openai.WithBaseURL("https://my-resource.openai.azure.com/openai/v1"),
	openai.WithMaxTokensField("max_tokens"),
)

Turning reasoning off, which Thinking cannot do here

goCompiles

req.ProviderOptions = map[string]any{"reasoning_effort": "minimal"}

req.ProviderOptions = map[string]any{"reasoning_effort": "minimal"}

Against the sandbox

goCompiles

p := openai.New("sandbox-key",
	openai.WithBaseURL("http://127.0.0.1:8099/openai/v1"))

p := openai.New("sandbox-key",
	openai.WithBaseURL("http://127.0.0.1:8099/openai/v1"))

Troubleshooting#

Turning reasoning off did not reduce cost

&Thinking{Enabled: false} is ignored by this adapter. Send reasoning_effort through ProviderOptions.

A 400 mentioning max_tokens

The host expects the older field name. Use WithMaxTokensField("max_tokens").

My generationConfig-style override lost other fields

The merge is shallow: a nested object you set replaces the whole object. Restate every sibling key skyl would have set — including messages and tools, which are the dangerous ones.

Streaming reports zero usage

The host must honour stream_options.include_usage. OpenAI itself does; many compatible hosts do not.