refactor(ai): simplify sseFraming and untrace inner request-lowering helpers (#53072)

This commit is contained in:
Aiden Cline
2026-10-04 16:27:06 -05:00
committed by GitHub
Unverified
parent 3e77d5c170
commit 238d4cf5ca
11 changed files with 67 additions and 70 deletions
@@ -584,7 +584,7 @@ const serverToolResultType = (name: string): AnthropicServerToolResultType | und
return undefined
}
const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (
const lowerServerToolResult = Effect.fnUntraced(function* (
part: ToolResultPart,
providerMetadataKey: string,
) {
@@ -657,7 +657,7 @@ const citationsFromMetadata = (metadata: MediaPart["metadata"]): AnthropicDocume
const isHttpUrl = (value: string) => /^https?:\/\//i.test(value.trim())
const lowerMedia = Effect.fn("AnthropicMessages.lowerMedia")(function* (
const lowerMedia = Effect.fnUntraced(function* (
part: MediaPart,
breakpoints?: Cache.Breakpoints,
) {
@@ -847,7 +847,7 @@ const splitsLocalToolResults = (messages: LLMRequest["messages"], index: number)
return pending.size > 0
}
const lowerNativeSystemUpdate = Effect.fn("AnthropicMessages.lowerNativeSystemUpdate")(function* (
const lowerNativeSystemUpdate = Effect.fnUntraced(function* (
message: LLMRequest["messages"][number],
breakpoints: Cache.Breakpoints,
) {
@@ -862,7 +862,7 @@ const lowerNativeSystemUpdate = Effect.fn("AnthropicMessages.lowerNativeSystemUp
}
})
const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
const lowerMessages = Effect.fnUntraced(function* (
request: LLMRequest,
breakpoints: Cache.Breakpoints,
) {
@@ -285,7 +285,7 @@ const lowerToolCall = (part: ToolCallPart, normalizeID: (id: string) => string):
},
})
const lowerToolResultContent = Effect.fn("BedrockConverse.lowerToolResultContent")(function* (
const lowerToolResultContent = Effect.fnUntraced(function* (
part: ToolResultPart,
documentNames: Set<string>,
) {
@@ -305,7 +305,7 @@ const lowerToolResultContent = Effect.fn("BedrockConverse.lowerToolResultContent
return content
})
const lowerToolResult = Effect.fn("BedrockConverse.lowerToolResult")(function* (
const lowerToolResult = Effect.fnUntraced(function* (
part: ToolResultPart,
documentNames: Set<string>,
normalizeID: (id: string) => string,
@@ -322,7 +322,7 @@ const lowerToolResult = Effect.fn("BedrockConverse.lowerToolResult")(function* (
// Keep Claude and Nova tool-result images inline; put other models' images beside the result.
const keepToolImagesInline = (id: string) => id.includes("anthropic.claude-") || id.includes("amazon.nova-")
const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
const lowerMessages = Effect.fnUntraced(function* (
request: LLMRequest,
breakpoints: BedrockCache.Breakpoints,
) {
+2 -2
View File
@@ -283,7 +283,7 @@ const lowerToolConfig = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
tool: (name) => ({ functionCallingConfig: { mode: "ANY" as const, allowedFunctionNames: [name] } }),
})
const lowerContentPart = Effect.fn("Gemini.lowerContentPart")(function* (part: TextPart | MediaPart) {
const lowerContentPart = Effect.fnUntraced(function* (part: TextPart | MediaPart) {
if (part.type === "text") return { text: part.text }
return yield* GeminiGenerateContent.mediaPart("Gemini", part.media)
})
@@ -302,7 +302,7 @@ const lowerToolCall = (part: ToolCallPart, omitIds: boolean, metadataKey: string
thoughtSignature: thoughtSignature(part.providerMetadata, metadataKey),
})
const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
const lowerMessages = Effect.fnUntraced(function* (request: LLMRequest) {
const contents: GeminiContent[] = []
const metadataKey = request.model.route.providerMetadataKey ?? String(request.model.provider)
const omitCallIds = omitsFunctionCallIds(request.model.id)
@@ -182,7 +182,7 @@ type StepResult = readonly [ParserState, ReadonlyArray<LLMEvent>]
// =============================================================================
// Request Body Construction
// =============================================================================
const mediaContent = Effect.fn("GoogleInteractions.mediaContent")(function* (asset: Media.Asset) {
const mediaContent = Effect.fnUntraced(function* (asset: Media.Asset) {
if (
asset.kind !== "image" &&
asset.kind !== "audio" &&
@@ -208,7 +208,7 @@ const signature = (metadata: ProviderMetadata | undefined, key: string) => {
: undefined
}
const lowerMessages = Effect.fn("GoogleInteractions.lowerMessages")(function* (request: LLMRequest) {
const lowerMessages = Effect.fnUntraced(function* (request: LLMRequest) {
const steps: InputStep[] = []
const key = request.model.route.providerMetadataKey ?? String(request.model.provider)
for (const message of request.messages) {
@@ -275,7 +275,7 @@ const lowerMessages = Effect.fn("GoogleInteractions.lowerMessages")(function* (r
return steps
})
const lowerToolResult = Effect.fn("GoogleInteractions.lowerToolResult")(function* (part: ToolResultPart) {
const lowerToolResult = Effect.fnUntraced(function* (part: ToolResultPart) {
if (part.result.type === "json" && ProviderShared.isRecord(part.result.value)) return part.result.value
if (part.result.type !== "content") return ProviderShared.toolResultText(part)
+5 -5
View File
@@ -226,7 +226,7 @@ const MistralEvent = Schema.StructWithRest(
type MistralEvent = Schema.Schema.Type<typeof MistralEvent>
const MistralStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(MistralEvent)])
const lowerMedia = Effect.fn("MistralChat.lowerMedia")(function* (part: MediaPart) {
const lowerMedia = Effect.fnUntraced(function* (part: MediaPart) {
const mime = part.media.mediaType.toLowerCase()
const url =
ProviderShared.mediaUrl(part.media) ??
@@ -236,7 +236,7 @@ const lowerMedia = Effect.fn("MistralChat.lowerMedia")(function* (part: MediaPar
return yield* ProviderShared.invalidRequest(`Mistral Chat does not support media type ${part.media.mediaType}`)
})
const lowerUser = Effect.fn("MistralChat.lowerUser")(function* (message: LLMRequest["messages"][number]) {
const lowerUser = Effect.fnUntraced(function* (message: LLMRequest["messages"][number]) {
const content: MistralUserContent[] = []
for (const part of message.content) {
if (part.type === "text") {
@@ -260,7 +260,7 @@ const lowerToolCall = (part: ToolCallPart, normalizeID: (id: string) => string):
function: { name: part.name, arguments: ProviderShared.encodeJson(part.input) },
})
const lowerAssistant = Effect.fn("MistralChat.lowerAssistant")(function* (
const lowerAssistant = Effect.fnUntraced(function* (
message: LLMRequest["messages"][number],
normalizeID: (id: string) => string,
prefix: boolean,
@@ -298,7 +298,7 @@ const lowerAssistant = Effect.fn("MistralChat.lowerAssistant")(function* (
}
})
const lowerToolResults = Effect.fn("MistralChat.lowerToolResults")(function* (
const lowerToolResults = Effect.fnUntraced(function* (
message: LLMRequest["messages"][number],
normalizeID: (id: string) => string,
) {
@@ -335,7 +335,7 @@ const lowerToolResults = Effect.fn("MistralChat.lowerToolResults")(function* (
return output
})
const lowerMessages = Effect.fn("MistralChat.lowerMessages")(function* (request: LLMRequest) {
const lowerMessages = Effect.fnUntraced(function* (request: LLMRequest) {
const normalizeID = MistralToolID.normalizer(request)
const messages: MistralMessage[] =
request.system.length === 0
+11 -10
View File
@@ -442,7 +442,7 @@ interface ReasoningStreamItem {
// =============================================================================
// Request Lowering
// =============================================================================
export const lowerTool = Effect.fn("OpenResponses.lowerTool")(function* (protocolName: string, tool: ToolDefinition) {
export const lowerTool = Effect.fnUntraced(function* (protocolName: string, tool: ToolDefinition) {
if (tool.native !== undefined)
return yield* ProviderShared.invalidRequest(`${protocolName} does not support provider-native tool ${tool.name}`)
return {
@@ -507,7 +507,10 @@ const lowerReasoning = (part: ReasoningPart, providerMetadataKey: string): OpenR
}
}
const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
const decodeImageDetail = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenResponsesInputImage.fields.detail))
const decodeMessageMetadata = ProviderShared.validateWith(Schema.decodeUnknownEffect(MessageMetadata))
const lowerMedia = Effect.fnUntraced(function* (
part: MediaPart,
request: LLMRequest,
adapter: ProviderAdapter,
@@ -516,9 +519,8 @@ const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
const media = part.media.inline()
const providerMedia = adapter.lowerMedia?.({ part, media, request })
if (providerMedia) return providerMedia
const detail = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenResponsesInputImage.fields.detail))(
part.providerMetadata?.[metadataKey(request.model)]?.detail,
)
const rawDetail = part.providerMetadata?.[metadataKey(request.model)]?.detail
const detail = rawDetail === undefined ? undefined : yield* decodeImageDetail(rawDetail)
const mime = part.media.mediaType.toLowerCase()
const url = ProviderShared.mediaUrl(part.media)
const location = url ?? (yield* ProviderShared.requireInlineMedia(adapter.name, part.media)).dataUrl
@@ -591,7 +593,7 @@ const lowerToolResultOutput = Effect.fnUntraced(function* (
const DEFAULT_EFFORT = "medium"
const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (
const lowerMessages = Effect.fnUntraced(function* (
request: LLMRequest,
adapter: ProviderAdapter,
) {
@@ -599,9 +601,8 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (
const providerMetadataKey = metadataKey(request.model)
for (const message of request.messages) {
const metadata = yield* ProviderShared.validateWith(
Schema.decodeUnknownEffect(Schema.UndefinedOr(MessageMetadata)),
)(message.providerMetadata?.[providerMetadataKey])
const rawMetadata = message.providerMetadata?.[providerMetadataKey]
const metadata = rawMetadata === undefined ? undefined : yield* decodeMessageMetadata(rawMetadata)
if (message.role === "system") {
const update = effortUpdate(message)
if (update) {
@@ -755,7 +756,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (
return input
})
export const lowerConversation = Effect.fn("OpenResponses.lowerConversation")(function* (
export const lowerConversation = Effect.fnUntraced(function* (
request: LLMRequest,
adapter: ProviderAdapter,
) {
+6 -6
View File
@@ -360,7 +360,7 @@ const lowerToolCall = (
extra_content: decodeExtraContent(part.providerMetadata?.[options.providerMetadataKey]?.extraContent),
})
const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart) {
const lowerMedia = Effect.fnUntraced(function* (part: MediaPart) {
// Chat Completions accepts PDFs, and no other documents, as inline `file` parts; file URLs are not supported.
if (part.media.mediaType.toLowerCase() === "application/pdf")
return {
@@ -406,7 +406,7 @@ const lowerReasoningDetail = (detail: ReasoningDetail) => {
const isKimiDetail = (detail: { readonly type: string }) => detail.type === "summary" || detail.type === "encrypted"
const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
const lowerUserMessage = Effect.fnUntraced(function* (
message: OpenAIChatRequestMessage,
options: LoweringOptions,
) {
@@ -430,7 +430,7 @@ const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
return { role: "user" as const, content }
})
const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(function* (
const lowerAssistantMessage = Effect.fnUntraced(function* (
message: OpenAIChatRequestMessage,
configuredField: string | undefined,
requireReasoning: boolean,
@@ -495,7 +495,7 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
return { ...result, [field]: reasoningText }
})
const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
const lowerToolMessages = Effect.fnUntraced(function* (
message: OpenAIChatRequestMessage,
options: LoweringOptions,
) {
@@ -532,7 +532,7 @@ const toolMessage = (toolCallID: string, text: string, cacheControl: OpenAIChatC
content: cacheControl === undefined ? text : [{ type: "text" as const, text, cache_control: cacheControl }],
})
const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
const lowerMessage = Effect.fnUntraced(function* (
message: OpenAIChatRequestMessage,
reasoningField: string | undefined,
requireReasoning: boolean,
@@ -544,7 +544,7 @@ const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
return (yield* lowerToolMessages(message, options)).messages
})
const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request: LLMRequest, options: LoweringOptions) {
const lowerMessages = Effect.fnUntraced(function* (request: LLMRequest, options: LoweringOptions) {
const system: OpenAIChatMessage[] =
request.system.length === 0
? []
@@ -157,7 +157,7 @@ const nativeImageTool = (tool: ToolDefinition) => {
return Schema.is(OpenAIResponsesImageGenerationTool)(native) ? native : undefined
}
const lowerTool = Effect.fn("OpenAIResponses.lowerTool")(function* (tool: ToolDefinition) {
const lowerTool = Effect.fnUntraced(function* (tool: ToolDefinition) {
const native = nativeImageToolInput(tool)
if (native !== undefined) {
if (Schema.is(OpenAIResponsesImageGenerationTool)(native)) return native
@@ -168,7 +168,7 @@ const lowerTool = Effect.fn("OpenAIResponses.lowerTool")(function* (tool: ToolDe
// Native namespaces hold only function tools, so deeper levels flatten into
// the leaf names the same way non-native protocols flatten the whole tree.
const lowerToolEntry = Effect.fn("OpenAIResponses.lowerToolEntry")(function* (tool: ToolEntry) {
const lowerToolEntry = Effect.fnUntraced(function* (tool: ToolEntry) {
if (tool.type === "tool") return yield* lowerTool(tool)
// OpenAI requires a namespace description; fall back to a generic one so a
// missing description never blocks the request.
+2 -2
View File
@@ -154,7 +154,7 @@ export const wrapSystemUpdate = (parts: ReadonlyArray<{ readonly text: string }>
* raw retrieved, tool, or web content into privileged updates: keep untrusted
* data in ordinary user/tool messages instead.
*/
export const systemUpdateText = Effect.fn("ProviderShared.systemUpdateText")(function* (
export const systemUpdateText = Effect.fnUntraced(function* (
route: string,
message: LLMRequest["messages"][number],
) {
@@ -167,7 +167,7 @@ export const systemUpdateText = Effect.fn("ProviderShared.systemUpdateText")(fun
})
/** Lower an unsupported privileged update into visible, in-order user text. */
export const wrappedSystemUpdate = Effect.fn("ProviderShared.wrappedSystemUpdate")(function* (
export const wrappedSystemUpdate = Effect.fnUntraced(function* (
route: string,
message: LLMRequest["messages"][number],
) {
@@ -76,7 +76,7 @@ function documentName(filename: string | undefined, names: Set<string>) {
return name
}
const mediaBase64 = Effect.fn("BedrockMedia.mediaBase64")(function* (part: MediaPart) {
const mediaBase64 = Effect.fnUntraced(function* (part: MediaPart) {
const media = yield* ProviderShared.requireInlineMedia("Bedrock Converse", part.media)
const bytes = yield* Effect.fromResult(Encoding.decodeBase64(media.base64)).pipe(
Effect.mapError((cause) =>
@@ -91,7 +91,7 @@ const mediaBase64 = Effect.fn("BedrockMedia.mediaBase64")(function* (part: Media
// document block. Image MIME types not in `IMAGE_FORMATS` (e.g. `image/svg+xml`)
// get an image-specific error so the caller knows it's a format-support issue,
// not a kind-detection issue.
export const lower = Effect.fn("BedrockMedia.lower")(function* (part: MediaPart, documentNames: Set<string>) {
export const lower = Effect.fnUntraced(function* (part: MediaPart, documentNames: Set<string>) {
const mime = part.media.mediaType.toLowerCase()
const imageFormat = IMAGE_FORMATS[mime as keyof typeof IMAGE_FORMATS]
if (imageFormat) {
+27 -31
View File
@@ -1,5 +1,5 @@
import { Effect, Stream } from "effect"
import { makeParser, type Event } from "effect/unstable/encoding/Sse"
import { makeParser } from "effect/unstable/encoding/Sse"
import { AIError, InvalidProviderOutputError } from "../schema/index.js"
/**
@@ -42,43 +42,39 @@ export const sseFraming = (
Stream.decodeText(),
Stream.mapAccumEffect(
() => {
const output: Event[] = []
const output: string[] = []
return {
output,
parser: makeParser((event) => {
if (event._tag === "Event") output.push(event)
if (
event._tag === "Event" &&
(events === undefined || events.has(event.event)) &&
event.data.length > 0 &&
// Some OpenAI-compatible proxies serialize an empty flush as a bare
// `data: null`, between events or after `[DONE]`. No protocol has a
// null event, so it carries nothing and must not abort the stream.
event.data !== "null" &&
// Vertex AI partner models (e.g. `xai/grok-4.6`) send their SSE
// keepalive comment as `data: : keepalive` while reasoning.
event.data !== ": keepalive" &&
(event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message"))
)
output.push(event.data)
}),
}
},
(state, chunk) =>
Effect.gen(function* () {
const error = state.parser.feed(chunk)
if (error)
return yield* new AIError({
reason: new InvalidProviderOutputError({
route: "sse",
message: error.message,
body: chunk,
cause: error,
}),
})
return [state, state.output.splice(0)] as const
}),
(state, chunk) => {
const error = state.parser.feed(chunk)
if (!error) return Effect.succeed([state, state.output.splice(0)] as const)
const reason = new InvalidProviderOutputError({
route: "sse",
message: error.message,
body: chunk,
cause: error,
})
return Effect.fail(new AIError({ reason }))
},
),
Stream.filter(
(event) =>
(events === undefined || events.has(event.event)) &&
event.data.length > 0 &&
// Some OpenAI-compatible proxies serialize an empty flush as a bare
// `data: null`, between events or after `[DONE]`. No protocol has a
// null event, so it carries nothing and must not abort the stream.
event.data !== "null" &&
// Vertex AI partner models (e.g. `xai/grok-4.6`) send their SSE
// keepalive comment as `data: : keepalive` while reasoning.
event.data !== ": keepalive" &&
(event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
),
Stream.map((event) => event.data),
)
/** Server-Sent Events framing. Used by every JSON-streaming HTTP provider. */