File size: 18,414 Bytes
eb1202c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 | import { Effect, Schema } from "effect"
import { Route } from "../route/client"
import { Auth } from "../route/auth"
import { Endpoint } from "../route/endpoint"
import { Framing } from "../route/framing"
import { Protocol } from "../route/protocol"
import {
LLMEvent,
Usage,
type FinishReason,
type JsonSchema,
type LLMRequest,
type MediaPart,
type ProviderMetadata,
type TextPart,
type ToolCallPart,
type ToolDefinition,
type ToolContent,
} from "../schema"
import { JsonObject, optionalArray, ProviderShared } from "./shared"
import { GeminiToolSchema } from "./utils/gemini-tool-schema"
import { Lifecycle } from "./utils/lifecycle"
import { ToolSchemaProjection } from "./utils/tool-schema"
const ADAPTER = "gemini"
const MEDIA_MIMES = new Set<string>(ProviderShared.MEDIA_MIMES)
export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
// =============================================================================
// Request Body Schema
// =============================================================================
const GeminiTextPart = Schema.Struct({
text: Schema.String,
thought: Schema.optional(Schema.Boolean),
thoughtSignature: Schema.optional(Schema.String),
})
const GeminiInlineDataPart = Schema.Struct({
inlineData: Schema.Struct({
mimeType: Schema.String,
data: Schema.String,
}),
})
const GeminiFunctionCallPart = Schema.Struct({
functionCall: Schema.Struct({
name: Schema.String,
args: Schema.Unknown,
}),
thoughtSignature: Schema.optional(Schema.String),
})
const GeminiFunctionResponsePart = Schema.Struct({
functionResponse: Schema.Struct({
name: Schema.String,
response: Schema.Unknown,
}),
})
const GeminiContentPart = Schema.Union([
GeminiTextPart,
GeminiInlineDataPart,
GeminiFunctionCallPart,
GeminiFunctionResponsePart,
])
const GeminiContent = Schema.Struct({
role: Schema.Literals(["user", "model"]),
parts: Schema.Array(GeminiContentPart),
})
type GeminiContent = Schema.Schema.Type<typeof GeminiContent>
const GeminiSystemInstruction = Schema.Struct({
parts: Schema.Array(Schema.Struct({ text: Schema.String })),
})
const GeminiFunctionDeclaration = Schema.Struct({
name: Schema.String,
description: Schema.String,
parameters: Schema.optional(JsonObject),
})
const GeminiTool = Schema.Struct({
functionDeclarations: Schema.Array(GeminiFunctionDeclaration),
})
const GeminiToolConfig = Schema.Struct({
functionCallingConfig: Schema.Struct({
mode: Schema.Literals(["AUTO", "NONE", "ANY"]),
allowedFunctionNames: optionalArray(Schema.String),
}),
})
const GeminiThinkingConfig = Schema.Struct({
thinkingBudget: Schema.optional(Schema.Number),
includeThoughts: Schema.optional(Schema.Boolean),
})
const GeminiGenerationConfig = Schema.Struct({
maxOutputTokens: Schema.optional(Schema.Number),
temperature: Schema.optional(Schema.Number),
topP: Schema.optional(Schema.Number),
topK: Schema.optional(Schema.Number),
stopSequences: optionalArray(Schema.String),
thinkingConfig: Schema.optional(GeminiThinkingConfig),
})
const GeminiBodyFields = {
contents: Schema.Array(GeminiContent),
systemInstruction: Schema.optional(GeminiSystemInstruction),
tools: optionalArray(GeminiTool),
toolConfig: Schema.optional(GeminiToolConfig),
generationConfig: Schema.optional(GeminiGenerationConfig),
}
const GeminiBody = Schema.Struct(GeminiBodyFields)
export type GeminiBody = Schema.Schema.Type<typeof GeminiBody>
const GeminiUsage = Schema.Struct({
cachedContentTokenCount: Schema.optional(Schema.Number),
thoughtsTokenCount: Schema.optional(Schema.Number),
promptTokenCount: Schema.optional(Schema.Number),
candidatesTokenCount: Schema.optional(Schema.Number),
totalTokenCount: Schema.optional(Schema.Number),
})
type GeminiUsage = Schema.Schema.Type<typeof GeminiUsage>
const GeminiCandidate = Schema.Struct({
content: Schema.optional(GeminiContent),
finishReason: Schema.optional(Schema.String),
})
const GeminiEvent = Schema.Struct({
candidates: optionalArray(GeminiCandidate),
usageMetadata: Schema.optional(GeminiUsage),
})
type GeminiEvent = Schema.Schema.Type<typeof GeminiEvent>
interface ParserState {
readonly finishReason?: string
readonly hasToolCalls: boolean
readonly nextToolCallId: number
readonly usage?: Usage
readonly lifecycle: Lifecycle.State
readonly reasoningSignature?: string
}
// =============================================================================
// Tool Schema Conversion
// =============================================================================
// Tool-schema conversion has two distinct concerns:
//
// 1. Sanitize — fix common authoring mistakes Gemini rejects: integer/number
// enums (must be strings), `required` entries that don't match a property,
// untyped arrays (`items` must be present), and `properties`/`required`
// keys on non-object scalars. Mirrors OpenCode's historical Gemini rules.
//
// 2. Project — lossy mapping from JSON Schema to Gemini's schema dialect:
// drop empty objects, derive `nullable: true` from `type: [..., "null"]`,
// coerce `const` to `[const]` enum, recurse properties/items, propagate
// only an allowlisted set of keys (description, required, format, type,
// properties, items, allOf, anyOf, oneOf, minLength). Anything outside the
// allowlist (e.g. `additionalProperties`, `$ref`) is silently dropped.
//
// Sanitize runs first, then project. The implementation lives in
// `utils/gemini-tool-schema` so this protocol keeps the same shape as the other
// provider protocols.
// =============================================================================
// Request Lowering
// =============================================================================
const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema) => ({
name: tool.name,
description: tool.description,
parameters: GeminiToolSchema.convert(inputSchema),
})
const lowerToolConfig = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
ProviderShared.matchToolChoice("Gemini", toolChoice, {
auto: () => ({ functionCallingConfig: { mode: "AUTO" as const } }),
none: () => ({ functionCallingConfig: { mode: "NONE" as const } }),
required: () => ({ functionCallingConfig: { mode: "ANY" as const } }),
tool: (name) => ({ functionCallingConfig: { mode: "ANY" as const, allowedFunctionNames: [name] } }),
})
const lowerUserPart = Effect.fn("Gemini.lowerUserPart")(function* (part: TextPart | MediaPart) {
if (part.type === "text") return { text: part.text }
const media = yield* ProviderShared.validateMedia("Gemini", part, MEDIA_MIMES)
return { inlineData: { mimeType: media.mime, data: media.base64 } }
})
const googleMetadata = (metadata: Record<string, unknown>): ProviderMetadata => ({ google: metadata })
const thoughtSignature = (providerMetadata: ProviderMetadata | undefined) => {
const google = providerMetadata?.google
return ProviderShared.isRecord(google) && typeof google.thoughtSignature === "string"
? google.thoughtSignature
: undefined
}
const lowerToolCall = (part: ToolCallPart) => ({
functionCall: { name: part.name, args: part.input },
thoughtSignature: thoughtSignature(part.providerMetadata),
})
const lowerMessages = Effect.fn("Gemini.lowerMessages")(function* (request: LLMRequest) {
const contents: GeminiContent[] = []
for (const message of request.messages) {
if (message.role === "system") {
const part = yield* ProviderShared.wrappedSystemUpdate("Gemini", message)
const previous = contents.at(-1)
if (previous?.role === "user")
contents[contents.length - 1] = { role: "user", parts: [...previous.parts, { text: part.text }] }
else contents.push({ role: "user", parts: [{ text: part.text }] })
continue
}
if (message.role === "user") {
const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
for (const part of message.content) {
if (!ProviderShared.supportsContent(part, ["text", "media"]))
return yield* ProviderShared.unsupportedContent("Gemini", "user", ["text", "media"])
parts.push(yield* lowerUserPart(part))
}
contents.push({ role: "user", parts })
continue
}
if (message.role === "assistant") {
const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
for (const part of message.content) {
if (!ProviderShared.supportsContent(part, ["text", "reasoning", "tool-call"]))
return yield* ProviderShared.unsupportedContent("Gemini", "assistant", ["text", "reasoning", "tool-call"])
if (part.type === "text") {
parts.push({ text: part.text })
continue
}
if (part.type === "reasoning") {
parts.push({ text: part.text, thought: true, thoughtSignature: thoughtSignature(part.providerMetadata) })
continue
}
if (part.type === "tool-call") {
parts.push(lowerToolCall(part))
continue
}
}
contents.push({ role: "model", parts })
continue
}
const parts: Array<Schema.Schema.Type<typeof GeminiContentPart>> = []
for (const part of message.content) {
if (!ProviderShared.supportsContent(part, ["tool-result"]))
return yield* ProviderShared.unsupportedContent("Gemini", "tool", ["tool-result"])
if (part.result.type !== "content") {
parts.push({
functionResponse: {
name: part.name,
response: {
name: part.name,
content: ProviderShared.toolResultText(part),
},
},
})
continue
}
const content: ReadonlyArray<ToolContent> = part.result.value
const text = content.filter((item) => item.type === "text").map((item) => item.text)
parts.push({
functionResponse: {
name: part.name,
response: {
name: part.name,
content: text.join("\n"),
},
},
})
for (const item of content) {
if (item.type === "text") continue
const media = yield* ProviderShared.validateToolFile("Gemini", item, MEDIA_MIMES)
parts.push({ inlineData: { mimeType: media.mime, data: media.base64 } })
}
}
contents.push({ role: "user", parts })
}
return contents
})
const geminiOptions = (request: LLMRequest) => request.providerOptions?.gemini
const thinkingConfig = (request: LLMRequest) => {
const value = geminiOptions(request)?.thinkingConfig
if (!ProviderShared.isRecord(value)) return undefined
const result = {
thinkingBudget: typeof value.thinkingBudget === "number" ? value.thinkingBudget : undefined,
includeThoughts: typeof value.includeThoughts === "boolean" ? value.includeThoughts : undefined,
}
return Object.values(result).some((item) => item !== undefined) ? result : undefined
}
const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMRequest) {
const toolsEnabled = request.tools.length > 0 && request.toolChoice?.type !== "none"
const generation = request.generation
const toolSchemaCompatibility = request.model.compatibility?.toolSchema
const generationConfig = {
maxOutputTokens: generation?.maxTokens,
temperature: generation?.temperature,
topP: generation?.topP,
topK: generation?.topK,
stopSequences: generation?.stop,
thinkingConfig: thinkingConfig(request),
}
return {
contents: yield* lowerMessages(request),
systemInstruction:
request.system.length === 0 ? undefined : { parts: [{ text: ProviderShared.joinText(request.system) }] },
tools: toolsEnabled
? [
{
functionDeclarations: request.tools.map((tool) =>
lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility)),
),
},
]
: undefined,
toolConfig: toolsEnabled && request.toolChoice ? yield* lowerToolConfig(request.toolChoice) : undefined,
generationConfig: Object.values(generationConfig).some((value) => value !== undefined)
? generationConfig
: undefined,
}
})
// =============================================================================
// Stream Parsing
// =============================================================================
// Gemini reports `promptTokenCount` (inclusive total) with a
// `cachedContentTokenCount` subset. `candidatesTokenCount` is *exclusive*
// of `thoughtsTokenCount` — visible-only, not a total — so we sum the two
// to produce the inclusive `outputTokens` the rest of the contract expects.
const mapUsage = (usage: GeminiUsage | undefined) => {
if (!usage) return undefined
const cached = usage.cachedContentTokenCount
const nonCached = ProviderShared.subtractTokens(usage.promptTokenCount, cached)
// `candidatesTokenCount` is visible-only; sum with thoughts to produce the
// inclusive `outputTokens` the contract expects. Only compute the total
// when the visible component is reported — otherwise we'd fabricate an
// inclusive number from a partial breakdown.
const outputTokens =
usage.candidatesTokenCount !== undefined ? usage.candidatesTokenCount + (usage.thoughtsTokenCount ?? 0) : undefined
return new Usage({
inputTokens: usage.promptTokenCount,
outputTokens,
nonCachedInputTokens: nonCached,
cacheReadInputTokens: cached,
reasoningTokens: usage.thoughtsTokenCount,
totalTokens: ProviderShared.totalTokens(usage.promptTokenCount, outputTokens, usage.totalTokenCount),
providerMetadata: { google: usage },
})
}
const mapFinishReason = (finishReason: string | undefined, hasToolCalls: boolean): FinishReason => {
if (finishReason === "STOP") return hasToolCalls ? "tool-calls" : "stop"
if (finishReason === "MAX_TOKENS") return "length"
if (
finishReason === "IMAGE_SAFETY" ||
finishReason === "RECITATION" ||
finishReason === "SAFETY" ||
finishReason === "BLOCKLIST" ||
finishReason === "PROHIBITED_CONTENT" ||
finishReason === "SPII"
)
return "content-filter"
if (finishReason === "MALFORMED_FUNCTION_CALL") return "error"
return "unknown"
}
const finish = (state: ParserState): ReadonlyArray<LLMEvent> =>
state.finishReason || state.usage
? (() => {
const events: LLMEvent[] = []
const lifecycle = state.reasoningSignature
? Lifecycle.reasoningEnd(
state.lifecycle,
events,
"reasoning-0",
googleMetadata({ thoughtSignature: state.reasoningSignature }),
)
: state.lifecycle
Lifecycle.finish(lifecycle, events, {
reason: mapFinishReason(state.finishReason, state.hasToolCalls),
usage: state.usage,
})
return events
})()
: []
const step = (state: ParserState, event: GeminiEvent) => {
const nextState = {
...state,
usage: event.usageMetadata ? (mapUsage(event.usageMetadata) ?? state.usage) : state.usage,
}
const candidate = event.candidates?.[0]
if (!candidate?.content)
return Effect.succeed([
{ ...nextState, finishReason: candidate?.finishReason ?? nextState.finishReason },
[],
] as const)
const events: LLMEvent[] = []
let hasToolCalls = nextState.hasToolCalls
let lifecycle = nextState.lifecycle
let nextToolCallId = nextState.nextToolCallId
let reasoningSignature = nextState.reasoningSignature
for (const part of candidate.content.parts) {
if ("thoughtSignature" in part && part.thoughtSignature && "thought" in part && part.thought)
reasoningSignature = part.thoughtSignature
if ("text" in part && part.text.length > 0) {
if (part.thought) {
lifecycle = Lifecycle.reasoningDelta(
lifecycle,
events,
"reasoning-0",
part.text,
part.thoughtSignature ? googleMetadata({ thoughtSignature: part.thoughtSignature }) : undefined,
)
continue
}
lifecycle = Lifecycle.reasoningEnd(
lifecycle,
events,
"reasoning-0",
reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
)
lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", part.text)
continue
}
if ("functionCall" in part) {
const input = part.functionCall.args
const id = `tool_${nextToolCallId++}`
lifecycle = Lifecycle.reasoningEnd(
lifecycle,
events,
"reasoning-0",
reasoningSignature ? googleMetadata({ thoughtSignature: reasoningSignature }) : undefined,
)
lifecycle = Lifecycle.stepStart(lifecycle, events)
events.push(
LLMEvent.toolCall({
id,
name: part.functionCall.name,
input,
providerMetadata: part.thoughtSignature
? googleMetadata({ thoughtSignature: part.thoughtSignature })
: undefined,
}),
)
hasToolCalls = true
}
}
return Effect.succeed([
{
...nextState,
hasToolCalls,
lifecycle,
nextToolCallId,
reasoningSignature,
finishReason: candidate.finishReason ?? nextState.finishReason,
},
events,
] as const)
}
// =============================================================================
// Protocol And Gemini Route
// =============================================================================
/**
* The Gemini protocol — request body construction, body schema, and the
* streaming-event state machine. Used by Google AI Studio Gemini and (once
* registered) Vertex Gemini.
*/
export const protocol = Protocol.make({
id: ADAPTER,
body: {
schema: GeminiBody,
from: fromRequest,
},
stream: {
event: Protocol.jsonEvent(GeminiEvent),
initial: () => ({ hasToolCalls: false, nextToolCallId: 0, lifecycle: Lifecycle.initial() }),
step,
onHalt: finish,
},
})
export const route = Route.make({
id: ADAPTER,
provider: "google",
protocol,
// Gemini's path embeds the model id and pins SSE framing at the URL level.
endpoint: Endpoint.path(({ request }) => `/models/${request.model.id}:streamGenerateContent?alt=sse`, {
baseURL: DEFAULT_BASE_URL,
}),
auth: Auth.none,
framing: Framing.sse,
})
export * as Gemini from "./gemini"
|