@neurocode-ai/llm 1.18.8

package/src/protocols/shared.ts

@@ -0,0 +1,326 @@
+import { Buffer } from "node:buffer"
+import { Effect, Schema, Stream } from "effect"
+import * as Sse from "effect/unstable/encoding/Sse"
+import { Headers, HttpClientRequest } from "effect/unstable/http"
+import {
+  InvalidProviderOutputReason,
+  InvalidRequestReason,
+  LLMError,
+  type ContentPart,
+  type LLMRequest,
+  type MediaPart,
+  type ToolFileContent,
+  type TextPart,
+  type ToolResultPart,
+} from "../schema"
+import { isRecord } from "../utils/record"
+export { isRecord }
+
+export const Json = Schema.fromJsonString(Schema.Unknown)
+export const decodeJson = Schema.decodeUnknownSync(Json)
+export const encodeJson = Schema.encodeSync(Json)
+const isJson = Schema.is(Schema.Json)
+export const JsonObject = Schema.Record(Schema.String, Schema.Unknown)
+export const optionalArray = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.Array(schema))
+export const optionalNull = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.NullOr(schema))
+
+/**
+ * Streaming tool-call accumulator. Adapters that build a tool call across
+ * multiple `tool-input-delta` chunks store the partial JSON input string here
+ * and finalize it with `parseToolInput` once the call completes.
+ */
+export interface ToolAccumulator {
+  readonly id: string
+  readonly name: string
+  readonly input: string
+}
+
+/**
+ * `Usage.totalTokens` policy shared by every route. Honors a provider-
+ * supplied total; otherwise falls back to `inputTokens + outputTokens` only
+ * when at least one is defined. Returns `undefined` when neither input nor
+ * output is known so routes don't publish a misleading `0`.
+ *
+ * Under the additive `LLM.Usage` contract, `inputTokens` and `outputTokens`
+ * are the non-cached input and visible output only. The provider-supplied
+ * `total` is the source of truth when present; the computed fallback
+ * under-counts cache and reasoning by design and exists mainly so
+ * Anthropic-style providers (which don't surface a total) still get a
+ * sensible aggregate on the input + output axes.
+ */
+export const totalTokens = (
+  inputTokens: number | undefined,
+  outputTokens: number | undefined,
+  total: number | undefined,
+) => {
+  if (total !== undefined) return total
+  if (inputTokens === undefined && outputTokens === undefined) return undefined
+  return (inputTokens ?? 0) + (outputTokens ?? 0)
+}
+
+/**
+ * Subtract `subtrahend` from `total`, clamping to zero if the provider
+ * reports a non-sensical breakdown (e.g. `cached_tokens > prompt_tokens`).
+ * Used by protocol mappers when deriving a non-overlapping breakdown field
+ * from a provider's inclusive total — `nonCachedInputTokens` from
+ * `inputTokens - cacheReadInputTokens - cacheWriteInputTokens`.
+ *
+ * If `total` is `undefined`, returns `undefined` (we don't fabricate
+ * counts). If `subtrahend` is `undefined`, returns `total` unchanged. The
+ * provider-native breakdown stays available on `Usage.native` for debugging.
+ */
+export const subtractTokens = (total: number | undefined, subtrahend: number | undefined): number | undefined => {
+  if (total === undefined) return undefined
+  if (subtrahend === undefined) return total
+  return Math.max(0, total - subtrahend)
+}
+
+/**
+ * Sum a list of optional token counts, returning `undefined` only when
+ * every value is `undefined` (so we don't fabricate a `0`). Used by
+ * protocol mappers to derive the inclusive `inputTokens` total from a
+ * provider that natively reports a non-overlapping breakdown
+ * (e.g. Anthropic, whose `input_tokens` is already non-cached only).
+ */
+export const sumTokens = (...values: ReadonlyArray<number | undefined>): number | undefined => {
+  if (values.every((value) => value === undefined)) return undefined
+  return values.reduce((acc: number, value) => acc + (value ?? 0), 0)
+}
+
+export const eventError = (route: string, message: string, raw?: string) =>
+  new LLMError({
+    module: "ProviderShared",
+    method: "stream",
+    reason: new InvalidProviderOutputReason({ route, message, raw }),
+  })
+
+export const parseJson = (route: string, input: string, message: string) =>
+  Effect.try({
+    try: () => decodeJson(input),
+    catch: () => eventError(route, message, input),
+  })
+
+/**
+ * Join the `text` field of a list of parts with newlines. Used by routes
+ * that flatten system / message content arrays into a single provider string
+ * (OpenAI Chat `system` content, OpenAI Responses `system` content, Gemini
+ * `systemInstruction.parts[].text`).
+ */
+export const joinText = (parts: ReadonlyArray<{ readonly text: string }>) => parts.map((part) => part.text).join("\n")
+
+const escapeSystemUpdateText = (text: string) =>
+  text.replaceAll("&", "&amp;").replaceAll("<", "&lt;").replaceAll(">", "&gt;")
+
+/**
+ * Stable fallback representation for chronological `Message.system(...)`
+ * updates on routes that do not support that privileged role natively. The
+ * wrapper remains visibly lower-authority user text, preserves the original
+ * temporal position, and XML-escapes content so it cannot close the wrapper.
+ */
+export const wrapSystemUpdate = (parts: ReadonlyArray<{ readonly text: string }>) =>
+  `<system-update>\n${escapeSystemUpdateText(joinText(parts))}\n</system-update>`
+
+/**
+ * Chronological system updates deliberately accept text only. Do not insert
+ * 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* (
+  route: string,
+  message: LLMRequest["messages"][number],
+) {
+  const content: TextPart[] = []
+  for (const part of message.content) {
+    if (!supportsContent(part, ["text"])) return yield* unsupportedContent(route, "system", ["text"])
+    content.push(part)
+  }
+  return content
+})
+
+/** Lower an unsupported privileged update into visible, in-order user text. */
+export const wrappedSystemUpdate = Effect.fn("ProviderShared.wrappedSystemUpdate")(function* (
+  route: string,
+  message: LLMRequest["messages"][number],
+) {
+  const content = yield* systemUpdateText(route, message)
+  return { type: "text" as const, text: wrapSystemUpdate(content), cache: content.at(-1)?.cache }
+})
+
+/**
+ * Parse the streamed JSON input of a tool call. Treats an empty string as
+ * `"{}"` — providers occasionally finish a tool call without ever emitting
+ * input deltas (e.g. zero-arg tools). The error message is uniform across
+ * routes: `Invalid JSON input for <route> tool call <name>`.
+ */
+export const parseToolInput = (route: string, name: string, raw: string) =>
+  parseJson(route, raw || "{}", `Invalid JSON input for ${route} tool call ${name}`)
+
+export const IMAGE_MIMES = ["image/png", "image/jpeg", "image/gif", "image/webp"] as const
+export const VIDEO_MIMES = ["video/mp4", "video/webm", "video/quicktime"] as const
+export const AUDIO_MIMES = ["audio/wav", "audio/mp3", "audio/aiff", "audio/aac", "audio/ogg", "audio/flac"] as const
+export const MEDIA_MIMES = [...IMAGE_MIMES, ...VIDEO_MIMES, ...AUDIO_MIMES] as const
+export const MAX_MEDIA_ENCODED_BYTES = 28 * 1024 * 1024
+export const MAX_MEDIA_DECODED_BYTES = 20 * 1024 * 1024
+
+const base64Pattern = /^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/
+
+export interface ValidatedMedia {
+  readonly mime: string
+  readonly base64: string
+  readonly dataUrl: string
+  readonly bytes: Uint8Array
+}
+
+export const validateMedia = Effect.fn("ProviderShared.validateMedia")(function* (
+  route: string,
+  part: MediaPart,
+  supportedMimes: ReadonlySet<string>,
+) {
+  const mime = part.mediaType.toLowerCase()
+  if (!supportedMimes.has(mime)) return yield* invalidRequest(`${route} does not support media type ${part.mediaType}`)
+
+  let base64: string
+  if (typeof part.data !== "string") {
+    if (part.data.byteLength > MAX_MEDIA_DECODED_BYTES)
+      return yield* invalidRequest(`${route} media exceeds the ${MAX_MEDIA_DECODED_BYTES} byte decoded limit`)
+    base64 = Buffer.from(part.data).toString("base64")
+  } else if (part.data.startsWith("data:")) {
+    const match = /^data:([^;,]+);base64,([A-Za-z0-9+/]*={0,2})$/s.exec(part.data)
+    if (!match) return yield* invalidRequest(`${route} media data URL must contain valid base64`)
+    if (match[1]!.toLowerCase() !== mime)
+      return yield* invalidRequest(`${route} media type ${part.mediaType} does not match data URL type ${match[1]}`)
+    base64 = match[2]!
+  } else {
+    base64 = part.data
+  }
+
+  if (Buffer.byteLength(base64, "utf8") > MAX_MEDIA_ENCODED_BYTES)
+    return yield* invalidRequest(`${route} media exceeds the ${MAX_MEDIA_ENCODED_BYTES} byte encoded limit`)
+  if (!base64 || base64.length % 4 !== 0 || !base64Pattern.test(base64))
+    return yield* invalidRequest(`${route} media must contain valid base64`)
+  const bytes = Buffer.from(base64, "base64")
+  if (bytes.byteLength > MAX_MEDIA_DECODED_BYTES)
+    return yield* invalidRequest(`${route} media exceeds the ${MAX_MEDIA_DECODED_BYTES} byte decoded limit`)
+  if (bytes.toString("base64") !== base64) return yield* invalidRequest(`${route} media must contain canonical base64`)
+  return { mime, base64, dataUrl: `data:${mime};base64,${base64}`, bytes } satisfies ValidatedMedia
+})
+
+export const validateToolFile = (route: string, part: ToolFileContent, supportedMimes: ReadonlySet<string>) =>
+  validateMedia(route, { type: "media", mediaType: part.mime, data: part.uri, filename: part.name }, supportedMimes)
+
+export const trimBaseUrl = (value: string) => value.replace(/\/+$/, "")
+
+export const toolResultText = (part: ToolResultPart) => {
+  if (part.result.type === "text") return String(part.result.value)
+  if (part.result.type === "error") {
+    const value = part.result.value
+    const prototype =
+      typeof value === "object" && value !== null && !Array.isArray(value) && Object.getPrototypeOf(value)
+    const structured = Array.isArray(value) || prototype === Object.prototype || prototype === null
+    return structured && isJson(value) ? encodeJson(value) : String(value)
+  }
+  return encodeJson(part.result.value)
+}
+
+export const errorText = (error: unknown) => {
+  if (error instanceof Error) return error.message
+  if (typeof error === "string") return error
+  if (typeof error === "number" || typeof error === "boolean" || typeof error === "bigint") return String(error)
+  if (error === null) return "null"
+  if (error === undefined) return "undefined"
+  return "Unknown stream error"
+}
+
+/**
+ * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
+ * decoder, and drops empty / `[DONE]` keep-alive events so the downstream
+ * `decodeChunk` sees one JSON string per element. The SSE channel emits a
+ * `Retry` control event on its error channel; we drop it here (we don't
+ * implement client-driven retries) so the public error channel stays
+ * `LLMError`.
+ */
+export const sseFraming = (bytes: Stream.Stream<Uint8Array, LLMError>): Stream.Stream<string, LLMError> =>
+  bytes.pipe(
+    Stream.decodeText(),
+    Stream.pipeThroughChannel(Sse.decode()),
+    Stream.catchTag("Retry", () => Stream.empty),
+    Stream.filter((event) => event.data.length > 0 && event.data !== "[DONE]"),
+    Stream.map((event) => event.data),
+  )
+
+/**
+ * Canonical invalid-request constructor. Lift one-line `const invalid =
+ * (message) => invalidRequest(message)` aliases out of every
+ * route so the error constructor lives in one place. If we ever extend
+ * `InvalidRequestReason` with route context or trace metadata, the change
+ * lands here.
+ */
+export const invalidRequest = (message: string) =>
+  new LLMError({
+    module: "ProviderShared",
+    method: "request",
+    reason: new InvalidRequestReason({ message }),
+  })
+
+export const matchToolChoice = <Auto, None, Required, Tool>(
+  route: string,
+  toolChoice: NonNullable<LLMRequest["toolChoice"]>,
+  cases: {
+    readonly auto: () => Auto
+    readonly none: () => None
+    readonly required: () => Required
+    readonly tool: (name: string) => Tool
+  },
+) =>
+  Effect.gen(function* () {
+    if (toolChoice.type === "auto") return cases.auto()
+    if (toolChoice.type === "none") return cases.none()
+    if (toolChoice.type === "required") return cases.required()
+    if (!toolChoice.name) return yield* invalidRequest(`${route} tool choice requires a tool name`)
+    return cases.tool(toolChoice.name)
+  })
+
+type ContentType = ContentPart["type"]
+
+const formatContentTypes = (types: ReadonlyArray<ContentType>) => {
+  if (types.length <= 1) return types[0] ?? ""
+  if (types.length === 2) return `${types[0]} and ${types[1]}`
+  return `${types.slice(0, -1).join(", ")}, and ${types.at(-1)}`
+}
+
+export const supportsContent = <const Type extends ContentType>(
+  part: ContentPart,
+  types: ReadonlyArray<Type>,
+): part is Extract<ContentPart, { readonly type: Type }> => (types as ReadonlyArray<ContentType>).includes(part.type)
+
+export const unsupportedContent = (
+  route: string,
+  role: LLMRequest["messages"][number]["role"],
+  types: ReadonlyArray<ContentType>,
+) => invalidRequest(`${route} ${role} messages only support ${formatContentTypes(types)} content for now`)
+
+/**
+ * Build a `validate` step from a Schema decoder. Replaces the per-route
+ * lambda body `(payload) => decode(payload).pipe(Effect.mapError((e) =>
+ * invalid(e.message)))`. Any decode error is translated into
+ * `LLMError` carrying the original parse-error message.
+ */
+export const validateWith =
+  <A, I, E extends { readonly message: string }>(decode: (input: I) => Effect.Effect<A, E>) =>
+  (payload: I) =>
+    decode(payload).pipe(Effect.mapError((error) => invalidRequest(error.message)))
+
+/**
+ * Build an HTTP POST with a JSON body. Sets `content-type: application/json`
+ * automatically after caller-supplied headers so routes cannot accidentally
+ * send JSON with a stale content type. The body is passed pre-encoded so
+ * routes can choose between
+ * `Schema.encodeSync(payload)` and `ProviderShared.encodeJson(payload)`.
+ */
+export const jsonPost = (input: { readonly url: string; readonly body: string; readonly headers?: Headers.Input }) =>
+  HttpClientRequest.post(input.url).pipe(
+    HttpClientRequest.setHeaders(Headers.set(Headers.fromInput(input.headers), "content-type", "application/json")),
+    HttpClientRequest.bodyText(input.body, "application/json"),
+  )
+
+export * as ProviderShared from "./shared"

package/src/protocols/utils/bedrock-auth.ts

@@ -0,0 +1,70 @@
+import { AwsV4Signer } from "aws4fetch"
+import { Effect } from "effect"
+import { Headers } from "effect/unstable/http"
+import { Auth, type AuthInput } from "../../route/auth"
+import { ProviderShared } from "../shared"
+
+/**
+ * AWS credentials for SigV4 signing. Bedrock also supports Bearer API key auth,
+ * which provider facades configure as route auth instead of SigV4. STS-vended
+ * credentials should be refreshed by the consumer (rebuild the model) before
+ * they expire; the route does not refresh.
+ */
+export interface Credentials {
+  readonly region: string
+  readonly accessKeyId: string
+  readonly secretAccessKey: string
+  readonly sessionToken?: string
+}
+
+const signRequest = (input: {
+  readonly url: string
+  readonly body: string
+  readonly headers: Headers.Headers
+  readonly credentials: Credentials
+}) =>
+  Effect.tryPromise({
+    try: async () => {
+      const signed = await new AwsV4Signer({
+        url: input.url,
+        method: "POST",
+        headers: Object.entries(input.headers),
+        body: input.body,
+        region: input.credentials.region,
+        accessKeyId: input.credentials.accessKeyId,
+        secretAccessKey: input.credentials.secretAccessKey,
+        sessionToken: input.credentials.sessionToken,
+        service: "bedrock",
+      }).sign()
+      return Object.fromEntries(signed.headers.entries())
+    },
+    catch: (error) =>
+      ProviderShared.invalidRequest(
+        `Bedrock Converse SigV4 signing failed: ${error instanceof Error ? error.message : String(error)}`,
+      ),
+  })
+
+/** Sign the exact JSON bytes with SigV4 using credentials configured on the route. */
+export const sigV4 = (credentials: Credentials | undefined) =>
+  Auth.custom((input: AuthInput) => {
+    return Effect.gen(function* () {
+      if (!credentials) {
+        return yield* ProviderShared.invalidRequest(
+          "Bedrock Converse requires either route bearer auth or AWS credentials configured on the route",
+        )
+      }
+      const headersForSigning = Headers.set(input.headers, "content-type", "application/json")
+      const signed = yield* signRequest({
+        url: input.url,
+        body: input.body,
+        headers: headersForSigning,
+        credentials,
+      })
+      return Headers.setAll(headersForSigning, signed)
+    })
+  })
+
+/** Bedrock route auth defaults to SigV4 and expects credentials from route configuration. */
+export const auth = sigV4(undefined)
+
+export * as BedrockAuth from "./bedrock-auth"

package/src/protocols/utils/bedrock-cache.ts

@@ -0,0 +1,37 @@
+import { Schema } from "effect"
+import type { CacheHint } from "../../schema"
+import { newBreakpoints, ttlBucket, type Breakpoints } from "./cache"
+
+// Bedrock cache markers are positional: emit a `cachePoint` block immediately
+// after the content the caller wants treated as a cacheable prefix. Bedrock
+// accepts optional `ttl: "5m" | "1h"` on cachePoint, mirroring Anthropic.
+export const CachePointBlock = Schema.Struct({
+  cachePoint: Schema.Struct({
+    type: Schema.tag("default"),
+    ttl: Schema.optional(Schema.Literals(["5m", "1h"])),
+  }),
+})
+export type CachePointBlock = Schema.Schema.Type<typeof CachePointBlock>
+
+// Bedrock-Claude enforces the same 4-breakpoint cap as the Anthropic Messages
+// API. Callers pass a shared counter through every `block()` call site so the
+// budget is respected across `system`, `messages`, and `tools`.
+export const BEDROCK_BREAKPOINT_CAP = 4
+
+export type { Breakpoints } from "./cache"
+export const breakpoints = () => newBreakpoints(BEDROCK_BREAKPOINT_CAP)
+
+const DEFAULT_5M: CachePointBlock = { cachePoint: { type: "default" } }
+const DEFAULT_1H: CachePointBlock = { cachePoint: { type: "default", ttl: "1h" } }
+
+export const block = (breakpoints: Breakpoints, cache: CacheHint | undefined): CachePointBlock | undefined => {
+  if (cache?.type !== "ephemeral" && cache?.type !== "persistent") return undefined
+  if (breakpoints.remaining <= 0) {
+    breakpoints.dropped += 1
+    return undefined
+  }
+  breakpoints.remaining -= 1
+  return ttlBucket(cache.ttlSeconds) === "1h" ? DEFAULT_1H : DEFAULT_5M
+}
+
+export * as BedrockCache from "./bedrock-cache"

package/src/protocols/utils/bedrock-media.ts

@@ -0,0 +1,90 @@
+import { Effect, Schema } from "effect"
+import type { MediaPart } from "../../schema"
+import { ProviderShared } from "../shared"
+
+// Bedrock Converse accepts image `format` as the file extension and
+// `source.bytes` as base64 in the JSON wire format.
+export const ImageFormat = Schema.Literals(["png", "jpeg", "gif", "webp"])
+export type ImageFormat = Schema.Schema.Type<typeof ImageFormat>
+
+export const ImageBlock = Schema.Struct({
+  image: Schema.Struct({
+    format: ImageFormat,
+    source: Schema.Struct({ bytes: Schema.String }),
+  }),
+})
+export type ImageBlock = Schema.Schema.Type<typeof ImageBlock>
+
+// Bedrock document blocks require a user-facing name so the model can refer to
+// the uploaded document.
+export const DocumentFormat = Schema.Literals(["pdf", "csv", "doc", "docx", "xls", "xlsx", "html", "txt", "md"])
+export type DocumentFormat = Schema.Schema.Type<typeof DocumentFormat>
+
+export const DocumentBlock = Schema.Struct({
+  document: Schema.Struct({
+    format: DocumentFormat,
+    name: Schema.String,
+    source: Schema.Struct({ bytes: Schema.String }),
+  }),
+})
+export type DocumentBlock = Schema.Schema.Type<typeof DocumentBlock>
+
+const IMAGE_FORMATS = {
+  "image/png": "png",
+  "image/jpeg": "jpeg",
+  "image/jpg": "jpeg",
+  "image/gif": "gif",
+  "image/webp": "webp",
+} as const satisfies Record<string, ImageFormat>
+
+const DOCUMENT_FORMATS = {
+  "application/pdf": "pdf",
+  "text/csv": "csv",
+  "application/msword": "doc",
+  "application/vnd.openxmlformats-officedocument.wordprocessingml.document": "docx",
+  "application/vnd.ms-excel": "xls",
+  "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet": "xlsx",
+  "text/html": "html",
+  "text/plain": "txt",
+  "text/markdown": "md",
+} as const satisfies Record<string, DocumentFormat>
+
+const documentBlock = (part: MediaPart, format: DocumentFormat, bytes: string): DocumentBlock => ({
+  document: {
+    format,
+    name: part.filename ?? `document.${format}`,
+    source: { bytes },
+  },
+})
+
+// Route by MIME. Known image/document formats lower into a typed block; anything
+// else fails with a clear error instead of silently degrading to a malformed
+// 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) {
+  const mime = part.mediaType.toLowerCase()
+  const imageFormat = IMAGE_FORMATS[mime as keyof typeof IMAGE_FORMATS]
+  if (imageFormat) {
+    const media = yield* ProviderShared.validateMedia(
+      "Bedrock Converse",
+      part,
+      new Set<string>(Object.keys(IMAGE_FORMATS)),
+    )
+    return { image: { format: imageFormat, source: { bytes: media.base64 } } } satisfies ImageBlock
+  }
+  if (mime.startsWith("image/"))
+    return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support image media type ${part.mediaType}`)
+  const documentFormat = DOCUMENT_FORMATS[mime as keyof typeof DOCUMENT_FORMATS]
+  if (documentFormat) {
+    const media = yield* ProviderShared.validateMedia(
+      "Bedrock Converse",
+      part,
+      new Set<string>(Object.keys(DOCUMENT_FORMATS)),
+    )
+    return documentBlock(part, documentFormat, media.base64)
+  }
+  return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support media type ${part.mediaType}`)
+})
+
+export * as BedrockMedia from "./bedrock-media"

package/src/protocols/utils/cache.ts

@@ -0,0 +1,16 @@
+// Shared helpers for provider cache-marker lowering. Anthropic and Bedrock
+// both enforce a 4-breakpoint cap per request and accept the same `5m`/`1h`
+// TTL buckets, so the counter and TTL mapping live here.
+
+export interface Breakpoints {
+  remaining: number
+  dropped: number
+}
+
+export const newBreakpoints = (cap: number): Breakpoints => ({ remaining: cap, dropped: 0 })
+
+// Returns `"1h"` for any `ttlSeconds >= 3600`, otherwise `undefined` (the
+// provider default 5m). Anthropic & Bedrock both treat anything shorter than
+// an hour as 5m.
+export const ttlBucket = (ttlSeconds: number | undefined): "1h" | undefined =>
+  ttlSeconds !== undefined && ttlSeconds >= 3600 ? "1h" : undefined

package/src/protocols/utils/gemini-tool-schema.ts

@@ -0,0 +1,99 @@
+import { isRecord } from "../../utils/record"
+
+// Gemini accepts a JSON Schema-like dialect for tool parameters, but rejects a
+// handful of common JSON Schema shapes. Keep this projection isolated so the
+// Gemini protocol file still reads like the other protocol modules.
+const SCHEMA_INTENT_KEYS = [
+  "type",
+  "properties",
+  "items",
+  "prefixItems",
+  "enum",
+  "const",
+  "$ref",
+  "additionalProperties",
+  "patternProperties",
+  "required",
+  "not",
+  "if",
+  "then",
+  "else",
+]
+
+const hasCombiner = (schema: unknown) =>
+  isRecord(schema) && (Array.isArray(schema.anyOf) || Array.isArray(schema.oneOf) || Array.isArray(schema.allOf))
+
+const hasSchemaIntent = (schema: unknown) =>
+  isRecord(schema) && (hasCombiner(schema) || SCHEMA_INTENT_KEYS.some((key) => key in schema))
+
+const sanitizeNode = (schema: unknown): unknown => {
+  if (!isRecord(schema)) return Array.isArray(schema) ? schema.map(sanitizeNode) : schema
+
+  const result: Record<string, unknown> = Object.fromEntries(
+    Object.entries(schema).map(([key, value]) => [
+      key,
+      key === "enum" && Array.isArray(value) ? value.map(String) : sanitizeNode(value),
+    ]),
+  )
+
+  if (Array.isArray(result.enum) && (result.type === "integer" || result.type === "number")) result.type = "string"
+
+  const properties = result.properties
+  if (result.type === "object" && isRecord(properties) && Array.isArray(result.required)) {
+    result.required = result.required.filter((field) => typeof field === "string" && field in properties)
+  }
+
+  if (result.type === "array" && !hasCombiner(result)) {
+    result.items = result.items ?? {}
+    if (isRecord(result.items) && !hasSchemaIntent(result.items)) result.items = { ...result.items, type: "string" }
+  }
+
+  if (typeof result.type === "string" && result.type !== "object" && !hasCombiner(result)) {
+    delete result.properties
+    delete result.required
+  }
+
+  return result
+}
+
+const emptyObjectSchema = (schema: Record<string, unknown>) =>
+  schema.type === "object" &&
+  (!isRecord(schema.properties) || Object.keys(schema.properties).length === 0) &&
+  !schema.additionalProperties
+
+const projectNode = (schema: unknown): Record<string, unknown> | undefined => {
+  if (!isRecord(schema)) return undefined
+  if (emptyObjectSchema(schema)) return undefined
+  return Object.fromEntries(
+    [
+      ["description", schema.description],
+      ["required", schema.required],
+      ["format", schema.format],
+      ["type", Array.isArray(schema.type) ? schema.type.filter((type) => type !== "null")[0] : schema.type],
+      ["nullable", Array.isArray(schema.type) && schema.type.includes("null") ? true : undefined],
+      ["enum", schema.const !== undefined ? [schema.const] : schema.enum],
+      [
+        "properties",
+        isRecord(schema.properties)
+          ? Object.fromEntries(Object.entries(schema.properties).map(([key, value]) => [key, projectNode(value)]))
+          : undefined,
+      ],
+      [
+        "items",
+        Array.isArray(schema.items)
+          ? schema.items.map(projectNode)
+          : schema.items === undefined
+            ? undefined
+            : projectNode(schema.items),
+      ],
+      ["allOf", Array.isArray(schema.allOf) ? schema.allOf.map(projectNode) : undefined],
+      ["anyOf", Array.isArray(schema.anyOf) ? schema.anyOf.map(projectNode) : undefined],
+      ["oneOf", Array.isArray(schema.oneOf) ? schema.oneOf.map(projectNode) : undefined],
+      ["minLength", schema.minLength],
+    ].filter((entry) => entry[1] !== undefined),
+  )
+}
+
+export const convert = (schema: unknown) => projectNode(sanitizeNode(schema))
+
+export * as GeminiToolSchema from "./gemini-tool-schema"