opencode-ide-plugin/cloud/function/src/gateway.ts

import { Hono, MiddlewareHandler } from "hono"
import { type ProviderMetadata, type LanguageModelUsage } from "ai"
import { createAnthropic } from "@ai-sdk/anthropic"
import { createOpenAI } from "@ai-sdk/openai"
import { createOpenAICompatible } from "@ai-sdk/openai-compatible"
import type { LanguageModelV2Prompt } from "@ai-sdk/provider"
import { type ChatCompletionCreateParamsBase } from "openai/resources/chat/completions"
import { Actor } from "@opencode/cloud-core/actor.js"
import { Database, eq, sql } from "@opencode/cloud-core/drizzle/index.js"
import { KeyTable } from "@opencode/cloud-core/schema/key.sql.js"
import { Billing } from "@opencode/cloud-core/billing.js"
import { Resource } from "@opencode/cloud-resource"

type Env = {}

const SUPPORTED_MODELS = {
  "anthropic/claude-sonnet-4": {
    input: 0.0000015,
    output: 0.000006,
    reasoning: 0.0000015,
    cacheRead: 0.0000001,
    cacheWrite: 0.0000001,
    model: () =>
      createAnthropic({
        apiKey: Resource.ANTHROPIC_API_KEY.value,
      })("claude-sonnet-4-20250514"),
  },
  "openai/gpt-4.1": {
    input: 0.0000015,
    output: 0.000006,
    reasoning: 0.0000015,
    cacheRead: 0.0000001,
    cacheWrite: 0.0000001,
    model: () =>
      createOpenAI({
        apiKey: Resource.OPENAI_API_KEY.value,
      })("gpt-4.1"),
  },
  "zhipuai/glm-4.5-flash": {
    input: 0,
    output: 0,
    reasoning: 0,
    cacheRead: 0,
    cacheWrite: 0,
    model: () =>
      createOpenAICompatible({
        name: "Zhipu AI",
        baseURL: "https://api.z.ai/api/paas/v4",
        apiKey: Resource.ZHIPU_API_KEY.value,
      })("glm-4.5-flash"),
  },
}

const GatewayAuth: MiddlewareHandler = async (c, next) => {
  const authHeader = c.req.header("authorization")

  if (!authHeader || !authHeader.startsWith("Bearer ")) {
    return c.json(
      {
        error: {
          message: "Missing API key.",
          type: "invalid_request_error",
          param: null,
          code: "unauthorized",
        },
      },
      401,
    )
  }

  const apiKey = authHeader.split(" ")[1]

  // Check against KeyTable
  const keyRecord = await Database.use((tx) =>
    tx
      .select({
        id: KeyTable.id,
        workspaceID: KeyTable.workspaceID,
      })
      .from(KeyTable)
      .where(eq(KeyTable.key, apiKey))
      .then((rows) => rows[0]),
  )

  if (!keyRecord) {
    return c.json(
      {
        error: {
          message: "Invalid API key.",
          type: "invalid_request_error",
          param: null,
          code: "unauthorized",
        },
      },
      401,
    )
  }

  c.set("keyRecord", keyRecord)
  await next()
}

const app = new Hono<{ Bindings: Env; Variables: { keyRecord?: { id: string; workspaceID: string } } }>()
  .get("/", (c) => c.text("Hello, world!"))
  .post("/v1/chat/completions", GatewayAuth, async (c) => {
    const keyRecord = c.get("keyRecord")!

    return await Actor.provide("system", { workspaceID: keyRecord.workspaceID }, async () => {
      try {
        // Check balance
        const customer = await Billing.get()
        if (customer.balance <= 0) {
          return c.json(
            {
              error: {
                message: "Insufficient balance",
                type: "insufficient_quota",
                param: null,
                code: "insufficient_quota",
              },
            },
            401,
          )
        }

        const body = await c.req.json<ChatCompletionCreateParamsBase>()
        const model = SUPPORTED_MODELS[body.model as keyof typeof SUPPORTED_MODELS]?.model()
        if (!model) throw new Error(`Unsupported model: ${body.model}`)

        const requestBody = transformOpenAIRequestToAiSDK()

        return body.stream ? await handleStream() : await handleGenerate()

        async function handleStream() {
          const result = await model.doStream({
            ...requestBody,
          })

          const encoder = new TextEncoder()
          const stream = new ReadableStream({
            async start(controller) {
              const id = `chatcmpl-${Date.now()}`
              const created = Math.floor(Date.now() / 1000)

              try {
                for await (const chunk of result.stream) {
                  console.log("!!! CHUNK !!! : " + chunk.type)
                  switch (chunk.type) {
                    case "text-delta": {
                      const data = {
                        id,
                        object: "chat.completion.chunk",
                        created,
                        model: body.model,
                        choices: [
                          {
                            index: 0,
                            delta: {
                              content: chunk.delta,
                            },
                            finish_reason: null,
                          },
                        ],
                      }
                      controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`))
                      break
                    }

                    case "reasoning-delta": {
                      const data = {
                        id,
                        object: "chat.completion.chunk",
                        created,
                        model: body.model,
                        choices: [
                          {
                            index: 0,
                            delta: {
                              reasoning_content: chunk.delta,
                            },
                            finish_reason: null,
                          },
                        ],
                      }
                      controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`))
                      break
                    }

                    case "tool-call": {
                      const data = {
                        id,
                        object: "chat.completion.chunk",
                        created,
                        model: body.model,
                        choices: [
                          {
                            index: 0,
                            delta: {
                              tool_calls: [
                                {
                                  index: 0,
                                  id: chunk.toolCallId,
                                  type: "function",
                                  function: {
                                    name: chunk.toolName,
                                    arguments: chunk.input,
                                  },
                                },
                              ],
                            },
                            finish_reason: null,
                          },
                        ],
                      }
                      controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`))
                      break
                    }

                    case "error": {
                      const data = {
                        id,
                        object: "chat.completion.chunk",
                        created,
                        model: body.model,
                        choices: [
                          {
                            index: 0,
                            delta: {},
                            finish_reason: "stop",
                          },
                        ],
                        error: {
                          message: typeof chunk.error === "string" ? chunk.error : chunk.error,
                          type: "server_error",
                        },
                      }
                      controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`))
                      controller.enqueue(encoder.encode("data: [DONE]\n\n"))
                      controller.close()
                      break
                    }

                    case "finish": {
                      const data = {
                        id,
                        object: "chat.completion.chunk",
                        created,
                        model: body.model,
                        choices: [
                          {
                            index: 0,
                            delta: {},
                            finish_reason:
                              {
                                stop: "stop",
                                length: "length",
                                "content-filter": "content_filter",
                                "tool-calls": "tool_calls",
                                error: "stop",
                                other: "stop",
                                unknown: "stop",
                              }[chunk.finishReason] || "stop",
                          },
                        ],
                        usage: {
                          prompt_tokens: chunk.usage.inputTokens,
                          completion_tokens: chunk.usage.outputTokens,
                          total_tokens: chunk.usage.totalTokens,
                          completion_tokens_details: {
                            reasoning_tokens: chunk.usage.reasoningTokens,
                          },
                          prompt_tokens_details: {
                            cached_tokens: chunk.usage.cachedInputTokens,
                          },
                        },
                      }
                      await trackUsage(body.model, chunk.usage, chunk.providerMetadata)
                      controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`))
                      controller.enqueue(encoder.encode("data: [DONE]\n\n"))
                      controller.close()
                      break
                    }

                    //case "stream-start":
                    //case "response-metadata":
                    case "text-start":
                    case "text-end":
                    case "reasoning-start":
                    case "reasoning-end":
                    case "tool-input-start":
                    case "tool-input-delta":
                    case "tool-input-end":
                    case "raw":
                    default:
                      // Log unknown chunk types for debugging
                      console.warn(`Unknown chunk type: ${(chunk as any).type}`)
                      break
                  }
                }
              } catch (error) {
                controller.error(error)
              }
            },
          })

          return new Response(stream, {
            headers: {
              "Content-Type": "text/plain; charset=utf-8",
              "Cache-Control": "no-cache",
              Connection: "keep-alive",
            },
          })
        }

        async function handleGenerate() {
          const response = await model.doGenerate({
            ...requestBody,
          })
          await trackUsage(body.model, response.usage, response.providerMetadata)
          return c.json({
            id: `chatcmpl-${Date.now()}`,
            object: "chat.completion" as const,
            created: Math.floor(Date.now() / 1000),
            model: body.model,
            choices: [
              {
                index: 0,
                message: {
                  role: "assistant" as const,
                  content: response.content?.find((c) => c.type === "text")?.text ?? "",
                  reasoning_content: response.content?.find((c) => c.type === "reasoning")?.text,
                  tool_calls: response.content
                    ?.filter((c) => c.type === "tool-call")
                    .map((toolCall) => ({
                      id: toolCall.toolCallId,
                      type: "function" as const,
                      function: {
                        name: toolCall.toolName,
                        arguments: toolCall.input,
                      },
                    })),
                },
                finish_reason:
                  (
                    {
                      stop: "stop",
                      length: "length",
                      "content-filter": "content_filter",
                      "tool-calls": "tool_calls",
                      error: "stop",
                      other: "stop",
                      unknown: "stop",
                    } as const
                  )[response.finishReason] || "stop",
              },
            ],
            usage: {
              prompt_tokens: response.usage?.inputTokens,
              completion_tokens: response.usage?.outputTokens,
              total_tokens: response.usage?.totalTokens,
              completion_tokens_details: {
                reasoning_tokens: response.usage?.reasoningTokens,
              },
              prompt_tokens_details: {
                cached_tokens: response.usage?.cachedInputTokens,
              },
            },
          })
        }

        function transformOpenAIRequestToAiSDK() {
          const prompt = transformMessages()
          const tools = transformTools()

          return {
            prompt,
            maxOutputTokens: body.max_tokens ?? body.max_completion_tokens ?? undefined,
            temperature: body.temperature ?? undefined,
            topP: body.top_p ?? undefined,
            frequencyPenalty: body.frequency_penalty ?? undefined,
            presencePenalty: body.presence_penalty ?? undefined,
            providerOptions: body.reasoning_effort
              ? {
                  anthropic: {
                    reasoningEffort: body.reasoning_effort,
                  },
                }
              : undefined,
            stopSequences: (typeof body.stop === "string" ? [body.stop] : body.stop) ?? undefined,
            responseFormat: (() => {
              if (!body.response_format) return { type: "text" as const }
              if (body.response_format.type === "json_schema")
                return {
                  type: "json" as const,
                  schema: body.response_format.json_schema.schema,
                  name: body.response_format.json_schema.name,
                  description: body.response_format.json_schema.description,
                }
              if (body.response_format.type === "json_object") return { type: "json" as const }
              throw new Error("Unsupported response format")
            })(),
            seed: body.seed ?? undefined,
            tools: tools.tools,
            toolChoice: tools.toolChoice,
          }

          function transformTools() {
            const { tools, tool_choice } = body

            if (!tools || tools.length === 0) {
              return { tools: undefined, toolChoice: undefined }
            }

            const aiSdkTools = tools.map((tool) => {
              return {
                type: tool.type,
                name: tool.function.name,
                description: tool.function.description,
                inputSchema: tool.function.parameters!,
              }
            })

            let aiSdkToolChoice
            if (tool_choice == null) {
              aiSdkToolChoice = undefined
            } else if (tool_choice === "auto") {
              aiSdkToolChoice = { type: "auto" as const }
            } else if (tool_choice === "none") {
              aiSdkToolChoice = { type: "none" as const }
            } else if (tool_choice === "required") {
              aiSdkToolChoice = { type: "required" as const }
            } else if (tool_choice.type === "function") {
              aiSdkToolChoice = {
                type: "tool" as const,
                toolName: tool_choice.function.name,
              }
            }

            return { tools: aiSdkTools, toolChoice: aiSdkToolChoice }
          }

          function transformMessages() {
            const { messages } = body
            const prompt: LanguageModelV2Prompt = []

            for (const message of messages) {
              switch (message.role) {
                case "system": {
                  prompt.push({
                    role: "system",
                    content: message.content as string,
                  })
                  break
                }

                case "user": {
                  if (typeof message.content === "string") {
                    prompt.push({
                      role: "user",
                      content: [{ type: "text", text: message.content }],
                    })
                  } else {
                    const content = message.content.map((part) => {
                      switch (part.type) {
                        case "text":
                          return { type: "text" as const, text: part.text }
                        case "image_url":
                          return {
                            type: "file" as const,
                            mediaType: "image/jpeg" as const,
                            data: part.image_url.url,
                          }
                        default:
                          throw new Error(`Unsupported content part type: ${(part as any).type}`)
                      }
                    })
                    prompt.push({
                      role: "user",
                      content,
                    })
                  }
                  break
                }

                case "assistant": {
                  const content: Array<
                    | { type: "text"; text: string }
                    | {
                        type: "tool-call"
                        toolCallId: string
                        toolName: string
                        input: any
                      }
                  > = []

                  if (message.content) {
                    content.push({
                      type: "text",
                      text: message.content as string,
                    })
                  }

                  if (message.tool_calls) {
                    for (const toolCall of message.tool_calls) {
                      content.push({
                        type: "tool-call",
                        toolCallId: toolCall.id,
                        toolName: toolCall.function.name,
                        input: JSON.parse(toolCall.function.arguments),
                      })
                    }
                  }

                  prompt.push({
                    role: "assistant",
                    content,
                  })
                  break
                }

                case "tool": {
                  prompt.push({
                    role: "tool",
                    content: [
                      {
                        type: "tool-result",
                        toolName: "placeholder",
                        toolCallId: message.tool_call_id,
                        output: {
                          type: "text",
                          value: message.content as string,
                        },
                      },
                    ],
                  })
                  break
                }

                default: {
                  throw new Error(`Unsupported message role: ${message.role}`)
                }
              }
            }

            return prompt
          }
        }

        async function trackUsage(model: string, usage: LanguageModelUsage, providerMetadata?: ProviderMetadata) {
          const modelData = SUPPORTED_MODELS[model as keyof typeof SUPPORTED_MODELS]
          if (!modelData) throw new Error(`Unsupported model: ${model}`)

          const inputTokens = usage.inputTokens ?? 0
          const outputTokens = usage.outputTokens ?? 0
          const reasoningTokens = usage.reasoningTokens ?? 0
          const cacheReadTokens = usage.cachedInputTokens ?? 0
          const cacheWriteTokens =
            providerMetadata?.["anthropic"]?.["cacheCreationInputTokens"] ??
            // @ts-expect-error
            providerMetadata?.["bedrock"]?.["usage"]?.["cacheWriteInputTokens"] ??
            0

          const inputCost = modelData.input * inputTokens
          const outputCost = modelData.output * outputTokens
          const reasoningCost = modelData.reasoning * reasoningTokens
          const cacheReadCost = modelData.cacheRead * cacheReadTokens
          const cacheWriteCost = modelData.cacheWrite * cacheWriteTokens
          const costInCents = (inputCost + outputCost + reasoningCost + cacheReadCost + cacheWriteCost) * 100

          await Billing.consume({
            model,
            inputTokens,
            outputTokens,
            reasoningTokens,
            cacheReadTokens,
            cacheWriteTokens,
            costInCents,
          })

          await Database.use((tx) =>
            tx
              .update(KeyTable)
              .set({ timeUsed: sql`now()` })
              .where(eq(KeyTable.id, keyRecord.id)),
          )
        }
      } catch (error: any) {
        return c.json({ error: { message: error.message } }, 500)
      }
    })
  })
  .all("*", (c) => c.text("Not Found"))

export type ApiType = typeof app

export default app