1.fix claude code 2.0.28 warmup request consume premium request, forcing small model if no tools are used

2.add bun idleTimeout = 0
3.feat: Compatible with Claude code JSONL file usage error scenarios, delay closeBlockIfOpen and map responses api to anthropic  support tool_use and fix spelling errors
4.feat: add configuration management with extra prompt handling and ensure config file creation

Author: caozhiyuan

src/lib/config.ts

@@ -0,0 +1,87 @@
+import consola from "consola"
+import fs from "node:fs"
+
+import { PATHS } from "./paths"
+
+export interface AppConfig {
+  extraPrompts?: Record<string, string>
+  smallModel?: string
+}
+
+const defaultConfig: AppConfig = {
+  extraPrompts: {
+    "gpt-5-codex": `
+## Tool use
+- You have access to many tools. If a tool exists to perform a specific task, you MUST use that tool instead of running a terminal command to perform that task.
+### Bash tool
+When using the Bash tool, follow these rules:
+- always run_in_background set to false, unless you are running a long-running command (e.g., a server or a watch command).
+### BashOutput tool
+When using the BashOutput tool, follow these rules:
+- Only Bash Tool run_in_background set to true, Use BashOutput to read the output later
+### TodoWrite tool
+When using the TodoWrite tool, follow these rules:
+- Skip using the TodoWrite tool for tasks with three or fewer steps.
+- Do not make single-step todo lists.
+- When you made a todo, update it after having performed one of the sub-tasks that you shared on the todo list.
+## Special user requests
+- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as 'date'), you should do so.
+`,
+  },
+  smallModel: "gpt-5-mini",
+}
+
+let cachedConfig: AppConfig | null = null
+
+function ensureConfigFile(): void {
+  try {
+    fs.accessSync(PATHS.CONFIG_PATH, fs.constants.R_OK | fs.constants.W_OK)
+  } catch {
+    fs.writeFileSync(
+      PATHS.CONFIG_PATH,
+      `${JSON.stringify(defaultConfig, null, 2)}\n`,
+      "utf8",
+    )
+    try {
+      fs.chmodSync(PATHS.CONFIG_PATH, 0o600)
+    } catch {
+      return
+    }
+  }
+}
+
+function readConfigFromDisk(): AppConfig {
+  ensureConfigFile()
+  try {
+    const raw = fs.readFileSync(PATHS.CONFIG_PATH, "utf8")
+    if (!raw.trim()) {
+      fs.writeFileSync(
+        PATHS.CONFIG_PATH,
+        `${JSON.stringify(defaultConfig, null, 2)}\n`,
+        "utf8",
+      )
+      return defaultConfig
+    }
+    return JSON.parse(raw) as AppConfig
+  } catch (error) {
+    consola.error("Failed to read config file, using default config", error)
+    return defaultConfig
+  }
+}
+
+export function getConfig(): AppConfig {
+  if (!cachedConfig) {
+    cachedConfig = readConfigFromDisk()
+  }
+  return cachedConfig
+}
+
+export function getExtraPromptForModel(model: string): string {
+  const config = getConfig()
+  return config.extraPrompts?.[model] ?? ""
+}
+
+export function getSmallModel(): string {
+  const config = getConfig()
+  return config.smallModel ?? "gpt-5-mini"
+}

src/lib/paths.ts

@@ -5,15 +5,18 @@ import path from "node:path"
 const APP_DIR = path.join(os.homedir(), ".local", "share", "copilot-api")
 
 const GITHUB_TOKEN_PATH = path.join(APP_DIR, "github_token")
+const CONFIG_PATH = path.join(APP_DIR, "config.json")
 
 export const PATHS = {
   APP_DIR,
   GITHUB_TOKEN_PATH,
+  CONFIG_PATH,
 }
 
 export async function ensurePaths(): Promise<void> {
   await fs.mkdir(PATHS.APP_DIR, { recursive: true })
   await ensureFile(PATHS.GITHUB_TOKEN_PATH)
+  await ensureFile(PATHS.CONFIG_PATH)
 }
 
 async function ensureFile(filePath: string): Promise<void> {

src/routes/messages/handler.ts

@@ -4,6 +4,7 @@ import consola from "consola"
 import { streamSSE } from "hono/streaming"
 
 import { awaitApproval } from "~/lib/approval"
+import { getSmallModel } from "~/lib/config"
 import { checkRateLimit } from "~/lib/rate-limit"
 import { state } from "~/lib/state"
 import {
@@ -42,6 +43,11 @@ export async function handleCompletion(c: Context) {
   const anthropicPayload = await c.req.json<AnthropicMessagesPayload>()
   consola.debug("Anthropic request payload:", JSON.stringify(anthropicPayload))
 
+  // fix claude code 2.0.28 warmup request consume premium request, forcing small model if no tools are used
+  if (!anthropicPayload.tools || anthropicPayload.tools.length === 0) {
+    anthropicPayload.model = getSmallModel()
+  }
+
   const useResponsesApi = shouldUseResponsesApi(anthropicPayload.model)
 
   if (state.manualApprove) {

src/routes/messages/responses-stream-translation.ts

@@ -176,8 +176,6 @@ const handleOutputItemDone = (
     state.blockHasDelta.add(blockIndex)
   }
 
-  closeBlockIfOpen(state, blockIndex, events)
-
   return events
 }
 
@@ -232,7 +230,6 @@ const handleFunctionCallArgumentsDone = (
     state.blockHasDelta.add(blockIndex)
   }
 
-  closeBlockIfOpen(state, blockIndex, events)
   state.functionCallStateByOutputIndex.delete(outputIndex)
   return events
 }
@@ -340,8 +337,6 @@ const handleOutputTextDone = (
     })
   }
 
-  closeBlockIfOpen(state, blockIndex, events)
-
   return events
 }
 
@@ -421,9 +416,7 @@ const messageStart = (
         usage: {
           input_tokens: inputTokens,
           output_tokens: 0,
-          ...(inputCachedTokens !== undefined && {
-            cache_creation_input_tokens: inputCachedTokens,
-          }),
+          cache_read_input_tokens: inputCachedTokens ?? 0,
         },
       },
     },
@@ -449,6 +442,7 @@ const openTextBlockIfNeeded = (
   }
 
   if (!state.openBlocks.has(blockIndex)) {
+    closeOpenBlocks(state, events)
     events.push({
       type: "content_block_start",
       index: blockIndex,
@@ -480,6 +474,7 @@ const openThinkingBlockIfNeeded = (
   }
 
   if (!state.openBlocks.has(blockIndex)) {
+    closeOpenBlocks(state, events)
     events.push({
       type: "content_block_start",
       index: blockIndex,
@@ -508,13 +503,20 @@ const closeBlockIfOpen = (
   state.blockHasDelta.delete(blockIndex)
 }
 
-const closeAllOpenBlocks = (
+const closeOpenBlocks = (
   state: ResponsesStreamState,
   events: Array<AnthropicStreamEventData>,
 ) => {
   for (const blockIndex of state.openBlocks) {
     closeBlockIfOpen(state, blockIndex, events)
   }
+}
+
+const closeAllOpenBlocks = (
+  state: ResponsesStreamState,
+  events: Array<AnthropicStreamEventData>,
+) => {
+  closeOpenBlocks(state, events)
 
   state.functionCallStateByOutputIndex.clear()
 }
@@ -562,6 +564,7 @@ const openFunctionCallBlock = (
   const { blockIndex } = functionCallState
 
   if (!state.openBlocks.has(blockIndex)) {
+    closeOpenBlocks(state, events)
     events.push({
       type: "content_block_start",
       index: blockIndex,

src/routes/messages/responses-translation.ts

@@ -1,5 +1,6 @@
 import consola from "consola"
 
+import { getExtraPromptForModel } from "~/lib/config"
 import {
   type ResponsesPayload,
   type ResponseInputContent,
@@ -60,8 +61,8 @@ export const translateAnthropicMessagesToResponsesPayload = (
   const responsesPayload: ResponsesPayload = {
     model: payload.model,
     input,
-    instructions: translateSystemPrompt(payload.system),
-    temperature: payload.temperature ?? null,
+    instructions: translateSystemPrompt(payload.system, payload.model),
+    temperature: 1, // reasoning high temperature fixed to 1
     top_p: payload.top_p ?? null,
     max_output_tokens: payload.max_tokens,
     tools: translatedTools,
@@ -277,36 +278,22 @@ const createFunctionCallOutput = (
 
 const translateSystemPrompt = (
   system: string | Array<AnthropicTextBlock> | undefined,
+  model: string,
 ): string | null => {
   if (!system) {
     return null
   }
 
-  const toolUsePrompt = `
-## Tool use
-- You have access to many tools. If a tool exists to perform a specific task, you MUST use that tool instead of running a terminal command to perform that task.
-### Bash tool
-When using the Bash tool, follow these rules:
-- always run_in_background set to false, unless you are running a long-running command (e.g., a server or a watch command).
-### BashOutput tool
-When using the BashOutput tool, follow these rules:
-- Only Bash Tool run_in_background set to true, Use BashOutput to read the output later
-### TodoWrite tool
-When using the TodoWrite tool, follow these rules:
-- Skip using the TodoWrite tool for tasks with three or fewer steps.
-- Do not make single-step todo lists.
-- When you made a todo, update it after having performed one of the sub-tasks that you shared on the todo list.
-## Special user requests
-- If the user makes a simple request (such as asking for the time) which you can fulfill by running a terminal command (such as ''date''), you should do so.`
+  const extraPrompt = getExtraPromptForModel(model)
 
   if (typeof system === "string") {
-    return system + toolUsePrompt
+    return system + extraPrompt
   }
 
   const text = system
     .map((block, index) => {
       if (index === 0) {
-        return block.text + toolUsePrompt
+        return block.text + extraPrompt
       }
       return block.text
     })
@@ -548,6 +535,9 @@ const mapResponsesStopReason = (
   const { status, incomplete_details: incompleteDetails } = response
 
   if (status === "completed") {
+    if (response.output.some((item) => item.type === "function_call")) {
+      return "tool_use"
+    }
     return "end_turn"
   }
 

src/routes/responses/handler.ts

@@ -52,29 +52,13 @@ export const handleResponses = async (c: Context) => {
   if (isStreamingRequested(payload) && isAsyncIterable(response)) {
     consola.debug("Forwarding native Responses stream")
     return streamSSE(c, async (stream) => {
-      const pingInterval = setInterval(async () => {
-        try {
-          await stream.writeSSE({
-            event: "ping",
-            data: JSON.stringify({ timestamp: Date.now() }),
-          })
-        } catch (error) {
-          consola.warn("Failed to send ping:", error)
-          clearInterval(pingInterval)
-        }
-      }, 3000)
-
-      try {
-        for await (const chunk of response) {
-          consola.debug("Responses stream chunk:", JSON.stringify(chunk))
-          await stream.writeSSE({
-            id: (chunk as { id?: string }).id,
-            event: (chunk as { event?: string }).event,
-            data: (chunk as { data?: string }).data ?? "",
-          })
-        }
-      } finally {
-        clearInterval(pingInterval)
+      for await (const chunk of response) {
+        consola.debug("Responses stream chunk:", JSON.stringify(chunk))
+        await stream.writeSSE({
+          id: (chunk as { id?: string }).id,
+          event: (chunk as { event?: string }).event,
+          data: (chunk as { data?: string }).data ?? "",
+        })
       }
     })
   }

src/start.ts

@@ -111,6 +111,9 @@ export async function runServer(options: RunServerOptions): Promise<void> {
   serve({
     fetch: server.fetch as ServerHandler,
     port: options.port,
+    bun: {
+      idleTimeout: 0,
+    },
   })
 }
 

tests/responses-stream-translation.test.ts

@@ -96,12 +96,7 @@ describe("translateResponsesStreamEvent tool calls", () => {
       partial_json: "[]}",
     })
 
-    const blockStop = events.find(
-      (event) => event.type === "content_block_stop",
-    )
-    expect(blockStop).toBeDefined()
-
-    expect(state.openBlocks.size).toBe(0)
+    expect(state.openBlocks.size).toBe(1)
     expect(state.functionCallStateByOutputIndex.size).toBe(0)
   })
 
@@ -139,12 +134,7 @@ describe("translateResponsesStreamEvent tool calls", () => {
         '{"todos":[{"content":"Review src/routes/responses/translation.ts"}]}',
     })
 
-    const blockStop = events.find(
-      (event) => event.type === "content_block_stop",
-    )
-    expect(blockStop).toBeDefined()
-
-    expect(state.openBlocks.size).toBe(0)
+    expect(state.openBlocks.size).toBe(1)
     expect(state.functionCallStateByOutputIndex.size).toBe(0)
   })
 })