Merge pull request #20 from caozhiyuan/all

czy

Author: cuipengfei

src/lib/api-config.ts

@@ -14,8 +14,9 @@ const USER_AGENT = `GitHubCopilotChat/${COPILOT_VERSION}`
 const API_VERSION = "2025-10-01"
 
 export const copilotBaseUrl = (state: State) =>
-  `https://api.${state.accountType}.githubcopilot.com`
-
+  state.accountType === "individual" ?
+    "https://api.githubcopilot.com"
+  : `https://api.${state.accountType}.githubcopilot.com`
 export const copilotHeaders = (state: State, vision: boolean = false) => {
   const headers: Record<string, string> = {
     Authorization: `Bearer ${state.copilotToken}`,
@@ -24,7 +25,7 @@ export const copilotHeaders = (state: State, vision: boolean = false) => {
     "editor-version": `vscode/${state.vsCodeVersion}`,
     "editor-plugin-version": EDITOR_PLUGIN_VERSION,
     "user-agent": USER_AGENT,
-    "openai-intent": "conversation-panel",
+    "openai-intent": "conversation-agent",
     "x-github-api-version": API_VERSION,
     "x-request-id": randomUUID(),
     "x-vscode-user-agent-library-version": "electron-fetch",

src/lib/config.ts

@@ -10,6 +10,7 @@ export interface AppConfig {
     string,
     "none" | "minimal" | "low" | "medium" | "high" | "xhigh"
   >
+  useFunctionApplyPatch?: boolean
 }
 
 const gpt5ExplorationPrompt = `## Exploration and reading files
@@ -28,6 +29,7 @@ const defaultConfig: AppConfig = {
   modelReasoningEfforts: {
     "gpt-5-mini": "low",
   },
+  useFunctionApplyPatch: true,
 }
 
 let cachedConfig: AppConfig | null = null

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

@@ -14,7 +14,6 @@ import {
 import {
   type AnthropicAssistantContentBlock,
   type AnthropicAssistantMessage,
-  type AnthropicMessage,
   type AnthropicMessagesPayload,
   type AnthropicResponse,
   type AnthropicTextBlock,
@@ -38,9 +37,9 @@ export function translateToOpenAI(
   return {
     model: modelId,
     messages: translateAnthropicMessagesToOpenAI(
-      payload.messages,
-      payload.system,
+      payload,
       modelId,
+      thinkingBudget,
     ),
     max_tokens: payload.max_tokens,
     stop: payload.stop_sequences,
@@ -86,32 +85,74 @@ function translateModelName(model: string): string {
 }
 
 function translateAnthropicMessagesToOpenAI(
-  anthropicMessages: Array<AnthropicMessage>,
-  system: string | Array<AnthropicTextBlock> | undefined,
+  payload: AnthropicMessagesPayload,
   modelId: string,
+  thinkingBudget: number | undefined,
 ): Array<Message> {
-  const systemMessages = handleSystemPrompt(system)
-
-  const otherMessages = anthropicMessages.flatMap((message) =>
+  const systemMessages = handleSystemPrompt(
+    payload.system,
+    modelId,
+    thinkingBudget,
+  )
+  const otherMessages = payload.messages.flatMap((message) =>
     message.role === "user" ?
       handleUserMessage(message)
     : handleAssistantMessage(message, modelId),
   )
-
+  if (modelId.startsWith("claude") && thinkingBudget) {
+    const reminder =
+      "<system-reminder>you MUST follow interleaved_thinking_protocol</system-reminder>"
+    const firstUserIndex = otherMessages.findIndex((m) => m.role === "user")
+    if (firstUserIndex !== -1) {
+      const userMessage = otherMessages[firstUserIndex]
+      if (typeof userMessage.content === "string") {
+        userMessage.content = reminder + "\n\n" + userMessage.content
+      } else if (Array.isArray(userMessage.content)) {
+        userMessage.content = [
+          { type: "text", text: reminder },
+          ...userMessage.content,
+        ] as Array<ContentPart>
+      }
+    }
+  }
   return [...systemMessages, ...otherMessages]
 }
 
 function handleSystemPrompt(
   system: string | Array<AnthropicTextBlock> | undefined,
+  modelId: string,
+  thinkingBudget: number | undefined,
 ): Array<Message> {
   if (!system) {
     return []
   }
 
+  let extraPrompt = ""
+  if (modelId.startsWith("claude") && thinkingBudget) {
+    extraPrompt = `
+<interleaved_thinking_protocol>
+ABSOLUTE REQUIREMENT - NON-NEGOTIABLE:
+The current thinking_mode is interleaved, Whenever you have the result of a function call, think carefully , MUST output a thinking block
+RULES:
+Tool result → thinking block (ALWAYS, no exceptions)
+This is NOT optional - it is a hard requirement
+The thinking block must contain substantive reasoning (minimum 3-5 sentences)
+Think about: what the results mean, what to do next, how to answer the user
+NEVER skip this step, even if the result seems simple or obvious
+</interleaved_thinking_protocol>`
+  }
+
   if (typeof system === "string") {
-    return [{ role: "system", content: system }]
+    return [{ role: "system", content: system + extraPrompt }]
   } else {
-    const systemText = system.map((block) => block.text).join("\n\n")
+    const systemText = system
+      .map((block, index) => {
+        if (index === 0) {
+          return block.text + extraPrompt
+        }
+        return block.text
+      })
+      .join("\n\n")
     return [{ role: "system", content: systemText }]
   }
 }

src/routes/messages/stream-translation.ts

@@ -218,6 +218,7 @@ function handleContent(
     delta.content === ""
     && delta.reasoning_opaque
     && delta.reasoning_opaque.length > 0
+    && state.thinkingBlockOpen
   ) {
     events.push(
       {
@@ -317,6 +318,15 @@ function handleThinkingText(
   events: Array<AnthropicStreamEventData>,
 ) {
   if (delta.reasoning_text && delta.reasoning_text.length > 0) {
+    // compatible with copilot API returning content->reasoning_text->reasoning_opaque in different deltas
+    // this is an extremely abnormal situation, probably a server-side bug
+    // only occurs in the claude model, with a very low probability of occurrence
+    if (state.contentBlockOpen) {
+      delta.content = delta.reasoning_text
+      delta.reasoning_text = undefined
+      return
+    }
+
     if (!state.thinkingBlockOpen) {
       events.push({
         type: "content_block_start",

src/routes/responses/handler.ts

@@ -4,6 +4,7 @@ import consola from "consola"
 import { streamSSE } from "hono/streaming"
 
 import { awaitApproval } from "~/lib/approval"
+import { getConfig } from "~/lib/config"
 import { createHandlerLogger } from "~/lib/logger"
 import { checkRateLimit } from "~/lib/rate-limit"
 import { state } from "~/lib/state"
@@ -70,6 +71,8 @@ export const handleResponses = async (c: Context) => {
   const payload = await c.req.json<ResponsesPayload>()
   logger.debug("Responses request payload:", JSON.stringify(payload))
 
+  useFunctionApplyPatch(payload)
+
   const selectedModel = state.models?.data.find(
     (model) => model.id === payload.model,
   )
@@ -146,3 +149,35 @@ const isAsyncIterable = <T>(value: unknown): value is AsyncIterable<T> =>
 
 const isStreamingRequested = (payload: ResponsesPayload): boolean =>
   Boolean(payload.stream)
+
+const useFunctionApplyPatch = (payload: ResponsesPayload): void => {
+  const config = getConfig()
+  const useFunctionApplyPatch = config.useFunctionApplyPatch ?? true
+  if (useFunctionApplyPatch) {
+    logger.debug("Using function tool apply_patch for responses")
+    if (Array.isArray(payload.tools)) {
+      const toolsArr = payload.tools
+      for (let i = 0; i < toolsArr.length; i++) {
+        const t = toolsArr[i]
+        if (t.type === "custom" && t.name === "apply_patch") {
+          toolsArr[i] = {
+            type: "function",
+            name: t.name,
+            description: "Use the `apply_patch` tool to edit files",
+            parameters: {
+              type: "object",
+              properties: {
+                input: {
+                  type: "string",
+                  description: "The entire contents of the apply_patch command",
+                },
+              },
+              required: ["input"],
+            },
+            strict: false,
+          }
+        }
+      }
+    }
+  }
+}

src/services/copilot/create-responses.ts

@@ -33,7 +33,7 @@ export interface ToolChoiceFunction {
   type: "function"
 }
 
-export type Tool = FunctionTool
+export type Tool = FunctionTool | Record<string, unknown>
 
 export interface FunctionTool {
   name: string