feat: add CLAUDE.md for project guidance and update token counting logic in translation functions

Author: cuipengfei

CLAUDE.md

@@ -0,0 +1,92 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Development Commands
+
+- **Build**: `bun run build` (uses tsdown)
+- **Development**: `bun run dev` (with file watching)
+- **Production**: `bun run start` (sets NODE_ENV=production)
+- **Lint**: `bun run lint` (uses @echristian/eslint-config with cache)
+- **Lint fix**: `bunx lint-staged` (fixes staged files)
+- **Typecheck**: `bun run typecheck` (runs TypeScript compiler)
+- **Test all**: `bun test`
+- **Test single file**: `bun test tests/[filename].test.ts`
+- **Package**: `bun run prepack` (builds before packaging)
+
+## Project Architecture
+
+### High-Level Structure
+This is a GitHub Copilot API proxy server that exposes Copilot as both OpenAI-compatible and Anthropic-compatible APIs. The server is built with Hono framework and uses Bun as the runtime.
+
+### Core Architecture Components
+
+**API Translation Layer** (`src/routes/messages/`):
+- Translates between Anthropic Messages API format and OpenAI Chat Completions format
+- Handles both streaming and non-streaming responses
+- Key files: `handler.ts`, `anthropic-types.ts`, `stream-translation.ts`, `non-stream-translation.ts`
+
+**Token Counting for Anthropic Models** (`src/lib/tokenizer.ts`):
+- Uses `gpt-tokenizer/model/gpt-4o` for token counting
+- Separates input tokens (all messages except last assistant message) from output tokens (last assistant message)
+- Filters out tool messages and extracts text content from multipart messages
+- Used by `/v1/messages/count_tokens` endpoint for Anthropic compatibility
+
+**GitHub Copilot Integration** (`src/services/`):
+- Authentication flow using device code OAuth
+- Token management and refresh
+- API requests to GitHub Copilot endpoints
+- Usage monitoring and quota tracking
+
+**Rate Limiting & Controls** (`src/lib/`):
+- Rate limiting between requests (`rate-limit.ts`)
+- Manual approval system for requests (`approval.ts`)
+- State management for server configuration (`state.ts`)
+
+### API Endpoints Structure
+
+**OpenAI Compatible**:
+- `/v1/chat/completions` - Chat completions
+- `/v1/models` - Available models
+- `/v1/embeddings` - Text embeddings
+
+**Anthropic Compatible**:
+- `/v1/messages` - Message completions (translates to/from OpenAI format)
+- `/v1/messages/count_tokens` - Token counting for Anthropic format
+
+**Monitoring**:
+- `/usage` - GitHub Copilot usage dashboard
+- `/token` - Current Copilot token info
+
+### Key Implementation Details
+
+**Anthropic Token Counting**:
+The `getTokenCount()` function in `src/lib/tokenizer.ts` implements token counting specifically for Anthropic compatibility:
+- Converts multipart content to text-only for counting
+- Splits messages into input (all except last assistant) and output (last assistant message only)
+- Uses GPT-4o tokenizer as the underlying counting mechanism
+- Returns `{input: number, output: number}` format
+
+**Message Translation**:
+- OpenAI → Anthropic: Converts chat completion responses to Anthropic message format
+- Anthropic → OpenAI: Converts Anthropic message requests to OpenAI chat completion format
+- Handles tool calls, system messages, and content blocks appropriately
+
+**Streaming Translation**:
+Real-time conversion of OpenAI SSE chunks to Anthropic streaming events, maintaining state for proper message reconstruction.
+
+## Code Style & Conventions
+
+- **TypeScript**: Strict mode enabled, avoid `any` types
+- **Imports**: Use `~/*` path aliases for `src/*` imports
+- **Error Handling**: Use explicit error classes from `src/lib/error.ts`
+- **Testing**: Place tests in `tests/` directory with `*.test.ts` naming
+- **Formatting**: Prettier with package.json plugin
+- **Linting**: @echristian/eslint-config with strict rules
+
+## Important Notes
+
+- Server uses GitHub Copilot as the underlying LLM provider
+- Rate limiting and manual approval features help avoid GitHub abuse detection
+- Token counting uses GPT-4o tokenizer regardless of the actual model being proxied
+- All API translations maintain compatibility with both OpenAI and Anthropic client libraries

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

@@ -313,9 +313,7 @@ export function translateToAnthropic(
     stop_reason: mapOpenAIStopReasonToAnthropic(stopReason),
     stop_sequence: null,
     usage: {
-      input_tokens:
-        (response.usage?.prompt_tokens ?? 0)
-        - (response.usage?.prompt_tokens_details?.cached_tokens ?? 0),
+      input_tokens: response.usage?.prompt_tokens ?? 0,
       output_tokens: response.usage?.completion_tokens ?? 0,
       ...(response.usage?.prompt_tokens_details?.cached_tokens
         !== undefined && {

src/routes/messages/stream-translation.ts

@@ -42,9 +42,7 @@ export function translateChunkToAnthropicEvents(
         stop_reason: null,
         stop_sequence: null,
         usage: {
-          input_tokens:
-            (chunk.usage?.prompt_tokens ?? 0)
-            - (chunk.usage?.prompt_tokens_details?.cached_tokens ?? 0),
+          input_tokens: chunk.usage?.prompt_tokens ?? 0,
           output_tokens: 0, // Will be updated in message_delta when finished
           ...(chunk.usage?.prompt_tokens_details?.cached_tokens
             !== undefined && {
@@ -159,9 +157,7 @@ export function translateChunkToAnthropicEvents(
           stop_sequence: null,
         },
         usage: {
-          input_tokens:
-            (chunk.usage?.prompt_tokens ?? 0)
-            - (chunk.usage?.prompt_tokens_details?.cached_tokens ?? 0),
+          input_tokens: chunk.usage?.prompt_tokens ?? 0,
           output_tokens: chunk.usage?.completion_tokens ?? 0,
           ...(chunk.usage?.prompt_tokens_details?.cached_tokens
             !== undefined && {

src/routes/models/route.ts

@@ -21,6 +21,7 @@ modelRoutes.get("/", async (c) => {
       created_at: new Date(0).toISOString(), // No date available from source
       owned_by: model.vendor,
       display_name: model.name,
+      capabilities: model.capabilities, // Include full capabilities from GitHub Copilot
     }))
 
     return c.json({