Fixed: #140 #133 #80

feat(translator): add token counting functionality for Gemini, Claude, and CLI - Introduced `TokenCount` handling across various Codex translators (Gemini, Claude, CLI) with respective implementations. - Added utility methods for token counting and formatting responses. - Integrated `tiktoken-go/tokenizer` library for tokenization. - Updated CodexExecutor with token counting logic to support multiple models including GPT-5 variants. - Refined go.mod and go.sum to include new dependencies. feat(runtime): add token counting functionality across executors - Implemented token counting in OpenAICompatExecutor, QwenExecutor, and IFlowExecutor. - Added utilities for token counting and response formatting using `tiktoken-go/tokenizer`. - Integrated token counting into translators for Gemini, Claude, and Gemini CLI. - Enhanced multiple model support, including GPT-5 variants, for token counting. docs: update environment variable instructions for multi-model support - Added details for setting `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, and `ANTHROPIC_DEFAULT_HAIKU_MODEL` for version 2.x.x. - Clarified usage of `ANTHROPIC_MODEL` and `ANTHROPIC_SMALL_FAST_MODEL` for version 1.x.x. - Expanded examples for setting environment variables across different models including Gemini, GPT-5, Claude, and Qwen3.
2026-02-02 20:40:52 +08:00 · 2025-10-26 04:51:23 +08:00
parent f6cf784cd1
commit a552a45b81
21 changed files with 588 additions and 22 deletions
--- a/README.md
+++ b/README.md
@@ -556,12 +556,17 @@ The server will relay the `loadCodeAssist`, `onboardUser`, and `countTokens` req

 ## Claude Code with multiple account load balancing

-Start CLI Proxy API server, and then set the `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_MODEL`, `ANTHROPIC_SMALL_FAST_MODEL` environment variables.
+Start CLI Proxy API server, and then set the `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_DEFAULT_HAIKU_MODEL` (or `ANTHROPIC_MODEL`, `ANTHROPIC_SMALL_FAST_MODEL` for version 1.x.x) environment variables.

 Using Gemini models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gemini-2.5-pro
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gemini-2.5-flash
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gemini-2.5-flash-lite
+# version 1.x.x
 export ANTHROPIC_MODEL=gemini-2.5-pro
 export ANTHROPIC_SMALL_FAST_MODEL=gemini-2.5-flash
 ```
@@ -570,6 +575,11 @@ Using OpenAI GPT 5 models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gpt-5-high
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gpt-5-medium
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-5-minimal
+# version 1.x.x
 export ANTHROPIC_MODEL=gpt-5
 export ANTHROPIC_SMALL_FAST_MODEL=gpt-5-minimal
 ```
@@ -578,6 +588,11 @@ Using OpenAI GPT 5 Codex models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gpt-5-codex-high
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gpt-5-codex-medium
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-5-codex-low
+# version 1.x.x
 export ANTHROPIC_MODEL=gpt-5-codex
 export ANTHROPIC_SMALL_FAST_MODEL=gpt-5-codex-low
 ```
@@ -586,6 +601,11 @@ Using Claude models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-1-20250805
+export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-5-20250929
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-3-5-haiku-20241022
+# version 1.x.x
 export ANTHROPIC_MODEL=claude-sonnet-4-20250514
 export ANTHROPIC_SMALL_FAST_MODEL=claude-3-5-haiku-20241022
 ```
@@ -594,6 +614,11 @@ Using Qwen models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder-flash
+# version 1.x.x
 export ANTHROPIC_MODEL=qwen3-coder-plus
 export ANTHROPIC_SMALL_FAST_MODEL=qwen3-coder-flash
 ```
@@ -602,6 +627,11 @@ Using iFlow models:
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# version 2.x.x
+export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-max
+export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-235b-a22b-instruct
+# version 1.x.x
 export ANTHROPIC_MODEL=qwen3-max
 export ANTHROPIC_SMALL_FAST_MODEL=qwen3-235b-a22b-instruct
 ```
--- a/README_CN.md
+++ b/README_CN.md
@@ -564,12 +564,17 @@ export CODE_ASSIST_ENDPOINT="http://127.0.0.1:8317"

 ## Claude Code 的使用方法

-启动 CLI Proxy API 服务器, 设置如下系统环境变量 `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_MODEL`, `ANTHROPIC_SMALL_FAST_MODEL`
+启动 CLI Proxy API 服务器, 设置如下系统环境变量 `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN`, `ANTHROPIC_DEFAULT_OPUS_MODEL`, `ANTHROPIC_DEFAULT_SONNET_MODEL`, `ANTHROPIC_DEFAULT_HAIKU_MODEL` (或 `ANTHROPIC_MODEL`, `ANTHROPIC_SMALL_FAST_MODEL` 对应 1.x.x 版本)

 使用 Gemini 模型：
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gemini-2.5-pro
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gemini-2.5-flash
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gemini-2.5-flash-lite
+# 1.x.x 版本
 export ANTHROPIC_MODEL=gemini-2.5-pro
 export ANTHROPIC_SMALL_FAST_MODEL=gemini-2.5-flash
 ```
@@ -578,6 +583,11 @@ export ANTHROPIC_SMALL_FAST_MODEL=gemini-2.5-flash
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gpt-5-high
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gpt-5-medium
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-5-minimal
+# 1.x.x 版本
 export ANTHROPIC_MODEL=gpt-5
 export ANTHROPIC_SMALL_FAST_MODEL=gpt-5-minimal
 ```
@@ -586,15 +596,24 @@ export ANTHROPIC_SMALL_FAST_MODEL=gpt-5-minimal
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=gpt-5-codex-high
+export ANTHROPIC_DEFAULT_SONNET_MODEL=gpt-5-codex-medium
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=gpt-5-codex-low
+# 1.x.x 版本
 export ANTHROPIC_MODEL=gpt-5-codex
 export ANTHROPIC_SMALL_FAST_MODEL=gpt-5-codex-low
 ```

-
 使用 Claude 模型：
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-4-1-20250805
+export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-4-5-20250929
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-3-5-haiku-20241022
+# 1.x.x 版本
 export ANTHROPIC_MODEL=claude-sonnet-4-20250514
 export ANTHROPIC_SMALL_FAST_MODEL=claude-3-5-haiku-20241022
 ```
@@ -603,6 +622,11 @@ export ANTHROPIC_SMALL_FAST_MODEL=claude-3-5-haiku-20241022
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-coder-flash
+# 1.x.x 版本
 export ANTHROPIC_MODEL=qwen3-coder-plus
 export ANTHROPIC_SMALL_FAST_MODEL=qwen3-coder-flash
 ```
@@ -611,6 +635,11 @@ export ANTHROPIC_SMALL_FAST_MODEL=qwen3-coder-flash
 ```bash
 export ANTHROPIC_BASE_URL=http://127.0.0.1:8317
 export ANTHROPIC_AUTH_TOKEN=sk-dummy
+# 2.x.x 版本
+export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen3-max
+export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen3-coder-plus
+export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen3-235b-a22b-instruct
+# 1.x.x 版本
 export ANTHROPIC_MODEL=qwen3-max
 export ANTHROPIC_SMALL_FAST_MODEL=qwen3-235b-a22b-instruct
 ```
--- a/go.mod
+++ b/go.mod
@@ -7,14 +7,15 @@ require (
 	github.com/gin-gonic/gin v1.10.1
 	github.com/go-git/go-git/v6 v6.0.0-20251009132922-75a182125145
 	github.com/google/uuid v1.6.0
-	github.com/joho/godotenv v1.5.1
 	github.com/jackc/pgx/v5 v5.7.6
+	github.com/joho/godotenv v1.5.1
 	github.com/klauspost/compress v1.17.4
 	github.com/minio/minio-go/v7 v7.0.66
 	github.com/sirupsen/logrus v1.9.3
 	github.com/skratchdot/open-golang v0.0.0-20200116055534-eef842397966
 	github.com/tidwall/gjson v1.18.0
 	github.com/tidwall/sjson v1.2.5
+	github.com/tiktoken-go/tokenizer v0.7.0
 	golang.org/x/crypto v0.43.0
 	golang.org/x/net v0.46.0
 	golang.org/x/oauth2 v0.30.0
@@ -32,6 +33,7 @@ require (
 	github.com/cloudwego/base64x v0.1.4 // indirect
 	github.com/cloudwego/iasm v0.2.0 // indirect
 	github.com/cyphar/filepath-securejoin v0.4.1 // indirect
+	github.com/dlclark/regexp2 v1.11.5 // indirect
 	github.com/dustin/go-humanize v1.0.1 // indirect
 	github.com/emirpasic/gods v1.18.1 // indirect
 	github.com/gabriel-vasile/mimetype v1.4.3 // indirect
--- a/go.sum
+++ b/go.sum
@@ -23,6 +23,8 @@ github.com/cyphar/filepath-securejoin v0.4.1/go.mod h1:Sdj7gXlvMcPZsbhwhQ33GguGL
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/dlclark/regexp2 v1.11.5 h1:Q/sSnsKerHeCkc/jSTNq1oCm7KiVgUMZRDUoRu0JQZQ=
+github.com/dlclark/regexp2 v1.11.5/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/dustin/go-humanize v1.0.1 h1:GzkhY7T5VNhEkwH0PVJgjz+fX1rhBrR7pRT3mDkpeCY=
 github.com/dustin/go-humanize v1.0.1/go.mod h1:Mu1zIs6XwVuF/gI1OepvI0qD18qycQx+mFykh5fBlto=
 github.com/elazarl/goproxy v1.7.2 h1:Y2o6urb7Eule09PjlhQRGNsqRfPmYI3KKQLFpCAV3+o=
@@ -147,6 +149,8 @@ github.com/tidwall/pretty v1.2.0 h1:RWIZEg2iJ8/g6fDDYzMpobmaoGh5OLl4AXtGUGPcqCs=
 github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU=
 github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY=
 github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28=
+github.com/tiktoken-go/tokenizer v0.7.0 h1:VMu6MPT0bXFDHr7UPh9uii7CNItVt3X9K90omxL54vw=
+github.com/tiktoken-go/tokenizer v0.7.0/go.mod h1:6UCYI/DtOallbmL7sSy30p6YQv60qNyU/4aVigPOx6w=
 github.com/twitchyliquid64/golang-asm v0.15.1 h1:SU5vSMR7hnwNxj24w34ZyCi/FmDZTkS4MhqMhdFk5YI=
 github.com/twitchyliquid64/golang-asm v0.15.1/go.mod h1:a1lVb/DtPvCB8fslRZhAngC2+aY1QWCk3Cedj/Gdt08=
 github.com/ugorji/go/codec v1.2.12 h1:9LC83zGrHhuUA9l16C9AHXAqEV/2wBQ4nkvumAE65EE=
--- a/internal/runtime/executor/codex_executor.go
+++ b/internal/runtime/executor/codex_executor.go
@@ -20,6 +20,7 @@ import (
 	log "github.com/sirupsen/logrus"
 	"github.com/tidwall/gjson"
 	"github.com/tidwall/sjson"
+	"github.com/tiktoken-go/tokenizer"

 	"github.com/gin-gonic/gin"
 	"github.com/google/uuid"
@@ -277,7 +278,180 @@ func (e *CodexExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 }

 func (e *CodexExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	return cliproxyexecutor.Response{Payload: []byte{}}, fmt.Errorf("not implemented")
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("codex")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+
+	modelForCounting := req.Model
+
+	if util.InArray([]string{"gpt-5", "gpt-5-minimal", "gpt-5-low", "gpt-5-medium", "gpt-5-high"}, req.Model) {
+		modelForCounting = "gpt-5"
+		body, _ = sjson.SetBytes(body, "model", "gpt-5")
+		switch req.Model {
+		case "gpt-5-minimal":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "minimal")
+		case "gpt-5-low":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "low")
+		case "gpt-5-medium":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "medium")
+		case "gpt-5-high":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "high")
+		default:
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "low")
+		}
+	} else if util.InArray([]string{"gpt-5-codex", "gpt-5-codex-low", "gpt-5-codex-medium", "gpt-5-codex-high"}, req.Model) {
+		modelForCounting = "gpt-5"
+		body, _ = sjson.SetBytes(body, "model", "gpt-5-codex")
+		switch req.Model {
+		case "gpt-5-codex-low":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "low")
+		case "gpt-5-codex-medium":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "medium")
+		case "gpt-5-codex-high":
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "high")
+		default:
+			body, _ = sjson.SetBytes(body, "reasoning.effort", "low")
+		}
+	}
+
+	body, _ = sjson.DeleteBytes(body, "previous_response_id")
+	body, _ = sjson.SetBytes(body, "stream", false)
+
+	enc, err := tokenizerForCodexModel(modelForCounting)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("codex executor: tokenizer init failed: %w", err)
+	}
+
+	count, err := countCodexInputTokens(enc, body)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("codex executor: token counting failed: %w", err)
+	}
+
+	usageJSON := fmt.Sprintf(`{"response":{"usage":{"input_tokens":%d,"output_tokens":0,"total_tokens":%d}}}`, count, count)
+	translated := sdktranslator.TranslateTokenCount(ctx, to, from, count, []byte(usageJSON))
+	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
+}
+
+func tokenizerForCodexModel(model string) (tokenizer.Codec, error) {
+	sanitized := strings.ToLower(strings.TrimSpace(model))
+	switch {
+	case sanitized == "":
+		return tokenizer.Get(tokenizer.Cl100kBase)
+	case strings.HasPrefix(sanitized, "gpt-5"):
+		return tokenizer.ForModel(tokenizer.GPT5)
+	case strings.HasPrefix(sanitized, "gpt-4.1"):
+		return tokenizer.ForModel(tokenizer.GPT41)
+	case strings.HasPrefix(sanitized, "gpt-4o"):
+		return tokenizer.ForModel(tokenizer.GPT4o)
+	case strings.HasPrefix(sanitized, "gpt-4"):
+		return tokenizer.ForModel(tokenizer.GPT4)
+	case strings.HasPrefix(sanitized, "gpt-3.5"), strings.HasPrefix(sanitized, "gpt-3"):
+		return tokenizer.ForModel(tokenizer.GPT35Turbo)
+	default:
+		return tokenizer.Get(tokenizer.Cl100kBase)
+	}
+}
+
+func countCodexInputTokens(enc tokenizer.Codec, body []byte) (int64, error) {
+	if enc == nil {
+		return 0, fmt.Errorf("encoder is nil")
+	}
+	if len(body) == 0 {
+		return 0, nil
+	}
+
+	root := gjson.ParseBytes(body)
+	var segments []string
+
+	if inst := strings.TrimSpace(root.Get("instructions").String()); inst != "" {
+		segments = append(segments, inst)
+	}
+
+	inputItems := root.Get("input")
+	if inputItems.IsArray() {
+		arr := inputItems.Array()
+		for i := range arr {
+			item := arr[i]
+			switch item.Get("type").String() {
+			case "message":
+				content := item.Get("content")
+				if content.IsArray() {
+					parts := content.Array()
+					for j := range parts {
+						part := parts[j]
+						if text := strings.TrimSpace(part.Get("text").String()); text != "" {
+							segments = append(segments, text)
+						}
+					}
+				}
+			case "function_call":
+				if name := strings.TrimSpace(item.Get("name").String()); name != "" {
+					segments = append(segments, name)
+				}
+				if args := strings.TrimSpace(item.Get("arguments").String()); args != "" {
+					segments = append(segments, args)
+				}
+			case "function_call_output":
+				if out := strings.TrimSpace(item.Get("output").String()); out != "" {
+					segments = append(segments, out)
+				}
+			default:
+				if text := strings.TrimSpace(item.Get("text").String()); text != "" {
+					segments = append(segments, text)
+				}
+			}
+		}
+	}
+
+	tools := root.Get("tools")
+	if tools.IsArray() {
+		tarr := tools.Array()
+		for i := range tarr {
+			tool := tarr[i]
+			if name := strings.TrimSpace(tool.Get("name").String()); name != "" {
+				segments = append(segments, name)
+			}
+			if desc := strings.TrimSpace(tool.Get("description").String()); desc != "" {
+				segments = append(segments, desc)
+			}
+			if params := tool.Get("parameters"); params.Exists() {
+				val := params.Raw
+				if params.Type == gjson.String {
+					val = params.String()
+				}
+				if trimmed := strings.TrimSpace(val); trimmed != "" {
+					segments = append(segments, trimmed)
+				}
+			}
+		}
+	}
+
+	textFormat := root.Get("text.format")
+	if textFormat.Exists() {
+		if name := strings.TrimSpace(textFormat.Get("name").String()); name != "" {
+			segments = append(segments, name)
+		}
+		if schema := textFormat.Get("schema"); schema.Exists() {
+			val := schema.Raw
+			if schema.Type == gjson.String {
+				val = schema.String()
+			}
+			if trimmed := strings.TrimSpace(val); trimmed != "" {
+				segments = append(segments, trimmed)
+			}
+		}
+	}
+
+	text := strings.Join(segments, "\n")
+	if text == "" {
+		return 0, nil
+	}
+
+	count, err := enc.Count(text)
+	if err != nil {
+		return 0, err
+	}
+	return int64(count), nil
 }

 func (e *CodexExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
--- a/internal/runtime/executor/iflow_executor.go
+++ b/internal/runtime/executor/iflow_executor.go
@@ -221,9 +221,24 @@ func (e *IFlowExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Au
 	return stream, nil
 }

-// CountTokens is not implemented for iFlow.
-func (e *IFlowExecutor) CountTokens(context.Context, *cliproxyauth.Auth, cliproxyexecutor.Request, cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	return cliproxyexecutor.Response{Payload: nil}, fmt.Errorf("not implemented")
+func (e *IFlowExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+
+	enc, err := tokenizerForModel(req.Model)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("iflow executor: tokenizer init failed: %w", err)
+	}
+
+	count, err := countOpenAIChatTokens(enc, body)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("iflow executor: token counting failed: %w", err)
+	}
+
+	usageJSON := buildOpenAIUsageJSON(count)
+	translated := sdktranslator.TranslateTokenCount(ctx, to, from, count, usageJSON)
+	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }

 // Refresh refreshes OAuth tokens and updates the stored API key.
--- a/internal/runtime/executor/openai_compat_executor.go
+++ b/internal/runtime/executor/openai_compat_executor.go
@@ -219,7 +219,29 @@ func (e *OpenAICompatExecutor) ExecuteStream(ctx context.Context, auth *cliproxy
 }

 func (e *OpenAICompatExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	return cliproxyexecutor.Response{Payload: []byte{}}, fmt.Errorf("not implemented")
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai")
+	translated := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+
+	modelForCounting := req.Model
+	if modelOverride := e.resolveUpstreamModel(req.Model, auth); modelOverride != "" {
+		translated = e.overrideModel(translated, modelOverride)
+		modelForCounting = modelOverride
+	}
+
+	enc, err := tokenizerForModel(modelForCounting)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("openai compat executor: tokenizer init failed: %w", err)
+	}
+
+	count, err := countOpenAIChatTokens(enc, translated)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("openai compat executor: token counting failed: %w", err)
+	}
+
+	usageJSON := buildOpenAIUsageJSON(count)
+	translatedUsage := sdktranslator.TranslateTokenCount(ctx, to, from, count, usageJSON)
+	return cliproxyexecutor.Response{Payload: []byte(translatedUsage)}, nil
 }

 // Refresh is a no-op for API-key based compatibility providers.
--- a/internal/runtime/executor/qwen_executor.go
+++ b/internal/runtime/executor/qwen_executor.go
@@ -207,7 +207,28 @@ func (e *QwenExecutor) ExecuteStream(ctx context.Context, auth *cliproxyauth.Aut
 }

 func (e *QwenExecutor) CountTokens(ctx context.Context, auth *cliproxyauth.Auth, req cliproxyexecutor.Request, opts cliproxyexecutor.Options) (cliproxyexecutor.Response, error) {
-	return cliproxyexecutor.Response{Payload: []byte{}}, fmt.Errorf("not implemented")
+	from := opts.SourceFormat
+	to := sdktranslator.FromString("openai")
+	body := sdktranslator.TranslateRequest(from, to, req.Model, bytes.Clone(req.Payload), false)
+
+	modelName := gjson.GetBytes(body, "model").String()
+	if strings.TrimSpace(modelName) == "" {
+		modelName = req.Model
+	}
+
+	enc, err := tokenizerForModel(modelName)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("qwen executor: tokenizer init failed: %w", err)
+	}
+
+	count, err := countOpenAIChatTokens(enc, body)
+	if err != nil {
+		return cliproxyexecutor.Response{}, fmt.Errorf("qwen executor: token counting failed: %w", err)
+	}
+
+	usageJSON := buildOpenAIUsageJSON(count)
+	translated := sdktranslator.TranslateTokenCount(ctx, to, from, count, usageJSON)
+	return cliproxyexecutor.Response{Payload: []byte(translated)}, nil
 }

 func (e *QwenExecutor) Refresh(ctx context.Context, auth *cliproxyauth.Auth) (*cliproxyauth.Auth, error) {
--- a/internal/runtime/executor/token_helpers.go
+++ b/internal/runtime/executor/token_helpers.go
@@ -0,0 +1,234 @@
+package executor
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/tidwall/gjson"
+	"github.com/tiktoken-go/tokenizer"
+)
+
+// tokenizerForModel returns a tokenizer codec suitable for an OpenAI-style model id.
+func tokenizerForModel(model string) (tokenizer.Codec, error) {
+	sanitized := strings.ToLower(strings.TrimSpace(model))
+	switch {
+	case sanitized == "":
+		return tokenizer.Get(tokenizer.Cl100kBase)
+	case strings.HasPrefix(sanitized, "gpt-5"):
+		return tokenizer.ForModel(tokenizer.GPT5)
+	case strings.HasPrefix(sanitized, "gpt-4.1"):
+		return tokenizer.ForModel(tokenizer.GPT41)
+	case strings.HasPrefix(sanitized, "gpt-4o"):
+		return tokenizer.ForModel(tokenizer.GPT4o)
+	case strings.HasPrefix(sanitized, "gpt-4"):
+		return tokenizer.ForModel(tokenizer.GPT4)
+	case strings.HasPrefix(sanitized, "gpt-3.5"), strings.HasPrefix(sanitized, "gpt-3"):
+		return tokenizer.ForModel(tokenizer.GPT35Turbo)
+	case strings.HasPrefix(sanitized, "o1"):
+		return tokenizer.ForModel(tokenizer.O1)
+	case strings.HasPrefix(sanitized, "o3"):
+		return tokenizer.ForModel(tokenizer.O3)
+	case strings.HasPrefix(sanitized, "o4"):
+		return tokenizer.ForModel(tokenizer.O4Mini)
+	default:
+		return tokenizer.Get(tokenizer.O200kBase)
+	}
+}
+
+// countOpenAIChatTokens approximates prompt tokens for OpenAI chat completions payloads.
+func countOpenAIChatTokens(enc tokenizer.Codec, payload []byte) (int64, error) {
+	if enc == nil {
+		return 0, fmt.Errorf("encoder is nil")
+	}
+	if len(payload) == 0 {
+		return 0, nil
+	}
+
+	root := gjson.ParseBytes(payload)
+	segments := make([]string, 0, 32)
+
+	collectOpenAIMessages(root.Get("messages"), &segments)
+	collectOpenAITools(root.Get("tools"), &segments)
+	collectOpenAIFunctions(root.Get("functions"), &segments)
+	collectOpenAIToolChoice(root.Get("tool_choice"), &segments)
+	collectOpenAIResponseFormat(root.Get("response_format"), &segments)
+	addIfNotEmpty(&segments, root.Get("input").String())
+	addIfNotEmpty(&segments, root.Get("prompt").String())
+
+	joined := strings.TrimSpace(strings.Join(segments, "\n"))
+	if joined == "" {
+		return 0, nil
+	}
+
+	count, err := enc.Count(joined)
+	if err != nil {
+		return 0, err
+	}
+	return int64(count), nil
+}
+
+// buildOpenAIUsageJSON returns a minimal usage structure understood by downstream translators.
+func buildOpenAIUsageJSON(count int64) []byte {
+	return []byte(fmt.Sprintf(`{"usage":{"prompt_tokens":%d,"completion_tokens":0,"total_tokens":%d}}`, count, count))
+}
+
+func collectOpenAIMessages(messages gjson.Result, segments *[]string) {
+	if !messages.Exists() || !messages.IsArray() {
+		return
+	}
+	messages.ForEach(func(_, message gjson.Result) bool {
+		addIfNotEmpty(segments, message.Get("role").String())
+		addIfNotEmpty(segments, message.Get("name").String())
+		collectOpenAIContent(message.Get("content"), segments)
+		collectOpenAIToolCalls(message.Get("tool_calls"), segments)
+		collectOpenAIFunctionCall(message.Get("function_call"), segments)
+		return true
+	})
+}
+
+func collectOpenAIContent(content gjson.Result, segments *[]string) {
+	if !content.Exists() {
+		return
+	}
+	if content.Type == gjson.String {
+		addIfNotEmpty(segments, content.String())
+		return
+	}
+	if content.IsArray() {
+		content.ForEach(func(_, part gjson.Result) bool {
+			partType := part.Get("type").String()
+			switch partType {
+			case "text", "input_text", "output_text":
+				addIfNotEmpty(segments, part.Get("text").String())
+			case "image_url":
+				addIfNotEmpty(segments, part.Get("image_url.url").String())
+			case "input_audio", "output_audio", "audio":
+				addIfNotEmpty(segments, part.Get("id").String())
+			case "tool_result":
+				addIfNotEmpty(segments, part.Get("name").String())
+				collectOpenAIContent(part.Get("content"), segments)
+			default:
+				if part.IsArray() {
+					collectOpenAIContent(part, segments)
+					return true
+				}
+				if part.Type == gjson.JSON {
+					addIfNotEmpty(segments, part.Raw)
+					return true
+				}
+				addIfNotEmpty(segments, part.String())
+			}
+			return true
+		})
+		return
+	}
+	if content.Type == gjson.JSON {
+		addIfNotEmpty(segments, content.Raw)
+	}
+}
+
+func collectOpenAIToolCalls(calls gjson.Result, segments *[]string) {
+	if !calls.Exists() || !calls.IsArray() {
+		return
+	}
+	calls.ForEach(func(_, call gjson.Result) bool {
+		addIfNotEmpty(segments, call.Get("id").String())
+		addIfNotEmpty(segments, call.Get("type").String())
+		function := call.Get("function")
+		if function.Exists() {
+			addIfNotEmpty(segments, function.Get("name").String())
+			addIfNotEmpty(segments, function.Get("description").String())
+			addIfNotEmpty(segments, function.Get("arguments").String())
+			if params := function.Get("parameters"); params.Exists() {
+				addIfNotEmpty(segments, params.Raw)
+			}
+		}
+		return true
+	})
+}
+
+func collectOpenAIFunctionCall(call gjson.Result, segments *[]string) {
+	if !call.Exists() {
+		return
+	}
+	addIfNotEmpty(segments, call.Get("name").String())
+	addIfNotEmpty(segments, call.Get("arguments").String())
+}
+
+func collectOpenAITools(tools gjson.Result, segments *[]string) {
+	if !tools.Exists() {
+		return
+	}
+	if tools.IsArray() {
+		tools.ForEach(func(_, tool gjson.Result) bool {
+			appendToolPayload(tool, segments)
+			return true
+		})
+		return
+	}
+	appendToolPayload(tools, segments)
+}
+
+func collectOpenAIFunctions(functions gjson.Result, segments *[]string) {
+	if !functions.Exists() || !functions.IsArray() {
+		return
+	}
+	functions.ForEach(func(_, function gjson.Result) bool {
+		addIfNotEmpty(segments, function.Get("name").String())
+		addIfNotEmpty(segments, function.Get("description").String())
+		if params := function.Get("parameters"); params.Exists() {
+			addIfNotEmpty(segments, params.Raw)
+		}
+		return true
+	})
+}
+
+func collectOpenAIToolChoice(choice gjson.Result, segments *[]string) {
+	if !choice.Exists() {
+		return
+	}
+	if choice.Type == gjson.String {
+		addIfNotEmpty(segments, choice.String())
+		return
+	}
+	addIfNotEmpty(segments, choice.Raw)
+}
+
+func collectOpenAIResponseFormat(format gjson.Result, segments *[]string) {
+	if !format.Exists() {
+		return
+	}
+	addIfNotEmpty(segments, format.Get("type").String())
+	addIfNotEmpty(segments, format.Get("name").String())
+	if schema := format.Get("json_schema"); schema.Exists() {
+		addIfNotEmpty(segments, schema.Raw)
+	}
+	if schema := format.Get("schema"); schema.Exists() {
+		addIfNotEmpty(segments, schema.Raw)
+	}
+}
+
+func appendToolPayload(tool gjson.Result, segments *[]string) {
+	if !tool.Exists() {
+		return
+	}
+	addIfNotEmpty(segments, tool.Get("type").String())
+	addIfNotEmpty(segments, tool.Get("name").String())
+	addIfNotEmpty(segments, tool.Get("description").String())
+	if function := tool.Get("function"); function.Exists() {
+		addIfNotEmpty(segments, function.Get("name").String())
+		addIfNotEmpty(segments, function.Get("description").String())
+		if params := function.Get("parameters"); params.Exists() {
+			addIfNotEmpty(segments, params.Raw)
+		}
+	}
+}
+
+func addIfNotEmpty(segments *[]string, value string) {
+	if segments == nil {
+		return
+	}
+	if trimmed := strings.TrimSpace(value); trimmed != "" {
+		*segments = append(*segments, trimmed)
+	}
+}
--- a/internal/translator/codex/claude/codex_claude_response.go
+++ b/internal/translator/codex/claude/codex_claude_response.go
@@ -354,3 +354,7 @@ func buildReverseMapFromClaudeOriginalShortToOriginal(original []byte) map[strin
 	}
 	return rev
 }
+
+func ClaudeTokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"input_tokens":%d}`, count)
+}
--- a/internal/translator/codex/claude/init.go
+++ b/internal/translator/codex/claude/init.go
@@ -12,8 +12,9 @@ func init() {
 		Codex,
 		ConvertClaudeRequestToCodex,
 		interfaces.TranslateResponse{
-			Stream:    ConvertCodexResponseToClaude,
-			NonStream: ConvertCodexResponseToClaudeNonStream,
+			Stream:     ConvertCodexResponseToClaude,
+			NonStream:  ConvertCodexResponseToClaudeNonStream,
+			TokenCount: ClaudeTokenCount,
 		},
 	)
 }
--- a/internal/translator/codex/gemini-cli/codex_gemini-cli_response.go
+++ b/internal/translator/codex/gemini-cli/codex_gemini-cli_response.go
@@ -6,6 +6,7 @@ package geminiCLI

 import (
 	"context"
+	"fmt"

 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/codex/gemini"
 	"github.com/tidwall/sjson"
@@ -54,3 +55,7 @@ func ConvertCodexResponseToGeminiCLINonStream(ctx context.Context, modelName str
 	strJSON, _ = sjson.SetRaw(json, "response", strJSON)
 	return strJSON
 }
+
+func GeminiCLITokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
+}
--- a/internal/translator/codex/gemini-cli/init.go
+++ b/internal/translator/codex/gemini-cli/init.go
@@ -12,8 +12,9 @@ func init() {
 		Codex,
 		ConvertGeminiCLIRequestToCodex,
 		interfaces.TranslateResponse{
-			Stream:    ConvertCodexResponseToGeminiCLI,
-			NonStream: ConvertCodexResponseToGeminiCLINonStream,
+			Stream:     ConvertCodexResponseToGeminiCLI,
+			NonStream:  ConvertCodexResponseToGeminiCLINonStream,
+			TokenCount: GeminiCLITokenCount,
 		},
 	)
 }
--- a/internal/translator/codex/gemini/codex_gemini_response.go
+++ b/internal/translator/codex/gemini/codex_gemini_response.go
@@ -8,6 +8,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"fmt"
 	"time"

 	"github.com/tidwall/gjson"
@@ -330,3 +331,7 @@ func mustMarshalJSON(v interface{}) string {
 	}
 	return string(data)
 }
+
+func GeminiTokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
+}
--- a/internal/translator/codex/gemini/init.go
+++ b/internal/translator/codex/gemini/init.go
@@ -12,8 +12,9 @@ func init() {
 		Codex,
 		ConvertGeminiRequestToCodex,
 		interfaces.TranslateResponse{
-			Stream:    ConvertCodexResponseToGemini,
-			NonStream: ConvertCodexResponseToGeminiNonStream,
+			Stream:     ConvertCodexResponseToGemini,
+			NonStream:  ConvertCodexResponseToGeminiNonStream,
+			TokenCount: GeminiTokenCount,
 		},
 	)
 }
--- a/internal/translator/openai/claude/init.go
+++ b/internal/translator/openai/claude/init.go
@@ -12,8 +12,9 @@ func init() {
 		OpenAI,
 		ConvertClaudeRequestToOpenAI,
 		interfaces.TranslateResponse{
-			Stream:    ConvertOpenAIResponseToClaude,
-			NonStream: ConvertOpenAIResponseToClaudeNonStream,
+			Stream:     ConvertOpenAIResponseToClaude,
+			NonStream:  ConvertOpenAIResponseToClaudeNonStream,
+			TokenCount: ClaudeTokenCount,
 		},
 	)
 }
--- a/internal/translator/openai/claude/openai_claude_response.go
+++ b/internal/translator/openai/claude/openai_claude_response.go
@@ -9,6 +9,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"fmt"
 	"strings"

 	"github.com/router-for-me/CLIProxyAPI/v6/internal/util"
@@ -630,3 +631,7 @@ func ConvertOpenAIResponseToClaudeNonStream(_ context.Context, _ string, origina
 	}
 	return string(responseJSON)
 }
+
+func ClaudeTokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"input_tokens":%d}`, count)
+}
--- a/internal/translator/openai/gemini-cli/init.go
+++ b/internal/translator/openai/gemini-cli/init.go
@@ -12,8 +12,9 @@ func init() {
 		OpenAI,
 		ConvertGeminiCLIRequestToOpenAI,
 		interfaces.TranslateResponse{
-			Stream:    ConvertOpenAIResponseToGeminiCLI,
-			NonStream: ConvertOpenAIResponseToGeminiCLINonStream,
+			Stream:     ConvertOpenAIResponseToGeminiCLI,
+			NonStream:  ConvertOpenAIResponseToGeminiCLINonStream,
+			TokenCount: GeminiCLITokenCount,
 		},
 	)
 }
--- a/internal/translator/openai/gemini-cli/openai_gemini_response.go
+++ b/internal/translator/openai/gemini-cli/openai_gemini_response.go
@@ -7,6 +7,7 @@ package geminiCLI

 import (
 	"context"
+	"fmt"

 	. "github.com/router-for-me/CLIProxyAPI/v6/internal/translator/openai/gemini"
 	"github.com/tidwall/sjson"
@@ -51,3 +52,7 @@ func ConvertOpenAIResponseToGeminiCLINonStream(ctx context.Context, modelName st
 	strJSON, _ = sjson.SetRaw(json, "response", strJSON)
 	return strJSON
 }
+
+func GeminiCLITokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
+}
--- a/internal/translator/openai/gemini/init.go
+++ b/internal/translator/openai/gemini/init.go
@@ -12,8 +12,9 @@ func init() {
 		OpenAI,
 		ConvertGeminiRequestToOpenAI,
 		interfaces.TranslateResponse{
-			Stream:    ConvertOpenAIResponseToGemini,
-			NonStream: ConvertOpenAIResponseToGeminiNonStream,
+			Stream:     ConvertOpenAIResponseToGemini,
+			NonStream:  ConvertOpenAIResponseToGeminiNonStream,
+			TokenCount: GeminiTokenCount,
 		},
 	)
 }
--- a/internal/translator/openai/gemini/openai_gemini_response.go
+++ b/internal/translator/openai/gemini/openai_gemini_response.go
@@ -9,6 +9,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"fmt"
 	"strconv"
 	"strings"

@@ -609,3 +610,7 @@ func ConvertOpenAIResponseToGeminiNonStream(_ context.Context, _ string, origina

 	return out
 }
+
+func GeminiTokenCount(ctx context.Context, count int64) string {
+	return fmt.Sprintf(`{"totalTokens":%d,"promptTokensDetails":[{"modality":"TEXT","tokenCount":%d}]}`, count, count)
+}