Merge pull request 'feat: add multi-key support with automatic rate-limit rotation' (#11) from renekv/ocgo:feat/multi-key-rotation into main

Reviewed-on: #11

Makefile

@@ -1,12 +1,14 @@
 .PHONY: build run test clean install release
 
 VERSION ?= $(shell git describe --tags --always --dirty 2>/dev/null || echo dev)
+EXE := $(shell go env GOEXE)
+OCGO_BIN := bin/ocgo$(EXE)
 GOBIN := $(shell go env GOBIN)
 GOPATH := $(shell go env GOPATH)
-INSTALL_DIR := $(if $(GOBIN),$(GOBIN),$(GOPATH)/bin)
+INSTALL_DIR := $(HOME)/.local/bin
 
 build:
-	go build -ldflags "-X main.version=$(VERSION)" -o bin/ocgo ./cmd/ocgo
+	go build -ldflags "-X main.version=$(VERSION)" -o $(OCGO_BIN) ./cmd/ocgo
 
 run:
 	go run ./cmd/ocgo
@@ -19,7 +21,7 @@ clean:
 
 install: build
 	mkdir -p "$(INSTALL_DIR)"
-	install -m 0755 bin/ocgo "$(INSTALL_DIR)/ocgo"
+	install -m 0755 "$(OCGO_BIN)" "$(INSTALL_DIR)/ocgo$(EXE)"
 
 release:
 	@[ -n "$(TAG)" ] || (echo "Usage: make release TAG=v0.1.0" && exit 1)

build.bat

@@ -0,0 +1,9 @@
+@echo off
+setlocal
+
+for /f "delims=" %%v in ('git describe --tags --always --dirty 2^>NUL') do set "OCGO_VERSION=%%v"
+if not defined OCGO_VERSION set "OCGO_VERSION=dev"
+
+if not exist bin mkdir bin
+go build -ldflags "-X main.version=%OCGO_VERSION%" -o bin\ocgo.exe .\cmd\ocgo
+exit /b %ERRORLEVEL%

cmd/ocgo/listener_pid_netstat.go

@@ -0,0 +1,40 @@
+package main
+
+import (
+	"errors"
+	"strconv"
+	"strings"
+)
+
+func parseWindowsNetstatPID(output string, port int) (int, error) {
+	wanted := strconv.Itoa(port)
+	for _, line := range strings.Split(output, "\n") {
+		fields := strings.Fields(line)
+		if len(fields) < 5 || !strings.EqualFold(fields[0], "tcp") {
+			continue
+		}
+		if !strings.EqualFold(fields[len(fields)-2], "listening") {
+			continue
+		}
+		if !netstatAddressUsesPort(fields[1], wanted) {
+			continue
+		}
+		pid, err := strconv.Atoi(fields[len(fields)-1])
+		if err == nil && pid > 0 {
+			return pid, nil
+		}
+	}
+	return 0, errors.New("no listener found")
+}
+
+func netstatAddressUsesPort(address, port string) bool {
+	address = strings.TrimSpace(address)
+	if address == "" {
+		return false
+	}
+	if strings.HasPrefix(address, "[") {
+		return strings.HasSuffix(address, "]:"+port)
+	}
+	idx := strings.LastIndex(address, ":")
+	return idx >= 0 && address[idx+1:] == port
+}

cmd/ocgo/listener_pid_unix.go

@@ -0,0 +1,27 @@
+//go:build !windows
+
+package main
+
+import (
+	"errors"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+func findListenerPID(port int) (int, error) {
+	if port == 0 {
+		return 0, errors.New("missing port")
+	}
+	out, err := exec.Command("lsof", "-nP", "-tiTCP:"+strconv.Itoa(port), "-sTCP:LISTEN").Output()
+	if err != nil {
+		return 0, err
+	}
+	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
+		pid, err := strconv.Atoi(strings.TrimSpace(line))
+		if err == nil && pid > 0 {
+			return pid, nil
+		}
+	}
+	return 0, errors.New("no listener found")
+}

cmd/ocgo/listener_pid_windows.go

@@ -0,0 +1,19 @@
+//go:build windows
+
+package main
+
+import (
+	"errors"
+	"os/exec"
+)
+
+func findListenerPID(port int) (int, error) {
+	if port == 0 {
+		return 0, errors.New("missing port")
+	}
+	out, err := exec.Command("netstat", "-ano", "-p", "tcp").Output()
+	if err != nil {
+		return 0, err
+	}
+	return parseWindowsNetstatPID(string(out), port)
+}

cmd/ocgo/main.go

@@ -32,9 +32,11 @@ const (
 var version = "dev"
 
 type Config struct {
-	APIKey string `json:"api_key"`
-	Host   string `json:"host"`
-	Port   int    `json:"port"`
+	APIKey  string   `json:"api_key,omitempty"`
+	APIKeys []string `json:"api_keys,omitempty"`
+	Host    string   `json:"host"`
+	Port    int      `json:"port"`
+	Model   string   `json:"model"`
 }
 
 type AnthropicRequest struct {
@@ -60,13 +62,18 @@ type ATool struct {
 }
 
 type OAIRequest struct {
-	Model       string       `json:"model"`
-	Messages    []OAIMessage `json:"messages"`
-	Stream      bool         `json:"stream,omitempty"`
-	MaxTokens   int          `json:"max_tokens,omitempty"`
-	Temperature *float64     `json:"temperature,omitempty"`
-	TopP        *float64     `json:"top_p,omitempty"`
-	Tools       []OAITool    `json:"tools,omitempty"`
+	Model         string            `json:"model"`
+	Messages      []OAIMessage      `json:"messages"`
+	Stream        bool              `json:"stream,omitempty"`
+	StreamOptions *OAIStreamOptions `json:"stream_options,omitempty"`
+	MaxTokens     int               `json:"max_tokens,omitempty"`
+	Temperature   *float64          `json:"temperature,omitempty"`
+	TopP          *float64          `json:"top_p,omitempty"`
+	Tools         []OAITool         `json:"tools,omitempty"`
+}
+
+type OAIStreamOptions struct {
+	IncludeUsage bool `json:"include_usage"`
 }
 
 type ResponsesRequest struct {
@@ -142,30 +149,38 @@ func main() {
 }
 
 func setupCmd() *cobra.Command {
-	var key string
+	var apiKeys string
 	cmd := &cobra.Command{
 		Use:   "setup",
-		Short: "Save your OpenCode Go API key",
+		Short: "Save your OpenCode Go API key(s)",
 		RunE: func(cmd *cobra.Command, args []string) error {
-			if strings.TrimSpace(key) == "" {
-				key = os.Getenv("OCGO_API_KEY")
+			if apiKeys == "" {
+				apiKeys, _ = cmd.Flags().GetString("api-key")
 			}
-			if strings.TrimSpace(key) == "" {
-				fmt.Print("OpenCode Go API key: ")
+			if strings.TrimSpace(apiKeys) == "" {
+				apiKeys = os.Getenv("OCGO_API_KEYS")
+			}
+			if strings.TrimSpace(apiKeys) == "" {
+				apiKeys = os.Getenv("OCGO_API_KEY")
+			}
+			if strings.TrimSpace(apiKeys) == "" {
+				fmt.Print("OpenCode Go API key(s) (comma-separated): ")
 				line, err := bufio.NewReader(os.Stdin).ReadString('\n')
 				if err != nil && line == "" {
 					return err
 				}
-				key = line
+				apiKeys = line
 			}
-			cfg := Config{APIKey: strings.TrimSpace(key), Host: defaultHost, Port: defaultPort}
-			if cfg.APIKey == "" {
-				return errors.New("API key cannot be empty")
+			keys := parseKeys(strings.TrimSpace(apiKeys))
+			if len(keys) == 0 {
+				return errors.New("at least one API key is required")
 			}
+			cfg := Config{APIKeys: keys, Host: defaultHost, Port: defaultPort}
 			return saveConfig(cfg)
 		},
 	}
-	cmd.Flags().StringVar(&key, "api-key", "", "OpenCode Go API key")
+	cmd.Flags().StringVar(&apiKeys, "api-keys", "", "OpenCode Go API key(s), comma-separated")
+	cmd.Flags().String("api-key", "", "OpenCode Go API key (single)")
 	return cmd
 }
 
@@ -179,7 +194,39 @@ func listCmd() *cobra.Command {
 }
 
 func knownModelIDs() []string {
-	return []string{"glm-5.1", "glm-5", "kimi-k2.6", "kimi-k2.5", "mimo-v2.5-pro", "mimo-v2.5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.7", "minimax-m2.5", "deepseek-v4-pro", "deepseek-v4-flash", "qwen3.6-plus", "qwen3.5-plus"}
+	return []string{"glm-5.1", "glm-5", "kimi-k2.6", "kimi-k2.5", "mimo-v2.5-pro", "mimo-v2.5", "mimo-v2-pro", "mimo-v2-omni", "minimax-m2.7", "minimax-m2.5", "deepseek-v4-pro[1M]", "deepseek-v4-flash[1M]", "qwen3.6-plus", "qwen3.5-plus"}
+}
+
+func isKnownModelID(model string) bool {
+	cleaned := cleanModelName(model)
+	for _, m := range knownModelIDs() {
+		if cleanModelName(m) == cleaned {
+			return true
+		}
+	}
+	return false
+}
+
+func cleanModelName(model string) string {
+	model = strings.TrimSuffix(model, "[1M]")
+	model = strings.TrimSuffix(model, "[128K]")
+	return model
+}
+
+func modelContextWindow(model string) int {
+	switch cleanModelName(model) {
+	case "deepseek-v4-pro", "deepseek-v4-flash":
+		return 1000000
+	default:
+		return 128000
+	}
+}
+
+func modelDisplayName(model string) string {
+	if modelContextWindow(model) >= 1000000 {
+		return model + "[1M]"
+	}
+	return model
 }
 
 func modelSupportsImages(model string) bool {
@@ -279,17 +326,22 @@ func launchCmd() *cobra.Command {
 
 func serveCmd() *cobra.Command {
 	var background bool
+	var model string
 	cmd := &cobra.Command{Use: "serve", Short: "Start local Anthropic-compatible proxy", RunE: func(cmd *cobra.Command, args []string) error {
 		if background {
-			return startBackground()
+			return startBackground(model)
 		}
 		cfg, err := loadConfig()
 		if err != nil {
 			return err
 		}
+		if model != "" {
+			cfg.Model = model
+		}
 		return runServer(cfg)
 	}}
 	cmd.Flags().BoolVarP(&background, "background", "b", false, "Run proxy in the background")
+	cmd.Flags().StringVarP(&model, "model", "m", "", "Upstream model ID (default: kimi-k2.6)")
 	return cmd
 }
 
@@ -330,6 +382,10 @@ func statusCmd() *cobra.Command {
 			fmt.Printf("Proxy is running on %s:%d (PID %d)\n", cfg.Host, cfg.Port, pid)
 			return
 		}
+		if pid, err := findListenerPID(cfg.Port); err == nil {
+			fmt.Printf("Proxy is running on %s:%d (PID %d, discovered from listener)\n", cfg.Host, cfg.Port, pid)
+			return
+		}
 		fmt.Printf("Proxy is running on %s:%d (no ocgo PID file)\n", cfg.Host, cfg.Port)
 	}}
 }
@@ -360,20 +416,13 @@ func proxyMessages(w http.ResponseWriter, r *http.Request, cfg Config) {
 		http.Error(w, "invalid request", http.StatusBadRequest)
 		return
 	}
-	or := convertRequest(ar)
+	or := convertRequest(ar, cfg.Model)
 	if err := validateImageSupport(or); err != nil {
 		http.Error(w, err.Error(), http.StatusBadRequest)
 		return
 	}
 	body, _ := json.Marshal(or)
-	req, err := http.NewRequestWithContext(r.Context(), http.MethodPost, openAIURL, bytes.NewReader(body))
-	if err != nil {
-		http.Error(w, err.Error(), http.StatusInternalServerError)
-		return
-	}
-	req.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := (&http.Client{Timeout: 10 * time.Minute}).Do(req)
+	resp, err := cfg.postWithRetry(r.Context(), openAIURL, "application/json", body)
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusBadGateway)
 		return
@@ -385,10 +434,10 @@ func proxyMessages(w http.ResponseWriter, r *http.Request, cfg Config) {
 		return
 	}
 	if ar.Stream {
-		streamAnthropic(w, resp.Body, or.Model)
+		streamAnthropic(w, resp.Body, modelDisplayName(or.Model))
 		return
 	}
-	writeAnthropicResponse(w, resp.Body, or.Model)
+	writeAnthropicResponse(w, resp.Body, modelDisplayName(or.Model))
 }
 
 func proxyChatCompletions(w http.ResponseWriter, r *http.Request, cfg Config) {
@@ -406,14 +455,18 @@ func proxyChatCompletions(w http.ResponseWriter, r *http.Request, cfg Config) {
 		http.Error(w, err.Error(), http.StatusBadRequest)
 		return
 	}
-	req, err := http.NewRequestWithContext(r.Context(), http.MethodPost, openAIURL, bytes.NewReader(body))
-	if err != nil {
-		http.Error(w, err.Error(), http.StatusInternalServerError)
-		return
+	// Apply default model if configured
+	if cfg.Model != "" {
+		dm := cleanModelName(cfg.Model)
+		var reqMap map[string]any
+		if json.Unmarshal(body, &reqMap) == nil {
+			if m, _ := reqMap["model"].(string); m == "" || !isKnownModelID(m) {
+				reqMap["model"] = dm
+				body, _ = json.Marshal(reqMap)
+			}
+		}
 	}
-	req.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := (&http.Client{Timeout: 10 * time.Minute}).Do(req)
+	resp, err := cfg.postWithRetry(r.Context(), openAIURL, "application/json", body)
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusBadGateway)
 		return
@@ -434,20 +487,13 @@ func proxyResponses(w http.ResponseWriter, r *http.Request, cfg Config) {
 		http.Error(w, "invalid request", http.StatusBadRequest)
 		return
 	}
-	or := responsesToChat(rr)
+	or := responsesToChat(rr, cfg.Model)
 	if err := validateImageSupport(or); err != nil {
 		http.Error(w, err.Error(), http.StatusBadRequest)
 		return
 	}
 	body, _ := json.Marshal(or)
-	req, err := http.NewRequestWithContext(r.Context(), http.MethodPost, openAIURL, bytes.NewReader(body))
-	if err != nil {
-		http.Error(w, err.Error(), http.StatusInternalServerError)
-		return
-	}
-	req.Header.Set("Authorization", "Bearer "+cfg.APIKey)
-	req.Header.Set("Content-Type", "application/json")
-	resp, err := (&http.Client{Timeout: 10 * time.Minute}).Do(req)
+	resp, err := cfg.postWithRetry(r.Context(), openAIURL, "application/json", body)
 	if err != nil {
 		http.Error(w, err.Error(), http.StatusBadGateway)
 		return
@@ -459,10 +505,10 @@ func proxyResponses(w http.ResponseWriter, r *http.Request, cfg Config) {
 		return
 	}
 	if rr.Stream {
-		streamResponses(w, resp.Body, or.Model)
+		streamResponses(w, resp.Body, modelDisplayName(or.Model))
 		return
 	}
-	writeResponsesResponse(w, resp.Body, or.Model)
+	writeResponsesResponse(w, resp.Body, modelDisplayName(or.Model))
 }
 
 func copyHeaders(dst, src http.Header) {
@@ -478,14 +524,19 @@ func prepareChatBody(body []byte) ([]byte, error) {
 	if json.Unmarshal(body, &req) != nil {
 		return body, nil
 	}
+	changed := requestStreamingUsage(req)
 	model, _ := req["model"].(string)
-	if !rawChatBodyHasImages(req) {
-		return body, nil
+	if clean := cleanModelName(model); clean != model {
+		req["model"] = clean
+		changed = true
+		model = clean
 	}
-	if !modelSupportsImages(model) {
-		return nil, unsupportedImageModelError(model)
+	if rawChatBodyHasImages(req) {
+		if !modelSupportsImages(model) {
+			return nil, unsupportedImageModelError(model)
+		}
+		changed = stripRawChatImageDetails(req) || changed
 	}
-	changed := stripRawChatImageDetails(req)
 	if !changed {
 		return body, nil
 	}
@@ -496,6 +547,23 @@ func prepareChatBody(body []byte) ([]byte, error) {
 	return out, nil
 }
 
+func requestStreamingUsage(req map[string]any) bool {
+	streaming, _ := req["stream"].(bool)
+	if !streaming {
+		return false
+	}
+	options, ok := req["stream_options"].(map[string]any)
+	if !ok {
+		options = map[string]any{}
+		req["stream_options"] = options
+	}
+	if enabled, _ := options["include_usage"].(bool); enabled {
+		return false
+	}
+	options["include_usage"] = true
+	return true
+}
+
 func rawChatBodyHasImages(req map[string]any) bool {
 	messages, _ := req["messages"].([]any)
 	for _, item := range messages {
@@ -555,12 +623,15 @@ func stripRawChatImageDetails(req map[string]any) bool {
 	return changed
 }
 
-func convertRequest(ar AnthropicRequest) OAIRequest {
-	model := ar.Model
-	if model == "" || strings.HasPrefix(model, "claude-") {
+func convertRequest(ar AnthropicRequest, defaultModel string) OAIRequest {
+	model := cleanModelName(ar.Model)
+	if model == "" || !isKnownModelID(model) {
+		model = cleanModelName(defaultModel)
+	}
+	if model == "" {
 		model = "kimi-k2.6"
 	}
-	out := OAIRequest{Model: model, Stream: ar.Stream, MaxTokens: ar.MaxTokens, Temperature: ar.Temperature, TopP: ar.TopP}
+	out := OAIRequest{Model: model, Stream: ar.Stream, StreamOptions: streamUsageOptions(ar.Stream), MaxTokens: ar.MaxTokens, Temperature: ar.Temperature, TopP: ar.TopP}
 	if sys := systemText(ar.System); sys != "" {
 		out.Messages = append(out.Messages, OAIMessage{Role: "system", Content: sys})
 	}
@@ -573,12 +644,15 @@ func convertRequest(ar AnthropicRequest) OAIRequest {
 	return out
 }
 
-func responsesToChat(rr ResponsesRequest) OAIRequest {
-	model := rr.Model
+func responsesToChat(rr ResponsesRequest, defaultModel string) OAIRequest {
+	model := cleanModelName(rr.Model)
+	if model == "" || !isKnownModelID(model) {
+		model = cleanModelName(defaultModel)
+	}
 	if model == "" {
 		model = "kimi-k2.6"
 	}
-	out := OAIRequest{Model: model, Stream: rr.Stream, MaxTokens: rr.MaxTokens, Temperature: rr.Temperature, TopP: rr.TopP}
+	out := OAIRequest{Model: model, Stream: rr.Stream, StreamOptions: streamUsageOptions(rr.Stream), MaxTokens: rr.MaxTokens, Temperature: rr.Temperature, TopP: rr.TopP}
 	if rr.Instructions != "" {
 		out.Messages = append(out.Messages, OAIMessage{Role: "system", Content: rr.Instructions})
 	}
@@ -591,6 +665,13 @@ func responsesToChat(rr ResponsesRequest) OAIRequest {
 	return out
 }
 
+func streamUsageOptions(streaming bool) *OAIStreamOptions {
+	if !streaming {
+		return nil
+	}
+	return &OAIStreamOptions{IncludeUsage: true}
+}
+
 func requestHasImages(or OAIRequest) bool {
 	for _, m := range or.Messages {
 		if contentHasImage(m.Content) {
@@ -923,6 +1004,98 @@ func blockText(raw json.RawMessage) string {
 	return b.String()
 }
 
+type tokenUsage struct {
+	InputTokens       int
+	OutputTokens      int
+	TotalTokens       int
+	CachedInputTokens int
+	Present           bool
+}
+
+func usageFromJSON(raw json.RawMessage) tokenUsage {
+	var fields map[string]any
+	if len(raw) == 0 || json.Unmarshal(raw, &fields) != nil {
+		return tokenUsage{}
+	}
+	return usageFromFields(fields)
+}
+
+func usageFromFields(fields map[string]any) tokenUsage {
+	if len(fields) == 0 {
+		return tokenUsage{}
+	}
+	u := tokenUsage{Present: true}
+	u.InputTokens = intField(fields, "prompt_tokens")
+	if u.InputTokens == 0 {
+		u.InputTokens = intField(fields, "input_tokens")
+	}
+	u.OutputTokens = intField(fields, "completion_tokens")
+	if u.OutputTokens == 0 {
+		u.OutputTokens = intField(fields, "output_tokens")
+	}
+	u.TotalTokens = intField(fields, "total_tokens")
+	if u.TotalTokens == 0 && (u.InputTokens > 0 || u.OutputTokens > 0) {
+		u.TotalTokens = u.InputTokens + u.OutputTokens
+	}
+	u.CachedInputTokens = cachedTokens(fields)
+	return u
+}
+
+func intField(fields map[string]any, name string) int {
+	v, ok := fields[name]
+	if !ok {
+		return 0
+	}
+	switch n := v.(type) {
+	case float64:
+		return int(n)
+	case int:
+		return n
+	case json.Number:
+		i, _ := n.Int64()
+		return int(i)
+	}
+	return 0
+}
+
+func cachedTokens(fields map[string]any) int {
+	for _, key := range []string{"prompt_tokens_details", "input_tokens_details"} {
+		if nested, ok := fields[key].(map[string]any); ok {
+			if n := intField(nested, "cached_tokens"); n > 0 {
+				return n
+			}
+		}
+	}
+	return intField(fields, "cached_tokens")
+}
+
+func anthropicUsage(u tokenUsage) map[string]int {
+	usage := map[string]int{"input_tokens": u.InputTokens, "output_tokens": u.OutputTokens}
+	if u.CachedInputTokens > 0 {
+		usage["cache_read_input_tokens"] = u.CachedInputTokens
+	}
+	return usage
+}
+
+func anthropicDeltaUsage(u tokenUsage) map[string]int {
+	usage := map[string]int{"output_tokens": u.OutputTokens}
+	if u.InputTokens > 0 {
+		usage["input_tokens"] = u.InputTokens
+	}
+	if u.CachedInputTokens > 0 {
+		usage["cache_read_input_tokens"] = u.CachedInputTokens
+	}
+	return usage
+}
+
+func responsesUsage(u tokenUsage) map[string]any {
+	usage := map[string]any{"input_tokens": u.InputTokens, "output_tokens": u.OutputTokens, "total_tokens": u.TotalTokens}
+	if u.CachedInputTokens > 0 {
+		usage["input_tokens_details"] = map[string]int{"cached_tokens": u.CachedInputTokens}
+	}
+	return usage
+}
+
 func streamAnthropic(w http.ResponseWriter, body io.Reader, model string) {
 	w.Header().Set("Content-Type", "text/event-stream")
 	flusher, _ := w.(http.Flusher)
@@ -936,6 +1109,7 @@ func streamAnthropic(w http.ResponseWriter, body io.Reader, model string) {
 	toolIndexes := map[int]int{}
 	var tools []streamedResponseToolCall
 	var reasoning strings.Builder
+	usage := tokenUsage{}
 	s := bufio.NewScanner(body)
 	for s.Scan() {
 		line := strings.TrimSpace(s.Text())
@@ -947,6 +1121,9 @@ func streamAnthropic(w http.ResponseWriter, body io.Reader, model string) {
 			break
 		}
 		chunk := parseOpenAIStreamChunk([]byte(data))
+		if chunk.Usage.Present {
+			usage = chunk.Usage
+		}
 		if chunk.ReasoningContent != "" {
 			reasoning.WriteString(chunk.ReasoningContent)
 		}
@@ -1010,7 +1187,8 @@ func streamAnthropic(w http.ResponseWriter, body io.Reader, model string) {
 	if len(tools) > 0 {
 		stopReason = "tool_use"
 	}
-	fmt.Fprintf(w, "event: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":%q,\"stop_sequence\":null},\"usage\":{\"output_tokens\":0}}\n\n", stopReason)
+	usageJSON, _ := json.Marshal(anthropicDeltaUsage(usage))
+	fmt.Fprintf(w, "event: message_delta\ndata: {\"type\":\"message_delta\",\"delta\":{\"stop_reason\":%q,\"stop_sequence\":null},\"usage\":%s}\n\n", stopReason, usageJSON)
 	fmt.Fprint(w, "event: message_stop\ndata: {\"type\":\"message_stop\"}\n\n")
 }
 
@@ -1036,6 +1214,7 @@ func writeAnthropicResponse(w http.ResponseWriter, body io.Reader, model string)
 				Content string `json:"content"`
 			} `json:"message"`
 		} `json:"choices"`
+		Usage json.RawMessage `json:"usage"`
 	}
 	_ = json.NewDecoder(body).Decode(&v)
 	text := ""
@@ -1043,7 +1222,7 @@ func writeAnthropicResponse(w http.ResponseWriter, body io.Reader, model string)
 		text = v.Choices[0].Message.Content
 	}
 	w.Header().Set("Content-Type", "application/json")
-	_ = json.NewEncoder(w).Encode(map[string]any{"id": "ocgo", "type": "message", "role": "assistant", "model": model, "content": []map[string]string{{"type": "text", "text": text}}, "stop_reason": "end_turn", "usage": map[string]int{"input_tokens": 0, "output_tokens": 0}})
+	_ = json.NewEncoder(w).Encode(map[string]any{"id": "ocgo", "type": "message", "role": "assistant", "model": model, "content": []map[string]string{{"type": "text", "text": text}}, "stop_reason": "end_turn", "usage": anthropicUsage(usageFromJSON(v.Usage))})
 }
 
 func streamResponses(w http.ResponseWriter, body io.Reader, model string) {
@@ -1060,6 +1239,7 @@ func streamResponses(w http.ResponseWriter, body io.Reader, model string) {
 	nextOutputIndex := 0
 	var text strings.Builder
 	var reasoning strings.Builder
+	usage := tokenUsage{}
 	toolIndexes := map[int]int{}
 	var tools []streamedResponseToolCall
 	s := bufio.NewScanner(body)
@@ -1073,6 +1253,9 @@ func streamResponses(w http.ResponseWriter, body io.Reader, model string) {
 			break
 		}
 		chunk := parseOpenAIStreamChunk([]byte(data))
+		if chunk.Usage.Present {
+			usage = chunk.Usage
+		}
 		if chunk.ReasoningContent != "" {
 			reasoning.WriteString(chunk.ReasoningContent)
 		}
@@ -1140,7 +1323,7 @@ func streamResponses(w http.ResponseWriter, body io.Reader, model string) {
 		writeResponseEvent(w, "response.output_item.done", map[string]any{"type": "response.output_item.done", "output_index": tool.OutputIndex, "item": item})
 		output = append(output, item)
 	}
-	writeResponseEvent(w, "response.completed", map[string]any{"type": "response.completed", "response": map[string]any{"id": id, "object": "response", "model": model, "status": "completed", "output": output}})
+	writeResponseEvent(w, "response.completed", map[string]any{"type": "response.completed", "response": map[string]any{"id": id, "object": "response", "model": model, "status": "completed", "output": output, "usage": responsesUsage(usage)}})
 }
 
 type streamedResponseToolCall struct {
@@ -1159,6 +1342,7 @@ type openAIStreamChunk struct {
 	Content          string
 	ReasoningContent string
 	ToolCalls        []openAIStreamToolCall
+	Usage            tokenUsage
 }
 
 func parseOpenAIStreamChunk(data []byte) openAIStreamChunk {
@@ -1177,13 +1361,16 @@ func parseOpenAIStreamChunk(data []byte) openAIStreamChunk {
 				} `json:"tool_calls"`
 			} `json:"delta"`
 		} `json:"choices"`
+		Usage json.RawMessage `json:"usage"`
 	}
 	_ = json.Unmarshal(data, &v)
+	out := openAIStreamChunk{Usage: usageFromJSON(v.Usage)}
 	if len(v.Choices) == 0 {
-		return openAIStreamChunk{}
+		return out
 	}
 	delta := v.Choices[0].Delta
-	out := openAIStreamChunk{Content: delta.Content, ReasoningContent: delta.ReasoningContent}
+	out.Content = delta.Content
+	out.ReasoningContent = delta.ReasoningContent
 	for _, tc := range delta.ToolCalls {
 		out.ToolCalls = append(out.ToolCalls, openAIStreamToolCall{Index: tc.Index, ID: tc.ID, Name: tc.Function.Name, Arguments: tc.Function.Arguments})
 	}
@@ -1204,6 +1391,7 @@ func writeResponsesResponse(w http.ResponseWriter, body io.Reader, model string)
 				ToolCalls        []OAIToolCall `json:"tool_calls"`
 			} `json:"message"`
 		} `json:"choices"`
+		Usage json.RawMessage `json:"usage"`
 	}
 	_ = json.NewDecoder(body).Decode(&v)
 	text := ""
@@ -1221,7 +1409,7 @@ func writeResponsesResponse(w http.ResponseWriter, body io.Reader, model string)
 		output = append(output, map[string]any{"id": "msg_ocgo", "type": "message", "role": "assistant", "content": []map[string]string{{"type": "output_text", "text": text}}})
 	}
 	w.Header().Set("Content-Type", "application/json")
-	_ = json.NewEncoder(w).Encode(map[string]any{"id": "resp_ocgo", "object": "response", "created_at": time.Now().Unix(), "model": model, "status": "completed", "output": output, "usage": map[string]int{"input_tokens": 0, "output_tokens": 0, "total_tokens": 0}})
+	_ = json.NewEncoder(w).Encode(map[string]any{"id": "resp_ocgo", "object": "response", "created_at": time.Now().Unix(), "model": model, "status": "completed", "output": output, "usage": responsesUsage(usageFromJSON(v.Usage))})
 }
 
 func countTokens(w http.ResponseWriter, r *http.Request) {
@@ -1232,7 +1420,7 @@ func ensureServer(base string) error {
 	if healthy(base) {
 		return nil
 	}
-	if err := startBackground(); err != nil {
+	if err := startBackground(""); err != nil {
 		return err
 	}
 	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
@@ -1250,7 +1438,7 @@ func startLaunchServer(base string) (*exec.Cmd, error) {
 	if healthy(base) {
 		return nil, nil
 	}
-	cmd, err := startServerProcess(false)
+	cmd, err := startServerProcess(false, "")
 	if err != nil {
 		return nil, err
 	}
@@ -1285,12 +1473,12 @@ func healthy(base string) bool {
 	return resp.StatusCode == 200
 }
 
-func startBackground() error {
-	_, err := startServerProcess(true)
+func startBackground(model string) error {
+	_, err := startServerProcess(true, model)
 	return err
 }
 
-func startServerProcess(detached bool) (*exec.Cmd, error) {
+func startServerProcess(detached bool, model string) (*exec.Cmd, error) {
 	bin, err := os.Executable()
 	if err != nil {
 		return nil, err
@@ -1299,6 +1487,9 @@ func startServerProcess(detached bool) (*exec.Cmd, error) {
 		return nil, err
 	}
 	args := []string{"serve"}
+	if model != "" {
+		args = append(args, "--model", model)
+	}
 	cmd := exec.Command(bin, args...)
 	logf, err := os.OpenFile(filepath.Join(configDir(), "ocgo.log"), os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0644)
 	if err != nil {
@@ -1319,6 +1510,76 @@ func startServerProcess(detached bool) (*exec.Cmd, error) {
 func configDir() string  { home, _ := os.UserHomeDir(); return filepath.Join(home, ".config", "ocgo") }
 func configFile() string { return filepath.Join(configDir(), "config.json") }
 func pidFile() string    { return filepath.Join(configDir(), "ocgo.pid") }
+func keyIndexFile() string { return filepath.Join(configDir(), "key-index") }
+
+func readKeyIndex() int {
+	b, err := os.ReadFile(keyIndexFile())
+	if err != nil {
+		return 0
+	}
+	var idx int
+	fmt.Sscanf(string(b), "%d", &idx)
+	return idx
+}
+
+func writeKeyIndex(idx int) error {
+	return os.WriteFile(keyIndexFile(), []byte(fmt.Sprint(idx)), 0600)
+}
+
+func (cfg Config) activeKeys() []string {
+	if len(cfg.APIKeys) > 0 {
+		return cfg.APIKeys
+	}
+	if cfg.APIKey != "" {
+		return []string{cfg.APIKey}
+	}
+	return nil
+}
+
+func (cfg Config) currentKey() string {
+	keys := cfg.activeKeys()
+	if len(keys) == 0 {
+		return ""
+	}
+	idx := readKeyIndex() % len(keys)
+	return keys[idx]
+}
+
+func (cfg Config) rotateKey() string {
+	keys := cfg.activeKeys()
+	if len(keys) <= 1 {
+		return cfg.currentKey()
+	}
+	idx := readKeyIndex()
+	idx = (idx + 1) % len(keys)
+	_ = writeKeyIndex(idx)
+	return keys[idx]
+}
+
+func (cfg Config) postWithRetry(ctx context.Context, url, contentType string, body []byte) (*http.Response, error) {
+	client := &http.Client{Timeout: 10 * time.Minute}
+	keys := cfg.activeKeys()
+	for i := 0; i < len(keys); i++ {
+		key := cfg.currentKey()
+		req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+		if err != nil {
+			return nil, err
+		}
+		req.Header.Set("Authorization", "Bearer "+key)
+		req.Header.Set("Content-Type", contentType)
+		resp, err := client.Do(req)
+		if err != nil {
+			return nil, err
+		}
+		if resp.StatusCode == http.StatusTooManyRequests && i < len(keys)-1 {
+			resp.Body.Close()
+			cfg.rotateKey()
+			continue
+		}
+		return resp, nil
+	}
+	panic("unreachable")
+}
 
 func codexConfigFile() string {
 	home, _ := os.UserHomeDir()
@@ -1382,6 +1643,7 @@ func writeCodexProfile(path, baseURL string) error {
 func writeCodexModelCatalog(path string) error {
 	models := make([]map[string]any, 0, len(knownModelIDs()))
 	for i, id := range knownModelIDs() {
+		cw := modelContextWindow(id)
 		models = append(models, map[string]any{
 			"slug":                             id,
 			"display_name":                     id,
@@ -1404,8 +1666,8 @@ func writeCodexModelCatalog(path string) error {
 			"truncation_policy":                map[string]any{"mode": "tokens", "limit": 10000},
 			"supports_parallel_tool_calls":     false,
 			"supports_image_detail_original":   false,
-			"context_window":                   128000,
-			"max_context_window":               128000,
+			"context_window":                   cw,
+			"max_context_window":               cw,
 			"auto_compact_token_limit":         nil,
 			"effective_context_window_percent": 95,
 			"experimental_supported_tools":     []any{},
@@ -1470,6 +1732,12 @@ func versionParts(v string) [3]int {
 }
 
 func saveConfig(cfg Config) error {
+	// Write api_key for backward compatibility with older ocgo versions
+	if len(cfg.APIKeys) > 0 {
+		cfg.APIKey = cfg.APIKeys[0]
+	} else {
+		cfg.APIKey = ""
+	}
 	if err := os.MkdirAll(configDir(), 0755); err != nil {
 		return err
 	}
@@ -1482,12 +1750,21 @@ func saveConfig(cfg Config) error {
 }
 
 func loadConfig() (Config, error) {
-	cfg := Config{Host: defaultHost, Port: defaultPort, APIKey: os.Getenv("OCGO_API_KEY")}
+	cfg := Config{Host: defaultHost, Port: defaultPort}
 	b, err := os.ReadFile(configFile())
 	if err == nil {
 		_ = json.Unmarshal(b, &cfg)
 	}
-	if cfg.APIKey == "" {
+	if envKey := os.Getenv("OCGO_API_KEY"); envKey != "" {
+		cfg.APIKey = envKey
+	}
+	if envKeys := os.Getenv("OCGO_API_KEYS"); envKeys != "" {
+		cfg.APIKeys = parseKeys(envKeys)
+	}
+	if len(cfg.APIKeys) == 0 && cfg.APIKey != "" {
+		cfg.APIKeys = []string{cfg.APIKey}
+	}
+	if len(cfg.APIKeys) == 0 {
 		return cfg, errors.New("missing API key; run: ocgo setup")
 	}
 	if cfg.Host == "" {
@@ -1499,6 +1776,17 @@ func loadConfig() (Config, error) {
 	return cfg, nil
 }
 
+func parseKeys(s string) []string {
+	parts := strings.Split(s, ",")
+	var keys []string
+	for _, p := range parts {
+		if k := strings.TrimSpace(p); k != "" {
+			keys = append(keys, k)
+		}
+	}
+	return keys
+}
+
 func readPID() (int, error) {
 	b, err := os.ReadFile(pidFile())
 	if err != nil {
@@ -1508,24 +1796,3 @@ func readPID() (int, error) {
 	_, err = fmt.Sscan(string(b), &pid)
 	return pid, err
 }
-
-func findListenerPID(port int) (int, error) {
-	if port == 0 {
-		return 0, errors.New("missing port")
-	}
-	out, err := exec.Command("lsof", "-nP", "-tiTCP:"+strconv.Itoa(port), "-sTCP:LISTEN").Output()
-	if err != nil {
-		return 0, err
-	}
-	for _, line := range strings.Split(strings.TrimSpace(string(out)), "\n") {
-		line = strings.TrimSpace(line)
-		if line == "" {
-			continue
-		}
-		pid, err := strconv.Atoi(line)
-		if err == nil && pid > 0 {
-			return pid, nil
-		}
-	}
-	return 0, errors.New("no listener found")
-}

cmd/ocgo/main_test.go

@@ -1,6 +1,7 @@
 package main
 
 import (
+	"encoding/json"
 	"net/http/httptest"
 	"os"
 	"path/filepath"
@@ -59,6 +60,36 @@ func TestWriteCodexProfileReplacesExistingSections(t *testing.T) {
 	}
 }
 
+func TestModelContextWindow(t *testing.T) {
+	if got := modelContextWindow("deepseek-v4-flash[1M]"); got != 1000000 {
+		t.Fatalf("deepseek-v4-flash[1M]: got %d, want 1000000", got)
+	}
+	if got := modelContextWindow("deepseek-v4-pro[1M]"); got != 1000000 {
+		t.Fatalf("deepseek-v4-pro[1M]: got %d, want 1000000", got)
+	}
+	if got := modelContextWindow("kimi-k2.6"); got != 128000 {
+		t.Fatalf("kimi-k2.6: got %d, want 128000", got)
+	}
+	if got := modelContextWindow("unknown-model"); got != 128000 {
+		t.Fatalf("unknown-model: got %d, want 128000", got)
+	}
+}
+
+func TestCleanModelName(t *testing.T) {
+	tests := []struct{ in, want string }{
+		{"deepseek-v4-flash[1M]", "deepseek-v4-flash"},
+		{"deepseek-v4-pro[1M]", "deepseek-v4-pro"},
+		{"deepseek-v4-flash", "deepseek-v4-flash"},
+		{"kimi-k2.6", "kimi-k2.6"},
+		{"kimi-k2.6[128K]", "kimi-k2.6"},
+	}
+	for _, tt := range tests {
+		if got := cleanModelName(tt.in); got != tt.want {
+			t.Errorf("cleanModelName(%q) = %q, want %q", tt.in, got, tt.want)
+		}
+	}
+}
+
 func TestWriteCodexModelCatalog(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "ocgo-models.json")
 	if err := writeCodexModelCatalog(path); err != nil {
@@ -69,7 +100,7 @@ func TestWriteCodexModelCatalog(t *testing.T) {
 		t.Fatal(err)
 	}
 	content := string(b)
-	for _, want := range []string{`"models"`, `"slug": "deepseek-v4-pro"`, `"context_window": 128000`, `"truncation_policy"`, `"supports_image_detail_original": false`, `"image"`} {
+	for _, want := range []string{`"models"`, `"slug": "deepseek-v4-pro[1M]"`, `"context_window": 1000000`, `"truncation_policy"`, `"supports_image_detail_original": false`, `"image"`} {
 		if !strings.Contains(content, want) {
 			t.Fatalf("missing %q in:\n%s", want, content)
 		}
@@ -186,7 +217,7 @@ func TestResponsesInputPreservesImages(t *testing.T) {
 
 func TestResponsesImageKeepsKimiModel(t *testing.T) {
 	req := ResponsesRequest{Model: "kimi-k2.6", Input: []byte(`[{"type":"message","role":"user","content":[{"type":"input_text","text":"describe this"},{"type":"input_image","image_url":"data:image/png;base64,abc"}]}]`)}
-	out := responsesToChat(req)
+	out := responsesToChat(req, "")
 	if out.Model != "kimi-k2.6" {
 		t.Fatalf("image request should keep Kimi model, got %q", out.Model)
 	}
@@ -197,7 +228,7 @@ func TestResponsesImageKeepsKimiModel(t *testing.T) {
 
 func TestResponsesImageRejectsUnsupportedModel(t *testing.T) {
 	req := ResponsesRequest{Model: "deepseek-v4-pro", Input: []byte(`[{"type":"message","role":"user","content":[{"type":"input_text","text":"describe this"},{"type":"input_image","image_url":"data:image/png;base64,abc"}]}]`)}
-	out := responsesToChat(req)
+	out := responsesToChat(req, "")
 	if err := validateImageSupport(out); err == nil || !strings.Contains(err.Error(), "deepseek-v4-pro") {
 		t.Fatalf("DeepSeek image request should be rejected, got %v", err)
 	}
@@ -223,6 +254,39 @@ func TestRawChatImageRejectsUnsupportedModel(t *testing.T) {
 	}
 }
 
+func TestRawChatStreamRequestsUsage(t *testing.T) {
+	body, err := prepareChatBody([]byte(`{"model":"kimi-k2.6","stream":true,"stream_options":{"foo":"bar"},"messages":[{"role":"user","content":"hello"}]}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var req map[string]any
+	if err := json.Unmarshal(body, &req); err != nil {
+		t.Fatal(err)
+	}
+	options, ok := req["stream_options"].(map[string]any)
+	if !ok {
+		t.Fatalf("missing stream options in %s", string(body))
+	}
+	if options["include_usage"] != true || options["foo"] != "bar" {
+		t.Fatalf("bad stream options: %+v", options)
+	}
+}
+
+func TestConvertedStreamingRequestsAskForUsage(t *testing.T) {
+	anthropic := convertRequest(AnthropicRequest{Model: "kimi-k2.6", Stream: true, Messages: []AMessage{{Role: "user", Content: []byte(`hello`)}}}, "")
+	if anthropic.StreamOptions == nil || !anthropic.StreamOptions.IncludeUsage {
+		t.Fatalf("anthropic conversion should request stream usage: %+v", anthropic.StreamOptions)
+	}
+	responses := responsesToChat(ResponsesRequest{Model: "kimi-k2.6", Stream: true, Input: []byte(`"hello"`)}, "")
+	if responses.StreamOptions == nil || !responses.StreamOptions.IncludeUsage {
+		t.Fatalf("responses conversion should request stream usage: %+v", responses.StreamOptions)
+	}
+	plain := responsesToChat(ResponsesRequest{Model: "kimi-k2.6", Input: []byte(`"hello"`)}, "")
+	if plain.StreamOptions != nil {
+		t.Fatalf("non-streaming conversion should not set stream options: %+v", plain.StreamOptions)
+	}
+}
+
 func TestAnthropicContentPreservesImages(t *testing.T) {
 	messages := contentToOpenAI(AMessage{Role: "user", Content: []byte(`[{"type":"text","text":"what is this?"},{"type":"image","source":{"type":"base64","media_type":"image/jpeg","data":"abc"}}]`)})
 	if len(messages) != 1 {
@@ -241,7 +305,7 @@ func TestAnthropicContentPreservesImages(t *testing.T) {
 }
 
 func TestAnthropicImageKeepsKimiModel(t *testing.T) {
-	out := convertRequest(AnthropicRequest{Model: "kimi-k2.6", Messages: []AMessage{{Role: "user", Content: []byte(`[{"type":"text","text":"what is this?"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"abc"}}]`)}}})
+	out := convertRequest(AnthropicRequest{Model: "kimi-k2.6", Messages: []AMessage{{Role: "user", Content: []byte(`[{"type":"text","text":"what is this?"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"abc"}}]`)}}}, "")
 	if out.Model != "kimi-k2.6" {
 		t.Fatalf("image request should keep Kimi model, got %q", out.Model)
 	}
@@ -251,7 +315,7 @@ func TestAnthropicImageKeepsKimiModel(t *testing.T) {
 }
 
 func TestAnthropicImageRejectsUnsupportedModel(t *testing.T) {
-	out := convertRequest(AnthropicRequest{Model: "deepseek-v4-pro", Messages: []AMessage{{Role: "user", Content: []byte(`[{"type":"text","text":"what is this?"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"abc"}}]`)}}})
+	out := convertRequest(AnthropicRequest{Model: "deepseek-v4-pro", Messages: []AMessage{{Role: "user", Content: []byte(`[{"type":"text","text":"what is this?"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"abc"}}]`)}}}, "")
 	if err := validateImageSupport(out); err == nil || !strings.Contains(err.Error(), "deepseek-v4-pro") {
 		t.Fatalf("DeepSeek image request should be rejected, got %v", err)
 	}
@@ -262,6 +326,115 @@ func contentString(v any) string {
 	return s
 }
 
+func TestParseWindowsNetstatPID(t *testing.T) {
+	output := strings.Join([]string{
+		"Proto  Local Address          Foreign Address        State           PID",
+		"TCP    127.0.0.1:3456       0.0.0.0:0              LISTENING       4321",
+		"TCP    [::1]:9999           [::]:0                 LISTENING       8765",
+		"TCP    127.0.0.1:34560      0.0.0.0:0              LISTENING       1111",
+	}, "\n")
+	pid, err := parseWindowsNetstatPID(output, 3456)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if pid != 4321 {
+		t.Fatalf("pid = %d, want 4321", pid)
+	}
+}
+
+func TestParseWindowsNetstatPIDMatchesIPv6(t *testing.T) {
+	output := "TCP    [::]:3456             [::]:0                 LISTENING       2468\n"
+	pid, err := parseWindowsNetstatPID(output, 3456)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if pid != 2468 {
+		t.Fatalf("pid = %d, want 2468", pid)
+	}
+}
+
+func TestWriteAnthropicResponseIncludesUsage(t *testing.T) {
+	body := strings.NewReader(`{"choices":[{"message":{"content":"done"}}],"usage":{"prompt_tokens":11,"completion_tokens":5,"total_tokens":16,"prompt_tokens_details":{"cached_tokens":4}}}`)
+	w := httptest.NewRecorder()
+	writeAnthropicResponse(w, body, "kimi-k2.6")
+	var out map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &out); err != nil {
+		t.Fatal(err)
+	}
+	usage, ok := out["usage"].(map[string]any)
+	if !ok {
+		t.Fatalf("missing usage: %+v", out)
+	}
+	if usage["input_tokens"] != float64(11) || usage["output_tokens"] != float64(5) || usage["cache_read_input_tokens"] != float64(4) {
+		t.Fatalf("bad anthropic usage: %+v", usage)
+	}
+}
+
+func TestWriteResponsesResponseIncludesUsage(t *testing.T) {
+	body := strings.NewReader(`{"choices":[{"message":{"content":"done"}}],"usage":{"prompt_tokens":7,"completion_tokens":3,"total_tokens":10,"input_tokens_details":{"cached_tokens":2}}}`)
+	w := httptest.NewRecorder()
+	writeResponsesResponse(w, body, "kimi-k2.6")
+	var out map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &out); err != nil {
+		t.Fatal(err)
+	}
+	usage, ok := out["usage"].(map[string]any)
+	if !ok {
+		t.Fatalf("missing usage: %+v", out)
+	}
+	if usage["input_tokens"] != float64(7) || usage["output_tokens"] != float64(3) || usage["total_tokens"] != float64(10) {
+		t.Fatalf("bad responses usage: %+v", usage)
+	}
+	details, ok := usage["input_tokens_details"].(map[string]any)
+	if !ok || details["cached_tokens"] != float64(2) {
+		t.Fatalf("bad cached details: %+v", usage["input_tokens_details"])
+	}
+}
+
+func TestParseOpenAIStreamChunkReadsUsageOnly(t *testing.T) {
+	chunk := parseOpenAIStreamChunk([]byte(`{"choices":[],"usage":{"prompt_tokens":8,"completion_tokens":4,"total_tokens":12}}`))
+	if !chunk.Usage.Present || chunk.Usage.InputTokens != 8 || chunk.Usage.OutputTokens != 4 || chunk.Usage.TotalTokens != 12 {
+		t.Fatalf("bad stream usage: %+v", chunk.Usage)
+	}
+	if chunk.Content != "" || len(chunk.ToolCalls) != 0 {
+		t.Fatalf("usage-only chunk should not include deltas: %+v", chunk)
+	}
+}
+
+func TestStreamAnthropicIncludesFinalUsage(t *testing.T) {
+	body := strings.NewReader(strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"hi"}}]}`,
+		`data: {"choices":[],"usage":{"prompt_tokens":7,"completion_tokens":3,"total_tokens":10}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n\n"))
+	w := httptest.NewRecorder()
+	streamAnthropic(w, body, "kimi-k2.6")
+	out := w.Body.String()
+	for _, want := range []string{`"input_tokens":7`, `"output_tokens":3`} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("missing %q in:\n%s", want, out)
+		}
+	}
+}
+
+func TestStreamResponsesIncludesCompletedUsage(t *testing.T) {
+	body := strings.NewReader(strings.Join([]string{
+		`data: {"choices":[{"delta":{"content":"hi"}}]}`,
+		`data: {"choices":[],"usage":{"prompt_tokens":7,"completion_tokens":3,"total_tokens":10}}`,
+		`data: [DONE]`,
+		``,
+	}, "\n\n"))
+	w := httptest.NewRecorder()
+	streamResponses(w, body, "kimi-k2.6")
+	out := w.Body.String()
+	for _, want := range []string{`event: response.completed`, `"input_tokens":7`, `"output_tokens":3`, `"total_tokens":10`} {
+		if !strings.Contains(out, want) {
+			t.Fatalf("missing %q in:\n%s", want, out)
+		}
+	}
+}
+
 func TestStreamAnthropicForwardsToolCalls(t *testing.T) {
 	reasoningContentCache.Lock()
 	reasoningContentCache.byCallID = map[string]string{}