AI Gateway Go 0.10.0 源码快照 + 旗舰版需求规划报告

M0-M7 已完成:核心网关(身份/RBAC/TOTP/OIDC/SAML/Provider/配额/路由/内容策略/审计/定价)+ 资源市场(MCP/Skills/数字员工)。
含 22 个 PostgreSQL 迁移、管理端/门户端前端源码、OpenAPI 契约、部署 compose。

Co-Authored-By: Claude <noreply@anthropic.com>
This commit is contained in:
ben
2026-08-12 11:45:54 +08:00
commit 5759c1862e
807 changed files with 114727 additions and 0 deletions
+139
View File
@@ -0,0 +1,139 @@
package gateway
import (
"context"
"errors"
"fmt"
"strconv"
"time"
"aigateway.local/core/internal/apikey"
"github.com/redis/go-redis/v9"
)
var ErrAdmissionUnavailable = errors.New("admission control unavailable")
type AdmissionReason string
const (
AdmissionAllowed AdmissionReason = ""
AdmissionRateLimit AdmissionReason = "rate_limit"
AdmissionMonthlyQuota AdmissionReason = "monthly_quota"
)
type AdmissionDecision struct {
Allowed bool
Reason AdmissionReason
Limit int64
Remaining int64
ResetAt time.Time
RetryAfter time.Duration
}
type AdmissionController interface {
Allow(context.Context, apikey.Principal, time.Time) (AdmissionDecision, error)
}
type RedisAdmissionController struct {
client *redis.Client
script *redis.Script
}
func NewRedisAdmissionController(client *redis.Client) *RedisAdmissionController {
return &RedisAdmissionController{client: client, script: redis.NewScript(admissionScript)}
}
func (c *RedisAdmissionController) Allow(ctx context.Context, principal apikey.Principal, now time.Time) (AdmissionDecision, error) {
if principal.RequestsPerMinute == 0 && principal.MonthlyRequestQuota == 0 {
return AdmissionDecision{Allowed: true}, nil
}
if c == nil || c.client == nil || principal.APIKeyID == "" {
return AdmissionDecision{}, ErrAdmissionUnavailable
}
now = now.UTC()
minuteReset := now.Truncate(time.Minute).Add(time.Minute)
monthReset := time.Date(now.Year(), now.Month()+1, 1, 0, 0, 0, 0, time.UTC)
minuteKey := fmt.Sprintf("gateway:limit:api-key:%s:minute:%d", principal.APIKeyID, now.Unix()/60)
monthKey := fmt.Sprintf("gateway:limit:api-key:%s:month:%s", principal.APIKeyID, now.Format("200601"))
result, err := c.script.Run(ctx, c.client, []string{minuteKey, monthKey},
principal.RequestsPerMinute, principal.MonthlyRequestQuota,
int64(minuteReset.Sub(now).Seconds())+2, int64(monthReset.Sub(now).Seconds())+86400,
).Slice()
if err != nil {
return AdmissionDecision{}, fmt.Errorf("%w: %v", ErrAdmissionUnavailable, err)
}
if len(result) != 3 {
return AdmissionDecision{}, ErrAdmissionUnavailable
}
code, err := redisInteger(result[0])
if err != nil {
return AdmissionDecision{}, ErrAdmissionUnavailable
}
minuteCount, err := redisInteger(result[1])
if err != nil {
return AdmissionDecision{}, ErrAdmissionUnavailable
}
monthCount, err := redisInteger(result[2])
if err != nil {
return AdmissionDecision{}, ErrAdmissionUnavailable
}
switch code {
case 0:
decision := AdmissionDecision{Allowed: true}
if principal.RequestsPerMinute > 0 {
decision.Limit = int64(principal.RequestsPerMinute)
decision.Remaining = max(decision.Limit-minuteCount, 0)
decision.ResetAt = minuteReset
}
return decision, nil
case 1:
return AdmissionDecision{
Allowed: false, Reason: AdmissionRateLimit, Limit: int64(principal.RequestsPerMinute),
Remaining: 0, ResetAt: minuteReset, RetryAfter: minuteReset.Sub(now),
}, nil
case 2:
return AdmissionDecision{
Allowed: false, Reason: AdmissionMonthlyQuota, Limit: principal.MonthlyRequestQuota,
Remaining: max(principal.MonthlyRequestQuota-monthCount, 0), ResetAt: monthReset, RetryAfter: monthReset.Sub(now),
}, nil
default:
return AdmissionDecision{}, ErrAdmissionUnavailable
}
}
func redisInteger(value any) (int64, error) {
switch number := value.(type) {
case int64:
return number, nil
case string:
return strconv.ParseInt(number, 10, 64)
case []byte:
return strconv.ParseInt(string(number), 10, 64)
default:
return 0, fmt.Errorf("unexpected redis integer %T", value)
}
}
const admissionScript = `
local rpm = tonumber(ARGV[1])
local monthly_quota = tonumber(ARGV[2])
local minute_count = tonumber(redis.call('GET', KEYS[1]) or '0')
local month_count = tonumber(redis.call('GET', KEYS[2]) or '0')
if monthly_quota > 0 and month_count >= monthly_quota then
return {2, minute_count, month_count}
end
if rpm > 0 and minute_count >= rpm then
return {1, minute_count, month_count}
end
if rpm > 0 then
minute_count = redis.call('INCR', KEYS[1])
if minute_count == 1 then redis.call('EXPIRE', KEYS[1], tonumber(ARGV[3])) end
end
if monthly_quota > 0 then
month_count = redis.call('INCR', KEYS[2])
if month_count == 1 then redis.call('EXPIRE', KEYS[2], tonumber(ARGV[4])) end
end
return {0, minute_count, month_count}
`
+30
View File
@@ -0,0 +1,30 @@
package gateway
import (
"testing"
"time"
)
func TestRedisInteger(t *testing.T) {
for _, value := range []any{int64(7), "7", []byte("7")} {
got, err := redisInteger(value)
if err != nil || got != 7 {
t.Fatalf("redisInteger(%T) = %d, %v", value, got, err)
}
}
if _, err := redisInteger(7.0); err == nil {
t.Fatal("unexpected redis number type accepted")
}
}
func TestAdmissionResetBoundaries(t *testing.T) {
now := time.Date(2026, time.August, 10, 13, 25, 40, 0, time.UTC)
minuteReset := now.Truncate(time.Minute).Add(time.Minute)
monthReset := time.Date(now.Year(), now.Month()+1, 1, 0, 0, 0, 0, time.UTC)
if minuteReset != time.Date(2026, time.August, 10, 13, 26, 0, 0, time.UTC) {
t.Fatal("minute boundary is incorrect")
}
if monthReset != time.Date(2026, time.September, 1, 0, 0, 0, 0, time.UTC) {
t.Fatal("month boundary is incorrect")
}
}
+221
View File
@@ -0,0 +1,221 @@
package gateway
import (
"bufio"
"encoding/json"
"io"
"net"
"net/http"
"sync"
"time"
"aigateway.local/core/internal/apikey"
auditpkg "aigateway.local/core/internal/audit"
"aigateway.local/core/internal/contentpolicy"
"aigateway.local/core/internal/pricing"
)
const auditRequestCaptureBytes = 64 << 10
type AuditRecorder interface {
Record(auditpkg.Event) bool
}
type auditSpan struct {
recorder AuditRecorder
started time.Time
event auditpkg.Event
mu sync.Mutex
capture *captureReadCloser
target string
pricing *pricing.Service
policies []contentpolicy.Match
redacted bool
}
func newAuditSpan(recorder AuditRecorder, principal apikey.Principal, request *http.Request, started time.Time) *auditSpan {
if recorder == nil {
return nil
}
event := auditpkg.Event{TenantID: principal.TenantID, RequestID: RequestID(request.Context()), Protocol: request.URL.Path, RecordedAt: time.Now().UTC()}
if principal.APIKeyID != "" {
id := principal.APIKeyID
event.APIKeyID = &id
}
return &auditSpan{recorder: recorder, started: started, event: event}
}
func (s *auditSpan) setModel(model string) {
if s != nil && model != "" {
s.mu.Lock()
s.event.Model = model
s.mu.Unlock()
}
}
func (s *auditSpan) setRoute(providerCode, targetModel string) {
if s == nil {
return
}
s.mu.Lock()
s.event.ProviderCode = providerCode
s.target = targetModel
s.mu.Unlock()
}
func (s *auditSpan) setUsage(usage TokenUsage) {
if s == nil {
return
}
s.mu.Lock()
s.event.PromptTokens = usage.Input
s.event.CompletionTokens = usage.Output
s.calculateCostLocked()
s.mu.Unlock()
}
func (s *auditSpan) setContentPolicy(result contentpolicy.Result) {
if s == nil {
return
}
s.mu.Lock()
s.policies = append(s.policies, result.Matches...)
s.redacted = result.Redacted
s.mu.Unlock()
}
func (s *auditSpan) captureBody(body io.ReadCloser) io.ReadCloser {
if s == nil || body == nil || body == http.NoBody {
return body
}
capture := &captureReadCloser{ReadCloser: body, limit: auditRequestCaptureBytes}
s.capture = capture
return capture
}
func (s *auditSpan) finish(status int) {
if s == nil {
return
}
s.mu.Lock()
if s.event.Model == "" && s.capture != nil {
var payload map[string]json.RawMessage
if json.Unmarshal(s.capture.buffer, &payload) == nil {
_ = json.Unmarshal(payload["model"], &s.event.Model)
}
}
s.calculateCostLocked()
s.event.StatusCode = status
s.event.LatencyMS = int(time.Since(s.started).Milliseconds())
s.event.RecordedAt = time.Now().UTC()
labels := map[string]any{}
if s.target != "" && s.target != s.event.Model {
labels["target_model"] = s.target
labels["routed"] = true
}
if len(s.policies) > 0 {
labels["content_policies"] = s.policies
}
if s.redacted {
labels["content_redacted"] = true
}
if s.event.PriceID != "" {
labels["price_id"] = s.event.PriceID
labels["currency"] = s.event.Currency
}
s.event.Labels = labels
event := s.event
s.mu.Unlock()
s.recorder.Record(event)
}
func (s *auditSpan) calculateCostLocked() {
if s.pricing == nil || s.event.ProviderCode == "" || s.event.Model == "" {
return
}
model := s.event.Model
if s.target != "" {
model = s.target
}
cost := s.pricing.Calculate(s.event.ProviderCode, model, s.event.PromptTokens, s.event.CompletionTokens, s.started.UTC())
s.event.CostMicrounits = cost.Microunits
s.event.PriceID = cost.PriceID
s.event.Currency = cost.Currency
}
type captureReadCloser struct {
io.ReadCloser
buffer []byte
limit int
}
func (r *captureReadCloser) Read(buffer []byte) (int, error) {
n, err := r.ReadCloser.Read(buffer)
if n > 0 && len(r.buffer) < r.limit {
remaining := r.limit - len(r.buffer)
r.buffer = append(r.buffer, buffer[:min(n, remaining)]...)
}
return n, err
}
type statusResponseWriter struct {
http.ResponseWriter
status int
}
func (w *statusResponseWriter) WriteHeader(status int) {
if w.status != 0 {
return
}
w.status = status
w.ResponseWriter.WriteHeader(status)
}
func (w *statusResponseWriter) Write(buffer []byte) (int, error) {
if w.status == 0 {
w.WriteHeader(http.StatusOK)
}
return w.ResponseWriter.Write(buffer)
}
func (w *statusResponseWriter) Status() int {
if w.status == 0 {
return http.StatusOK
}
return w.status
}
func (w *statusResponseWriter) Flush() {
if w.status == 0 {
w.WriteHeader(http.StatusOK)
}
if flusher, ok := w.ResponseWriter.(http.Flusher); ok {
flusher.Flush()
}
}
func (w *statusResponseWriter) Unwrap() http.ResponseWriter { return w.ResponseWriter }
func (w *statusResponseWriter) Hijack() (net.Conn, *bufio.ReadWriter, error) {
if hijacker, ok := w.ResponseWriter.(http.Hijacker); ok {
return hijacker.Hijack()
}
return nil, nil, http.ErrNotSupported
}
func (w *statusResponseWriter) Push(target string, options *http.PushOptions) error {
if pusher, ok := w.ResponseWriter.(http.Pusher); ok {
return pusher.Push(target, options)
}
return http.ErrNotSupported
}
func (w *statusResponseWriter) ReadFrom(reader io.Reader) (int64, error) {
if w.status == 0 {
w.WriteHeader(http.StatusOK)
}
if readerFrom, ok := w.ResponseWriter.(io.ReaderFrom); ok {
return readerFrom.ReadFrom(reader)
}
return io.Copy(struct{ io.Writer }{w.ResponseWriter}, reader)
}
@@ -0,0 +1,141 @@
package gateway
import (
"context"
"io"
"log/slog"
"net/http"
"net/http/httptest"
"os"
"strings"
"sync/atomic"
"testing"
"time"
"aigateway.local/core/internal/audit"
"aigateway.local/core/internal/contentpolicy"
"aigateway.local/core/internal/pricing"
provideropenai "aigateway.local/core/internal/provider/openai"
"github.com/jackc/pgx/v5/pgxpool"
)
func TestContentPolicyAndPricingIntegration(t *testing.T) {
databaseURL := os.Getenv("CONTENT_PRICING_TEST_DATABASE_URL")
if databaseURL == "" {
t.Skip("CONTENT_PRICING_TEST_DATABASE_URL is not set")
}
ctx := context.Background()
pool, err := pgxpool.New(ctx, databaseURL)
if err != nil {
t.Fatal(err)
}
defer pool.Close()
const blockID = "00000000-0000-4000-8000-000000001601"
const priceID = "00000000-0000-4000-8000-000000001602"
_, err = pool.Exec(ctx, `INSERT INTO gateway.content_policies(id,name,action,priority,rules,enabled) VALUES($1,'integration block','block',2000,'[{"name":"forbidden","pattern":"forbidden","replacement":"[BLOCKED]"}]',true) ON CONFLICT(id) DO UPDATE SET enabled=true`, blockID)
if err != nil {
t.Fatal(err)
}
_, err = pool.Exec(ctx, `INSERT INTO gateway.model_prices(id,provider_code,model_pattern,input_microunits_per_million,output_microunits_per_million,currency,effective_from,enabled) VALUES($1,'environment','gpt-test',2000000,8000000,'USD',clock_timestamp()-interval '1 hour',true) ON CONFLICT(id) DO UPDATE SET enabled=true`, priceID)
if err != nil {
t.Fatal(err)
}
defer func() {
_, _ = pool.Exec(context.Background(), `DELETE FROM gateway.outbox_events WHERE event_type='content_policy.matched' AND aggregate_id IN ('m3-content-price-redact','m3-content-price-block')`)
_, _ = pool.Exec(context.Background(), `DELETE FROM gateway.content_policies WHERE id=$1`, blockID)
_, _ = pool.Exec(context.Background(), `DELETE FROM gateway.model_prices WHERE id=$1`, priceID)
}()
var upstreamCalls atomic.Int64
bodySeen := make(chan string, 1)
upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
upstreamCalls.Add(1)
body, _ := io.ReadAll(r.Body)
bodySeen <- string(body)
w.Header().Set("Content-Type", "application/json")
_, _ = io.WriteString(w, `{"id":"ok","usage":{"prompt_tokens":1000,"completion_tokens":500,"total_tokens":1500}}`)
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-key")
engine := contentpolicy.NewEngine(pool, time.Minute, slog.Default())
if err := engine.Reload(ctx); err != nil {
t.Fatal(err)
}
prices := pricing.NewService(pool, time.Minute, slog.Default())
if err := prices.Reload(ctx); err != nil {
t.Fatal(err)
}
recorder := audit.NewRecorder(pool, slog.Default(), 100, 1, 10*time.Millisecond)
recordCtx, cancel := context.WithCancel(ctx)
stopped := make(chan struct{})
go func() { recorder.Run(recordCtx); close(stopped) }()
defer func() { cancel(); <-stopped }()
proxy := NewProxy(adapter, "test-key", 1<<20, slog.Default())
proxy.SetAuditRecorder(recorder)
proxy.SetContentPolicyEngine(engine)
proxy.SetPricingService(prices)
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
proxy.ServeHTTP(w, r.WithContext(WithRequestID(r.Context(), r.Header.Get("X-Request-ID"))))
}))
defer server.Close()
req, _ := http.NewRequest(http.MethodPost, server.URL+"/v1/chat/completions", strings.NewReader(`{"model":"gpt-test","messages":[{"role":"user","content":"token sk-1234567890123456"}]}`))
req.Header.Set("Authorization", "Bearer test-key")
req.Header.Set("X-Request-ID", "m3-content-price-redact")
response, err := http.DefaultClient.Do(req)
if err != nil {
t.Fatal(err)
}
_, _ = io.ReadAll(response.Body)
_ = response.Body.Close()
if response.StatusCode != 200 || response.Header.Get("X-Gateway-Content-Redacted") != "true" {
t.Fatalf("unexpected response %d headers=%v", response.StatusCode, response.Header)
}
select {
case body := <-bodySeen:
if strings.Contains(body, "sk-1234567890123456") || !strings.Contains(body, "[REDACTED_API_KEY]") {
t.Fatalf("upstream saw unsafe body %s", body)
}
case <-time.After(time.Second):
t.Fatal("upstream did not receive request")
}
blocked, _ := http.NewRequest(http.MethodPost, server.URL+"/v1/responses", strings.NewReader(`{"model":"gpt-test","input":"forbidden"}`))
blocked.Header.Set("Authorization", "Bearer test-key")
blocked.Header.Set("X-Request-ID", "m3-content-price-block")
blockedResponse, err := http.DefaultClient.Do(blocked)
if err != nil {
t.Fatal(err)
}
_, _ = io.ReadAll(blockedResponse.Body)
_ = blockedResponse.Body.Close()
if blockedResponse.StatusCode != http.StatusUnprocessableEntity || upstreamCalls.Load() != 1 {
t.Fatalf("block failed status=%d calls=%d", blockedResponse.StatusCode, upstreamCalls.Load())
}
deadline := time.Now().Add(2 * time.Second)
for {
var cost *int64
var labels string
err = pool.QueryRow(ctx, `SELECT cost_microunits,labels::text FROM gateway.audit_events WHERE request_id='m3-content-price-redact' ORDER BY recorded_at DESC LIMIT 1`).Scan(&cost, &labels)
if err == nil && cost != nil && *cost == 6000 && strings.Contains(labels, "content_redacted") {
break
}
if time.Now().After(deadline) {
t.Fatalf("audit cost/redaction missing cost=%v labels=%s err=%v", cost, labels, err)
}
time.Sleep(20 * time.Millisecond)
}
deadline = time.Now().Add(2 * time.Second)
for {
var count int
err = pool.QueryRow(ctx, `SELECT count(*) FROM gateway.outbox_events WHERE event_type='content_policy.matched' AND aggregate_id IN ('m3-content-price-redact','m3-content-price-block')`).Scan(&count)
if err == nil && count == 2 {
break
}
if time.Now().After(deadline) {
t.Fatalf("content policy notification outbox missing count=%d err=%v", count, err)
}
time.Sleep(20 * time.Millisecond)
}
}
+14
View File
@@ -0,0 +1,14 @@
package gateway
import "context"
type requestIDKey struct{}
func WithRequestID(ctx context.Context, requestID string) context.Context {
return context.WithValue(ctx, requestIDKey{}, requestID)
}
func RequestID(ctx context.Context) string {
value, _ := ctx.Value(requestIDKey{}).(string)
return value
}
+430
View File
@@ -0,0 +1,430 @@
package gateway
import (
"context"
"crypto/subtle"
"encoding/json"
"errors"
"fmt"
"log/slog"
"net"
"net/http"
"net/http/httputil"
"strconv"
"strings"
"sync"
"time"
"aigateway.local/core/internal/apikey"
"aigateway.local/core/internal/contentpolicy"
"aigateway.local/core/internal/pricing"
"aigateway.local/core/internal/provider"
)
type Proxy struct {
resolver AdapterResolver
auth apikey.KeyAuthenticator
maxBody int64
logger *slog.Logger
transport *http.Transport
proxies sync.Map
circuits sync.Map
admission AdmissionController
tokenQuota TokenQuotaController
resilience ResiliencePolicy
audit AuditRecorder
policies *contentpolicy.Engine
pricing *pricing.Service
}
type cachedProxy struct {
key string
proxy *httputil.ReverseProxy
}
var (
ErrProviderNotFound = errors.New("requested provider is not available")
ErrProviderUnavailable = errors.New("provider configuration is unavailable")
)
type ResolvedAdapter struct {
Code string
Revision int64
Adapter provider.Adapter
Capabilities map[provider.Capability]bool
}
type AdapterResolver interface {
Resolve(providerCode string) (ResolvedAdapter, error)
}
func NewProxy(adapter provider.Adapter, apiKey string, maxBody int64, logger *slog.Logger) *Proxy {
return NewProxyWithAuthenticator(adapter, staticKeyAuthenticator(apiKey), maxBody, logger)
}
func NewProxyWithAuthenticator(adapter provider.Adapter, authenticator apikey.KeyAuthenticator, maxBody int64, logger *slog.Logger) *Proxy {
return NewDynamicProxy(staticAdapterResolver{adapter: adapter}, authenticator, maxBody, logger)
}
func NewDynamicProxy(resolver AdapterResolver, authenticator apikey.KeyAuthenticator, maxBody int64, logger *slog.Logger) *Proxy {
transport := &http.Transport{
Proxy: http.ProxyFromEnvironment,
DialContext: (&net.Dialer{Timeout: 5 * time.Second, KeepAlive: 30 * time.Second}).DialContext,
ForceAttemptHTTP2: true,
MaxIdleConns: 512,
MaxIdleConnsPerHost: 256,
IdleConnTimeout: 90 * time.Second,
TLSHandshakeTimeout: 5 * time.Second,
ResponseHeaderTimeout: 60 * time.Second,
ExpectContinueTimeout: time.Second,
}
return &Proxy{resolver: resolver, auth: authenticator, maxBody: maxBody, logger: logger, transport: transport, resilience: DefaultResiliencePolicy()}
}
func (p *Proxy) SetAdmissionController(controller AdmissionController) {
p.admission = controller
}
func (p *Proxy) SetTokenQuotaController(controller TokenQuotaController) {
p.tokenQuota = controller
}
func (p *Proxy) SetResiliencePolicy(policy ResiliencePolicy) {
p.resilience = policy
p.transport.ResponseHeaderTimeout = policy.ResponseHeaderTimeout
}
func (p *Proxy) SetAuditRecorder(recorder AuditRecorder) { p.audit = recorder }
func (p *Proxy) SetContentPolicyEngine(engine *contentpolicy.Engine) { p.policies = engine }
func (p *Proxy) SetPricingService(service *pricing.Service) { p.pricing = service }
func (p *Proxy) ServeHTTP(writer http.ResponseWriter, request *http.Request) {
started := time.Now()
if !isSupportedPath(request.URL.Path, request.Method) {
writeOpenAIError(writer, http.StatusNotFound, "invalid_request_error", "unsupported gateway endpoint")
return
}
principal, err := p.authorized(request)
if err != nil {
if errors.Is(err, apikey.ErrStore) {
writeOpenAIError(writer, http.StatusServiceUnavailable, "authentication_unavailable", "API key service is unavailable")
return
}
writer.Header().Set("WWW-Authenticate", "Bearer")
writeOpenAIError(writer, http.StatusUnauthorized, "authentication_error", "invalid API key")
return
}
span := newAuditSpan(p.audit, principal, request, started)
if span != nil {
span.pricing = p.pricing
}
if span != nil {
statusWriter := &statusResponseWriter{ResponseWriter: writer}
writer = statusWriter
defer func() { span.finish(statusWriter.Status()) }()
}
if p.admission != nil && principal.APIKeyID != "" {
decision, err := p.admission.Allow(request.Context(), principal, time.Now())
if err != nil {
writeOpenAIError(writer, http.StatusServiceUnavailable, "rate_limit_unavailable", "rate limit service is unavailable")
return
}
writeAdmissionHeaders(writer.Header(), decision)
if !decision.Allowed {
writer.Header().Set("Retry-After", strconv.FormatInt(max(int64(decision.RetryAfter.Seconds()), 1), 10))
if decision.Reason == AdmissionMonthlyQuota {
writeOpenAIError(writer, http.StatusTooManyRequests, "insufficient_quota", "monthly request quota exceeded")
} else {
writeOpenAIError(writer, http.StatusTooManyRequests, "rate_limit_error", "request rate limit exceeded")
}
return
}
}
if request.ContentLength > p.maxBody {
writeOpenAIError(writer, http.StatusRequestEntityTooLarge, "invalid_request_error", "request body is too large")
return
}
if p.policies != nil {
policyResult, policyErr := p.policies.Apply(request, p.maxBody, principal.APIKeyID)
if errors.Is(policyErr, contentpolicy.ErrBodyTooLarge) {
writeOpenAIError(writer, http.StatusRequestEntityTooLarge, "invalid_request_error", "request body is too large")
return
}
if policyErr != nil {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "request body could not be inspected")
return
}
span.setContentPolicy(policyResult)
if policyResult.Blocked {
writeOpenAIError(writer, http.StatusUnprocessableEntity, "content_policy_violation", "request was blocked by content policy")
return
}
if policyResult.Redacted {
writer.Header().Set("X-Gateway-Content-Redacted", "true")
}
}
routeResolver, routingEnabled := p.resolver.(ModelRouteResolver)
routingEnabled = routingEnabled && routeResolver.ModelRoutingEnabled()
var modelPayload modelRequest
if routingEnabled {
modelPayload, err = readModelRequest(request, p.maxBody)
if errors.Is(err, errRequestBodyTooLarge) {
writeOpenAIError(writer, http.StatusRequestEntityTooLarge, "invalid_request_error", "request body is too large")
return
}
if err != nil {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "request body could not be read")
return
}
}
var usage *usageSession
if p.tokenQuota != nil && principal.APIKeyID != "" {
estimate := int64(0)
if principal.MonthlyTokenQuota > 0 {
estimate, err = prepareTokenBudget(request, p.maxBody)
if errors.Is(err, errRequestBodyTooLarge) {
writeOpenAIError(writer, http.StatusRequestEntityTooLarge, "invalid_request_error", "request body is too large")
return
}
if err != nil {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "request body could not be read")
return
}
}
reservation, reserveErr := p.tokenQuota.Reserve(request.Context(), principal, estimate, time.Now())
if reserveErr != nil && principal.MonthlyTokenQuota > 0 {
writeOpenAIError(writer, http.StatusServiceUnavailable, "token_quota_unavailable", "token quota service is unavailable")
return
}
if reserveErr == nil && !reservation.Allowed {
writeTokenQuotaHeaders(writer.Header(), reservation)
writer.Header().Set("Retry-After", strconv.FormatInt(max(int64(time.Until(reservation.ResetAt).Seconds()), 1), 10))
writeOpenAIError(writer, http.StatusTooManyRequests, "insufficient_quota", "monthly token quota exceeded")
return
}
if reserveErr == nil && reservation.CounterKey != "" {
writeTokenQuotaHeaders(writer.Header(), reservation)
usage = &usageSession{controller: p.tokenQuota, reservation: reservation, log: p.logger}
request = withUsageSession(request, usage)
defer usage.finish(TokenUsage{})
}
}
if span != nil {
if usage == nil {
usage = &usageSession{log: p.logger}
request = withUsageSession(request, usage)
defer usage.finish(TokenUsage{})
}
usage.onFinish = span.setUsage
}
if request.Body != nil && request.Method != http.MethodGet {
if request.Header.Get("Idempotency-Key") != "" && request.GetBody == nil {
if _, err := prepareTokenBudget(request, p.maxBody); err != nil {
if errors.Is(err, errRequestBodyTooLarge) {
writeOpenAIError(writer, http.StatusRequestEntityTooLarge, "invalid_request_error", "request body is too large")
} else {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "request body could not be read")
}
return
}
}
request.Body = http.MaxBytesReader(writer, request.Body, p.maxBody)
}
providerCode := strings.ToLower(strings.TrimSpace(request.Header.Get("X-Gateway-Provider")))
var resolved ResolvedAdapter
if routingEnabled && modelPayload.model != "" {
span.setModel(modelPayload.model)
tenantID := ""
if principal.TenantID != nil {
tenantID = *principal.TenantID
}
route, routeErr := routeResolver.ResolveModelRoute(ModelRouteQuery{
ProviderCode: providerCode, Model: modelPayload.model, Endpoint: request.URL.Path,
APIKeyID: principal.APIKeyID, TenantID: tenantID, Seed: RequestID(request.Context()),
})
if routeErr != nil {
writeOpenAIError(writer, http.StatusServiceUnavailable, "provider_unavailable", "model routing configuration is unavailable")
return
}
if route.Known && !route.Matched {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "model route is not available for this request")
return
}
if route.Matched {
resolved = route.ResolvedAdapter
if err := modelPayload.rewrite(request, route.TargetModel); err != nil {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "request model could not be rewritten")
return
}
writer.Header().Set("X-Gateway-Model", route.TargetModel)
}
}
if resolved.Adapter == nil {
resolved, err = p.resolver.Resolve(providerCode)
}
if err != nil {
if errors.Is(err, ErrProviderNotFound) {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "requested provider is not available")
return
}
writeOpenAIError(writer, http.StatusServiceUnavailable, "provider_unavailable", "provider configuration is unavailable")
return
}
capability := capabilityForPath(request.URL.Path)
if capability != "" && !resolved.Capabilities[capability] {
writeOpenAIError(writer, http.StatusBadRequest, "invalid_request_error", "provider does not support this endpoint")
return
}
request.Header.Del("X-Gateway-Provider")
writer.Header().Set("X-Gateway-Provider", resolved.Code)
span.setRoute(resolved.Code, writer.Header().Get("X-Gateway-Model"))
request.Body = span.captureBody(request.Body)
p.proxyFor(resolved).ServeHTTP(writer, request)
}
func (p *Proxy) proxyFor(resolved ResolvedAdapter) *httputil.ReverseProxy {
target := resolved.Adapter.Target()
key := fmt.Sprintf("%s:%d:%s", resolved.Code, resolved.Revision, target.String())
if cached, ok := p.proxies.Load(resolved.Code); ok {
entry := cached.(cachedProxy)
if entry.key == key {
return entry.proxy
}
}
reverseProxy := httputil.NewSingleHostReverseProxy(target)
originalDirector := reverseProxy.Director
reverseProxy.Director = func(request *http.Request) {
originalDirector(request)
request.Host = target.Host
resolved.Adapter.Prepare(request)
}
reverseProxy.FlushInterval = -1
circuitValue, _ := p.circuits.LoadOrStore(resolved.Code, newCircuitBreaker(p.resilience))
reverseProxy.Transport = &resilientTransport{
base: p.transport, circuit: circuitValue.(*circuitBreaker), maxRetries: p.resilience.MaxRetries, backoff: p.resilience.RetryBackoff,
}
reverseProxy.ModifyResponse = func(response *http.Response) error {
if session := usageSessionFrom(response.Request); session != nil {
if response.StatusCode >= http.StatusOK && response.StatusCode < http.StatusMultipleChoices {
session.fallback = session.reservation.Reserved
response.Body = newUsageReadCloser(response.Body, response.Header.Get("Content-Type"), session)
} else {
session.fallback = 0
session.finish(TokenUsage{})
}
}
return nil
}
reverseProxy.ErrorHandler = func(writer http.ResponseWriter, request *http.Request, err error) {
if session := usageSessionFrom(request); session != nil {
session.fallback = 0
session.finish(TokenUsage{})
}
p.logger.Error("upstream request failed", "request_id", RequestID(request.Context()), "provider", resolved.Code, "error", err)
if errors.Is(err, ErrCircuitOpen) {
writer.Header().Set("Retry-After", "30")
writeOpenAIError(writer, http.StatusServiceUnavailable, "provider_unavailable", "provider circuit is temporarily open")
return
}
writeOpenAIError(writer, http.StatusBadGateway, "upstream_error", "upstream service is unavailable")
}
p.proxies.Store(resolved.Code, cachedProxy{key: key, proxy: reverseProxy})
return reverseProxy
}
func (p *Proxy) authorized(request *http.Request) (apikey.Principal, error) {
presented := strings.TrimSpace(request.Header.Get("X-Gateway-API-Key"))
if presented == "" {
authorization := strings.TrimSpace(request.Header.Get("Authorization"))
if len(authorization) > len("Bearer ") && strings.EqualFold(authorization[:len("Bearer ")], "Bearer ") {
presented = strings.TrimSpace(authorization[len("Bearer "):])
}
}
if p.auth == nil {
return apikey.Principal{}, apikey.ErrInvalid
}
if authenticator, ok := p.auth.(apikey.PrincipalAuthenticator); ok {
return authenticator.AuthenticatePrincipal(request.Context(), presented)
}
return apikey.Principal{}, p.auth.Authenticate(request.Context(), presented)
}
func writeAdmissionHeaders(header http.Header, decision AdmissionDecision) {
if decision.Limit <= 0 || decision.ResetAt.IsZero() {
return
}
header.Set("X-RateLimit-Limit", strconv.FormatInt(decision.Limit, 10))
header.Set("X-RateLimit-Remaining", strconv.FormatInt(max(decision.Remaining, 0), 10))
header.Set("X-RateLimit-Reset", strconv.FormatInt(decision.ResetAt.Unix(), 10))
}
type staticKeyAuthenticator string
func (a staticKeyAuthenticator) Authenticate(_ context.Context, presented string) error {
key := string(a)
if key == "" {
return nil
}
if len(presented) != len(key) || subtle.ConstantTimeCompare([]byte(presented), []byte(key)) != 1 {
return apikey.ErrInvalid
}
return nil
}
type staticAdapterResolver struct{ adapter provider.Adapter }
func (r staticAdapterResolver) Resolve(code string) (ResolvedAdapter, error) {
if r.adapter == nil || code != "" && code != "environment" {
return ResolvedAdapter{}, ErrProviderNotFound
}
capabilities := make(map[provider.Capability]bool)
for _, capability := range r.adapter.Capabilities() {
capabilities[capability] = true
}
return ResolvedAdapter{Code: "environment", Adapter: r.adapter, Capabilities: capabilities}, nil
}
func capabilityForPath(path string) provider.Capability {
switch path {
case "/v1/models":
return provider.CapabilityModels
case "/v1/chat/completions":
return provider.CapabilityChat
case "/v1/responses":
return provider.CapabilityResponses
case "/v1/embeddings":
return provider.CapabilityEmbeddings
case "/v1/messages":
return provider.CapabilityMessages
default:
return ""
}
}
func isSupportedPath(path, method string) bool {
switch path {
case "/v1/models":
return method == http.MethodGet
case "/v1/chat/completions", "/v1/responses", "/v1/embeddings", "/v1/messages":
return method == http.MethodPost
default:
return false
}
}
func writeOpenAIError(writer http.ResponseWriter, status int, errorType, message string) {
writer.Header().Set("Content-Type", "application/json")
writer.WriteHeader(status)
_ = json.NewEncoder(writer).Encode(map[string]any{
"error": map[string]any{"message": message, "type": errorType, "param": nil, "code": nil},
})
}
var errInvalidAdapter = errors.New("invalid provider adapter")
func ValidateAdapter(adapter provider.Adapter) error {
if adapter == nil || adapter.Target() == nil || adapter.Name() == "" {
return errInvalidAdapter
}
return nil
}
+252
View File
@@ -0,0 +1,252 @@
package gateway
import (
"context"
"io"
"log/slog"
"net/http"
"net/http/httptest"
"strings"
"testing"
"time"
"aigateway.local/core/internal/apikey"
auditpkg "aigateway.local/core/internal/audit"
"aigateway.local/core/internal/provider"
provideropenai "aigateway.local/core/internal/provider/openai"
)
type principalAuthenticator struct{ principal apikey.Principal }
func (a principalAuthenticator) Authenticate(context.Context, string) error { return nil }
func (a principalAuthenticator) AuthenticatePrincipal(context.Context, string) (apikey.Principal, error) {
return a.principal, nil
}
type fixedAdmission struct{ decision AdmissionDecision }
func (a fixedAdmission) Allow(context.Context, apikey.Principal, time.Time) (AdmissionDecision, error) {
return a.decision, nil
}
type recordingTokenQuota struct {
reservation TokenReservation
estimate int64
actual int64
}
type fixedRoutingResolver struct {
adapter ResolvedAdapter
}
type recordingAudit struct{ event auditpkg.Event }
func (r *recordingAudit) Record(event auditpkg.Event) bool { r.event = event; return true }
func (r fixedRoutingResolver) Resolve(string) (ResolvedAdapter, error) { return r.adapter, nil }
func (r fixedRoutingResolver) ModelRoutingEnabled() bool { return true }
func (r fixedRoutingResolver) ResolveModelRoute(query ModelRouteQuery) (ModelRouteResult, error) {
if query.Model == "public-chat" {
return ModelRouteResult{ResolvedAdapter: r.adapter, TargetModel: "upstream-chat", Matched: true, Known: true}, nil
}
return ModelRouteResult{}, nil
}
func (q *recordingTokenQuota) Reserve(_ context.Context, _ apikey.Principal, estimate int64, _ time.Time) (TokenReservation, error) {
q.estimate = estimate
return q.reservation, nil
}
func (q *recordingTokenQuota) Commit(_ context.Context, _ TokenReservation, actual int64) error {
q.actual = actual
return nil
}
func TestProxyRejectsInvalidKey(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
t.Fatal("upstream must not be called")
}))
defer upstream.Close()
adapter, err := provideropenai.New(upstream.URL, "upstream-secret")
if err != nil {
t.Fatal(err)
}
proxy := NewProxy(adapter, "gateway-secret", 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
request := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
request.Header.Set("Authorization", "Bearer wrong")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusUnauthorized {
t.Fatalf("got %d, want %d", response.Code, http.StatusUnauthorized)
}
}
func TestProxyRejectsKnownOversizedBody(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
t.Fatal("upstream must not be called")
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
proxy := NewProxy(adapter, "gateway-secret", 4, slog.New(slog.NewTextHandler(io.Discard, nil)))
request := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader("12345"))
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusRequestEntityTooLarge {
t.Fatalf("got %d, want %d", response.Code, http.StatusRequestEntityTooLarge)
}
}
func TestProxyReplacesClientAuthorization(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
if got := request.Header.Get("Authorization"); got != "Bearer upstream-secret" {
t.Fatalf("unexpected upstream authorization: %q", got)
}
writer.Header().Set("Content-Type", "application/json")
_, _ = writer.Write([]byte(`{"ok":true}`))
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
proxy := NewProxy(adapter, "gateway-secret", 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
request := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusOK {
t.Fatalf("got %d, want %d", response.Code, http.StatusOK)
}
}
func TestProxyRejectsRateLimitedPrincipalBeforeUpstream(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
t.Fatal("upstream must not be called")
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
proxy := NewProxyWithAuthenticator(adapter, principalAuthenticator{principal: apikey.Principal{
APIKeyID: "key-1", Scopes: []string{"gateway:invoke"}, RequestsPerMinute: 2,
}}, 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
proxy.SetAdmissionController(fixedAdmission{decision: AdmissionDecision{
Allowed: false, Reason: AdmissionRateLimit, Limit: 2, Remaining: 0,
ResetAt: time.Now().Add(time.Minute), RetryAfter: 30 * time.Second,
}})
request := httptest.NewRequest(http.MethodGet, "/v1/models", nil)
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusTooManyRequests {
t.Fatalf("got %d, want %d", response.Code, http.StatusTooManyRequests)
}
if response.Header().Get("X-RateLimit-Limit") != "2" || response.Header().Get("Retry-After") == "" {
t.Fatalf("missing rate limit headers: %#v", response.Header())
}
if !strings.Contains(response.Body.String(), `"type":"rate_limit_error"`) {
t.Fatalf("unexpected body: %s", response.Body.String())
}
}
func TestProxyReconcilesReservedTokensWithUpstreamUsage(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, _ *http.Request) {
writer.Header().Set("Content-Type", "application/json")
_, _ = writer.Write([]byte(`{"choices":[],"usage":{"prompt_tokens":7,"completion_tokens":5,"total_tokens":12}}`))
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
quota := &recordingTokenQuota{reservation: TokenReservation{
Allowed: true, CounterKey: "tokens", Reserved: 24, Limit: 1000, Remaining: 976, ResetAt: time.Now().Add(time.Hour),
}}
proxy := NewProxyWithAuthenticator(adapter, principalAuthenticator{principal: apikey.Principal{
APIKeyID: "key-1", Scopes: []string{"gateway:invoke"}, MonthlyTokenQuota: 1000,
}}, 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
proxy.SetTokenQuotaController(quota)
request := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{"model":"test","max_tokens":20}`))
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusOK {
t.Fatalf("got %d, want %d: %s", response.Code, http.StatusOK, response.Body.String())
}
if quota.estimate <= 20 {
t.Fatalf("request budget did not include input tokens: %d", quota.estimate)
}
if quota.actual != 12 {
t.Fatalf("committed %d tokens, want 12", quota.actual)
}
}
func TestProxyRejectsTokenReservationBeforeUpstream(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
t.Fatal("upstream must not be called")
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
quota := &recordingTokenQuota{reservation: TokenReservation{
Allowed: false, Limit: 100, Remaining: 4, ResetAt: time.Now().Add(time.Hour),
}}
proxy := NewProxyWithAuthenticator(adapter, principalAuthenticator{principal: apikey.Principal{
APIKeyID: "key-1", Scopes: []string{"gateway:invoke"}, MonthlyTokenQuota: 100,
}}, 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
proxy.SetTokenQuotaController(quota)
request := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(`{"model":"test","max_output_tokens":20}`))
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusTooManyRequests || !strings.Contains(response.Body.String(), `"type":"insufficient_quota"`) {
t.Fatalf("unexpected response %d: %s", response.Code, response.Body.String())
}
if response.Header().Get("X-TokenLimit-Limit") != "100" {
t.Fatalf("missing token quota headers: %#v", response.Header())
}
}
func TestProxyRewritesModelAliasBeforeUpstream(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, request *http.Request) {
body, _ := io.ReadAll(request.Body)
if !strings.Contains(string(body), `"model":"upstream-chat"`) {
t.Fatalf("model alias was not rewritten: %s", body)
}
writer.Header().Set("Content-Type", "application/json")
_, _ = writer.Write([]byte(`{"choices":[]}`))
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
resolved := ResolvedAdapter{Code: "routed", Adapter: adapter, Capabilities: map[provider.Capability]bool{provider.CapabilityChat: true}}
proxy := NewDynamicProxy(fixedRoutingResolver{adapter: resolved}, principalAuthenticator{principal: apikey.Principal{Scopes: []string{"gateway:invoke"}}}, 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
request := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{"model":"public-chat","messages":[]}`))
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusOK || response.Header().Get("X-Gateway-Model") != "upstream-chat" || response.Header().Get("X-Gateway-Provider") != "routed" {
t.Fatalf("unexpected routed response %d %#v: %s", response.Code, response.Header(), response.Body.String())
}
}
func TestProxyRecordsAuditWithoutBufferingWholeResponse(t *testing.T) {
upstream := httptest.NewServer(http.HandlerFunc(func(writer http.ResponseWriter, _ *http.Request) {
writer.Header().Set("Content-Type", "application/json")
_, _ = writer.Write([]byte(`{"choices":[],"usage":{"prompt_tokens":9,"completion_tokens":4,"total_tokens":13}}`))
}))
defer upstream.Close()
adapter, _ := provideropenai.New(upstream.URL, "upstream-secret")
recorder := &recordingAudit{}
proxy := NewProxyWithAuthenticator(adapter, principalAuthenticator{principal: apikey.Principal{
APIKeyID: "11111111-1111-4111-8111-111111111111", Scopes: []string{"gateway:invoke"},
}}, 1024, slog.New(slog.NewTextHandler(io.Discard, nil)))
proxy.SetAuditRecorder(recorder)
request := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(`{"model":"audit-model","messages":[]}`))
request.Header.Set("Authorization", "Bearer gateway-secret")
response := httptest.NewRecorder()
proxy.ServeHTTP(response, request)
if response.Code != http.StatusOK {
t.Fatalf("unexpected response %d: %s", response.Code, response.Body.String())
}
if recorder.event.Model != "audit-model" || recorder.event.ProviderCode != "environment" || recorder.event.StatusCode != 200 || recorder.event.PromptTokens != 9 || recorder.event.CompletionTokens != 4 {
t.Fatalf("unexpected audit event: %#v", recorder.event)
}
}
+46
View File
@@ -0,0 +1,46 @@
package gateway
import (
"bytes"
"encoding/json"
"errors"
"io"
"net/http"
)
var errRequestBodyTooLarge = errors.New("request body is too large")
func prepareTokenBudget(request *http.Request, maxBody int64) (int64, error) {
if request.Body == nil || request.Method == http.MethodGet {
return 0, nil
}
limited := io.LimitReader(request.Body, maxBody+1)
body, err := io.ReadAll(limited)
if err != nil {
return 0, err
}
_ = request.Body.Close()
if int64(len(body)) > maxBody {
return 0, errRequestBodyTooLarge
}
restoreRequestBody(request, body)
inputEstimate := max(int64((len(body)+3)/4), 1)
if request.URL.Path == "/v1/embeddings" {
return inputEstimate, nil
}
outputBudget := int64(1024)
decoder := json.NewDecoder(bytes.NewReader(body))
decoder.UseNumber()
var payload map[string]any
if decoder.Decode(&payload) == nil {
for _, field := range []string{"max_output_tokens", "max_completion_tokens", "max_tokens"} {
if value := jsonInt64(payload[field]); value > 0 {
outputBudget = value
break
}
}
}
// A per-request ceiling prevents malicious payloads from reserving unbounded Redis counters.
return min(inputEstimate+outputBudget, int64(100_000_000)), nil
}
+149
View File
@@ -0,0 +1,149 @@
package gateway
import (
"context"
"errors"
"io"
"net/http"
"sync"
"time"
)
var ErrCircuitOpen = errors.New("provider circuit is open")
type ResiliencePolicy struct {
ResponseHeaderTimeout time.Duration
MaxRetries int
RetryBackoff time.Duration
CircuitThreshold int
CircuitOpenDuration time.Duration
}
func DefaultResiliencePolicy() ResiliencePolicy {
return ResiliencePolicy{
ResponseHeaderTimeout: 60 * time.Second, MaxRetries: 2, RetryBackoff: 50 * time.Millisecond,
CircuitThreshold: 5, CircuitOpenDuration: 30 * time.Second,
}
}
type circuitBreaker struct {
mu sync.Mutex
failures int
threshold int
openFor time.Duration
openUntil time.Time
halfOpenRun bool
}
func newCircuitBreaker(policy ...ResiliencePolicy) *circuitBreaker {
settings := DefaultResiliencePolicy()
if len(policy) > 0 {
settings = policy[0]
}
return &circuitBreaker{threshold: settings.CircuitThreshold, openFor: settings.CircuitOpenDuration}
}
func (c *circuitBreaker) allow(now time.Time) bool {
c.mu.Lock()
defer c.mu.Unlock()
if c.openUntil.IsZero() {
return true
}
if now.Before(c.openUntil) || c.halfOpenRun {
return false
}
c.halfOpenRun = true
return true
}
func (c *circuitBreaker) success() {
c.mu.Lock()
c.failures = 0
c.openUntil = time.Time{}
c.halfOpenRun = false
c.mu.Unlock()
}
func (c *circuitBreaker) failure(now time.Time) {
c.mu.Lock()
defer c.mu.Unlock()
c.halfOpenRun = false
c.failures++
if c.failures >= c.threshold || !c.openUntil.IsZero() {
c.openUntil = now.Add(c.openFor)
}
}
type resilientTransport struct {
base http.RoundTripper
circuit *circuitBreaker
maxRetries int
backoff time.Duration
}
func (t *resilientTransport) RoundTrip(request *http.Request) (*http.Response, error) {
if !t.circuit.allow(time.Now()) {
return nil, ErrCircuitOpen
}
replayable := request.Method == http.MethodGet || request.Method == http.MethodHead ||
request.Header.Get("Idempotency-Key") != "" && request.GetBody != nil
attempts := 1
if replayable {
attempts += t.maxRetries
}
var response *http.Response
var err error
for attempt := 0; attempt < attempts; attempt++ {
current := request
if attempt > 0 {
if waitErr := waitBackoff(request.Context(), t.backoff*time.Duration(attempt)); waitErr != nil {
err = waitErr
break
}
current = request.Clone(request.Context())
if request.Body != nil && request.Body != http.NoBody {
body, bodyErr := request.GetBody()
if bodyErr != nil {
err = bodyErr
break
}
current.Body = body
}
}
response, err = t.base.RoundTrip(current)
if !retryableResult(response, err) || attempt == attempts-1 {
break
}
if response != nil {
_, _ = io.CopyN(io.Discard, response.Body, 4096)
_ = response.Body.Close()
}
}
if retryableResult(response, err) {
t.circuit.failure(time.Now())
} else {
t.circuit.success()
}
return response, err
}
func retryableResult(response *http.Response, err error) bool {
if err != nil {
return !errors.Is(err, context.Canceled) && !errors.Is(err, context.DeadlineExceeded)
}
return response != nil && (response.StatusCode == http.StatusBadGateway || response.StatusCode == http.StatusServiceUnavailable || response.StatusCode == http.StatusGatewayTimeout)
}
func waitBackoff(ctx context.Context, duration time.Duration) error {
if duration <= 0 {
return nil
}
timer := time.NewTimer(duration)
defer timer.Stop()
select {
case <-ctx.Done():
return ctx.Err()
case <-timer.C:
return nil
}
}
+61
View File
@@ -0,0 +1,61 @@
package gateway
import (
"errors"
"io"
"net/http"
"strings"
"testing"
"time"
)
type roundTripFunc func(*http.Request) (*http.Response, error)
func (f roundTripFunc) RoundTrip(request *http.Request) (*http.Response, error) { return f(request) }
func TestResilientTransportRetriesReplayableRequest(t *testing.T) {
attempts := 0
transport := &resilientTransport{base: roundTripFunc(func(*http.Request) (*http.Response, error) {
attempts++
if attempts < 3 {
return &http.Response{StatusCode: http.StatusServiceUnavailable, Body: io.NopCloser(strings.NewReader("busy")), Header: make(http.Header)}, nil
}
return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader("ok")), Header: make(http.Header)}, nil
}), circuit: newCircuitBreaker(), maxRetries: 2}
request, _ := http.NewRequest(http.MethodGet, "https://provider.example/v1/models", nil)
response, err := transport.RoundTrip(request)
if err != nil || response.StatusCode != http.StatusOK || attempts != 3 {
t.Fatalf("response=%v err=%v attempts=%d", response, err, attempts)
}
}
func TestResilientTransportDoesNotRetryUnsafePost(t *testing.T) {
attempts := 0
transport := &resilientTransport{base: roundTripFunc(func(*http.Request) (*http.Response, error) {
attempts++
return nil, errors.New("network failure")
}), circuit: newCircuitBreaker(), maxRetries: 2}
request, _ := http.NewRequest(http.MethodPost, "https://provider.example/v1/chat/completions", strings.NewReader("{}"))
_, _ = transport.RoundTrip(request)
if attempts != 1 {
t.Fatalf("unsafe POST attempted %d times", attempts)
}
}
func TestCircuitOpensAndAllowsSingleProbe(t *testing.T) {
circuit := newCircuitBreaker()
circuit.threshold = 2
circuit.openFor = time.Millisecond
now := time.Now()
circuit.failure(now)
circuit.failure(now)
if circuit.allow(now) {
t.Fatal("open circuit allowed request")
}
if !circuit.allow(now.Add(2 * time.Millisecond)) {
t.Fatal("circuit did not allow half-open probe")
}
if circuit.allow(now.Add(2 * time.Millisecond)) {
t.Fatal("circuit allowed concurrent half-open probe")
}
}
+76
View File
@@ -0,0 +1,76 @@
package gateway
import (
"bytes"
"encoding/json"
"io"
"net/http"
)
type ModelRouteQuery struct {
ProviderCode string
Model string
Endpoint string
APIKeyID string
TenantID string
Seed string
}
type ModelRouteResult struct {
ResolvedAdapter
TargetModel string
Matched bool
Known bool
}
type ModelRouteResolver interface {
ModelRoutingEnabled() bool
ResolveModelRoute(ModelRouteQuery) (ModelRouteResult, error)
}
type modelRequest struct {
body map[string]json.RawMessage
model string
}
func readModelRequest(request *http.Request, maxBody int64) (modelRequest, error) {
if request.Body == nil || request.Method == http.MethodGet {
return modelRequest{}, nil
}
body, err := io.ReadAll(io.LimitReader(request.Body, maxBody+1))
if err != nil {
return modelRequest{}, err
}
_ = request.Body.Close()
if int64(len(body)) > maxBody {
return modelRequest{}, errRequestBodyTooLarge
}
restoreRequestBody(request, body)
var payload map[string]json.RawMessage
if json.Unmarshal(body, &payload) != nil {
return modelRequest{}, nil
}
var model string
_ = json.Unmarshal(payload["model"], &model)
return modelRequest{body: payload, model: model}, nil
}
func (m modelRequest) rewrite(request *http.Request, target string) error {
if len(m.body) == 0 || target == "" || target == m.model {
return nil
}
encodedModel, _ := json.Marshal(target)
m.body["model"] = encodedModel
body, err := json.Marshal(m.body)
if err != nil {
return err
}
restoreRequestBody(request, body)
return nil
}
func restoreRequestBody(request *http.Request, body []byte) {
request.Body = io.NopCloser(bytes.NewReader(body))
request.ContentLength = int64(len(body))
request.GetBody = func() (io.ReadCloser, error) { return io.NopCloser(bytes.NewReader(body)), nil }
}
+129
View File
@@ -0,0 +1,129 @@
package gateway
import (
"context"
"errors"
"fmt"
"strconv"
"time"
"aigateway.local/core/internal/apikey"
"github.com/redis/go-redis/v9"
)
var ErrTokenQuotaUnavailable = errors.New("token quota unavailable")
type TokenReservation struct {
Allowed bool
APIKeyID string
CounterKey string
Reserved int64
Limit int64
Remaining int64
ResetAt time.Time
}
type TokenQuotaController interface {
Reserve(context.Context, apikey.Principal, int64, time.Time) (TokenReservation, error)
Commit(context.Context, TokenReservation, int64) error
}
type RedisTokenQuotaController struct {
client *redis.Client
reserveScript *redis.Script
commitScript *redis.Script
}
func NewRedisTokenQuotaController(client *redis.Client) *RedisTokenQuotaController {
return &RedisTokenQuotaController{
client: client, reserveScript: redis.NewScript(tokenReserveScript), commitScript: redis.NewScript(tokenCommitScript),
}
}
func (c *RedisTokenQuotaController) Reserve(ctx context.Context, principal apikey.Principal, estimate int64, now time.Time) (TokenReservation, error) {
if principal.APIKeyID == "" {
return TokenReservation{Allowed: true}, nil
}
if estimate < 0 {
estimate = 0
}
// Accounts without a monthly token quota never need a reservation and must
// not create a pointless monthly counter key in Redis.
if principal.MonthlyTokenQuota == 0 {
return TokenReservation{Allowed: true}, nil
}
if c == nil || c.client == nil {
return TokenReservation{}, ErrTokenQuotaUnavailable
}
now = now.UTC()
reset := time.Date(now.Year(), now.Month()+1, 1, 0, 0, 0, 0, time.UTC)
key := apikey.MonthlyTokenUsageKey(principal.APIKeyID, now)
result, err := c.reserveScript.Run(ctx, c.client, []string{key}, estimate, principal.MonthlyTokenQuota, int64(reset.Sub(now).Seconds())+86400).Slice()
if err != nil || len(result) != 2 {
if principal.MonthlyTokenQuota == 0 {
return TokenReservation{Allowed: true}, nil
}
return TokenReservation{}, fmt.Errorf("%w: %v", ErrTokenQuotaUnavailable, err)
}
allowed, err := redisInteger(result[0])
if err != nil {
return TokenReservation{}, ErrTokenQuotaUnavailable
}
current, err := redisInteger(result[1])
if err != nil {
return TokenReservation{}, ErrTokenQuotaUnavailable
}
return TokenReservation{
Allowed: allowed == 1, APIKeyID: principal.APIKeyID, CounterKey: key, Reserved: estimate,
Limit: principal.MonthlyTokenQuota, Remaining: max(principal.MonthlyTokenQuota-current, 0), ResetAt: reset,
}, nil
}
func (c *RedisTokenQuotaController) Commit(ctx context.Context, reservation TokenReservation, actual int64) error {
if c == nil || c.client == nil || reservation.CounterKey == "" || !reservation.Allowed {
return nil
}
if actual < 0 {
actual = 0
}
if _, err := c.commitScript.Run(ctx, c.client, []string{reservation.CounterKey}, actual-reservation.Reserved).Result(); err != nil {
return fmt.Errorf("%w: %v", ErrTokenQuotaUnavailable, err)
}
return nil
}
const tokenReserveScript = `
local estimate = tonumber(ARGV[1])
local quota = tonumber(ARGV[2])
local current = tonumber(redis.call('GET', KEYS[1]) or '0')
if quota > 0 and current + estimate > quota then
return {0, current}
end
if estimate > 0 then
current = redis.call('INCRBY', KEYS[1], estimate)
if current == estimate then redis.call('EXPIRE', KEYS[1], tonumber(ARGV[3])) end
elseif redis.call('EXISTS', KEYS[1]) == 0 then
redis.call('SET', KEYS[1], 0, 'EX', tonumber(ARGV[3]))
end
return {1, current}
`
const tokenCommitScript = `
local delta = tonumber(ARGV[1])
local current = tonumber(redis.call('GET', KEYS[1]) or '0')
local updated = current + delta
if updated < 0 then updated = 0 end
redis.call('SET', KEYS[1], updated, 'KEEPTTL')
return updated
`
func writeTokenQuotaHeaders(header mapHeader, reservation TokenReservation) {
if reservation.Limit <= 0 || reservation.ResetAt.IsZero() {
return
}
header.Set("X-TokenLimit-Limit", strconv.FormatInt(reservation.Limit, 10))
header.Set("X-TokenLimit-Remaining", strconv.FormatInt(max(reservation.Remaining, 0), 10))
header.Set("X-TokenLimit-Reset", strconv.FormatInt(reservation.ResetAt.Unix(), 10))
}
type mapHeader interface{ Set(string, string) }
+261
View File
@@ -0,0 +1,261 @@
package gateway
import (
"bytes"
"context"
"encoding/json"
"io"
"log/slog"
"mime"
"net/http"
"strings"
"sync"
"time"
)
// maxUsageDocumentBytes bounds the buffered tail of a non-streaming response.
// The window is kept from the end of the body (where providers place the
// "usage" object), so token accounting stays correct for responses far larger
// than this bound while memory use stays bounded per in-flight response.
const maxUsageDocumentBytes = 2 << 20
type usageSession struct {
controller TokenQuotaController
reservation TokenReservation
fallback int64
onFinish func(TokenUsage)
once sync.Once
log *slog.Logger
}
type TokenUsage struct {
Input int64
Output int64
Total int64
}
type usageSessionContextKey struct{}
func withUsageSession(request *http.Request, session *usageSession) *http.Request {
return request.WithContext(context.WithValue(request.Context(), usageSessionContextKey{}, session))
}
func usageSessionFrom(request *http.Request) *usageSession {
session, _ := request.Context().Value(usageSessionContextKey{}).(*usageSession)
return session
}
func (s *usageSession) finish(usage TokenUsage) {
if s == nil {
return
}
s.once.Do(func() {
usage.Total = max(usage.Total, usage.Input+usage.Output)
if usage.Total <= 0 {
usage.Total = s.fallback
}
if s.controller != nil {
ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
if err := s.controller.Commit(ctx, s.reservation, usage.Total); err != nil && s.log != nil {
// The reservation already counted towards the quota, so a
// failed reconciliation silently leaves the counter slightly
// off. Surface it instead of dropping it.
s.log.Warn("token quota commit failed", "error", err)
}
cancel()
}
if s.onFinish != nil {
s.onFinish(usage)
}
})
}
type usageReadCloser struct {
io.ReadCloser
collector *usageCollector
session *usageSession
}
func newUsageReadCloser(body io.ReadCloser, contentType string, session *usageSession) io.ReadCloser {
mediaType, _, _ := mime.ParseMediaType(contentType)
return &usageReadCloser{ReadCloser: body, collector: &usageCollector{sse: mediaType == "text/event-stream"}, session: session}
}
func (r *usageReadCloser) Read(buffer []byte) (int, error) {
n, err := r.ReadCloser.Read(buffer)
if n > 0 {
r.collector.feed(buffer[:n])
}
if err == io.EOF {
r.session.finish(r.collector.usage())
}
return n, err
}
func (r *usageReadCloser) Close() error {
r.session.finish(r.collector.usage())
return r.ReadCloser.Close()
}
type usageCollector struct {
sse bool
pending []byte
doc []byte
input int64
output int64
total int64
}
func (c *usageCollector) feed(chunk []byte) {
if !c.sse {
c.doc = append(c.doc, chunk...)
// Keep only a bounded tail window. The "usage" member lives at the end
// of a non-streaming response, so dropping the head (never the tail)
// preserves accounting for arbitrarily large bodies at a fixed memory
// cost instead of truncating usage away.
if len(c.doc) > maxUsageDocumentBytes {
c.doc = append([]byte(nil), c.doc[len(c.doc)-maxUsageDocumentBytes:]...)
}
return
}
c.pending = append(c.pending, chunk...)
for {
index := bytes.IndexByte(c.pending, '\n')
if index < 0 {
if len(c.pending) > maxUsageDocumentBytes {
c.pending = c.pending[:0]
}
return
}
line := strings.TrimSpace(string(c.pending[:index]))
c.pending = c.pending[index+1:]
if strings.HasPrefix(line, "data:") {
payload := strings.TrimSpace(strings.TrimPrefix(line, "data:"))
if payload != "" && payload != "[DONE]" {
c.consumeJSON([]byte(payload))
}
}
}
}
func (c *usageCollector) tokens() int64 {
return c.usage().Total
}
func (c *usageCollector) usage() TokenUsage {
if c.sse && len(c.pending) > 0 {
line := strings.TrimSpace(string(c.pending))
if strings.HasPrefix(line, "data:") {
c.consumeJSON([]byte(strings.TrimSpace(strings.TrimPrefix(line, "data:"))))
}
c.pending = nil
}
if !c.sse && len(c.doc) > 0 {
c.consumeJSON(c.doc) // fast path: whole valid JSON object
c.consumeUsageObject(c.doc) // tail extraction: covers truncated bodies
}
return TokenUsage{Input: c.input, Output: c.output, Total: max(c.total, c.input+c.output)}
}
// consumeUsageObject extracts the trailing `"usage"` object from a possibly
// truncated response document. It scans for the final `"usage"` member and
// decodes the JSON object that immediately follows — where OpenAI-compatible
// providers place token accounting in non-streaming responses — so usage is
// still counted when the buffered window starts mid-object.
func (c *usageCollector) consumeUsageObject(doc []byte) {
const key = `"usage"`
index := bytes.LastIndex(doc, []byte(key))
if index < 0 {
return
}
rest := doc[index+len(key):]
colon := bytes.IndexByte(rest, ':')
if colon < 0 {
return
}
rest = bytes.TrimSpace(rest[colon+1:])
if len(rest) == 0 || rest[0] != '{' {
return
}
depth := 0
end := -1
inString := false
escaped := false
for i := 0; i < len(rest); i++ {
ch := rest[i]
if inString {
if escaped {
escaped = false
} else if ch == '\\' {
escaped = true
} else if ch == '"' {
inString = false
}
continue
}
switch ch {
case '"':
inString = true
case '{':
depth++
case '}':
depth--
if depth == 0 {
end = i + 1
}
}
if end > 0 {
break
}
}
if end > 0 {
c.consumeJSON(rest[:end])
}
}
func (c *usageCollector) consumeJSON(payload []byte) {
decoder := json.NewDecoder(bytes.NewReader(payload))
decoder.UseNumber()
var value any
if decoder.Decode(&value) == nil {
c.walk(value, false)
}
}
func (c *usageCollector) walk(value any, inUsage bool) {
switch typed := value.(type) {
case map[string]any:
for key, child := range typed {
usage := inUsage || key == "usage"
if usage {
switch key {
case "input_tokens", "prompt_tokens":
c.input = max(c.input, jsonInt64(child))
case "output_tokens", "completion_tokens":
c.output = max(c.output, jsonInt64(child))
case "total_tokens":
c.total = max(c.total, jsonInt64(child))
}
}
c.walk(child, usage)
}
case []any:
for _, child := range typed {
c.walk(child, inUsage)
}
}
}
func jsonInt64(value any) int64 {
switch number := value.(type) {
case json.Number:
parsed, _ := number.Int64()
return max(parsed, 0)
case float64:
return max(int64(number), 0)
case int64:
return max(number, 0)
default:
return 0
}
}
+21
View File
@@ -0,0 +1,21 @@
package gateway
import "testing"
func TestUsageCollectorReadsOpenAIJSON(t *testing.T) {
collector := &usageCollector{}
collector.feed([]byte(`{"usage":{"prompt_tokens":11,"completion_tokens":7,"total_tokens":18}}`))
if got := collector.tokens(); got != 18 {
t.Fatalf("got %d, want 18", got)
}
}
func TestUsageCollectorReadsChunkedSSEAndAnthropicUsage(t *testing.T) {
collector := &usageCollector{sse: true}
collector.feed([]byte("event: message_start\ndata: {\"message\":{\"usage\":{\"input_tokens\":13}}}\n\n"))
collector.feed([]byte("event: message_delta\ndata: {\"usage\":{\"output_"))
collector.feed([]byte("tokens\":9}}\n\ndata: [DONE]\n\n"))
if got := collector.tokens(); got != 22 {
t.Fatalf("got %d, want 22", got)
}
}