Merge pull request #13 from RuoJi6/codex/fix-llm-strict-rate-limit

fix: enforce strict LLM HTTP request rate limits

agent/provider.go

@@ -250,14 +250,30 @@ func (c Config) Provider() string {
 	return "anthropic"
 }
 
-// NewProvider builds an llm.Provider from the config. When a rate is set, the
-// limiter lives on the single provider instance — so planner + all workers +
-// main agent (which share this provider) are bounded by one shared rate limit.
+// NewProvider builds a provider with its own HTTP request limiter. Every HTTP
+// attempt, including SDK retries, counts toward both rolling rate windows.
 func (c Config) NewProvider() (llm.Provider, error) {
+	return c.NewProviderWithLimiter(&RequestLimiter{})
+}
+
+// NewProviderWithLimiter reuses quota history across provider rebuilds. Callers
+// sharing a profile must retain the same limiter outside their provider cache.
+func (c Config) NewProviderWithLimiter(limiter *RequestLimiter) (llm.Provider, error) {
 	client, err := quotaAwareHTTPClient(c.Proxy, c.SessionHeaderKey)
 	if err != nil {
 		return nil, err
 	}
+	return c.newProvider(client, limiter)
+}
+
+func (c Config) newProvider(client *http.Client, limiter *RequestLimiter) (llm.Provider, error) {
+	if limiter == nil {
+		limiter = &RequestLimiter{}
+	}
+	if err := limiter.configure(c.RatePerSecond, c.RatePerMinute); err != nil {
+		return nil, err
+	}
+	client.Transport = rateLimitedTransport{base: client.Transport, limiter: limiter}
 	lc := llm.Config{
 		Format:     c.Format,
 		BaseURL:    c.BaseURL,
@@ -277,9 +293,8 @@ func (c Config) NewProvider() (llm.Provider, error) {
 	lc.RetryInterval = c.Retry.ConnectInterval
 	lc.EmptyResponseRetries = c.Retry.EmptyAttempts
 	lc.EmptyResponseInterval = c.Retry.EmptyInterval
-	if c.RatePerSecond > 0 || c.RatePerMinute > 0 {
-		lc.RateLimit = &llm.RateLimit{PerSecond: c.RatePerSecond, PerMinute: c.RatePerMinute}
-	}
+	// Norma's limiter allows a full-window burst and excludes establishment
+	// retries. Enforce limits on HTTP attempts instead; do not also enable it.
 	return llm.NewProvider(lc)
 }
 

agent/request_limiter.go

@@ -0,0 +1,136 @@
+package agent
+
+import (
+	"context"
+	"fmt"
+	"math"
+	"net/http"
+	"sync"
+	"time"
+)
+
+// RequestLimiter counts HTTP attempts in rolling second/minute windows. Keep it
+// outside provider caches so rebuilding a provider does not reset its quota.
+// Its zero value is ready for use; limits are set by NewProviderWithLimiter.
+type RequestLimiter struct {
+	mu      sync.Mutex
+	windows [2]requestWindow
+	sent    []time.Time
+	changed chan struct{}
+}
+
+type requestWindow struct {
+	period time.Duration
+	limit  float64
+}
+
+func (l *RequestLimiter) configure(perSecond, perMinute float64) error {
+	var windows [2]requestWindow
+	for i, rate := range [2]float64{perSecond, perMinute} {
+		if math.IsNaN(rate) || math.IsInf(rate, 0) {
+			return fmt.Errorf("LLM request rate must be finite")
+		}
+		if rate <= 0 {
+			continue
+		}
+		period := [2]time.Duration{time.Second, time.Minute}[i]
+		// Whole HTTP attempts cannot exceed a fractional window capacity.
+		limit := math.Floor(rate)
+		if rate < 1 {
+			// Preserve fractional rates such as 0.5/s as one attempt per 2s.
+			d := math.Ceil(float64(period) / rate)
+			if d >= float64(math.MaxInt64) {
+				return fmt.Errorf("LLM request rate is too small")
+			}
+			period, limit = time.Duration(d), 1
+		}
+		windows[i] = requestWindow{period: period, limit: limit}
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.windows = windows
+	if l.changed != nil {
+		close(l.changed)
+	}
+	l.changed = make(chan struct{})
+	return nil
+}
+
+// wait counts a request only when it may actually proceed. Waiters recheck both
+// windows under the same lock, so concurrency and cancellation cannot overbook
+// capacity. Requests in (now-period, now] count toward each rolling window.
+func (l *RequestLimiter) wait(ctx context.Context) error {
+	if l == nil {
+		return ctx.Err()
+	}
+	for {
+		l.mu.Lock()
+		if err := ctx.Err(); err != nil {
+			l.mu.Unlock()
+			return err
+		}
+		now := time.Now()
+		// Retain at least a minute across limit changes, including while limits
+		// are disabled. Fractional rates may need a longer history.
+		retention := time.Minute
+		for _, w := range l.windows {
+			if w.period > retention {
+				retention = w.period
+			}
+		}
+		first := 0
+		// Keep the last attempt even when older than retention: a subsequent
+		// switch to a fractional rate may need to space requests further apart.
+		for first+1 < len(l.sent) && !l.sent[first].After(now.Add(-retention)) {
+			first++
+		}
+		l.sent = l.sent[first:]
+		var delay time.Duration
+		for _, w := range l.windows {
+			if w.limit == 0 {
+				continue
+			}
+			start := len(l.sent)
+			for start > 0 && l.sent[start-1].After(now.Add(-w.period)) {
+				start--
+			}
+			count := len(l.sent) - start
+			if float64(count) >= w.limit {
+				// If a limit was lowered, enough old requests must expire to
+				// leave room for this attempt, not just the oldest request.
+				d := l.sent[len(l.sent)-int(w.limit)].Add(w.period).Sub(now)
+				if d > delay {
+					delay = d
+				}
+			}
+		}
+		if delay <= 0 {
+			l.sent = append(l.sent, now)
+			l.mu.Unlock()
+			return nil
+		}
+		changed := l.changed
+		l.mu.Unlock()
+		timer := time.NewTimer(delay)
+		select {
+		case <-ctx.Done():
+			timer.Stop()
+			return ctx.Err()
+		case <-changed:
+			timer.Stop()
+		case <-timer.C:
+		}
+	}
+}
+
+type rateLimitedTransport struct {
+	base    http.RoundTripper
+	limiter *RequestLimiter
+}
+
+func (t rateLimitedTransport) RoundTrip(req *http.Request) (*http.Response, error) {
+	if err := t.limiter.wait(req.Context()); err != nil {
+		return nil, err
+	}
+	return t.base.RoundTrip(req)
+}

agent/request_limiter_test.go

@@ -0,0 +1,446 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"io"
+	"math"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"testing/synctest"
+	"time"
+
+	"github.com/Autumn-27/norma/llm"
+)
+
+func TestRequestLimiterRollingWindows(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		rps, rpm float64
+		capacity int
+		window   time.Duration
+	}{
+		{"second", 10, 0, 10, time.Second},
+		{"minute_45", 0, 45, 45, time.Minute},
+		{"fractional_second", .5, 0, 1, 2 * time.Second},
+		{"fractional_minute", 0, .5, 1, 2 * time.Minute},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			synctest.Test(t, func(t *testing.T) {
+				l := &RequestLimiter{}
+				if err := l.configure(tc.rps, tc.rpm); err != nil {
+					t.Fatal(err)
+				}
+				start := time.Now()
+				for i := 0; i < tc.capacity; i++ {
+					if err := l.wait(context.Background()); err != nil {
+						t.Fatal(err)
+					}
+				}
+				ctx, cancel := context.WithTimeout(context.Background(), tc.window-time.Nanosecond)
+				defer cancel()
+				if err := l.wait(ctx); !errors.Is(err, context.DeadlineExceeded) {
+					t.Fatalf("extra attempt before window expired: %v", err)
+				}
+				if err := l.wait(context.Background()); err != nil {
+					t.Fatal(err)
+				}
+				if elapsed := time.Since(start); elapsed != tc.window {
+					t.Fatalf("next attempt at %s, want %s; cancelled wait must not consume quota", elapsed, tc.window)
+				}
+			})
+		})
+	}
+}
+
+func TestRequestLimiterBothWindowsAndConcurrentCallers(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		l := &RequestLimiter{}
+		if err := l.configure(3, 5); err != nil {
+			t.Fatal(err)
+		}
+		var mu sync.Mutex
+		var sent []time.Time
+		var wg sync.WaitGroup
+		start := time.Now()
+		for i := 0; i < 12; i++ {
+			wg.Add(1)
+			go func() {
+				defer wg.Done()
+				if err := l.wait(context.Background()); err != nil {
+					t.Error(err)
+					return
+				}
+				mu.Lock()
+				sent = append(sent, time.Now())
+				mu.Unlock()
+			}()
+		}
+		wg.Wait()
+		if elapsed := time.Since(start); elapsed != 2*time.Minute {
+			t.Fatalf("12 requests under 3/s and 5/min took %s, want 2m", elapsed)
+		}
+		assertRollingRequestCounts(t, sent, time.Second, 3)
+		assertRollingRequestCounts(t, sent, time.Minute, 5)
+	})
+}
+
+func assertRollingRequestCounts(t *testing.T, sent []time.Time, window time.Duration, limit int) {
+	t.Helper()
+	for _, end := range sent {
+		count := 0
+		for _, ts := range sent {
+			if ts.After(end.Add(-window)) && !ts.After(end) {
+				count++
+			}
+		}
+		if count > limit {
+			t.Fatalf("%d attempts in rolling %s, limit %d", count, window, limit)
+		}
+	}
+}
+
+func TestRequestLimiterReconfiguration(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		l := &RequestLimiter{}
+		if err := l.configure(0, 5); err != nil {
+			t.Fatal(err)
+		}
+		for i := 0; i < 5; i++ {
+			if err := l.wait(context.Background()); err != nil {
+				t.Fatal(err)
+			}
+		}
+		start := time.Now()
+		if err := l.configure(0, 2); err != nil {
+			t.Fatal(err)
+		}
+		done := make(chan error, 1)
+		go func() { done <- l.wait(context.Background()) }()
+		synctest.Wait()
+		select {
+		case err := <-done:
+			t.Fatalf("lowering limit reset history: %v", err)
+		default:
+		}
+		// A change must wake waiting requests rather than leave them sleeping
+		// until the old minute window expires.
+		if err := l.configure(0, 6); err != nil {
+			t.Fatal(err)
+		}
+		if err := <-done; err != nil {
+			t.Fatal(err)
+		}
+		if time.Since(start) != 0 {
+			t.Fatal("waiter did not wake on limit increase")
+		}
+		if err := l.configure(0, 2); err != nil {
+			t.Fatal(err)
+		}
+		if err := l.wait(context.Background()); err != nil {
+			t.Fatal(err)
+		}
+		if time.Since(start) != time.Minute {
+			t.Fatal("lower limit did not retain previous attempts")
+		}
+	})
+}
+
+func TestRequestLimiterDisabledAndCancelled(t *testing.T) {
+	l := &RequestLimiter{}
+	if err := l.configure(0, 0); err != nil {
+		t.Fatal(err)
+	}
+	for i := 0; i < 100; i++ {
+		if err := l.wait(context.Background()); err != nil {
+			t.Fatal(err)
+		}
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if err := l.wait(ctx); !errors.Is(err, context.Canceled) {
+		t.Fatalf("got %v", err)
+	}
+	if len(l.sent) != 100 {
+		t.Fatal("cancelled request consumed quota")
+	}
+	for _, rate := range []float64{math.NaN(), math.Inf(1), math.Inf(-1), math.SmallestNonzeroFloat64} {
+		if err := l.configure(rate, 0); err == nil {
+			t.Errorf("accepted invalid rate %v", rate)
+		}
+	}
+}
+
+// Exercise the actual Norma retry loops for all formats and both completion
+// modes. In-memory HTTP responses let synctest advance a full minute instantly.
+func TestProviderRateLimitCountsHTTPRetries(t *testing.T) {
+	for _, format := range []string{"openai", "anthropic", "openai-responses"} {
+		for _, streaming := range []bool{false, true} {
+			mode := "complete"
+			if streaming {
+				mode = "stream"
+			}
+			t.Run(format+"/"+mode, func(t *testing.T) {
+				synctest.Test(t, func(t *testing.T) {
+					var sent []time.Time
+					client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+						sent = append(sent, time.Now())
+						return &http.Response{StatusCode: 500, Header: make(http.Header),
+							Body: io.NopCloser(strings.NewReader(`{"error":{"message":"temporary error"}}`))}, nil
+					})}
+					cfg := ConfigFrom(format, "test-model", "http://example.invalid", "test-key", "")
+					cfg.RatePerSecond, cfg.RatePerMinute = 1, 2
+					cfg.Retry.ConnectInterval = time.Millisecond
+					prov, err := cfg.newProvider(client, &RequestLimiter{})
+					if err != nil {
+						t.Fatal(err)
+					}
+					start := time.Now()
+					req := llm.CompletionRequest{Messages: []llm.Message{llm.UserText("ping")}}
+					if streaming {
+						for _, streamErr := range prov.Stream(context.Background(), req) {
+							err = streamErr
+						}
+					} else {
+						_, _, _, err = prov.Complete(context.Background(), req)
+					}
+					if err == nil {
+						t.Fatal("expected upstream 500")
+					}
+					if len(sent) != 4 {
+						t.Fatalf("HTTP attempts=%d, want initial+3 retries", len(sent))
+					}
+					assertRollingRequestCounts(t, sent, time.Second, 1)
+					assertRollingRequestCounts(t, sent, time.Minute, 2)
+					if elapsed := time.Since(start); elapsed != time.Minute+time.Second {
+						t.Fatalf("retry attempts took %s, want 1m1s", elapsed)
+					}
+				})
+			})
+		}
+	}
+}
+
+func TestProviderRebuildPreservesRateLimit(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		l := &RequestLimiter{}
+		var sent []time.Time
+		cfg := ConfigFrom("openai", "test-model", "http://example.invalid", "test-key", "")
+		cfg.RatePerMinute = 1
+		cfg.Retry.ConnectAttempts = -1
+		for i := 0; i < 2; i++ {
+			client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+				sent = append(sent, time.Now())
+				return &http.Response{StatusCode: 500, Header: make(http.Header), Body: io.NopCloser(strings.NewReader("error"))}, nil
+			})}
+			prov, err := cfg.newProvider(client, l)
+			if err != nil {
+				t.Fatal(err)
+			}
+			prov.Complete(context.Background(), llm.CompletionRequest{Messages: []llm.Message{llm.UserText("ping")}})
+		}
+		if len(sent) != 2 || sent[1].Sub(sent[0]) != time.Minute {
+			t.Fatalf("provider rebuild reset quota: %v", sent)
+		}
+	})
+}
+
+func TestProviderRateLimitSuccessfulRequests(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		rps, rpm float64
+		calls    int
+		window   time.Duration
+		limit    int
+	}{
+		{"minute_45_requests_60", 0, 45, 60, time.Minute, 45},
+		{"second_10_requests_15", 10, 0, 15, time.Second, 10},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			synctest.Test(t, func(t *testing.T) {
+				var sent []time.Time
+				client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+					sent = append(sent, time.Now())
+					return &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(
+						`{"choices":[{"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}]}`))}, nil
+				})}
+				cfg := ConfigFrom("openai", "test-model", "http://example.invalid", "test-key", "")
+				cfg.RatePerSecond, cfg.RatePerMinute = tc.rps, tc.rpm
+				prov, err := cfg.newProvider(client, &RequestLimiter{})
+				if err != nil {
+					t.Fatal(err)
+				}
+				start := time.Now()
+				for i := 0; i < tc.calls; i++ {
+					_, _, _, err := prov.Complete(context.Background(), llm.CompletionRequest{Messages: []llm.Message{llm.UserText("ping")}})
+					if err != nil {
+						t.Fatal(err)
+					}
+				}
+				if len(sent) != tc.calls || time.Since(start) != tc.window {
+					t.Fatalf("sent %d attempts in %s, want %d in %s", len(sent), time.Since(start), tc.calls, tc.window)
+				}
+				assertRollingRequestCounts(t, sent, tc.window, tc.limit)
+			})
+		})
+	}
+}
+
+func TestProviderRateLimitCountsEmptyResponseRetries(t *testing.T) {
+	for _, streaming := range []bool{false, true} {
+		name := "complete"
+		if streaming {
+			name = "stream"
+		}
+		t.Run(name, func(t *testing.T) {
+			synctest.Test(t, func(t *testing.T) {
+				var sent []time.Time
+				client := &http.Client{Transport: roundTripperFunc(func(r *http.Request) (*http.Response, error) {
+					sent = append(sent, time.Now())
+					body := `{"choices":[{"message":{"role":"assistant","content":""},"finish_reason":"stop"}]}`
+					if streaming {
+						body = "data: {\"choices\":[{\"delta\":{},\"finish_reason\":\"stop\"}]}\n\ndata: [DONE]\n\n"
+					}
+					return &http.Response{StatusCode: 200, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(body))}, nil
+				})}
+				cfg := ConfigFrom("openai", "test-model", "http://example.invalid", "test-key", "")
+				cfg.RatePerMinute = 1
+				cfg.Retry.EmptyInterval = time.Millisecond
+				prov, err := cfg.newProvider(client, &RequestLimiter{})
+				if err != nil {
+					t.Fatal(err)
+				}
+				req := llm.CompletionRequest{Messages: []llm.Message{llm.UserText("ping")}}
+				if streaming {
+					for range prov.Stream(context.Background(), req) {
+					}
+				} else {
+					prov.Complete(context.Background(), req)
+				}
+				if len(sent) != 3 {
+					t.Fatalf("attempts=%d, want initial + 2 empty retries", len(sent))
+				}
+				assertRollingRequestCounts(t, sent, time.Minute, 1)
+			})
+		})
+	}
+}
+
+func TestRequestLimiterUsesRollingHistory(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		l := &RequestLimiter{}
+		if err := l.configure(3, 0); err != nil {
+			t.Fatal(err)
+		}
+		start := time.Now()
+		for i := 0; i < 2; i++ {
+			if err := l.wait(context.Background()); err != nil {
+				t.Fatal(err)
+			}
+		}
+		time.Sleep(750 * time.Millisecond)
+		if err := l.wait(context.Background()); err != nil {
+			t.Fatal(err)
+		}
+		time.Sleep(350 * time.Millisecond)
+		for i := 0; i < 2; i++ {
+			if err := l.wait(context.Background()); err != nil {
+				t.Fatal(err)
+			}
+		}
+		// Only the two t=0 attempts expired. The one at t=.75 still counts.
+		if err := l.wait(context.Background()); err != nil {
+			t.Fatal(err)
+		}
+		if time.Since(start) != 1750*time.Millisecond {
+			t.Fatal("window reset lost an unexpired attempt")
+		}
+	})
+}
+
+func TestRateLimitedTransportCancellationDoesNotSend(t *testing.T) {
+	synctest.Test(t, func(t *testing.T) {
+		l := &RequestLimiter{}
+		if err := l.configure(0, 1); err != nil {
+			t.Fatal(err)
+		}
+		calls := 0
+		tr := rateLimitedTransport{limiter: l, base: roundTripperFunc(func(*http.Request) (*http.Response, error) {
+			calls++
+			return &http.Response{StatusCode: 200, Body: http.NoBody}, nil
+		})}
+		req, err := http.NewRequest("POST", "http://example.invalid", nil)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if _, err := tr.RoundTrip(req); err != nil {
+			t.Fatal(err)
+		}
+		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+		defer cancel()
+		if _, err := tr.RoundTrip(req.WithContext(ctx)); !errors.Is(err, context.DeadlineExceeded) {
+			t.Fatalf("got %v", err)
+		}
+		if calls != 1 {
+			t.Fatalf("cancelled waiter sent HTTP request: %d", calls)
+		}
+	})
+}
+
+func TestProviderRateLimitPublicFactory(t *testing.T) {
+	for _, format := range []string{"openai", "anthropic", "openai-responses"} {
+		for _, streaming := range []bool{false, true} {
+			mode := "complete"
+			if streaming {
+				mode = "stream"
+			}
+			t.Run(format+"/"+mode, func(t *testing.T) {
+				var requests atomic.Int32
+				upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+					requests.Add(1)
+					w.WriteHeader(http.StatusInternalServerError)
+					io.WriteString(w, `{"error":{"message":"temporary error"}}`)
+				}))
+				defer upstream.Close()
+				cfg := ConfigFrom(format, "test-model", upstream.URL, "test-key", "")
+				cfg.RatePerSecond, cfg.RatePerMinute = 1, 1
+				cfg.Retry.ConnectAttempts = -1
+				limiter := &RequestLimiter{}
+				req := llm.CompletionRequest{Messages: []llm.Message{llm.UserText("ping")}}
+				for i := 0; i < 2; i++ {
+					// A new public provider factory call must keep the old quota.
+					prov, err := cfg.NewProviderWithLimiter(limiter)
+					if err != nil {
+						t.Fatal(err)
+					}
+					ctx := context.Background()
+					if i == 1 {
+						var cancel context.CancelFunc
+						ctx, cancel = context.WithTimeout(ctx, 50*time.Millisecond)
+						defer cancel()
+					}
+					if streaming {
+						for _, streamErr := range prov.Stream(ctx, req) {
+							err = streamErr
+						}
+					} else {
+						_, _, _, err = prov.Complete(ctx, req)
+					}
+					if err == nil {
+						t.Fatal("expected upstream error or quota wait cancellation")
+					}
+					if i == 1 && !errors.Is(err, context.DeadlineExceeded) {
+						t.Fatalf("got %v, want quota wait deadline", err)
+					}
+				}
+				if got := requests.Load(); got != 1 {
+					t.Fatalf("provider rebuild sent %d HTTP requests under 1/min", got)
+				}
+			})
+		}
+	}
+}

intercept/intercept.go

@@ -305,6 +305,8 @@ const (
 	settingJudgeProfileID        = "llm_judge_profile_id"
 	settingJudgePrompt           = "llm_judge_prompt"
 	settingJudgeTimeoutSecs      = "llm_judge_timeout_seconds"
+	settingJudgeRetryCount       = "llm_judge_retry_count"
+	settingJudgeRetryIntervalMS  = "llm_judge_retry_interval_ms"
 	settingJudgeFailAction       = "llm_judge_fail_action"
 	settingJudgeAskTimeoutSecs   = "llm_judge_ask_timeout_seconds"
 	settingJudgeAskTimeoutAction = "llm_judge_ask_timeout_action"
@@ -327,7 +329,9 @@ type JudgeConfig struct {
 	Enabled                bool   `json:"enabled"`
 	ProfileID              int64  `json:"profile_id"` // 0 = follow active/default
 	Prompt                 string `json:"prompt"`
-	TimeoutSeconds         int    `json:"timeout_seconds"`
+	TimeoutSeconds         int    `json:"timeout_seconds"` // per attempt
+	RetryCount             int    `json:"retry_count"`     // additional attempts; 0 = disabled
+	RetryIntervalMS        int    `json:"retry_interval_ms"`
 	FailAction             string `json:"fail_action"` // allow|ask|deny
 	AskTimeoutSeconds      int    `json:"ask_timeout_seconds"`
 	AskTimeoutAction       string `json:"ask_timeout_action"`       // allow|deny
@@ -342,6 +346,8 @@ func (i *Interceptor) judgeConfig() JudgeConfig {
 		Enabled:                i.db.GetBool(settingJudgeEnabled, false),
 		ProfileID:              int64(i.getSettingInt(settingJudgeProfileID, 0)),
 		TimeoutSeconds:         i.getSettingInt(settingJudgeTimeoutSecs, defaultJudgeTimeoutSecs),
+		RetryCount:             min(max(i.getSettingInt(settingJudgeRetryCount, 0), 0), MaxJudgeRetryCount),
+		RetryIntervalMS:        min(max(i.getSettingInt(settingJudgeRetryIntervalMS, defaultJudgeRetryIntervalMS), 0), MaxJudgeRetryIntervalMS),
 		FailAction:             i.getSettingChoice(settingJudgeFailAction, defaultJudgeFailAction, "allow", "ask", "deny"),
 		AskTimeoutSeconds:      i.getSettingInt(settingJudgeAskTimeoutSecs, defaultJudgeAskTimeoutSecs),
 		AskTimeoutAction:       i.getSettingChoice(settingJudgeAskTimeoutAction, defaultJudgeAskTimeoutAction, "allow", "deny"),
@@ -395,6 +401,9 @@ func (i *Interceptor) GetJudgeConfig() JudgeConfig { return i.judgeConfig() }
 // SetJudgeConfig persists the judge configuration. An empty Prompt clears the
 // override (the built-in template is used again).
 func (i *Interceptor) SetJudgeConfig(c JudgeConfig) error {
+	if err := c.ValidateRetry(); err != nil {
+		return err
+	}
 	if err := i.db.SetBool(settingJudgeEnabled, c.Enabled); err != nil {
 		return err
 	}
@@ -413,6 +422,12 @@ func (i *Interceptor) SetJudgeConfig(c JudgeConfig) error {
 	if err := i.db.SetSetting(settingJudgeTimeoutSecs, strconv.Itoa(c.TimeoutSeconds)); err != nil {
 		return err
 	}
+	if err := i.db.SetSetting(settingJudgeRetryCount, strconv.Itoa(c.RetryCount)); err != nil {
+		return err
+	}
+	if err := i.db.SetSetting(settingJudgeRetryIntervalMS, strconv.Itoa(c.RetryIntervalMS)); err != nil {
+		return err
+	}
 	if err := i.db.SetSetting(settingJudgeFailAction, c.FailAction); err != nil {
 		return err
 	}
@@ -432,7 +447,8 @@ func (i *Interceptor) SetJudgeConfig(c JudgeConfig) error {
 // returns (Decision, true) when the judge produced a terminal verdict, or
 // (Decision{}, false) when the fallback is disabled or not wired (caller then
 // keeps the current behavior: allow). On model error or an unparseable reply it
-// falls back to the configured FailAction. Ask verdicts carry the human-approval
+// retries up to RetryCount times, then falls back to the configured FailAction.
+// Ask verdicts carry the human-approval
 // timeout so the existing HandleAsk consumes them unchanged.
 func (i *Interceptor) Judge(ctx context.Context, tool string, arguments json.RawMessage) (Decision, bool) {
 	cfg := i.judgeConfig()
@@ -444,13 +460,13 @@ func (i *Interceptor) Judge(ctx context.Context, tool string, arguments json.Raw
 	}
 
 	cctx := ctx
+	cancel := func() {}
 	if cfg.TimeoutSeconds > 0 {
-		var cancel context.CancelFunc
 		cctx, cancel = context.WithTimeout(ctx, time.Duration(cfg.TimeoutSeconds)*time.Second)
-		defer cancel()
 	}
 
 	input, contextErr := BuildReviewInput(cctx, tool, arguments)
+	cancel() // Capture once; each model attempt gets its own timeout below.
 	cfg.Prompt = EffectiveJudgePrompt(cfg.Prompt)
 	var out Decision
 	var err error
@@ -461,7 +477,7 @@ func (i *Interceptor) Judge(ctx context.Context, tool string, arguments json.Raw
 		out = Decision{Action: "ask", ModelFallback: true, Message: "审查上下文不完整，需要人工确认：" + contextErr.Error()}
 	} else {
 		modelInput, _ = json.Marshal(input)
-		out, err = rv(cctx, cfg.ProfileID, cfg.Prompt, input)
+		out, err = reviewWithRetry(ctx, cfg, rv, input)
 	}
 	if err != nil {
 		out = Decision{ProfileID: out.ProfileID, ModelFallback: true, Action: cfg.FailAction, Message: "模型审批失败,按失败策略处理: " + err.Error()}

intercept/judge_retry.go

@@ -0,0 +1,67 @@
+package intercept
+
+import (
+	"context"
+	"fmt"
+	"time"
+)
+
+const (
+	MaxJudgeRetryCount          = 20
+	MaxJudgeRetryIntervalMS     = 3600_000
+	defaultJudgeRetryIntervalMS = 1000
+)
+
+// ValidateRetry rejects invalid settings before any judge configuration is saved.
+func (c JudgeConfig) ValidateRetry() error {
+	if c.RetryCount < 0 || c.RetryCount > MaxJudgeRetryCount {
+		return fmt.Errorf("retry_count 必须在 0 到 %d 之间", MaxJudgeRetryCount)
+	}
+	if c.RetryIntervalMS < 0 || c.RetryIntervalMS > MaxJudgeRetryIntervalMS {
+		return fmt.Errorf("retry_interval_ms 必须在 0 到 %d 之间", MaxJudgeRetryIntervalMS)
+	}
+	return nil
+}
+
+// reviewWithRetry retries failed classifications, including invalid verdicts.
+// The captured input stays identical across attempts. A valid deny/ask verdict
+// is final, and parent cancellation interrupts both requests and the retry wait.
+// Provider-level connection/empty-response retries still run within each attempt.
+func reviewWithRetry(ctx context.Context, cfg JudgeConfig, rv Reviewer, input ReviewInput) (out Decision, err error) {
+	for attempt := 0; attempt <= cfg.RetryCount; attempt++ {
+		if ctx.Err() != nil {
+			return out, ctx.Err()
+		}
+		if attempt > 0 && cfg.RetryIntervalMS > 0 {
+			timer := time.NewTimer(time.Duration(cfg.RetryIntervalMS) * time.Millisecond)
+			select {
+			case <-ctx.Done():
+				timer.Stop()
+				return out, ctx.Err()
+			case <-timer.C:
+			}
+		}
+		if ctx.Err() != nil {
+			return out, ctx.Err()
+		}
+		attemptCtx := ctx
+		cancel := func() {}
+		if cfg.TimeoutSeconds > 0 {
+			attemptCtx, cancel = context.WithTimeout(ctx, time.Duration(cfg.TimeoutSeconds)*time.Second)
+		}
+		out, err = rv(attemptCtx, cfg.ProfileID, cfg.Prompt, input)
+		if attemptCtx.Err() != nil {
+			err = attemptCtx.Err()
+		}
+		cancel()
+		if err == nil {
+			switch out.Action {
+			case "allow", "ask", "deny":
+				return out, nil
+			default:
+				err = fmt.Errorf("模型输出无法解析")
+			}
+		}
+	}
+	return out, err
+}

intercept/judge_retry_test.go

@@ -0,0 +1,247 @@
+package intercept
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"errors"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/Autumn-27/artex/db"
+)
+
+func retryTestInterceptor(t *testing.T, settings map[string]string) *Interceptor {
+	t.Helper()
+	d, err := sql.Open("sqlite", ":memory:")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = d.Close() })
+	if _, err := d.Exec(`CREATE TABLE settings (key TEXT PRIMARY KEY, value TEXT)`); err != nil {
+		t.Fatal(err)
+	}
+	for key, value := range settings {
+		if _, err := d.Exec(`INSERT INTO settings(key,value) VALUES (?,?)`, key, value); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return New(&db.DB{DB: d})
+}
+
+func TestJudgeRetryConfigDefaultsAndBounds(t *testing.T) {
+	for _, tc := range []struct {
+		name            string
+		settings        map[string]string
+		count, interval int
+	}{
+		{"legacy", nil, 0, 1000},
+		{"configured", map[string]string{settingJudgeRetryCount: "3", settingJudgeRetryIntervalMS: "25"}, 3, 25},
+		{"immediate", map[string]string{settingJudgeRetryCount: "1", settingJudgeRetryIntervalMS: "0"}, 1, 0},
+		{"malformed", map[string]string{settingJudgeRetryCount: "bad", settingJudgeRetryIntervalMS: "bad"}, 0, 1000},
+		{"negative", map[string]string{settingJudgeRetryCount: "-1", settingJudgeRetryIntervalMS: "-1"}, 0, 0},
+		{"excessive", map[string]string{settingJudgeRetryCount: "999", settingJudgeRetryIntervalMS: "999999999"}, 20, 3600000},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			cfg := retryTestInterceptor(t, tc.settings).GetJudgeConfig()
+			if cfg.RetryCount != tc.count || cfg.RetryIntervalMS != tc.interval {
+				t.Fatalf("config=%+v", cfg)
+			}
+		})
+	}
+}
+
+func TestJudgeRetriesBeforeApplyingFailurePolicy(t *testing.T) {
+	for _, tc := range []struct {
+		name, failAction, verdict string
+		failures, wantCalls       int
+		wantAction                string
+		wantFallback              bool
+	}{
+		{"recover", "deny", "allow", 2, 3, "allow", false},
+		{"valid_deny_is_final", "allow", "deny", 0, 1, "deny", false},
+		{"valid_ask_is_final", "allow", "ask", 0, 1, "ask", false},
+		{"exhaust_allow", "allow", "allow", 9, 3, "allow", true},
+		{"exhaust_ask", "ask", "allow", 9, 3, "ask", true},
+		{"exhaust_deny", "deny", "allow", 9, 3, "deny", true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			i := retryTestInterceptor(t, map[string]string{
+				settingJudgeEnabled: "true", settingJudgeRetryCount: "2", settingJudgeRetryIntervalMS: "0", settingJudgeFailAction: tc.failAction,
+			})
+			calls := 0
+			var first ReviewInput
+			i.SetReviewer(func(ctx context.Context, id int64, prompt string, input ReviewInput) (Decision, error) {
+				calls++
+				if ctx.Err() != nil || prompt == "" {
+					t.Fatal("invalid request context")
+				}
+				if calls == 1 {
+					first = input
+				} else if !reflect.DeepEqual(input, first) {
+					t.Fatal("review input changed on retry")
+				}
+				out := Decision{Action: tc.verdict, ProfileID: 42, Message: "review result"}
+				if calls <= tc.failures {
+					return out, errors.New("temporary model failure")
+				}
+				return out, nil
+			})
+			got, ok := i.Judge(t.Context(), "Bash", json.RawMessage(`{"command":"echo ok"}`))
+			if !ok || calls != tc.wantCalls || got.Action != tc.wantAction || got.ModelFallback != tc.wantFallback || got.ProfileID != 42 {
+				t.Fatalf("calls=%d decision=%+v ok=%v", calls, got, ok)
+			}
+			if got.RuleID != 0 || len(got.ModelInput) == 0 || got.ModelInputDigest == "" || got.ConfigDigest == "" {
+				t.Fatalf("missing audit data: %+v", got)
+			}
+			if got.Action == "ask" && (!got.TimeoutEnabled || got.TimeoutSeconds != 300 || got.TimeoutAction != "deny") {
+				t.Fatal("human approval timeout lost")
+			}
+			if tc.wantFallback && !strings.Contains(got.Message, "temporary model failure") {
+				t.Fatal("last failure lost")
+			}
+		})
+	}
+}
+
+func TestJudgeRetriesInvalidVerdict(t *testing.T) {
+	for _, invalid := range []string{"", "invalid"} {
+		calls := 0
+		out, err := reviewWithRetry(t.Context(), JudgeConfig{RetryCount: 1}, func(context.Context, int64, string, ReviewInput) (Decision, error) {
+			calls++
+			if calls == 1 {
+				return Decision{Action: invalid}, nil
+			}
+			return Decision{Action: "deny"}, nil
+		}, ReviewInput{})
+		if err != nil || out.Action != "deny" || calls != 2 {
+			t.Fatalf("calls=%d out=%+v err=%v", calls, out, err)
+		}
+	}
+}
+
+func TestJudgeLegacyConfigDoesNotRetry(t *testing.T) {
+	i := retryTestInterceptor(t, map[string]string{settingJudgeEnabled: "true", settingJudgeFailAction: "deny"})
+	calls := 0
+	i.SetReviewer(func(context.Context, int64, string, ReviewInput) (Decision, error) {
+		calls++
+		return Decision{}, errors.New("failed")
+	})
+	got, _ := i.Judge(t.Context(), "Bash", json.RawMessage(`{"command":"echo ok"}`))
+	if calls != 1 || got.Action != "deny" || !got.ModelFallback {
+		t.Fatalf("calls=%d out=%+v", calls, got)
+	}
+}
+
+func TestJudgeRetryTimeoutGetsFreshContext(t *testing.T) {
+	calls := 0
+	var previous context.Context
+	out, err := reviewWithRetry(t.Context(), JudgeConfig{RetryCount: 1, TimeoutSeconds: 1}, func(ctx context.Context, _ int64, _ string, _ ReviewInput) (Decision, error) {
+		calls++
+		if calls == 1 {
+			previous = ctx
+			<-ctx.Done()
+			return Decision{}, ctx.Err()
+		}
+		if previous.Err() != context.DeadlineExceeded || ctx.Err() != nil {
+			t.Fatal("retry reused expired context")
+		}
+		if deadline, ok := ctx.Deadline(); !ok || time.Until(deadline) < 500*time.Millisecond {
+			t.Fatal("retry has no fresh timeout")
+		}
+		return Decision{Action: "allow"}, nil
+	}, ReviewInput{})
+	if calls != 2 || err != nil || out.Action != "allow" {
+		t.Fatalf("calls=%d out=%+v err=%v", calls, out, err)
+	}
+	if previous.Err() == nil {
+		t.Fatal("attempt context was not released")
+	}
+}
+
+func TestJudgeRetryCancellationInterruptsWait(t *testing.T) {
+	ctx, cancel := context.WithCancel(t.Context())
+	defer cancel()
+	started := make(chan struct{})
+	done := make(chan error, 1)
+	calls := 0
+	go func() {
+		_, err := reviewWithRetry(ctx, JudgeConfig{RetryCount: 20, RetryIntervalMS: 3600000}, func(context.Context, int64, string, ReviewInput) (Decision, error) {
+			calls++
+			close(started)
+			return Decision{}, errors.New("failed")
+		}, ReviewInput{})
+		done <- err
+	}()
+	<-started
+	cancel()
+	select {
+	case err := <-done:
+		if !errors.Is(err, context.Canceled) || calls != 1 {
+			t.Fatalf("calls=%d err=%v", calls, err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("cancellation did not interrupt retry wait")
+	}
+}
+
+func TestJudgeRetryDoesNotCallCanceledOrDisabledReviewer(t *testing.T) {
+	ctx, cancel := context.WithCancel(t.Context())
+	cancel()
+	_, err := reviewWithRetry(ctx, JudgeConfig{RetryCount: 2}, func(context.Context, int64, string, ReviewInput) (Decision, error) {
+		t.Fatal("called after cancellation")
+		return Decision{}, nil
+	}, ReviewInput{})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatal(err)
+	}
+	i := retryTestInterceptor(t, nil)
+	i.SetReviewer(func(context.Context, int64, string, ReviewInput) (Decision, error) {
+		t.Fatal("called disabled reviewer")
+		return Decision{}, nil
+	})
+	if _, ok := i.Judge(t.Context(), "Bash", json.RawMessage(`{}`)); ok {
+		t.Fatal("disabled judge returned verdict")
+	}
+}
+
+func TestJudgeRetryWaitsBetweenFailures(t *testing.T) {
+	var first time.Time
+	calls := 0
+	_, err := reviewWithRetry(t.Context(), JudgeConfig{RetryCount: 1, RetryIntervalMS: 20}, func(context.Context, int64, string, ReviewInput) (Decision, error) {
+		calls++
+		if calls == 1 {
+			first = time.Now()
+			return Decision{}, errors.New("failed")
+		}
+		if time.Since(first) < 20*time.Millisecond {
+			t.Fatal("retry skipped configured interval")
+		}
+		return Decision{Action: "allow"}, nil
+	}, ReviewInput{})
+	if err != nil || calls != 2 {
+		t.Fatalf("calls=%d err=%v", calls, err)
+	}
+}
+
+func TestJudgeInvalidArgumentsDoNotCallReviewer(t *testing.T) {
+	i := retryTestInterceptor(t, map[string]string{settingJudgeEnabled: "true", settingJudgeRetryCount: "2"})
+	i.SetReviewer(func(context.Context, int64, string, ReviewInput) (Decision, error) {
+		t.Fatal("invalid input must go straight to human review")
+		return Decision{}, nil
+	})
+	got, ok := i.Judge(t.Context(), "Bash", json.RawMessage(`{broken`))
+	if !ok || got.Action != "ask" || !got.ModelFallback {
+		t.Fatalf("out=%+v ok=%v", got, ok)
+	}
+}
+
+func TestJudgeRejectsInvalidRetryBeforeWritingSettings(t *testing.T) {
+	for _, cfg := range []JudgeConfig{{RetryCount: -1}, {RetryCount: 21}, {RetryIntervalMS: -1}, {RetryIntervalMS: 3600001}} {
+		if err := New(nil).SetJudgeConfig(cfg); err == nil {
+			t.Fatalf("accepted invalid retry: %+v", cfg)
+		}
+	}
+}

server/intercept.go

@@ -474,6 +474,10 @@ func (s *Server) interceptSetJudgeConfig(w http.ResponseWriter, r *http.Request)
 		writeErr(w, 400, "incomplete_script_action 必须是 ask 或 model")
 		return
 	}
+	if err := req.ValidateRetry(); err != nil {
+		writeErr(w, 400, err.Error())
+		return
+	}
 	if err := s.m.interceptor.SetJudgeConfig(req); err != nil {
 		writeErr(w, 500, err.Error())
 		return

server/intercept_config_test.go

@@ -1,13 +1,18 @@
 package server
 
 import (
+	"database/sql"
+	"database/sql/driver"
+	"encoding/json"
 	"net/http/httptest"
 	"os"
 	"strings"
+	"sync"
 	"testing"
 
 	"github.com/Autumn-27/artex/db"
 	"github.com/Autumn-27/artex/intercept"
+	"modernc.org/sqlite"
 )
 
 func TestJudgeConfigIncompleteScriptDefaultAndOverride(t *testing.T) {
@@ -20,7 +25,7 @@ func TestJudgeConfigIncompleteScriptDefaultAndOverride(t *testing.T) {
 		t.Fatal(err)
 	}
 	t.Cleanup(func() { _ = pg.Close() })
-	keys := []string{"llm_judge_enabled", "llm_judge_profile_id", "llm_judge_prompt", "llm_judge_timeout_seconds", "llm_judge_fail_action", "llm_judge_ask_timeout_seconds", "llm_judge_ask_timeout_action", "llm_judge_incomplete_script_action"}
+	keys := []string{"llm_judge_enabled", "llm_judge_profile_id", "llm_judge_prompt", "llm_judge_timeout_seconds", "llm_judge_retry_count", "llm_judge_retry_interval_ms", "llm_judge_fail_action", "llm_judge_ask_timeout_seconds", "llm_judge_ask_timeout_action", "llm_judge_incomplete_script_action"}
 	for _, key := range keys {
 		value, exists, err := pg.GetSetting(key)
 		if err != nil {
@@ -47,14 +52,80 @@ func TestJudgeConfigIncompleteScriptDefaultAndOverride(t *testing.T) {
 		{`,"incomplete_script_action":"model"`, "model", 200},
 		{`,"incomplete_script_action":"invalid"`, "model", 400},
 	} {
-		req := httptest.NewRequest("PUT", "/api/intercept/judge-config", strings.NewReader(`{"fail_action":"allow","ask_timeout_action":"deny"`+tc.option+`}`))
+		req := httptest.NewRequest("PUT", "/api/intercept/judge-config", strings.NewReader(`{"fail_action":"allow","ask_timeout_action":"deny","retry_count":2,"retry_interval_ms":17`+tc.option+`}`))
 		w := httptest.NewRecorder()
 		s.interceptSetJudgeConfig(w, req)
 		if w.Code != tc.code {
 			t.Fatalf("option=%s status=%d body=%s", tc.option, w.Code, w.Body.String())
 		}
+		if cfg := s.m.interceptor.GetJudgeConfig(); cfg.RetryCount != 2 || cfg.RetryIntervalMS != 17 {
+			t.Fatalf("retry configuration not persisted: %+v", cfg)
+		}
 		if got := s.m.interceptor.GetJudgeConfig().IncompleteScriptAction; got != tc.want {
 			t.Fatalf("option=%s action=%s want=%s", tc.option, got, tc.want)
 		}
 	}
 }
+
+func TestJudgeConfigRejectsInvalidRetryBeforeSaving(t *testing.T) {
+	// No manager is needed: invalid input must be rejected before database writes.
+	s := &Server{}
+	for _, fields := range []string{
+		`"retry_count":-1`, `"retry_count":21`, `"retry_count":1.5`,
+		`"retry_interval_ms":-1`, `"retry_interval_ms":3600001`, `"retry_interval_ms":1.5`,
+	} {
+		t.Run(fields, func(t *testing.T) {
+			req := httptest.NewRequest("PUT", "/api/intercept/judge", strings.NewReader(`{"fail_action":"allow","ask_timeout_action":"deny",`+fields+`}`))
+			w := httptest.NewRecorder()
+			s.interceptSetJudgeConfig(w, req)
+			if w.Code != 400 {
+				t.Fatalf("status=%d body=%s", w.Code, w.Body.String())
+			}
+		})
+	}
+}
+
+var judgeConfigSQLiteNow sync.Once
+
+func TestJudgeConfigRetryRoundTrip(t *testing.T) {
+	// Exercise production settings SQL without requiring an external PG service.
+	judgeConfigSQLiteNow.Do(func() {
+		sqlite.MustRegisterDeterministicScalarFunction("now", 0, func(*sqlite.FunctionContext, []driver.Value) (driver.Value, error) {
+			return "2026-01-01 00:00:00", nil
+		})
+	})
+	sqldb, err := sql.Open("sqlite", ":memory:")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { _ = sqldb.Close() })
+	if _, err := sqldb.Exec(`CREATE TABLE settings (key TEXT PRIMARY KEY, value TEXT, updated_at TEXT)`); err != nil {
+		t.Fatal(err)
+	}
+	pg := &db.DB{DB: sqldb}
+	s := &Server{m: &Manager{pg: pg, interceptor: intercept.New(pg)}}
+	for _, tc := range []struct{ count, interval int }{{2, 17}, {20, 3600000}, {0, 0}} {
+		cfg := s.m.interceptor.GetJudgeConfig()
+		cfg.RetryCount, cfg.RetryIntervalMS = tc.count, tc.interval
+		body, err := json.Marshal(cfg)
+		if err != nil {
+			t.Fatal(err)
+		}
+		w := httptest.NewRecorder()
+		s.interceptSetJudgeConfig(w, httptest.NewRequest("PUT", "/api/intercept/judge", strings.NewReader(string(body))))
+		if w.Code != 200 {
+			t.Fatalf("status=%d body=%s", w.Code, w.Body.String())
+		}
+		// Construct a fresh interceptor to prove persistence, rather than cache reuse.
+		s.m.interceptor = intercept.New(pg)
+		w = httptest.NewRecorder()
+		s.interceptGetJudgeConfig(w, httptest.NewRequest("GET", "/api/intercept/judge", nil))
+		var got intercept.JudgeConfig
+		if err := json.Unmarshal(w.Body.Bytes(), &got); err != nil {
+			t.Fatal(err)
+		}
+		if got.RetryCount != tc.count || got.RetryIntervalMS != tc.interval {
+			t.Fatalf("saved=%+v got=%+v", cfg, got)
+		}
+	}
+}

server/request_limiter_test.go

@@ -0,0 +1,15 @@
+package server
+
+import "testing"
+
+func TestProfileRequestLimiterSurvivesCacheInvalidation(t *testing.T) {
+	s := &Server{}
+	limiter := s.requestLimiterForProfile(1)
+	if limiter == s.requestLimiterForProfile(2) || limiter == s.requestLimiterForProfile(0) {
+		t.Fatal("different profiles must have independent quotas")
+	}
+	s.invalidateProfileAgents()
+	if got := s.requestLimiterForProfile(1); got != limiter {
+		t.Fatal("clearing providers reset the profile's rate limiter")
+	}
+}

server/server.go

@@ -103,6 +103,9 @@ type Server struct {
 	provCacheMu   sync.Mutex
 	provByProfile map[int64]*provEntry
 	provCacheGen  uint64
+	// Quota histories survive provider/agent cache invalidation. Key 0 belongs
+	// to the non-persisted global config; persisted profiles use their DB IDs.
+	requestLimiters map[int64]*agent.RequestLimiter
 
 	// llmHealth is the process-wide circuit-breaker state for LLM failover (轮询).
 	// It deliberately lives OUTSIDE the provider caches: rebuilding the chain
@@ -446,7 +449,7 @@ func (s *Server) applyLLM(cfg agent.Config) error {
 	}
 	if prov == nil {
 		var err error
-		prov, err = cfg.NewProvider()
+		prov, err = cfg.NewProviderWithLimiter(s.requestLimiterForProfile(0))
 		if err != nil {
 			return err
 		}
@@ -581,31 +584,51 @@ func (s *Server) providerForProfile(id int64) (llm.Provider, agent.Config, bool)
 	if !ok {
 		return nil, agent.Config{}, false
 	}
-	prov, err := cfg.NewProvider()
-	if err != nil {
-		log.Printf("[engine] build provider for LLM profile %d failed: %v", id, err)
-		return nil, agent.Config{}, false
-	}
-	// Wrap with recorder, tagged with this profile's name.
-	if p, _ := s.m.pg.ProfileByID(id); p != nil {
-		prov = llmrec.Wrap(prov, s.m.PG(), cfg.Model, p.Name, cfg.ThinkingType, cfg.ReasoningEffort, s.m.LLMRecordEnabled)
-		prov = bindSideProvider(prov, cfg, id, p.Name)
-	}
+	profile, _ := s.m.pg.ProfileByID(id)
 	s.provCacheMu.Lock()
 	if generation != s.provCacheGen {
 		s.provCacheMu.Unlock()
 		return s.providerForProfile(id)
 	}
 	if e := s.provByProfile[id]; e != nil { // lost the race → keep the winner
-		prov, cfg = e.prov, e.cfg
-	} else {
-		s.provByProfile[id] = &provEntry{prov: prov, cfg: cfg}
-		log.Printf("[engine] built provider for LLM profile %d (%s / %s)", id, cfg.Provider(), cfg.Model)
+		s.provCacheMu.Unlock()
+		return e.prov, e.cfg, true
 	}
+	// Building the HTTP client does not perform network I/O. Keep this step
+	// under the generation lock so a stale build cannot overwrite the retained
+	// limiter's settings after a newer configuration has already been installed.
+	prov, err := cfg.NewProviderWithLimiter(s.requestLimiterForProfileLocked(id))
+	if err != nil {
+		s.provCacheMu.Unlock()
+		log.Printf("[engine] build provider for LLM profile %d failed: %v", id, err)
+		return nil, agent.Config{}, false
+	}
+	if profile != nil {
+		prov = llmrec.Wrap(prov, s.m.PG(), cfg.Model, profile.Name, cfg.ThinkingType, cfg.ReasoningEffort, s.m.LLMRecordEnabled)
+		prov = bindSideProvider(prov, cfg, id, profile.Name)
+	}
+	s.provByProfile[id] = &provEntry{prov: prov, cfg: cfg}
+	log.Printf("[engine] built provider for LLM profile %d (%s / %s)", id, cfg.Provider(), cfg.Model)
 	s.provCacheMu.Unlock()
 	return prov, cfg, true
 }
 
+func (s *Server) requestLimiterForProfile(id int64) *agent.RequestLimiter {
+	s.provCacheMu.Lock()
+	defer s.provCacheMu.Unlock()
+	return s.requestLimiterForProfileLocked(id)
+}
+
+func (s *Server) requestLimiterForProfileLocked(id int64) *agent.RequestLimiter {
+	if s.requestLimiters == nil {
+		s.requestLimiters = make(map[int64]*agent.RequestLimiter)
+	}
+	if s.requestLimiters[id] == nil {
+		s.requestLimiters[id] = &agent.RequestLimiter{}
+	}
+	return s.requestLimiters[id]
+}
+
 // chatAgentForProfile returns a ChatAgent built from a specific LLM profile, cached
 // per profile id. Returns nil if the profile is missing or has no API key.
 func (s *Server) chatAgentForProfile(id int64) *agent.ChatAgent {

web/src/app/(main)/function/tasks/page.tsx

@@ -1222,7 +1222,8 @@ const TaskRow = React.memo(function TaskRow({
       </TableCell>
       <TableCell className="text-center text-xs tabular-nums">
         {(() => {
-          const f = task.distinct_findings ?? task.findings;
+          // Keep report severity counts visible while a merged folder awaits its own assessment.
+          const f = task.findings ?? task.distinct_findings;
           const total = f ? f.critical + f.high + f.medium + f.low : 0;
           if (!f || (total === 0 && !task.distinct_findings?.total))
             return <span className="text-muted-foreground">0</span>;
@@ -1230,7 +1231,7 @@ const TaskRow = React.memo(function TaskRow({
           return (
             <span
               className="font-medium whitespace-nowrap"
-              title={`独立漏洞：严重 / 高 / 中 / 低；上报 ${task.distinct_findings?.reports ?? total} 条；待评估 ${task.distinct_findings?.unassessed ?? 0} 个`}
+              title={`${task.findings ? "原始上报" : "独立漏洞"}：严重 / 高 / 中 / 低；上报 ${task.distinct_findings?.reports ?? total} 条；待评估 ${task.distinct_findings?.unassessed ?? 0} 个`}
             >
               {seg(f.critical, "text-rose-600 dark:text-rose-400")}
               <span className="text-muted-foreground">/</span>

web/src/app/(main)/system/intercept/page.tsx

@@ -21,6 +21,7 @@ import {
   DialogTitle,
 } from "@/components/ui/dialog";
 import { Input } from "@/components/ui/input";
+import { Field as FormField, FieldDescription, FieldGroup, FieldLabel } from "@/components/ui/field";
 import { Label } from "@/components/ui/label";
 import { Separator } from "@/components/ui/separator";
 import { Switch } from "@/components/ui/switch";
@@ -200,6 +201,8 @@ const defaultJudge = (): JudgeConfig => ({
   profile_id: 0,
   prompt: "",
   timeout_seconds: 15,
+  retry_count: 0,
+  retry_interval_ms: 1000,
   fail_action: "allow",
   ask_timeout_seconds: 300,
   ask_timeout_action: "deny",
@@ -207,7 +210,7 @@ const defaultJudge = (): JudgeConfig => ({
 });
 
 function JudgeCard() {
-  const [cfg, setCfg] = React.useState<JudgeConfig>(defaultJudge());
+  const [cfg, setCfg] = React.useState<JudgeConfig>(defaultJudge);
   const [profiles, setProfiles] = React.useState<LLMProfile[]>([]);
   const [loading, setLoading] = React.useState(true);
   const [saving, setSaving] = React.useState(false);
@@ -337,7 +340,7 @@ function JudgeCard() {
         <div className="grid gap-4 lg:grid-cols-5">
           {/* 左:提示词编辑器(直接展开,主区域) */}
           <Card className="lg:col-span-3">
-            <CardContent className="flex h-full flex-col gap-2 p-4">
+            <CardContent className="flex min-h-0 flex-1 flex-col gap-2 p-4">
               <div className="flex items-center justify-between">
                 <div>
                   <p className="text-sm font-medium">审批提示词</p>
@@ -348,7 +351,7 @@ function JudgeCard() {
                 </Button>
               </div>
               <Textarea
-                className="min-h-[22rem] flex-1 resize-none font-mono text-xs leading-relaxed"
+                className="field-sizing-fixed min-h-[22rem] max-h-[none] flex-1 resize-none font-mono text-xs leading-relaxed"
                 value={cfg.prompt}
                 onChange={(e) => patch({ prompt: e.target.value })}
                 placeholder="留空使用内置模板"
@@ -376,7 +379,7 @@ function JudgeCard() {
                     </SelectContent>
                   </Select>
                 </Field>
-                <Field label="模型判定超时（秒）">
+                <Field label="每次模型判定超时（秒）">
                   <Input
                     type="number"
                     min={1}
@@ -387,7 +390,48 @@ function JudgeCard() {
                     }}
                   />
                 </Field>
-                <Field label="模型失败时（出错 / 超时 / 无法解析）">
+                <FieldGroup>
+                  <FormField>
+                    <FieldLabel htmlFor="judge-retry-count">失败重试次数</FieldLabel>
+                    <Input
+                      id="judge-retry-count"
+                      type="number"
+                      min={0}
+                      max={20}
+                      step={1}
+                      value={cfg.retry_count}
+                      onChange={(e) => {
+                        const n = Number(e.target.value);
+                        if (Number.isInteger(n) && n >= 0 && n <= 20) patch({ retry_count: n });
+                      }}
+                      aria-describedby="judge-retry-count-hint"
+                    />
+                    <FieldDescription id="judge-retry-count-hint">
+                      首次审批失败后额外重试 0–20 次，0 表示不重试。出错、超时或裁决无法解析时重试；正常放行、转人工或拦截后立即结束。
+                    </FieldDescription>
+                  </FormField>
+                  <FormField data-disabled={cfg.retry_count === 0}>
+                    <FieldLabel htmlFor="judge-retry-interval">重试间隔（毫秒）</FieldLabel>
+                    <Input
+                      id="judge-retry-interval"
+                      type="number"
+                      min={0}
+                      max={3600000}
+                      step={1}
+                      value={cfg.retry_interval_ms}
+                      disabled={cfg.retry_count === 0}
+                      onChange={(e) => {
+                        const n = Number(e.target.value);
+                        if (Number.isInteger(n) && n >= 0 && n <= 3600000) patch({ retry_interval_ms: n });
+                      }}
+                      aria-describedby="judge-retry-interval-hint"
+                    />
+                    <FieldDescription id="judge-retry-interval-hint">
+                      默认 1000 毫秒，0 表示立即重试。每次审批均有独立超时，全部尝试失败后执行下方失败策略。模型配置中的建连和空响应重试仍在每次审批内生效。
+                    </FieldDescription>
+                  </FormField>
+                </FieldGroup>
+                <Field label="模型失败时（重试耗尽后）">
                   <Select value={cfg.fail_action} onValueChange={(v) => patch({ fail_action: v as JudgeConfig["fail_action"] })}>
                     <SelectTrigger><SelectValue /></SelectTrigger>
                     <SelectContent>

web/src/components/http-code-block.tsx

@@ -111,6 +111,40 @@ function HighlightedBody({ body, format }: { body: string; format: BodyFormat })
   return body;
 }
 
+// Empty grid rows have no text, so native selection copying drops their newlines.
+// Map the selected DOM boundaries back to the raw packet instead.
+function selectedPacketText(container: HTMLElement, value: string, lines: string[]): string | null {
+  const selection = window.getSelection();
+  if (!selection || selection.isCollapsed || selection.rangeCount !== 1) return null;
+  const range = selection.getRangeAt(0);
+  if (!container.contains(range.startContainer) || !container.contains(range.endContainer)) return null;
+
+  let start: number | null = null;
+  let end: number | null = null;
+  let offset = 0;
+  container.querySelectorAll("[data-line] code").forEach((code, index) => {
+    if (range.intersectsNode(code)) {
+      const selected = document.createRange();
+      selected.selectNodeContents(code);
+      if (range.compareBoundaryPoints(Range.START_TO_START, selected) > 0) {
+        selected.setStart(range.startContainer, range.startOffset);
+      }
+      if (range.compareBoundaryPoints(Range.END_TO_END, selected) < 0) {
+        selected.setEnd(range.endContainer, range.endOffset);
+      }
+      const prefix = document.createRange();
+      prefix.selectNodeContents(code);
+      prefix.setEnd(selected.startContainer, selected.startOffset);
+      start ??= offset + prefix.toString().length;
+      prefix.setEnd(selected.endContainer, selected.endOffset);
+      end = offset + prefix.toString().length;
+    }
+    const lineEnd = offset + lines[index].length;
+    offset = lineEnd + (value[lineEnd] === "\r" ? 2 : 1);
+  });
+  return start !== null && end !== null ? value.slice(start, end) : null;
+}
+
 export function HttpCodeBlock({ raw }: { raw: string }) {
   const [wrapLines, setWrapLines] = React.useState(true);
   const [copied, setCopied] = React.useState(false);
@@ -183,6 +217,12 @@ export function HttpCodeBlock({ raw }: { raw: string }) {
         aria-readonly="true"
         tabIndex={0}
         className="max-h-[calc(100vh-15rem)] min-w-0 overflow-auto bg-background py-3 font-mono text-xs outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-ring"
+        onCopy={(event) => {
+          const text = selectedPacketText(event.currentTarget, value, lines);
+          if (text === null) return;
+          event.preventDefault();
+          event.clipboardData.setData("text/plain", text);
+        }}
         onKeyDown={(event) => {
           if (!(event.ctrlKey || event.metaKey) || event.key.toLowerCase() !== "a") return;
           event.preventDefault();

web/src/lib/mock/handler.ts

@@ -2511,6 +2511,8 @@ function route(m: string, path: string, seg: string[], q: URLSearchParams, b: Re
       profile_id: 0,
       prompt: "",
       timeout_seconds: 15,
+      retry_count: 0,
+      retry_interval_ms: 1000,
       fail_action: "allow",
       ask_timeout_seconds: 300,
       ask_timeout_action: "deny",

web/src/lib/types.ts

@@ -1361,7 +1361,9 @@ export interface JudgeConfig {
   enabled: boolean;
   profile_id: number; // 0 = 跟随激活/默认配置
   prompt: string; // 判定提示词;GET 未设置时后端回填内置模板全文
-  timeout_seconds: number; // 模型调用超时
+  timeout_seconds: number; // 每次模型审批调用的超时
+  retry_count: number; // 审批失败后的额外重试次数，0 = 不重试
+  retry_interval_ms: number; // 审批重试间隔（毫秒），0 = 立即重试
   fail_action: "allow" | "ask" | "deny"; // 模型出错/超时/不可解析时的回退
   ask_timeout_seconds: number; // 模型判 ask 转人工后的审批等待超时
   ask_timeout_action: "allow" | "deny"; // 审批超时后的默认动作