fix: add configurable model approval retries

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)
+		}
+	}
+}

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);
@@ -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/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