fix(agent): 模型输出 - 提前校验超限并有限纠正

backend/internal/app/cloud_agent_pi_model.go

@@ -125,7 +125,13 @@ func (s *Service) cloudAgentPiModel(ctx context.Context, userID, runID string, p
 		if !retryable {
 			return nil, err
 		}
-		if errors.Is(err, errCloudAgentTruncatedToolArguments) && correction == nil {
+		var outputLimitErr *cloudAgentPiOutputLimitError
+		if errors.As(err, &outputLimitErr) {
+			correction = []map[string]any{cloudAgentRuntimeMessage(cloudAgentRuntimeContext{
+				Kind: cloudAgentContextInvalidOutput, Detail: outputLimitErr.Detail,
+				MaxToolCalls: cloudAgentMaxToolCalls, MaxOutputBytes: cloudAgentMaxOutputBytes,
+			})}
+		} else if errors.Is(err, errCloudAgentTruncatedToolArguments) {
 			correction = []map[string]any{cloudAgentRuntimeMessage(cloudAgentRuntimeContext{Kind: cloudAgentContextTruncatedArguments})}
 		}
 	}
@@ -168,6 +174,15 @@ var errCloudAgentAwaitingApproval = errors.New("run is waiting for approval")
 // 调用未执行，可以带纠偏上下文重做同一步。
 var errCloudAgentTruncatedToolArguments = errors.New("truncated tool arguments")
 
+// cloudAgentPiOutputLimitError 标记整批未执行的输出超限，向有限重试提供具体纠正原因。
+type cloudAgentPiOutputLimitError struct {
+	Detail string
+}
+
+func (e *cloudAgentPiOutputLimitError) Error() string {
+	return "模型输出超出限制：" + e.Detail
+}
+
 // runCloudAgentModelStep 调度并等待一次模型步骤；第二个返回值表示失败是否值得重试。
 func (s *Service) runCloudAgentModelStep(ctx context.Context, userID, runID string, messages []map[string]any, thinkingLevel string, correction ...map[string]any) (any, bool, error) {
 	compactionSummary := cloudAgentPiCompactionSummaryRequest(messages)
@@ -326,6 +341,16 @@ func (s *Service) runCloudAgentModelStep(ctx context.Context, userID, runID stri
 	if err := json.Unmarshal([]byte(task.ResultJSON), &result); err != nil {
 		return nil, false, fmt.Errorf("decode model result: %w", err)
 	}
+	if !compactionSummary {
+		if violation := cloudAgentOutputViolation(result.Text, len(result.ToolCalls)); violation != "" {
+			// 在保存 Calls 或交给运行时执行前整批拒绝；不能截取前几项造成半批写入。
+			// 释放本步任务后沿用现有重试上限，模型调用仍正常计费，不执行任何工具。
+			if err := s.finishCloudAgentPiModelStep(userID, runID, task.ID, "", ""); err != nil {
+				return nil, false, err
+			}
+			return nil, true, &cloudAgentPiOutputLimitError{Detail: violation}
+		}
+	}
 	finishText, finishReasoning := result.Text, result.Reasoning
 	if compactionSummary {
 		finishText, finishReasoning = "", ""

backend/internal/app/cloud_agent_pi_output_limit_test.go

@@ -0,0 +1,171 @@
+package app
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"gorm.io/gorm"
+	"yingce/backend/internal/database"
+	"yingce/backend/internal/model"
+)
+
+// piOutputLimitFixture 复用现有测试库入口；PostgreSQL 只写独立临时 schema。
+// 模型结果由测试直接填入任务，不启动 Pi、HTTP 上游或真实生成。
+func piOutputLimitFixture(t *testing.T) (*Service, *gorm.DB, string) {
+	t.Helper()
+	var s *Service
+	var db *gorm.DB
+	if os.Getenv("CANVAS_TEST_POSTGRES_DSN") == "" {
+		s, db, _, _ = creationTestService(t)
+	} else {
+		s, db = newProjectNameConflictPostgresService(t)
+		s.dataDir = t.TempDir()
+		sqlDB, err := db.DB()
+		if err != nil {
+			t.Fatal(err)
+		}
+		t.Cleanup(func() { _ = sqlDB.Close() })
+		if err := db.AutoMigrate(database.Models()...); err != nil {
+			t.Fatal(err)
+		}
+		profile := mustEncodeModelCapabilityConfig(t, DefaultModelCapabilityConfigForModel("chat-completion", "text-test"))
+		for _, row := range []any{
+			&model.ModelChannel{ID: "channel", Scope: model.ChannelScopeSystem, Enabled: true},
+			&model.ChannelModel{ID: "cm", ChannelID: "channel", ModelKey: "text-test", Capability: "text", Protocol: model.ChannelInterfaceChatCompletion, CapabilityConfigJSON: profile, BillingMode: "fixed_request", UnitPriceMicrocredits: 100, PriceConfigured: true, Enabled: true},
+			&model.ChannelModelPriceTier{ID: "tier", ChannelModelID: "cm", SelectorKey: "{}", SelectorJSON: "{}", BillingMode: "fixed_request", UnitPriceMicrocredits: 100, PriceConfigured: true, Enabled: true},
+			&model.CreditAccount{UserID: "user", AvailableMicrocredits: 10000},
+		} {
+			if err := db.Create(row).Error; err != nil {
+				t.Fatal(err)
+			}
+		}
+	}
+	s.disablePiRuntime, s.legacyCloudAgentRootTask = true, false
+	if err := os.WriteFile(filepath.Join(s.dataDir, ".settings-key"), make([]byte, 32), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if err := db.Create(&model.CanvasProject{ID: "agent-canvas", UserID: "user", PayloadJSON: `{"nodes":[],"connections":[]}`}).Error; err != nil {
+		t.Fatal(err)
+	}
+	root, err := s.CreateCloudAgentRun("user", agentTestRequest(), "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return s, db, root.ID
+}
+
+// TestPiOutputLimitCorrection 验证真实模型桥的边界、纠正上下文与有限重试。
+// 超限批次不得进入 Calls 或执行；最终失败也必须清理活动任务且给出可读原因。
+func TestPiOutputLimitCorrection(t *testing.T) {
+	for _, tc := range []struct {
+		name     string
+		counts   []int
+		fail     bool
+		longText bool
+	}{
+		{"eight", []int{8}, false, false},
+		{"nine-then-eight", []int{9, 8}, false, false},
+		{"thirty-one-then-one", []int{31, 1}, false, false},
+		{"bounded-failure", []int{31, 31, 31, 31}, true, false},
+		{"oversized-text", []int{0, 1}, false, true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			s, db, runID := piOutputLimitFixture(t)
+			ctx, cancel := context.WithTimeout(context.Background(), 35*time.Second)
+			defer cancel()
+			type outcome struct {
+				result any
+				err    error
+			}
+			done := make(chan outcome, 1)
+			go func() {
+				r, err := s.cloudAgentPiModel(ctx, "user", runID, map[string]json.RawMessage{"messages": json.RawMessage(`[{"role":"user","content":"继续"}]`), "thinkingLevel": json.RawMessage(`"off"`)})
+				done <- outcome{r, err}
+			}()
+			previous := ""
+			for attempt, count := range tc.counts {
+				var task model.Task
+				for task.ID == "" {
+					_, state := agentInterjectionState(t, s, runID)
+					if state.ActiveTaskID != "" && state.ActiveTaskID != previous {
+						if len(state.Calls) != 0 {
+							t.Fatal("rejected calls entered the checkpoint")
+						}
+						if err := db.First(&task, "id = ?", state.ActiveTaskID).Error; err != nil {
+							t.Fatal(err)
+						}
+						break
+					}
+					select {
+					case got := <-done:
+						t.Fatalf("model bridge ended before expected task: %v", got.err)
+					case <-ctx.Done():
+						t.Fatal("model task not scheduled")
+					case <-time.After(20 * time.Millisecond):
+					}
+				}
+				if attempt > 0 && (!strings.Contains(task.InputJSON, "invalid_output") || !strings.Contains(task.InputJSON, "maxToolCalls")) {
+					t.Fatal("retry missing output-limit correction")
+				}
+				calls := make([]cloudAgentCall, count)
+				for i := range calls {
+					calls[i].ID = fmt.Sprintf("call-%d-%d", attempt, i)
+					calls[i].Function.Name = "canvas_get_state"
+					calls[i].Function.Arguments = "{}"
+				}
+				text := ""
+				if tc.longText && attempt == 0 {
+					text = strings.Repeat("x", cloudAgentMaxOutputBytes+1)
+				}
+				raw, err := json.Marshal(map[string]any{"text": text, "toolCalls": calls})
+				if err != nil {
+					t.Fatal(err)
+				}
+				if err := db.Model(&model.Task{}).Where("id = ?", task.ID).Updates(map[string]any{"status": model.TaskStatusSucceeded, "result_json": string(raw)}).Error; err != nil {
+					t.Fatal(err)
+				}
+				previous = task.ID
+			}
+			select {
+			case got := <-done:
+				if tc.fail {
+					if got.err == nil || !strings.Contains(got.err.Error(), "31 个工具调用") || strings.Contains(got.err.Error(), "checkpoint") {
+						t.Fatalf("wrong terminal error: %v", got.err)
+					}
+				} else if got.err != nil {
+					t.Fatal(got.err)
+				}
+			case <-ctx.Done():
+				t.Fatal("model bridge did not finish")
+			}
+			_, state := agentInterjectionState(t, s, runID)
+			want := tc.counts[len(tc.counts)-1]
+			if tc.fail {
+				want = 0
+			}
+			if state.ActiveTaskID != "" || len(state.Calls) != want || state.CallIndex != 0 {
+				t.Fatal("output admission left invalid state")
+			}
+			for _, message := range state.Canonical.Messages {
+				if strings.Contains(stringField(message, "content"), "invalid_output") {
+					t.Fatal("retry correction leaked into canonical history")
+				}
+			}
+			for _, event := range state.Events {
+				if event.Type == "tool_completed" || event.Type == "canvas_updated" || event.Type == "approval_requested" {
+					t.Fatal("model response admission executed a tool")
+				}
+			}
+			var tasks int64
+			if err := db.Model(&model.Task{}).Where("agent_run_id = ? AND operation = ?", runID, cloudAgentStepOperation).Count(&tasks).Error; err != nil || tasks != int64(len(tc.counts)) {
+				t.Fatalf("unexpected retry count: %d (%v)", tasks, err)
+			}
+		})
+	}
+}

backend/internal/prompts/agent-system-policy.md

@@ -1,6 +1,6 @@
 ---
 id: cloud-agent-system
-version: 13
+version: 14
 ---
 
 # Cloud Agent 系统行为策略
@@ -38,6 +38,8 @@ version: 13
 
 ## 输出
 
+每次模型回复的工具调用数量不得超过本轮执行上下文的 `maxToolCalls`（当前为 8）。任务需要更多调用时分批处理，收到这一批的真实执行结果后再继续；不要削减用户目标。该限制是工具调用数量，不是单个 `canvas_apply_ops` 内的操作数量，后者仍遵守该工具自己的参数契约。
+
 默认先给结论，再说明与任务有关的画布变化和必要的下一步；纯问答无需套用操作汇报模板。用户指定输出格式时按其格式回复。没有真实写入、生成或资产就绪结果时，明确说明仍未完成。
 
 ## 运行状态处理

backend/internal/prompts/agent_policy_test.go

@@ -10,12 +10,15 @@ func TestLoadAgentPoliciesUsesDocumentMetadata(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if system.ID != "cloud-agent-system" || system.Version != 13 || media.ID != "cloud-agent-media" || media.Version != 5 {
+	if system.ID != "cloud-agent-system" || system.Version != 14 || media.ID != "cloud-agent-media" || media.Version != 5 {
 		t.Fatalf("unexpected policy metadata: system=%+v media=%+v", system, media)
 	}
 	if strings.Contains(system.Text, "id: cloud-agent-system") || !strings.HasPrefix(system.Text, "# Cloud Agent") {
 		t.Fatalf("metadata leaked into compiled policy body: %q", system.Text)
 	}
+	if !strings.Contains(system.Text, "maxToolCalls") || !strings.Contains(system.Text, "收到这一批的真实执行结果后再继续") {
+		t.Fatal("system policy lost the tool-call batching instruction")
+	}
 	for _, phrase := range []string{
 		"auto 只豁免其他画布修改，不豁免媒体生成",
 		"所有权限模式都必须等待界面独立审批",

docs/content/docs/backend/http-api.mdx

@@ -5,6 +5,12 @@ description: 信封、错误码与 OpenAPI
 
 # HTTP API 合同
 
+## Agent 模型输出限制
+
+Pi 模型桥在保存待执行调用之前，复用单步输出限制校验：一次回复最多 8 个工具调用，正文最多 32000 字节。超限时整批未执行，释放本步模型任务并附带 `invalid_output` 及具体上限要求模型分批纠正；与既有异常重试共用首次失败后最多 3 次的上限，不截取部分调用执行。纠正请求仍是正常计费的文本模型任务，不创建媒体任务。
+
+系统行为策略版本升至 14，明确工具调用数量与单个 `canvas_apply_ops` 操作数量的区别。已有会话保留其冻结策略快照；部署后用新会话验证新提示词，不修改历史检查点。
+
 ## 注册邮箱等价检查
 
 公开注册与邮箱绑定的唯一性检查将 `gmail.com` 和 `googlemail.com` 视为同一域，移除本地部分的点及 `+` 后缀。例如 `a.b+x@googlemail.com` 与 `ab@gmail.com` 不允许创建两个账号。其他域保留点和 `+`；现有域名白名单规则保持不变。

docs/content/docs/progress/pending-test.mdx

@@ -5,6 +5,13 @@ description: 当前项目已实现功能的手工验证清单与未确认边界
 
 # 待测试
 
+## Pi 模型输出超限
+
+- 已修复：工具调用数量或正文超限时，在写入待执行调用之前整批拒绝，沿用有限重试传入具体纠正上下文，不执行前几个调用、不提高上限。
+- [x] PostgreSQL 临时 schema 在修改前复现“工具调用游标无效”；修改后验证 8 个调用允许、9/31 个调用纠正后恢复、持续超限最多 3 次重试及超长正文纠正。策略文档专项测试通过；未调用真实供应商。
+- [ ] 部署后在新会话验证大型任务按批执行；观察文本模型纠正次数及费用，失败时不声称已执行。
+- 本项独立于画布大事件修复，不改变审批和画布写入流程。
+
 ## 复制生图节点后节点消失
 
 - 已修复：副本不再继承原节点的 `generationEffectKeys`，避免普通保存把它误判为未确认的生成结果。图片内容、提示词和参考字段仍保留，原节点不受影响。