Add files via upload

internal/multiagent/eino_adk_run_loop.go

@@ -230,54 +230,61 @@ attemptLoop:
 				continue
 			}
 			if ev.Err != nil {
-				canRetry := attempt+1 < maxToolCallRecoveryAttempts
-
-				if canRetry && isRecoverableToolCallArgumentsJSONError(ev.Err) {
-					if logger != nil {
-						logger.Warn("eino: recoverable tool-call JSON error from model/API", zap.Error(ev.Err), zap.Int("attempt", attempt))
-					}
-					retryHints = append(retryHints, toolCallArgumentsJSONRetryHint())
-					if progress != nil {
-						progress("eino_recovery", toolCallArgumentsJSONRecoveryTimelineMessage(attempt), map[string]interface{}{
-							"conversationId": conversationID,
-							"source":         "eino",
-							"einoRetry":      attempt,
-							"runIndex":       attempt + 1,
-							"maxRuns":        maxToolCallRecoveryAttempts,
-							"reason":         "invalid_tool_arguments_json",
-						})
-					}
-					continue attemptLoop
-				}
-
-				if canRetry && isRecoverableToolExecutionError(ev.Err) {
-					if logger != nil {
-						logger.Warn("eino: recoverable tool execution error, will retry with corrective hint",
-							zap.Error(ev.Err), zap.Int("attempt", attempt))
-					}
+				// context.Canceled 是唯一应当直接终止编排的错误（用户关闭页面、主动停止等）。
+				if errors.Is(ev.Err, context.Canceled) {
 					flushAllPendingAsFailed(ev.Err)
-					retryHints = append(retryHints, toolExecutionRetryHint())
 					if progress != nil {
-						progress("eino_recovery", toolExecutionRecoveryTimelineMessage(attempt), map[string]interface{}{
+						progress("error", ev.Err.Error(), map[string]interface{}{
 							"conversationId": conversationID,
 							"source":         "eino",
-							"einoRetry":      attempt,
-							"runIndex":       attempt + 1,
-							"maxRuns":        maxToolCallRecoveryAttempts,
-							"reason":         "tool_execution_error",
 						})
 					}
-					continue attemptLoop
+					return nil, ev.Err
 				}
 
+				canRetry := attempt+1 < maxToolCallRecoveryAttempts
+				if !canRetry {
+					// 重试次数已耗尽，终止。
+					flushAllPendingAsFailed(ev.Err)
+					if progress != nil {
+						progress("error", ev.Err.Error(), map[string]interface{}{
+							"conversationId": conversationID,
+							"source":         "eino",
+						})
+					}
+					return nil, ev.Err
+				}
+
+				// 区分错误类型以选择最合适的纠错提示，但无论哪种都执行重试（default-soft）。
+				var hint *schema.Message
+				var reason, timelineMsg string
+				if isRecoverableToolCallArgumentsJSONError(ev.Err) {
+					hint = toolCallArgumentsJSONRetryHint()
+					reason = "invalid_tool_arguments_json"
+					timelineMsg = toolCallArgumentsJSONRecoveryTimelineMessage(attempt)
+				} else {
+					hint = toolExecutionRetryHint()
+					reason = "tool_execution_error"
+					timelineMsg = toolExecutionRecoveryTimelineMessage(attempt)
+				}
+
+				if logger != nil {
+					logger.Warn("eino: recoverable error, will retry with corrective hint",
+						zap.Error(ev.Err), zap.Int("attempt", attempt), zap.String("reason", reason))
+				}
 				flushAllPendingAsFailed(ev.Err)
+				retryHints = append(retryHints, hint)
 				if progress != nil {
-					progress("error", ev.Err.Error(), map[string]interface{}{
+					progress("eino_recovery", timelineMsg, map[string]interface{}{
 						"conversationId": conversationID,
 						"source":         "eino",
+						"einoRetry":      attempt,
+						"runIndex":       attempt + 1,
+						"maxRuns":        maxToolCallRecoveryAttempts,
+						"reason":         reason,
 					})
 				}
-				return nil, ev.Err
+				continue attemptLoop
 			}
 			if ev.AgentName != "" && progress != nil {
 				iterEinoAgent := orchestratorName

internal/multiagent/tool_error_middleware.go

@@ -41,62 +41,27 @@ func softRecoveryToolCallMiddleware() compose.InvokableToolMiddleware {
 
 // isSoftRecoverableToolError determines whether a tool execution error should be
 // silently converted to a tool-result message rather than crashing the graph.
+//
+// Design: default-soft (blacklist). Almost every tool execution error should be
+// fed back to the LLM so it can self-correct or choose an alternative tool.
+// Only a small set of "truly fatal" conditions (user cancellation) should
+// propagate as hard errors that terminate the orchestration graph.
+// This avoids the fragile whitelist approach where every new error pattern
+// would need to be explicitly enumerated.
 func isSoftRecoverableToolError(err error) bool {
 	if err == nil {
 		return false
 	}
 
-	// 用户取消 — 不应重试，让 hard error 传播以终止编排。
+	// 用户主动取消 — 唯一应当终止编排的情况，不应重试。
 	if errors.Is(err, context.Canceled) {
 		return false
 	}
 
-	// 工具执行超时 — 转为 soft error 让 LLM 知晓并选择替代方案，而非全局重试。
-	if errors.Is(err, context.DeadlineExceeded) {
-		return true
-	}
-
-	s := strings.ToLower(err.Error())
-
-	// JSON unmarshal/parse failures — the model generated truncated or malformed arguments.
-	if isJSONRelatedError(s) {
-		return true
-	}
-
-	// Sub-agent type not found (from deep/task_tool.go)
-	if strings.Contains(s, "subagent type") && strings.Contains(s, "not found") {
-		return true
-	}
-
-	// Tool not found in ToolsNode indexes
-	if strings.Contains(s, "tool") && strings.Contains(s, "not found") {
-		return true
-	}
-
-	return false
-}
-
-// isJSONRelatedError checks whether an error string indicates a JSON parsing problem.
-func isJSONRelatedError(lower string) bool {
-	if !strings.Contains(lower, "json") {
-		return false
-	}
-	jsonIndicators := []string{
-		"unexpected end of json",
-		"unmarshal",
-		"invalid character",
-		"cannot unmarshal",
-		"invalid tool arguments",
-		"failed to unmarshal",
-		"must be in json format",
-		"unexpected eof",
-	}
-	for _, ind := range jsonIndicators {
-		if strings.Contains(lower, ind) {
-			return true
-		}
-	}
-	return false
+	// 其他所有工具执行错误（超时、命令不存在、JSON 解析失败、工具未找到、
+	// 权限不足、网络不可达……）一律转为 soft error，让 LLM 看到错误信息
+	// 后自行决策：换工具、调整参数、或向用户说明。
+	return true
 }
 
 // buildSoftRecoveryMessage creates a bilingual error message that the LLM can act on.

internal/multiagent/tool_error_middleware_test.go

@@ -53,7 +53,12 @@ func TestIsSoftRecoverableToolError(t *testing.T) {
 		{
 			name:     "unrelated network error",
 			err:      errors.New("connection refused"),
-			expected: false,
+			expected: true, // default-soft: non-cancel errors are recoverable
+		},
+		{
+			name:     "tool binary not installed",
+			err:      errors.New("[LocalFunc] failed to invoke tool, toolName=grep, err=ripgrep (rg) is not installed or not in PATH"),
+			expected: true,
 		},
 		{
 			name:     "context cancelled",
@@ -131,15 +136,16 @@ func TestSoftRecoveryToolCallMiddleware_PropagatesNonRecoverable(t *testing.T) {
 		return nil, origErr
 	}
 	wrapped := mw(next)
-	_, err := wrapped(context.Background(), &compose.ToolInput{
+	out, err := wrapped(context.Background(), &compose.ToolInput{
 		Name:      "test_tool",
 		Arguments: `{}`,
 	})
-	if err == nil {
-		t.Fatal("expected error to propagate for non-recoverable errors")
+	// Default-soft: non-cancel errors are converted to tool-result messages.
+	if err != nil {
+		t.Fatalf("expected nil error (soft recovery), got: %v", err)
 	}
-	if err != origErr {
-		t.Fatalf("expected original error, got: %v", err)
+	if out == nil || out.Result == "" {
+		t.Fatal("expected non-empty recovery message")
 	}
 }
 

internal/multiagent/tool_execution_retry.go

@@ -2,74 +2,42 @@ package multiagent
 
 import (
 	"fmt"
-	"strings"
 
 	"github.com/cloudwego/eino/schema"
 )
 
-// isRecoverableToolExecutionError detects tool-level execution errors that can be
-// recovered by retrying with a corrective hint. These errors originate from eino
-// framework internals (e.g. task_tool.go, tool_node.go) when the LLM produces
-// invalid tool calls such as non-existent sub-agent types, malformed JSON arguments,
-// or unregistered tool names.
-func isRecoverableToolExecutionError(err error) bool {
-	if err == nil {
-		return false
-	}
-	s := strings.ToLower(err.Error())
-
-	// Sub-agent type not found (from deep/task_tool.go)
-	if strings.Contains(s, "subagent type") && strings.Contains(s, "not found") {
-		return true
-	}
-
-	// Tool not found in toolsNode indexes (from compose/tool_node.go, when UnknownToolsHandler is nil)
-	if strings.Contains(s, "tool") && strings.Contains(s, "not found") {
-		return true
-	}
-
-	// Invalid tool arguments JSON (from einomcp/mcp_tools.go or eino internals)
-	if strings.Contains(s, "invalid tool arguments json") {
-		return true
-	}
-
-	// Failed to unmarshal task tool input json (from deep/task_tool.go)
-	if strings.Contains(s, "failed to unmarshal") && strings.Contains(s, "json") {
-		return true
-	}
-
-	// Generic tool call stream/invoke failure wrapping the above
-	if (strings.Contains(s, "failed to stream tool call") || strings.Contains(s, "failed to invoke tool")) &&
-		(strings.Contains(s, "not found") || strings.Contains(s, "json") || strings.Contains(s, "unmarshal")) {
-		return true
-	}
-
-	return false
-}
-
 // toolExecutionRetryHint returns a user message appended to the conversation to prompt
-// the LLM to correct its tool call after a tool execution error.
+// the LLM to adjust after a tool execution error (tool not found, binary missing,
+// runtime failure, network error, etc.).
 func toolExecutionRetryHint() *schema.Message {
-	return schema.UserMessage(`[System] Your previous tool call failed because:
-- The tool or sub-agent name you used does not exist, OR
+	return schema.UserMessage(`[System] Your previous tool call failed. Possible causes:
+- The tool or sub-agent name does not exist (typo or unregistered name).
 - The tool call arguments were not valid JSON.
+- The tool's underlying binary is not installed or not in PATH.
+- The tool encountered a runtime error (timeout, network failure, permission denied, etc.).
 
-Please carefully review the available tools and sub-agents listed in your context, use only exact registered names (case-sensitive), and ensure all arguments are well-formed JSON objects. Then retry your action.
+Please review the error message above, check available tools, and either:
+1. Retry with corrected arguments or a different tool, OR
+2. Inform the user about the limitation and proceed with an alternative approach.
 
 [系统提示] 上一次工具调用失败，可能原因：
-- 你使用的工具名或子代理名称不存在；
-- 工具调用参数不是合法 JSON。
+- 工具名或子代理名称不存在（拼写错误或未注册）；
+- 工具调用参数不是合法 JSON；
+- 工具依赖的底层二进制程序未安装或不在 PATH 中；
+- 工具运行时遇到错误（超时、网络故障、权限不足等）。
 
-请仔细检查上下文中列出的可用工具和子代理名称（须完全匹配、区分大小写），确保所有参数均为合法的 JSON 对象，然后重新执行。`)
+请根据上述错误信息检查可用工具，然后：
+1. 修正参数或改用其他工具重试，或者
+2. 告知用户当前限制并采用替代方案继续。`)
 }
 
 // toolExecutionRecoveryTimelineMessage returns a message for the eino_recovery event
 // displayed in the UI timeline when a tool execution error triggers a retry.
 func toolExecutionRecoveryTimelineMessage(attempt int) string {
 	return fmt.Sprintf(
-		"工具调用执行失败（工具/子代理名称不存在或参数 JSON 无效）。已向对话追加纠错提示并要求模型重新生成。"+
+		"工具调用执行失败。已向对话追加纠错提示并要求模型调整策略。"+
 			"当前为第 %d/%d 轮完整运行。\n\n"+
-			"Tool call execution failed (unknown tool/sub-agent name or invalid JSON arguments). "+
+			"Tool call execution failed. "+
 			"A corrective hint was appended. This is full run %d of %d.",
 		attempt+1, maxToolCallRecoveryAttempts, attempt+1, maxToolCallRecoveryAttempts,
 	)