feat: improve 'no tools used' error handling with grace retry (#10196)

Co-authored-by: ellipsis-dev[bot] <65095814+ellipsis-dev[bot]@users.noreply.github.com>

Author: 2 people

src/api/providers/gemini.ts

@@ -318,21 +318,6 @@ export class GeminiHandler extends BaseProvider implements SingleCompletionHandl
 				}
 			}
 
-			// If we had reasoning but no content, emit a placeholder text to prevent "Empty assistant response" errors.
-			// This typically happens when the model hits max output tokens while reasoning.
-			if (hasReasoning && !hasContent) {
-				let message = t("common:errors.gemini.thinking_complete_no_output")
-				if (finishReason === "MAX_TOKENS") {
-					message = t("common:errors.gemini.thinking_complete_truncated")
-				} else if (finishReason === "SAFETY") {
-					message = t("common:errors.gemini.thinking_complete_safety")
-				} else if (finishReason === "RECITATION") {
-					message = t("common:errors.gemini.thinking_complete_recitation")
-				}
-
-				yield { type: "text", text: message }
-			}
-
 			if (finalResponse?.responseId) {
 				// Capture responseId so Task.addToApiConversationHistory can store it
 				// alongside the assistant message in api_history.json.

src/core/task/Task.ts

@@ -310,6 +310,7 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 	consecutiveMistakeCount: number = 0
 	consecutiveMistakeLimit: number
 	consecutiveMistakeCountForApplyDiff: Map<string, number> = new Map()
+	consecutiveNoToolUseCount: number = 0
 	toolUsage: ToolUsage = {}
 
 	// Checkpoints
@@ -1982,6 +1983,9 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 
 		this.abort = true
 
+		// Reset consecutive error counters on abort (manual intervention)
+		this.consecutiveNoToolUseCount = 0
+
 		// Force final token usage update before abort event
 		this.emitFinalTokenUsageUpdate()
 
@@ -2236,7 +2240,6 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 			} else {
 				// Use the task's locked protocol, NOT the current settings (fallback to xml if not set)
 				nextUserContent = [{ type: "text", text: formatResponse.noToolsUsed(this._taskToolProtocol ?? "xml") }]
-				this.consecutiveMistakeCount++
 			}
 		}
 	}
@@ -3250,12 +3253,24 @@ export class Task extends EventEmitter<TaskEvents> implements TaskLike {
 					)
 
 					if (!didToolUse) {
+						// Increment consecutive no-tool-use counter
+						this.consecutiveNoToolUseCount++
+
+						// Only show error and count toward mistake limit after 2 consecutive failures
+						if (this.consecutiveNoToolUseCount >= 2) {
+							await this.say("error", "MODEL_NO_TOOLS_USED")
+							// Only count toward mistake limit after second consecutive failure
+							this.consecutiveMistakeCount++
+						}
+
 						// Use the task's locked protocol for consistent behavior
 						this.userMessageContent.push({
 							type: "text",
 							text: formatResponse.noToolsUsed(this._taskToolProtocol ?? "xml"),
 						})
-						this.consecutiveMistakeCount++
+					} else {
+						// Reset counter when tools are used successfully
+						this.consecutiveNoToolUseCount = 0
 					}
 
 					// Push to stack if there's content OR if we're paused waiting for a subtask.

webview-ui/src/components/chat/ChatRow.tsx

@@ -1259,7 +1259,22 @@ export const ChatRowContent = ({
 						</div>
 					)
 				case "error":
-					return <ErrorRow type="error" message={t("chat:error")} errorDetails={message.text || undefined} />
+					// Check if this is a model response error based on marker strings from backend
+					const isNoToolsUsedError = message.text === "MODEL_NO_TOOLS_USED"
+
+					if (isNoToolsUsedError) {
+						return (
+							<ErrorRow
+								type="error"
+								title={t("chat:modelResponseIncomplete")}
+								message={t("chat:modelResponseErrors.noToolsUsed")}
+								errorDetails={t("chat:modelResponseErrors.noToolsUsedDetails")}
+							/>
+						)
+					}
+
+					// Fallback for generic errors
+					return <ErrorRow type="error" message={message.text || t("chat:error")} />
 				case "completion_result":
 					return (
 						<>

webview-ui/src/i18n/locales/ca/chat.json

@@ -285,6 +285,11 @@
 	},
 	"taskCompleted": "Task Completed",
 	"error": "Error",
+	"modelResponseIncomplete": "Model Response Incomplete",
+	"modelResponseErrors": {
+		"noToolsUsed": "The model failed to use any tools in its response. This typically happens when the model provides only text/reasoning without calling the required tools to complete the task.",
+		"noToolsUsedDetails": "The model provided text/reasoning but did not call any of the required tools. This usually indicates the model misunderstood the task or is having difficulty determining which tool to use. The model has been automatically prompted to retry with proper tool usage."
+	},
 	"errorDetails": {
 		"title": "Error Details",
 		"copyToClipboard": "Copy to Clipboard",