[FEATURE] Preserve tool execution failures in the agent loop

[alectimison-maker]

Is your feature request related to a problem? Please describe.

The agent loop appears to lose the real error when a registered tool throws.

executeToolCalls() catches non-abort failures, logs them, and returns no result for that tool:

NativeMindExtension/entrypoints/sidepanel/utils/agent/index.ts

Lines 467 to 512 in 70a2acc

	async executeToolCalls(
	toolCalls: PromptBasedToolNameAndParams<T>[],
	taskScopeToolCalls: PromptBasedToolNameAndParamsDistributed[],
	loopImages: ImageDataWithId[] = [],
	taskMessageModifier: TaskMessageModifier,
	eventBus: EventEmitter<AgentToolCallExecuteHooks>,
	) {
	toolCalls = this.deduplicateToolCalls(toolCalls)
	const currentLoopToolResults: (AgentToolExecuteResult & { toolName: T })[] = []
	for (const chunk of toolCalls) {
	const toolName = chunk.toolName as T
	const tool = this.tools[toolName]
	if (tool) {
	const params = chunk.params
	this.log.debug('Tool call start', chunk)
	const abortController = this.createAbortController()
	try {
	const executedResults = await tool.execute({
	params,
	taskScopeToolCalls,
	agentStorage: this.agentStorage,
	historyManager: this.historyManager,
	loopImages,
	abortSignal: abortController.signal,
	taskMessageModifier,
	hooks: eventBus,
	})
	for (const result of executedResults) {
	currentLoopToolResults.push({ ...result, toolName })
	}
	this.log.debug('Tool call executed', toolName, executedResults)
	}
	catch (e) {
	if (e instanceof AbortError) {
	this.log.debug('Tool call aborted', toolName)
	break
	}
	this.log.error('Tool call error', toolName, e)
	}
	taskMessageModifier.makeAllTaskDone()
	}
	else {
	this.log.warn('Tool not found', chunk)
	}
	}
	return currentLoopToolResults

The caller then treats an empty result list as Tool not found, even though the tool was present:

NativeMindExtension/entrypoints/sidepanel/utils/agent/index.ts

Lines 346 to 375 in 70a2acc

	if (currentLoopToolCalls.length > 0) {
	if (agentMessage.content || agentMessage.reasoning) {
	// create a new group if there are some plain messages
	taskMessageModifier = this.makeTaskMessageGroupProxy(abortController.signal)
	}
	this.log.debug('Executing tool calls', currentLoopToolCalls)
	const toolExecuteResults = await this.executeToolCalls(currentLoopToolCalls, taskScopeToolCalls, loopImages, taskMessageModifier, eventBus)
	this.log.debug('Tool calls executed', currentLoopToolCalls, toolExecuteResults)
	const toolResults = toolExecuteResults.filter((r) => r.type === 'tool-result')
	const handOffResults = toolExecuteResults.filter((r) => r.type === 'hand-off')
	if (handOffResults.length > 0) {
	const handoffResult = handOffResults[0]
	if (handoffResult) {
	// This feature is in beta, not used yet
	this.log.debug('Hand-off detected', handoffResult)
	const subAgent = new Agent({ tools: this.tools, agentStorage: this.agentStorage, historyManager: this.historyManager, maxIterations: this.maxIterations })
	abortController.signal.addEventListener('abort', () => subAgent.stop())
	loopMessages.push({ role: 'user', content: handoffResult.userPrompt })
	const lastMsg = await subAgent.run(this.overrideSystemPrompt([...baseMessages, ...loopMessages], handoffResult.overrideSystemPrompt))
	this.log.debug('Sub-agent finished', lastMsg)
	if (lastMsg?.content) loopMessages.push(lastMsg)
	}
	}
	else if (toolResults.length) {
	loopMessages.push({ role: 'user', content: this.buildExtendedUserMessage(iteration + 1, originalUserMessageText, toolResults) })
	}
	else {
	const errorResult = TagBuilder.fromStructured('error', { message: `Tool not found, available tools are: ${Object.keys(this.tools).join(', ')}` })
	loopMessages.push({ role: 'user', content: renderPrompt`${errorResult}` })
	}

That can give the model misleading recovery context and hide the actionable failure from the user. This is a source-based inference from commit 70a2acc1425749913748c92309254a8f4e69aa8a; I have not reproduced it in a running extension.

Describe the solution you'd like

Convert a non-abort tool exception into a structured error tool result associated with the original toolName. Reserve Tool not found for the branch where the requested tool is genuinely absent. Keep the existing abort behavior unchanged.

Suggested acceptance criteria:

• A registered tool that throws produces a tool-specific error result for the next agent iteration.
• An unknown tool still produces the available-tools guidance.
• AbortError stops execution without being converted into a normal tool failure.
• Unit tests cover all three paths and multiple tool calls where one fails.

Describe alternatives you've considered

• Rethrowing the exception would terminate the whole agent loop rather than allow recovery.
• Showing only a UI toast would still withhold the failure reason from the model.

Additional context

If maintainers agree with the desired error-result shape, I would be happy to prepare a focused PR after that guidance.