Images uploaded in a previous turn aren't available to image_gen_oai/gemini_image_gen tools in subsequent turns, even with resendFiles enabled

[Odrec]

What happened?

When using an agent with image generation/editing tools (image_gen_oai, gemini_image_gen), images uploaded in a previous conversation turn are not available to the image tools in subsequent turns, even though resendFiles defaults to true for agents.

Root cause:

In packages/api/src/agents/resources.ts, the categorizeFileForToolResources function only adds images to tool_resources.image_edit if the file ID is present in requestFileSet:

if (
  requestFileSet.has(file.file_id) &&
  file.type.startsWith('image') &&
  file.height &&
  file.width
) {
  addFileToResource({
    file,
    resourceType: EToolResources.image_edit,
    tool_resources,
    processedResourceFiles,
  });
}

requestFileSet only contains file IDs from the current request's attachments. When resendFiles is enabled, BaseClient.addPreviousAttachments() re-attaches files from previous messages as vision context (image_urls), but these resent files are never added to requestFileSet, so they are not categorized into tool_resources.image_edit.

This means the image editing/generation tools (image_gen_oai, gemini_image_gen) cannot access images from previous turns, even though the LLM can "see" them in the conversation.

Expected behavior:

When resendFiles is true (which is the default for agents, set in agents/initialize.js), images from previous turns should also be included in tool_resources.image_edit so that image tools can access them across conversation turns.

Version Information

LibreChat: ghcr.io/danny-avila/librechat:v0.8.3 (db02d9011e1e)
RAG API: ghcr.io/danny-avila/librechat-rag-api-dev:latest
CodeSandbox Bundler: ghcr.io/librechat-ai/codesandbox-client/bundler:latest
Exporter: ghcr.io/virtuos/librechat_exporter:2.0.9

Steps to Reproduce

1. User sets up an agent with image generation tools and resend files activated in the parameters of the model (gpt-5.2 from OpenAI API in my case)
2. User uploads an image and asks the agent to edit it
3. Agent responds asking a clarifying question (e.g., which model to use)
4. User answers in a new message (no image attached)
5. The image tools can no longer access the original image — the agent says it can't find the image and asks the user to re-upload it

What browsers are you seeing the problem on?

No response

Relevant log output

No relevant logs for this issue

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Thanks for the detailed report! This is working as designe, image_edit tool resources are intentionally scoped to the current request's attachments, unlike execute_code and file_search which have full cross-turn support when resendFiles is enabled. The image file IDs from previous turns are never wired into tool_resources.image_edit, so while the LLM can "see" resent images as vision context, the image tools don't receive them.

I've flagged this as a feature request. It's a valid enhancement and I'll take it into consideration as soon as time allows. Will especially take this into consideration when we improve the overall file experience.