Files are often not available after the first message

[maxesse]

Description

When I upload multiple files to kubecoderun through Librechat, the files are often not retrievable past the first execution, which causes issues as it can't process the files.
I dug into the logs a bit, these are the findings summarised by claude:

Actual Root Cause: Multi-file uploads create separate sessions

Each file upload creates a separate KubeCodeRun session because entity_id is null:

When the code executes, the orchestrator picks one session to reuse (RAQnp5FoSN4l - "Reusing session from file reference") and only pushes that session's files to the execution pod. The second file (5321947.csv) lives in a completely different session (tglCu-MvWqV5) and never gets pushed.

The first execution succeeded in reading both files (likely because both file IDs were explicitly referenced in the initial exec request and both got pushed). But from the second execution onward, only the "reused session's" files get pushed to each new pool pod — so 5321947.csv is gone.

All subsequent executions ran under RAQnp5FoSN4lCHwU2ddU6 on 8 different pool pods, all failing to find 5321947.csv.

The Possible Fix

Librechat is uploading each file individually with entity_id: null. To fix this, it needs to either:

1. Pass an entity_id (e.g., the conversation ID) on each /upload call so both files land in the same session
2. Upload both files in a single /upload call (the API supports count > 1)
3. Pass all file IDs in the exec request and have KubeCodeRun merge files across sessions (would require an API change)

Is it something I can fix with a config change, or do you think it needs to be fixed in Kubecoderun?

Thanks - it is an amazing product, much better than the old unmaintained official CI.

[ronniekolk]

@aron-muon
We can confirm this issue and have conducted systematic testing to reproduce it .

Our Analysis Confirms the Root Cause

Through different testing, we've observed the exact behavior described:

Evidence from Systematic Testing:

• Container hostname changes with every code execution: pool-py-566c3056 → pool-py-0f2b41b6 → pool-py-a2a66a28
• Files disappear selectively between executions
• Pattern: First execution works, subsequent executions miss files from "other sessions"
• Multi-file uploads create isolated sessions without proper entity_id grouping

Impact on Data Analysis Workflows:

This significantly impacts data science and analytics use cases where users:

• Upload multiple CSV/datasets
• Expect files to persist across analysis steps
• Need to correlate data from multiple sources over hours/days

Additional Technical Insights:

1. Session Isolation: Each file upload creates separate KubeCodeRun sessions due to entity_id: null
2. Worker Pod Isolation: New execution pods only receive files from the "reused" session
3. No Cross-Session File Access: Files in other sessions become inaccessible

Our idea:

**1. For e.g. LibreChat **

// LibreChat file upload handler
entity_id: conversationId // Link files to conversation