[BUG] GitHub Integration Cannot Properly Read File Contents – Regression in Project Knowledge

⚠️ I’m not sure if this is the right place to report this, but this is a very serious issue.

I'm on the Claude MAX plan ($100/month) and have been using Claude for a long time without such problems — until now. For over a week, it’s been impossible to work properly due to a critical failure in GitHub integration.

What’s especially strange is that when I upload a file as an attachment, Claude reads it perfectly and adds functions exactly as requested. But when using GitHub integration, the same file is either unreadable or reconstructed incorrectly, with missing or broken content.

Previously there was a syncing issue (files not showing at all) — that was fixed. But now, even though the files appear in the list, Claude cannot see or interpret their contents correctly, which results in incomplete or invalid code being generated.

➡️ This completely breaks the development workflow and makes it impossible to create or maintain projects as before.

I kindly ask your engineering team to urgently test this and apply a fix, as this is a core feature of the MAX plan and critical to my work. This is the first time I've experienced such persistent problems after months of reliable use.

1. PERSISTENT "RETRIEVING" STATUS BUG:

• Project knowledge shows permanent "Retrieving" status
• Shows low capacity usage (7%) but indexing never completes
• Repository is a standard TypeScript/React project from GitHub
• Status remains stuck for hours without any progress
• Project becomes partially unusable despite showing some indexed content

1. FRAGMENTED FILE SEARCH BUG:

• project_knowledge_search returns incomplete file fragments
• Cannot retrieve complete source code of files
• Returns random code snippets from different parts of files
• Makes comprehensive code analysis impossible
• Specifically affects TypeScript/React component files

EXPECTED BEHAVIOR:

• Indexing should complete and show "Ready" status
• File search should return complete, continuous file content

ACTUAL BEHAVIOR:

• Permanent "Retrieving" status despite low capacity usage
• Fragmented, incomplete file retrieval

IMPACT:

• Severely limits code analysis capabilities
• Makes project knowledge unreliable for development assistance
• User cannot get complete file reviews or comprehensive code help

BROWSER: [Your browser]
PROJECT TYPE: TypeScript/React GitHub repository

I’ve investigated the issue more deeply and confirmed that it occurs in projects that try to index more than 7% of their content. In such cases, Claude only sees fragmented parts of the files instead of their full content.

Sorry, I'm not sure I understand? Is this an issue with Claude Code and/or using Claude Code from with a Github Action? Or is this an issue with Github's first party solution. If the former, you'll have to report that to Github.

@DanielCyran +1, encountering the exact same problem.

+1 this is a massive issue, the github connection in Claude was previously very reliable, but it's become completely unusable in the last week or two.

Just to confirm, is this report about Claude Code or about Claude.ai's GitHub integration? It sounds like it's the latter. I'll pass this report along internally.

In general, the file limit is far too small – and this likely applies not just to GitHub integration. Here’s the response I received from support:

Thank you for your patience while I investigated this issue further with our technical team. I now have a much clearer picture of what you're experiencing.
I've just received clarification from our technical team: this is actually expected behavior, not a bug. Here's why- GitHub Integration uses RAG by chunking your repository automatically into smaller pieces for storage and retrieval, which is why you're seeing fragments like "5 relevant sections" and partial code snippets. When you do a search in your project when it is in retrieval mode, it find all relevant pieces, but cannot access the entire repository.
Direct uploads bypass RAG- this is why you get perfect, complete files when you upload directly - the system accesses the entire file without chunking.
The fragmented approach works well for document search but creates exactly the frustrating experience you're having with code files where you need complete context and structure.
For your immediate workflow: Continue using direct file uploads to get complete, properly structured code files. The GitHub integration is better suited for general code discussion rather than reproducing entire files.
I've documented your feedback about this limitation for our product team, as this architectural difference significantly impacts code development workflows that reply on seeing entire code files at once.                                

_________________________


 Hi again Daniel,
I completely understand how frustrating the 6% threshold is for your current development workflow. I agree that that'd be a good idea to implement to improve the experience, and would be happy to bring your request for higher project knowledge limits for Github to the attention of our product team. Though I can't guarantee they will make any immediate changes to address this, I can definitely get this seen by the team.
Unfortunately for the short term, I can't manually increase the project knowledge capacity limit to 15% or any other percentage. These limits are built into the fundamental architecture of the Project system. The limits cannot be manually adjusted, even by our engineering team without making significant updates first. The 6-7% threshold where you're experiencing issues is the system working as designed - performance degrading once file retrieval is turned on and smaller code is retrieved instead of everything being in context.
For now, I would recommend continuing your hybrid approach (the GitHub integration for general context, direct uploads for complete files) as it remains the best workaround within the current technical constraints.
Please let me know if you need any additional help optimizing your workflow.
Best,
Georgia

The current context limit in Claude AI is too low for working with larger programming projects. We kindly request an increase in this limit to enable full understanding of the project structure and code dependencies.

Since this issue is related to claude.ai and not Claude Code, I recommend going through the support channels for claude.ai since they'll be able to help you more effectively: https://support.anthropic.com/en/articles/9015913-how-to-get-support

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.