Uncapped Problems Output Can Overwhelm Session Context

[hannesrudolph]

App Version

v3.21.0 (accc5fb)

API Provider

Not Applicable / Other

Model Used

N/A

🔁 Steps to Reproduce

The core of this issue is architectural rather than a specific sequence of user actions. Any scenario that generates a very large number of "problems" in VSCode will trigger it.

For example, a failed pnpm install in a large monorepo can generate 60,000+ problems.

The underlying issue is that the extension feeds the entire content of the "Problems" panel into the session context without any limits.

💥 Outcome Summary

Expected: The number of problems sent to the chat session should be capped at a reasonable limit to prevent context overflow, even if thousands of problems are generated by the build process or language servers.

Actual: The problems output is not capped. A large number of generated problems (e.g., from a failed dependency install) can flood the session context, causing it to grow excessively (e.g., 800k+ tokens). This can make the session unresponsive, slow, and expensive.

📄 Relevant Logs or Errors (Optional)

N/A - The issue is the volume of correct error reporting, not a specific error message.

[daniel-lxs]

After reviewing the attempted fix in #4861, it seems the core issue is more complex than just limiting the number of problems. Since individual problems can vary greatly in size (some might include long stack traces, others just brief messages), counting problems doesn't effectively prevent context overflow.

A possible approach could be:

• Instead of limiting by problem count, consider limiting by total content size (characters or estimated tokens)
• Track the cumulative size while building the problems output
• Include complete problems up to the limit (avoiding partial problem truncation)
• Show a clear indication of how many problems were omitted

For example:

const MAX_PROBLEMS_CONTENT_SIZE = 50000; // characters
let currentSize = 0;
let includedCount = 0;
let totalCount = 0;

// Build output while tracking size
for (const problem of allProblems) {
  const problemText = formatProblem(problem);
  if (currentSize + problemText.length > MAX_PROBLEMS_CONTENT_SIZE && includedCount > 0) {
    break; // Stop before exceeding limit
  }
  result += problemText;
  currentSize += problemText.length;
  includedCount++;
  totalCount++;
}

if (totalCount > includedCount) {
  result += `\n\n... ${totalCount - includedCount} more problems omitted to prevent context overflow`;
}

This would ensure the context stays manageable regardless of individual problem sizes.