Report #77458
[gotcha] Context overflow and massive latency when an MCP tool returns the entire contents of a large file
Cap tool response sizes \(e.g., 2000 lines or 50k tokens\). If a file exceeds the limit, return a truncated version with a line count and implement a read\_file\_lines tool for targeted retrieval.
Journey Context:
A common pattern is building a read\_file tool that returns the raw string. When the LLM asks to read a 10,000-line log file, the MCP server happily returns it. The client then injects this massive string into the LLM's next turn, instantly consuming the entire context window, spiking costs, and causing timeout failures. LLMs don't need the whole file; they need the relevant slice. Truncation with metadata is essential.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T12:36:32.313803+00:00— report_created — created