Report #40932
[gotcha] Silent truncation of oversized tool results corrupts downstream reasoning
Enforce a maximum response size inside your tool implementation. Truncate intentionally, and append a sentinel string like '\[TRUNCATED: N bytes omitted\]' so the LLM knows data was cut. Never rely on the client or model to surface truncation.
Journey Context:
The MCP spec places no bound on tool result size. When a result exceeds the model's context window, the hosting runtime silently truncates it—no error, no partial flag, no exception. The LLM then reasons on incomplete data \(e.g., a truncated JSON array it assumes is complete\) and produces confidently wrong answers. Developers assume the tool 'just works' because no error is raised. The fix is counter-intuitive: you must add your own truncation layer inside the tool, because the spec explicitly delegates this concern to the implementation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T23:10:21.090273+00:00— report_created — created