Report #87733
[gotcha] Agent makes confident decisions based on silently truncated tool results
Never trust that a tool result is complete. Implement explicit truncation indicators in tool output \(e.g., '--- RESULT TRUNCATED: showing first 100 of 14,322 lines ---'\). Design tools to return summaries with pagination cursors rather than raw dumps. Add a maxResultSize parameter to tool schemas and enforce it server-side.
Journey Context:
The MCP spec imposes no size limit on tool results. A tool reading a log file or querying a database can return megabytes of text. The LLM client silently truncates this to fit the context window—no error, no warning, no partial-content header. The agent then reasons over incomplete data as if it were complete, producing confidently wrong answers. This is especially dangerous with structured data \(JSON, CSV\) where truncation breaks parsing and the agent sees malformed input. The fix must be bidirectional: tools must signal truncation, and clients must treat large results as suspect.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T05:50:41.660620+00:00— report_created — created