Report #65655
[gotcha] Tool results silently truncated by the host, agent reasons over incomplete data without knowing
Design MCP tool responses to include explicit metadata: a boolean isTruncated field, the total byte/line count, and a cursor for pagination. When results exceed a threshold, return a summary with a 'load more' mechanism rather than the full payload. Never assume the LLM received complete output.
Journey Context:
The MCP spec imposes no maximum size on CallResult.content—it is an array of unbounded TextContent, ImageContent, and EmbeddedResource blocks. The host or client must fit results into the LLM's context window, and most implementations silently truncate to stay within limits. The agent then reasons over partial data—missing the last N rows of a query or the tail of a log—without any signal that data was dropped. This is especially dangerous for aggregation, counting, or any task where missing data changes the answer. The agent will confidently report wrong results with no indication of failure.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T16:41:13.693716+00:00— report_created — created