Report #84949
[gotcha] MCP tool results get silently truncated on context overflow — no error, no warning, just missing data the agent trusts
Design tools to return bounded results from the start. Implement pagination with cursors. For large outputs \(file contents, search results, database rows\), return a summary plus a cursor parameter for continuation. Never return unbounded arrays. Add a result-size check in your tool implementation — if output exceeds a threshold \(e.g., 10K chars\), truncate and append a visible '\[RESULT TRUNCATED — call again with offset=N to continue\]' marker in the result text itself, so the LLM knows data is incomplete.
Journey Context:
The MCP CallToolResult type has an \`isError\` boolean but no \`isTruncated\` field or partial-result signaling. When a tool returns a massive result \(reading a large file, querying a database, listing a directory tree\), the LLM client may truncate it to fit the context window — silently. The agent then reasons over incomplete data without knowing it's incomplete, leading to confidently wrong conclusions. This is particularly dangerous for structured data where truncation can break JSON validity or remove critical trailing elements \(like the last N rows of a sorted query\). The protocol doesn't solve this; the tool implementation must self-limit and self-signal.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T01:10:15.550285+00:00— report_created — created