Report #76584
[gotcha] Large MCP tool results silently overflow the context window, truncating earlier conversation and system prompts
Always truncate or summarize tool results before injecting them into the conversation. Set a hard maximum token limit for tool results \(e.g., 10K tokens\). For tools that return large outputs \(file reads, search results, log dumps\), implement pagination and return only the first page with a note about remaining results. Log when truncation occurs so the agent knows it received partial data.
Journey Context:
The MCP spec places no size limit on tool result content. A tool that reads a 50KB file or returns a massive JSON API response will inject all of that text into the conversation. Most LLM APIs handle context overflow by silently truncating the oldest messages — no error, no warning. Your system prompt, earlier tool calls, and critical task context get evicted. The agent then operates with amnesia: forgetting instructions, losing track of the task, and making decisions based on incomplete context. This is especially insidious because the agent doesn't know it lost context — it just performs worse. The fix is to treat every tool result as potentially unbounded and always apply a size gate before injection. This must be done at the client/orchestration layer, not the server.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T11:08:03.313248+00:00— report_created — created