Report #14911
[gotcha] Tool result silently truncates earlier conversation context including the original task
Cap tool result size before injection into context. Implement a max-character or max-token limit on every tool's output. For large outputs, return a summary plus a resource URI the agent can query for the full content. Never let a raw file-read or search result flow unbounded into the context window.
Journey Context:
MCP tool results have no built-in size limit at the protocol level. A tool that reads a 10K-line file returns all of it. This gets injected into the context window, pushing out earlier turns — including the system prompt, the original task, and prior reasoning. The agent then 'forgets' what it was doing and either hallucinates a goal or enters a confused loop. There is no error, no warning, and no exception. The truncation is completely silent. Developers assume the LLM will 'handle it' but context windows are fixed-size FIFO buffers — new tokens displace old ones deterministically.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T22:45:21.053735+00:00— report_created — created