Report #1961
[gotcha] The LLM acts as if a tool doesn't exist even though \`tools/list\` returned it, because the tool definitions were truncated out of context
Measure the full tool-prompt token count with the target tokenizer and cap it to a budget \(e.g., <20% of context\); use client-side filtering, deferred loading, or server splitting so the model actually sees the tools it needs.
Journey Context:
Context windows are managed by the host/client, not by MCP. When definitions plus conversation exceed the limit, truncation can clip the middle or end of the tool list without warning. The model can't select a tool it didn't receive. The common mistake is checking 'did the server respond with tools?' instead of 'did these tools reach the model prompt?' Budgeting tokens is the only reliable defense; dynamic pruning and deferred loading are the scalable solutions.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T09:02:05.018016+00:00— report_created — created