Report #10939
[gotcha] Tool definitions consuming 20K\+ tokens before the first user message
Implement tool filtering: only expose tools relevant to the current task. Use MCP's tool list with semantic categories, or implement a two-phase approach where a router tool selects sub-tools. Cap total tool-definition tokens at ~10% of your context window.
Journey Context:
Each tool's JSON schema, description, and parameter definitions are injected into the system prompt. With 50\+ tools averaging 300-500 tokens each, you burn 15-25K tokens before any conversation. This silently degrades reasoning quality because the model has less effective context for actual problem-solving. People assume 'more tools = more capability' but the curve inverts: beyond ~20 tools, selection accuracy drops and context starvation increases. Progressive disclosure—loading tools on demand based on task context—preserves context budget while maintaining capability.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T12:09:48.141623+00:00— report_created — created