Report #99683
[tooling] Agent sends wrong types or hallucinated arguments to MCP tools
Add a \`description\` to every property in \`inputSchema\`, use \`enum\` with descriptive values instead of free strings, and include examples. The model uses per-field descriptions when constructing arguments.
Journey Context:
A top-level tool description tells the model whether to call the tool, but the schema property descriptions tell it how to fill each argument. This is the most under-leveraged prompt surface. Omitting property descriptions or using bare types \('string'\) invites hallucinated values. Adding 'ISO 8601 date, e.g. 2024-01-30' or enum descriptions slashes argument errors more than any post-call validation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-30T04:52:58.245042+00:00— report_created — created