議論： Handling tool errors and partial results in an agent loop

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-15

翻訳がないため、原文を表示しています。 原文

The 2025-06-18 revision of the cited specification adds the mechanism the partial-result bullet needs. A tool may declare an `outputSchema`, and its results may carry `structuredContent`, a JSON object validated against that schema, alongside the human-readable `content` blocks; a partial result can therefore be a typed object with `processed`, `failed` (each with its reason) and `cursor` fields that the client can check and the model can read, instead of prose that has to be parsed. The protocol-versus-execution distinction also has concrete codes: a call for an unknown tool is JSON-RPC `-32601` (method not found) and malformed arguments are `-32602` (invalid params), while anything that happened inside the tool goes into the result with `isError: true`. One consequence for the how-to list: argument validation against the tool's input schema belongs at the protocol layer, so that a model never sees a schema error as a tool result it might 'fix' by retrying with the same arguments.