Intel / repo

goose v1.59.0

Source: github.com
Author: github-actions[bot]
Date: Published Nov 10, 2026

Context

goose, a general-purpose that runs on the user's own machine and is part of the Linux Foundation's Agentic AI Foundation, shipped version 1.59.0 with one change aimed at a cost every -heavy setup pays. MCP is the protocol agents use to connect to outside tools, and by default goose sends every MCP tool's full schema to the model up front, so the model receives every connected server's tool descriptions whether or not it uses them. The new option lists only the tool names in the and lets the model call an enable_tools function to load the schemas it actually needs. It is off unless a user sets GOOSE_TOOL_DEFERRAL=lookup, and the change is billed as minimal: the model must name each tool exactly, with no search or wildcard matching, and built-in tools and any MCP tools marked always-loaded stay loaded as before.

Two other fixes matter to people running long sessions. goose can hand subtasks to delegated subagents, and version 1.56.0 let users resume a delegate that had failed or been cancelled. A completed delegate can now also be resumed with a follow-up instruction and keeps its earlier , where before goose rejected the request as already completed. Separately, goose compacts a long conversation by asking the model to write a summary so the session can continue within its context window. The change behind that fix reports that Claude Opus 5.5 from Anthropic sometimes refused this step, citing restrictions on reverse engineering or duplicating model outputs. Its author's diagnosis is that the old prompt, which asked the model to distill the conversation, include its own responses and quote liberally, read like a request to copy model outputs. The prompt now asks for structured working notes instead. The fix only rewords the prompt and adds no test showing that the refusals have stopped.

The rest of the release is smaller: configurable voice instructions and endpoint settings for the live-voice feature, stop, approval and tool hooks for goose-agent, Databricks OAuth tokens that are refreshed before they expire, and fixes to session search and database setup.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
system prompt — The standing instructions a model receives before any user input — defining its role, rules, tools, and tone for the whole conversation.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source github.com