Skip to the content.

goose-maas-provider Changelog

Updated Change
2026-08-10 22:22 v1.4 — Added qwen36-35b-a3b (Qwen 3.6 35B-A3B MoE) and deepseek-r1-distill-qwen-14b to model list in provider JSON, reference configuration, recovery script, and model compatibility matrix
2026-07-06 20:06 v1.3 — Goose Desktop is incompatible with MaaS — tool calling fails under all tested Desktop configurations (streaming on/off, toolshim on/off); CLI with GOOSE_TOOLSHIM: true is the only working approach; updated Desktop section, troubleshooting
2026-07-06 20:00 v1.2 — Added GOOSE_TOOLSHIM: true requirement; supports_streaming: false for Desktop; documented Desktop vs CLI differences
2026-07-06 19:39 v1.1 — Breaking: reasoning models (gpt-oss-120b, qwen3-14b, deepseek-r1-*) are fundamentally incompatible with Goose’s OpenAI streaming parser — goose drops tool_calls that follow reasoning_content chunks. Changed default model to llama-scout-17b. Added all available models to provider JSON. Updated model compatibility matrix, troubleshooting, recovery script. Config settings (reasoning: false, preserves_thinking: false) are necessary but NOT sufficient for reasoning models.
2026-07-06 19:25 v1.0 — Initial skill; extracted MaaS-specific content from goose-litellm-provider, added critical reasoning model fixes (reasoning: false, preserves_thinking: false), documented failure modes with evidence from real sessions, API key keyring handling, diagnostic tests