简介
本版本移除了一些过时的策略，减少针对特定模型的行为改变：上下文与输出上限、推理档位改从 models.dev 和用户配置读取，不再按模型名推断，也不再默认发送 temperature / top_p；/variants 可在配置中直接按档位名启用，子 agent 继承父会话的档位；GPT 模型也不再自动开启 Codex 风格的提示词和工具集，所有模型默认使用同一套（需要时可设置 MIMOCODE_CODEX_MODE=true）。
此外，内置联网搜索改为通过 MIMOCODE_ENABLE_EXA=1 启用（也可接入搜索 MCP）；小米 API key 与 Token Plan key 可同时登录；记忆搜索支持中文子串召回；并修复了额度耗尽仍重试、压缩失败残留等稳定性问题。
Highlights
This release removes some outdated policies and reduces model-specific behavior changes: context/output limits and reasoning levels now come from models.dev and your config rather than model-name heuristics, and temperature/top_p are no longer sent by default; reasoning levels for /variants can be enabled by name in config, and subagents inherit the parent session's level; GPT models also no longer switch to a Codex-like prompt and toolset automatically — every model uses the same defaults (set MIMOCODE_CODEX_MODE=true to opt in).
Also: built-in web search is now opt-in via MIMOCODE_ENABLE_EXA=1 (or add a search MCP server), Xiaomi API and Token Plan keys can be signed in together, memory search recalls Chinese substrings, and reliability fixes cover retries on exhausted quota and leftovers from failed compactions.
What's Changed
- fix(provider): stop rewriting adapter npm for mimo-named model ids by @yanyihan-xiaomi in #2717
- fix(cli): unpack with the Windows system tar, not the first tar on PATH by @agenthandler in #2721
- test(provider): cover Anthropic routing for MiMo models by @winhok in #2692
- docs(readme): refresh desktop screenshots and download links by @yanyihan-xiaomi in #2722
- test(workflow): slim global concurrency ceiling to 2x2 by @yanyihan-xiaomi in #2723
- feat(mcp): drive always-direct MCP tools from config by @wqymi in #2734
- fix(session): stop retrying exhausted quota by @peipeiwang-xiaomi in #2732
- feat(session): accept a caller-supplied session id on create by @yanyihan-xiaomi in #2738
- fix(provider): store xiaomi API and token-plan credentials under separate providers by @yanyihan-xiaomi in #2739
- fix(session): leave no boundary behind a failed compaction by @yanyihan-xiaomi in #2743
- fix(websearch): drop Xiaomi gateway search and make Exa search opt-in by @yanyihan-xiaomi in #2744
- fix(workflow): restore the DuckDuckGo fallback in research prompts by @yanyihan-xiaomi in #2745
- feat(provider): drive reasoning variants from models.dev, let config name levels, and pass the parent variant to subagents by @yanyihan-xiaomi in #2768
- fix(memory): support Chinese substring recall and migrate existing indexes by @peipeiwang-xiaomi in #2765
- fix(tool): publish applyable file-tool diffs and trim only for display by @yanyihan-xiaomi in #2771
- fix(history): stop over-fetching search rows now that each part has one index row by @yanyihan-xiaomi in #2772
- fix(share): remove cloud session sharing and third-party uploads by @peipeiwang-xiaomi in #2770
- feat(session): identify synthetic reminders by tag and inject a per-window session environment reminder by @yanyihan-xiaomi in #2776
- docs(security): document that non-interactive entry points skip workspace trust by @yanyihan-xiaomi in #2780
- docs(skill): give compose-next Report and Journey log explicit empty-state lines by @yanyihan-xiaomi in #2790
- feat(harness): make Codex mode opt-in and unify the default system prompt by @yanyihan-xiaomi in #2795
- docs(skill): compose-next spec presentation and sub-spec split for large features by @yanyihan-xiaomi in #2797
- fix(session): bound empty-step loops for checkpoint writers and stop logging request bodies by @wqymi in #2798
- feat(provider): take limits, sampling, effort and step termination from the model by @yanyihan-xiaomi in #2799
- fix(sdk): regenerate SDK and make MCP alwaysDirectTools mutable by @yanyihan-xiaomi in #2801
- docs: group builtin skills by category and sync docs with changes since 0.1.16 by @yanyihan-xiaomi in #2802
- chore: bump version to 0.1.17 by @yanyihan-xiaomi in #2803
New Contributors
- @agenthandler made their first contribution in #2721
- @winhok made their first contribution in #2692
Full Changelog: v0.1.16...v0.1.17