{"full_name":"EvoScientist/EvoScientist","owner":"EvoScientist","name":"EvoScientist","shard":"1","url":"https://github.com/EvoScientist/EvoScientist","description":"🔬 Harness Vibe Research with Self-evolving AI Scientists","homepage":"https://EvoScientist.ai/","topics":["ai-agent","ai4science","multi-agent-system","vibe-research"],"tag_pages":["ai-agent","ai4science","multi-agent-system"],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/257286936?v=4","created_at":"2026-01-26T11:19:26Z","pushed_at":"2026-11-14T15:36:31Z","metrics":{"stars":4988,"forks":370,"watchers":47,"issues":284,"prs":535,"releases":46,"issues_open":57,"issues_closed":227,"prs_open":21,"prs_closed":80,"prs_merged":434,"discussions":3},"growth":{"day":{"stars":9,"forks":1,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[4942,4945,4960,4965,4970,4979,4988],"forks":[351,354,362,365,370,369,370],"watchers":[null,47,47,47,47,47,47],"issues_open":[null,47,57,60,57,57,57],"prs_open":[null,11,17,18,19,21,21]},"boards":{},"same_owner":[],"issues":[{"number":512,"title":"[Tech RoadMap]: One-line install, desktop app and concurrent workspaces","comments":6,"created_at":"2026-09-25T16:53:58Z"},{"number":629,"title":"TUI: with two sub-agents waiting for approval, an approved command can run twice","comments":5,"created_at":"2026-10-13T19:44:12Z"},{"number":580,"title":"Follow-ups to the MCP get_tools time limit from #561","comments":4,"created_at":"2026-10-06T10:01:54Z"},{"number":715,"title":"feat(sandbox): close the remaining gaps in the cd, rm and kill checks","comments":3,"created_at":"2026-10-28T18:03:34Z"},{"number":413,"title":"[langgraph dev] Unpin the workspace and run as an OS-managed service","comments":3,"created_at":"2026-08-07T02:43:11Z"},{"number":496,"title":"Server-backend surfaces still build the in-process agent (double MCP load, wasted cold start)","comments":2,"created_at":"2026-09-22T11:48:24Z"},{"number":484,"title":"[Feature] Windows installer with a lightweight WebUI desktop shell","comments":2,"created_at":"2026-09-16T16:56:34Z"},{"number":833,"title":"Sandbox: after \"cd\", a second system path in the same word is not checked","comments":1,"created_at":"2026-11-13T16:30:04Z"}],"discussions":[{"number":760,"title":"Would EvoScientist be a named pilot for an open tool that records why AI-generated code changed?","comments":0,"category":"General","created_at":"2026-11-08T04:34:25Z"},{"number":288,"title":"Question about evolution rounds and human involvement behind reported results","comments":0,"category":"Q&A","created_at":"2026-06-16T01:15:59Z"},{"number":53,"title":"🚀 Feature Request: Remote SSH GPU Support for Experiment Execution","comments":2,"category":"General","created_at":"2026-03-16T07:10:29Z"}],"latest_release":{"name":"v0.4.3","tag":"v0.4.3","published_at":"2026-11-13T22:45:43Z","body":"# EvoScientist v0.4.3 — New Claude & Gemini Models, Projects That Follow Their Folder & ccproxy in the Background Server\n\nG'Day! v0.4.3 brings **three new models: Claude Fable 5.2, Gemini 4 Carbon and Gemini 4 Hydrogen**, and tool selection works again on Gemini through OpenRouter. **A project now follows its folder**: `EvoSci workspace move` carries a moved or renamed project over. **The background server runs the ccproxy behind Claude Pro/Max sign-in**, and `EvoSci server` now does what `EvoSci service` and `EvoSci ccproxy` did. The server also provides what the next WebUI release needs to **review AutoSkills drafts**. Channels now send the media files the agent writes or reads, keep same-named attachments apart and keep the blank line between a reply's messages. The file tools and the shell sandbox read more path spellings the same way, Windows drive and Git Bash paths included, and a run that fails on the server ends its turn with the error. Under the hood: deepagents 0.7.38, langchain 1.5 and langchain-core 1.8.\n\n---\n\n### 🔌 Providers & models (#768, #804, #807)\n\n- **Claude Fable 5.2** — `claude-fable-5-2` on the Anthropic API and `anthropic/claude-fable-5.2` on OpenRouter. With preserved thinking, Claude refuses a request when a thinking block it replays belongs to a conversation that has changed (\"Invalid `signature` in `thinking` block\"). That request is now sent again instead of failing the turn. The first-contact counters no longer rewrite `USER_PROFILE.md` once you have introduced yourself, and the rubric grader no longer mistakes a reply for reasoning. Known: Fable 5.2 on OpenRouter cannot be the rubric grader (set `auxiliary_model`); with more than 42 tools, the tool selector's first request on Fable 5.2 is refused once until langchain-anthropic lists the model; Claude sign-in through ccproxy is not tested with it. Thanks @din0s and @jfilipiuk for the reviews!\n- **Gemini 4 Carbon and Gemini 4 Hydrogen** — `gemini-4-carbon` / `gemini-4-hydrogen` on Google, `google/gemini-4-carbon` / `google/gemini-4-hydrogen` on OpenRouter, 1M context. Both were tested on OpenRouter as the main and the auxiliary model: rubric grading, goal drafting, tool selection, memory workers, AutoSkills and a scheduled task with a rubric. Gemini 4 Hydrogen is now preselected when you pick Google GenAI in `EvoSci onboard`.\n- **Tool selection on Gemini through OpenRouter** — with more than 42 tools (MCP-heavy setups), tool selection now returns on Gemini models through OpenRouter; before, the turn showed nothing. The \"Adaptive Selected Tools\" panel is back in the TUI and the terminal, and as the `tool_selection` line of `--output-format stream-json`. Known: Gemini's reasoning text on OpenRouter appears only when the route returns it; the panel appears only when the selection differs from the last one shown.\n- **A missing Anthropic key fails at once** — with the native Anthropic provider and no API key, a turn now stops at once with \"No Anthropic API key is set\" and how to set it, instead of `Connection error.` after the SDK's retries. The model still builds without a key, so the first start of the WebUI and the desktop app is unaffected. Thanks @Gennifer-Dominic!\n\n### 🧠 AutoSkills & memory (#765)\n\n- **AutoSkills review for the WebUI** — the background server now serves what the WebUI's review page needs: the drafts of every project, their files and a diff, approve / reject / Undo, and Run now. Each draft records who decided it and when. AutoSkills settings apply without a restart. The review page itself comes with the next WebUI release. Known: Undo is refused for an approved update; Run now's \"already running\" check does not count runs the schedule started. Thanks @din0s and @jfilipiuk for the reviews!\n- **Linked skill folders** — a skill folder that is a symbolic link inside a skills folder is now listed under its own name, and an approved AutoSkills update is written to the copy the agent actually loads. Known: when the global skills folder co","is_prerelease":false},"readme":true,"readme_zh":true,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}