ARIS ships 83 skills for one research workflow
ARIS ships 83 Markdown skills that a coding agent can follow. They cover literature search, idea selection, experiment planning, code review, result analysis, paper writing, rebuttals, talks, and a persistent research wiki. Claude Code is the main executor in the documented path, while Codex provides an independent review pass. There are also routes for Codex-native use, Cursor, Copilot CLI, OpenClaw, and other hosts.
This is a large methods kit rather than one Python application. The commit we measured contained 712 files, roughly 50,251 lines of source, and a 38.1 MB checkout. That breadth is useful when a lab wants common records such as experiment plans and review state. It also raises the reading cost. A new user must decide which workflow owns a task, which model may judge it, and which files the agent may change.
Auto-proceed can run 4 review rounds without a checkpoint
AUTO_PROCEED defaults to true, while human checkpoints default to false. The workflow can search for ideas, implement experiments, monitor results, revise a paper, and repeat its review cycle. It stops at a score threshold or after 4 rounds by default, and proposed experiments above 4 GPU-hours are skipped for manual follow-up. Those bounds help, but they do not validate the code or the scientific claim.
The optional overnight permission block allows model review calls plus Write and Edit without asking each time. Effort can rise from a 0.4x lite mode to a documented 5 to 8x beast mode. Open issue 451 describes the practical cost: a short request may invoke an Astra-backed skill before the user realizes it. Use explicit skill selection, fixed compute limits, and a branch whose generated changes can be reviewed line by line.
What happened when we ran it
Our sandbox tested commit 341f914 in the Python project under mcp-servers/feishu-bridge/. It had 3 CPUs, 8 GB of RAM, Python 3.12 on Debian, no secrets, and no elevated privileges. Installation succeeded in 19 seconds, pulled 43 packages, and left a 96 MB environment. The build completed in 5 seconds. Pip-audit found 0 known vulnerabilities in that installed dependency set.
The test step failed after 4 seconds with exit code 5. Pytest reported no tests ran in 0.00s, which means 0 passed and 0 failed out of 0. The repository had a tests directory and 4 CI workflow files, yet the command in this measured subproject discovered nothing. The log does not show a missing package, import failure, or broken assertion, so claiming any of those causes would be guesswork.
This run says something narrow and useful. The Feishu bridge's environment installs and its build path works in a fresh unprivileged container. It does not show that notifications reach Lark, that replies resume a waiting research loop, or that the other 82 skills work together. Those paths need credentials, model access, external services, and tests that our command never found.
The 96 MB bridge is the easy part of setup
The 96 MB environment belongs to a small Python HTTP service with a lark-oapi requirement. It reads a Feishu app ID and secret, can target a user's open ID, and listens on port 5000 by default. The integration guide sends interactive users through Lark app creation, 5 permissions, event configuration, and an external feishu-claude-code bridge. Push-only mode is simpler and needs a group webhook.
ARIS adds setup across its 83-skill catalog. The recommended project install creates one symlink per skill and a manifest of managed entries. Existing linked skills update with git pull, while added or removed skills need the installer again. Paper generation requires LaTeX, latexmk, and pdfinfo. Review skills need a second model route. Experiment automation may need GPU and SSH details. The absence of a Dockerfile leaves teams to define their own reproducible runtime.
A September 29 push followed the v0.4.28 release
GitHub showed 16,820 stars and 68 combined open issues and pull requests on September 29, 2026. The open list contained 55 issues and 13 pull requests. The repository was pushed that day, one day after v0.4.28 shipped. Recent changes fixed paper-verification behavior, concurrent watchdog writes, reviewer configuration, and proxy handling. This is active maintenance, not a repository coasting on old attention.
Fast movement has a cost. Open issue 433 argues that the Codex mirror's many specialized skills remain eligible for implicit selection, adding context overhead and routing competition. The README recommends project-local and selective installation, which is the sensible response today. Install the research stages you can name and supervise. ARIS earns a pilot for a lab that wants structured overnight progress, but 0 discovered bridge tests are too little evidence for unattended use on the only copy of valuable research work.