ARIS (Auto-Research-In-Sleep) is wanshuiyin’s MIT-licensed set of markdown-defined skills that turns Claude Code, Codex CLI, or another coding agent into an autonomous ML research loop, with a reviewer model from a different family critiquing every artifact at gated passes.
ARIS is this category’s first member whose judge is required to be a different model family: the executor drives and an independent reviewer (Codex MCP by default) demands revisions, an arrangement its technical report builds against the category’s shared failure mode, the plausible unsupported success.
What it is #
A methodology shipped as 87 markdown skills (83 in the standalone CLI bundle) with no framework and no lock-in, created 2026-03-10. The technical report (arXiv 2605.03042, May 2026, by Ruofeng Yang, Yongcan Li, and Shuai Li) describes three layers: an execution layer with more than 65 reusable markdown skills, MCP model integrations, and a persistent research wiki; a review layer where a reviewer from a different model family critiques intermediate artifacts and requests revisions; and deployment experience from long unattended runs. The skills cover the research arc end to end: idea discovery, experiment queues (including SSH multi-seed job queues), paper writing, citation audits, integrity forensics, and Overleaf sync. Adapters exist for Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw, and DeepSeek Harness. The family widened on November 8 with SuperARIS, a sibling MIT project that puts ARIS workflows in the hands of a persistent agent team (a coordinating main agent, Codex on experiments, Gemini on ideas and writing, or an all-Codex team), with its own repository and bilingual demo previews. Beyond the skills, the project ships a standalone Rust CLI (ARIS-Code, v0.4.28 on 2026-09-28, mid a 24-release train) and Claude Code, Codex CLI, and DeepSeek Harness plugins.
Status #
Active: 17,367 stars as of 2026-10-21, 1,498 forks, created 2026-03-10, last pushed 2026-11-19 (the November 17-19 batch: the community-QR refresh, PR #501’s watchdog tmux/screen liveness timeout fix, and PR #502’s anchored verify-papers arXiv ID match), following November 14’s install-script fixes (Windows symlink creation in bash environments, confirmation defaults matched with warnings on Windows copies, and prompt documentation), November 11’s SSH experiment-sync fix reusing validated environments, November 8’s SuperARIS introduction, November 3’s News entry featuring the Absent Author pattern, November 1-2’s bilingual News announcement of the model worker with a Goal Pulse reminder, a Claude Code Goal Pulse activation route, and session recovery connected to the HERO handoff skill, and October 31’s optional model-worker MCP server that offloads file authoring and writing skills, as of 2026-11-20.
October 2026 widened the loop three ways: a /lean-formalize skill that drives a mathematical statement to a checked Lean proof (October 7), Grok and Gemini/Antigravity MCP reviewer bridges that route reviews through existing CLI logins (October 5), and, on October 19, selectable reviewers inside the complete Claude Code and Codex skill packs, a /citation-crosscheck skill that diffs local BibTeX entries against freshly fetched records, and shared bounded fan-out rules capping parallel worker batches.
Late October added two community search skills, /keenable-search (keyless web search and page fetch, October 23) and /firecrawl-search (October 24), lifting the mainline count from 85 to 87, plus a refresh of the OpenAI model defaults and llm-chat request compatibility (October 24).
The technical report carries Hugging Face’s #1 Paper of the Day badge on the repository’s README, the only independent-attention signal in its footprint beside the PaperWeekly feature.
The Western discussion footprint is zero: a Hacker News search for the repository name returns no stories as of 2026-10-10, and adoption runs through a PaperWeekly feature and the Chinese-language community, the channel pattern DeepAnalyze shows; the Agon paper’s own count puts ARIS at 79 roles and 1,157.4 KiB of prompts, roughly four times Agon’s surface.
Strengths #
- The cross-model rule attacks a failure the kernel columns never face: a reviewer from a different family is less likely to inherit the executor’s framing, which is where the report locates unsupported claims.
- The skill set covers the whole arc, including integrity tooling (citation audit, fabrication forensics) most research loops leave to the human.
- No lock-in: the same markdown skills port across eight harnesses, so the loop survives any single vendor’s churn.
- The maintenance cadence is exceptional: 24 CLI releases between May and September 2026, including a same-week bridge when codex-cli 0.154 removed the reviewer’s entry point.
Cautions #
- The report’s deployment claims are self-published, and no independent replication surfaced in this run’s searches.
- The judge is still a model: cross-model review raises the bar but is not a kernel, a grader, or a benchmark.
- The prompt surface is large (79 roles, 1,157.4 KiB by the Agon paper’s count), which makes behavior harder to inspect than Agon’s 18 roles.
- The zero-story Hacker News footprint says the 17.2k stars measure attention in one language community, not broad engineering adoption.
- The loop depends on moving harness internals, as the codex-cli 0.154 breakage demonstrated.
Pricing #
Free and MIT; there is no paid tier, so pricing does not apply. Costs are your own model subscriptions or API keys (Claude executor plus a Codex reviewer by default, with ModelScope-hosted and local-model combinations documented) and your compute.
Compared to #
- Agon: the Claude Code plugin running producer-critic factories; ARIS is the skills-based loop whose reviewers come from other model families and which ports across harnesses.
- Karpathy Autoresearch: the founding keep-or-revert loop whose judge is a measured loss; ARIS’s judge is another model’s critique.
- OpenResearch: the workspace that turns your own agents into researchers with an archived evidence tree; ARIS adds adversarial review gates to the same idea.
Bottom line #
Recommended for ML researchers who want an unattended experiment loop with cross-model review, on the harness they already use, and for studying review-gate design against plausible unsupported successes. Not for anyone who needs a machine verifier behind results, a small inspectable prompt surface, or an independently replicated evaluation.
Changes #
- 2026-10-10 - Created.
- 2026-10-21 - Rotation re-verification: counted the main-branch skill set at 85 (from 82), recorded the October additions (the lean-formalize skill, the Grok and Gemini/Antigravity reviewer bridges, selectable reviewers, citation-crosscheck, bounded fan-out rules), moved the last push to October 20, and noted the report’s Hugging Face #1 Paper of the Day badge.
- 2026-10-25 - Volatile re-verification: recorded the two new community search skills (keenable-search and firecrawl-search, lifting the mainline count from 85 to 87), the October 24 OpenAI model-defaults refresh, and the push-date move to October 24.
- 2026-11-08 - Rotation re-verification: all eight references re-fetched and live (one API retry passed a rate limit), the skills directory still 91 raw entries (87 carrying SKILL.md by the counting rule), v0.4.28 still the latest release, and no default-branch commit after the November 6 WeChat refresh.
- 2026-11-11 - Volatile re-verification: recorded the November 8 SuperARIS introduction (the sibling team-of-agents project with its own repository) and the November 11 SSH experiment-sync fix, with the push date moved and v0.4.28 still the latest release.
- 2026-11-15 - Volatile re-verification: recorded the November 13-14 install-script fix batch (Windows symlink handling, confirmation defaults matched with warnings on Windows copies, install-script prompt docs), the push date moved to November 14, v0.4.28 still the latest release.
- 2026-11-20 - Volatile re-verification: recorded the November 17-19 batch (the community-QR refresh, PR #501’s watchdog tmux/screen liveness timeout fix, and PR #502’s anchored verify-papers arXiv ID match), the push date moved to November 19, v0.4.28 still the latest release.
See also #
- Agon - the producer-critic plugin that measured ARIS’s prompt surface
- Karpathy Autoresearch - the keep-or-revert ancestor of the loop family
- DeepAnalyze - the other thin-Western-footprint member, judged by itself where ARIS is judged by another model
- Automated Research Feature Matrix - the category comparison this note joins
References #
https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep - repository and README: skill set, harness adapters, executor-reviewer split, release train, HF #1 Paper of the Day badge (fetched 200, 2026-10-25)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep - stars, forks, created and pushed dates, MIT license, and topics for the as-of status (fetched 200, 2026-10-29 and 2026-11-20)
https://raw.githubusercontent.com/wanshuiyin/Auto-claude-code-research-in-sleep/main/README.md - the README source: skills, adapter list, ARIS-Code v0.4.28 details, the codex-cli 0.154 bridge note, the October what’s-new entries, and the SuperARIS family section (fetched 200, 2026-10-25 and 2026-11-11)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/contents/skills - the skills directory listing grounding the 87-skill count (91 entries minus the shared-references docs and the three Codex mirrors) (fetched 200, 2026-10-25)
https://arxiv.org/abs/2605.03042 - the technical report: cross-model adversarial collaboration, the plausible-unsupported-success failure mode, and the three-layer architecture (fetched 200, 2026-10-21)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/releases?per_page=5 - the release train through v0.4.28 (2026-09-28), still the latest (fetched 200, 2026-11-15)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/releases?per_page=3 - v0.4.28 (2026-09-28) still the latest release at the 2026-11-20 re-fetch (fetched 200, 2026-11-20)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/commits?since=2026-11-03T00:00:00Z&per_page=20 - the November 6 WeChat-group QR refresh behind that push-date move (documentation only), following the November 1-3 News commits (fetched 200, 2026-11-07)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/commits?since=2026-11-08T00:00:00Z&per_page=10 - the November 8 SuperARIS promotion and the November 11 SSH experiment-sync fix behind that push-date move (fetched 200, 2026-11-11)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/commits?since=2026-11-11T00:00:00Z&per_page=10 - the November 13-14 install-fix commits behind this push-date move (fetched 200, 2026-11-15)
https://api.github.com/repos/wanshuiyin/Auto-claude-code-research-in-sleep/commits?since=2026-11-15T00:00:00Z&per_page=10 - the November 17-19 batch behind this push-date move: the community-QR refresh, the PR #501 watchdog tmux/screen liveness timeout fix, and the PR #502 anchored verify-papers arXiv ID match (fetched 200, 2026-11-20)
https://api.github.com/repos/wanshuiyin/SuperARIS - the sibling project: created 2026-11-08, MIT, 77 stars as of 2026-11-11 (fetched 200, 2026-11-11)
https://hn.algolia.com/api/v1/search?query=claude-code-research-in-sleep - the zero-story Hacker News footprint (fetched 200, 2026-10-21)