# autoresearch

> Claude Autoresearch Skill — Autonomous goal-directed iteration for Claude Code. Inspired by Karpathy's autoresearch. Modify → Verify → Keep/Discard → Repeat forever.

- **URL**: https://www.freshcrate.ai/projects/autoresearch
- **Author**: uditgoenka
- **Category**: AI Agents
- **Latest version**: `v2.1.2` (2026-05-23)
- **License**: MIT
- **Source**: https://github.com/uditgoenka/autoresearch
- **Homepage**: https://udit.co/projects/autoresearch
- **Language**: Shell
- **GitHub**: 3,922 stars, 295 forks
- **Registry**: github
- **Tags**: `ai`, `autonomous-agent`, `autoresearch`, `claude`, `claude-code`, `iteration`, `karpathy`, `productivity`, `shell`

## Description

Claude Autoresearch Skill — Autonomous goal-directed iteration for Claude Code. Inspired by Karpathy's autoresearch. Modify → Verify → Keep/Discard → Repeat forever.

## Recent releases

| Version | Date | Urgency | Changes |
| --- | --- | --- | --- |
| `v2.1.2` | 2026-05-23 | High | ## What's New  ### `/autoresearch:improve` — Research What to Build Next  New subcommand for product companies. Fills the "what should we build next?" gap in a family that previously only covered "how to build it right."  **How it works:** 1. **Product Context** — resolves from learn summary, README, package.json, or auto-discovers 2. **Research Loop** — 5 categories (ICP challenges, competitor gaps, market trends, UX & experience, revenue & growth) with saturation-based termination 3. **Feature |
| `v2.1.1` | 2026-05-22 | High | ## What's New  **9 auto-firing hooks** that ship as part of the Claude Code plugin. Zero configuration — they activate on `npx skills add uditgoenka/autoresearch`.  ### Safety Gates (PreToolUse)  \| Hook \| What it does \| \|------\|-------------\| \| `scout-block` \| Blocks vendor dirs, `.git/`, `__pycache__/`, `dist/`, `build/`, `coverage/` — prevents context bloat. Loads `.ckignore` for per-project customization. Smart Bash argument parsing prevents false positives on string literals. \| \| `privacy-bl |
| `v2.0.04` | 2026-05-06 | High | ## What's New  ### Reliable Skill Triggering The autoresearch skill description across all 5 distributions (Claude Code, Claude Plugin, OpenCode, Agents, Codex) has been changed from passive "Use when user types..." to imperative "ALWAYS activate...MUST...BLOCKING". This matches the pattern used by reliably-triggering skills like `cook` and `fix`, ensuring autoresearch subcommands activate consistently without requiring users to add emphatic language like "strictly trigger" or "must must trigger |
| `v2.0.03` | 2026-05-02 | High | ## What's Changed * Release v2.0.03 by @uditgoenka in https://github.com/uditgoenka/autoresearch/pull/85   **Full Changelog**: https://github.com/uditgoenka/autoresearch/compare/v2.0.2...v2.0.03 |
| `v2.0.0` | 2026-04-28 | High | ## v2.0.0 — Multi-Platform GA Release  Promotes v2.0.0-beta to stable. **Claude Code + OpenCode + Codex** all fully supported with strict YAML compliance, security-hardened scripts, and complete command metadata.  ### What's New in v2.0.0  - **11 subcommands:** plan, debug, fix, security, ship, scenario, predict, learn, reason, probe (new) - **3 platforms:** Claude Code, OpenCode, OpenAI Codex - **`/autoresearch:probe`** — adversarial multi-persona requirement interrogation engine (new in this r |
| `v1.9.12` | 2026-04-15 | High | ## What's fixed  `/autoresearch` and its 9 subcommands (`plan`, `debug`, `fix`, `security`, `ship`, `scenario`, `predict`, `learn`, `reason`) now dispatch reliably on the **first** invocation. Previously the skill failed to trigger ~80% of the time, requiring repeated attempts (often 10+) before Claude Code would activate it.  ## Root cause  Claude Code routes slash commands through a **description-based fuzzy matcher**. None of the skill or command descriptions contained the literal `/autoresea |
| `v2.0.0-beta.0.2` | 2026-04-06 | Medium | ## What's New  ### OpenAI Codex Support Autoresearch now runs on **three platforms**: Claude Code, OpenCode, and OpenAI Codex.  **Codex-specific:** - Skills installed to `.agents/skills/autoresearch/` (Codex standard path) - Invoked via `$autoresearch` mention syntax (not slash commands) - Subcommands are keywords: `$autoresearch plan`, `$autoresearch debug`, `$autoresearch security`, etc. - Includes `agents/openai.yaml` for Codex UI metadata (display name, brand color, implicit invocation) - Gl |
| `v2.0.0-beta.0.1` | 2026-04-06 | Medium | ## 🧪 Beta Release — OpenCode Support  > **This is a beta release.** OpenCode support is new and needs community testing before stable promotion.  ### What's New  **Full OpenCode port of autoresearch** — all 10 commands, all 12 reference workflows, adapted for OpenCode's tool and command conventions.  #### OpenCode Support - **10 commands:** `/autoresearch`, `/autoresearch_plan`, `/autoresearch_debug`, `/autoresearch_fix`, `/autoresearch_security`, `/autoresearch_ship`, `/autoresearch_scenario`, |
| `v1.9.11` | 2026-04-06 | High | ## What's New  ### Metric Extraction Validation (PR #63) - **Mandatory numeric validation** — extracted values must match `^-?[0-9]+\.?[0-9]*$` before any decision logic runs - **`metric-error` status** — new iteration status for non-numeric extraction failures - **Two-consecutive-error halt** — stops the loop (even unbounded) when the verify pipeline is confirmed broken - **Diagnostic output** — shows raw verify output on failure so the problem is visible - **Whitespace trim** — strips leading/ |
| `v1.9.1` | 2026-04-06 | Medium | ## What's New  ### AGENTS.md — Universal AI Agent Onboarding  Added `AGENTS.md` to the project root so that any AI coding agent (Claude Code, Codex, OpenCode, Gemini CLI, etc.) can immediately discover and use all 10 autoresearch commands without reading the full README or skill files.  **Covers:** - Installation (Claude Code plugin, Codex plugin, manual copy) - All 10 commands with usage examples - Configuration fields (Goal, Scope, Metric, Verify, Guard, Iterations) - Per-command flag referenc |

## Dependency audit

- **Score**: 100/100
- **Total deps**: 0
- **Resolved**: 0
- **Unresolved**: 0
- **License conflicts**: 0
- **Warnings**: 0
- **Scanned**: 2026-05-25

## Citation

- HTML: https://www.freshcrate.ai/projects/autoresearch
- Markdown: https://www.freshcrate.ai/projects/autoresearch.md
- Dependencies JSON: https://www.freshcrate.ai/api/projects/autoresearch/deps

_Generated by freshcrate.ai. Indexes github releases for AI-agent ecosystem packages._