Skip to content

armgit / cua

Public Imported from GitHub
0

Scale computer-use 2.0 with open-source drivers, cross-OS sandboxes, and benchmarks for training, evaluation, and data generation.

Fix repo README

ddupontfa080a57ddac
diff --git a/README.md b/README.md
index 932e06f220..3dfe9a86d0 100644
--- a/README.md
+++ b/README.md
@@ -1,261 +1,228 @@
-# Snorkel TerminalBench+ Edition 2 Data Sample
+<div align="center">
+ <a href="https://cua.ai" target="_blank" rel="noopener noreferrer">
+ <picture>
+ <source media="(prefers-color-scheme: dark)" alt="Cua logo" width="150" srcset="img/logo_white.svg">
+ <source media="(prefers-color-scheme: light)" alt="Cua logo" width="150" srcset="img/logo_black.svg">
+ <img alt="Cua logo" width="150" src="img/logo_black.svg">
+ </picture>
+ </a>
+
+ <p align="center">Build, benchmark, and deploy agents that use computers</p>
+
+ <p align="center">
+ <a href="https://cua.ai" target="_blank" rel="noopener noreferrer"><img src="https://img.shields.io/badge/cua.ai-0ea5e9" alt="cua.ai"></a>
+ <a href="https://discord.gg/mVnXXpdE85" target="_blank" rel="noopener noreferrer"><img src="https://img.shields.io/badge/Discord-Join%20Server-10b981?logo=discord&logoColor=white" alt="Discord"></a>
+ <a href="https://x.com/trycua" target="_blank" rel="noopener noreferrer"><img src="https://img.shields.io/twitter/follow/trycua?style=social" alt="Twitter"></a>
+ <a href="https://cua.ai/docs" target="_blank" rel="noopener noreferrer"><img src="https://img.shields.io/badge/Docs-0ea5e9.svg" alt="Documentation"></a>
+ <br>
+<a href="https://trendshift.io/repositories/13685" target="_blank"><img src="https://trendshift.io/api/badge/repositories/13685" alt="trycua%2Fcua | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+ </p>
+
+</div>
+
+## Choose Your Path
+
+<div align="center">
+ <table>
+ <tr>
+ <td colspan="3" align="center">
+ <a href="#cua-driver---background-computer-use-on-macos-and-windows-with-linux-pre-release">
+ <img src="img/card-cua-driver.png" alt="Cua Driver — The background computer-use agent" width="888">
+ </a>
+ </td>
+ </tr>
+ <tr>
+ <td align="center">
+ <a href="#cua---agentic-ui-automation--code-execution">
+ <img src="img/card-cua-sandbox.png" alt="Cua &amp; Cua Sandbox" width="280">
+ </a>
+ </td>
+ <td align="center">
+ <a href="#cua-bench---benchmarks--rl-environments">
+ <img src="img/card-cua-bench.png" alt="Cua Bench" width="280">
+ </a>
+ </td>
+ <td align="center">
+ <a href="#lume---macos-virtualization">
+ <img src="img/card-cua-lume.png" alt="Lume" width="280">
+ </a>
+ </td>
+ </tr>
+ </table>
+</div>
-**Version:** February 2026 (v2.0)
+---
-## Disclaimer
+## Cua Driver - Background computer-use on macOS and Windows, with Linux pre-release
-Disclaimer: These materials and data samples are provided solely for evaluation and discussion purposes and do not constitute a delivery under any existing agreement between Snorkel AI, Inc. and the recipient. All rights, title, and interest in these materials and data samples remain the exclusive property of Snorkel AI, Inc. The recipient may not copy, distribute, or use these materials or data samples for any commercial or competitive purpose without Snorkel AI Inc.'s prior written consent or a separately executed license agreement. The materials and data samples are provided "AS IS" and without warranties of any kind. By accessing or using the materials or data samples, the recipient agrees to these terms.
+Drive native desktop apps **in the background** — agents click, type, and verify without stealing the cursor or focus. Use the same CLI and MCP server on macOS and Windows from Claude Code, Cursor, Codex, OpenClaw, and custom clients. Linux support is available as a pre-release backend while platform testing is still in progress.
-## Overview
+**macOS / Linux pre-release**
-Terminal-Bench is a benchmark designed to evaluate large language model (LLM) agents in realistic, interactive computing environments, where they must execute complex, multi-step workflows using command-line tools and programming languages. It covers a diverse range of domains including software engineering, data science, machine learning, DevOps, database management, security, and scientific computing.
+```sh
+/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/trycua/cua/main/libs/cua-driver/scripts/install.sh)"
+```
-This dataset represents samples from **Snorkel TerminalBench+ Edition 2**, the second entry in the Snorkel Agentic Coding Data Series. Snorkel is a leading contributor to Terminal-Bench 2.0. More details on our contributions: https://snorkel.ai/blog/evaluating-coding-agent-capabilities-with-terminal-bench-snorkels-role-in-building-the-next-generation-benchmark/
+**Windows (PowerShell)**
-## What's New in Edition 2
+```powershell
+irm https://raw.githubusercontent.com/trycua/cua/main/libs/cua-driver/scripts/install.ps1 | iex
+```
-Edition 1 (released December 2025) provided 6,000 expert-authored tasks testing individual competencies in isolation. Edition 2 moves beyond single-skill evaluation to composite, multi-step tasks that mirror real engineering workflows — requiring agents to chain competencies across milestones in larger, more complex environments. Together, the two editions form a sequenced training curriculum. Edition 2 is one of several editions planned for release throughout 2026. The Terminal-Bench+ series will deliver 15,000+ expert-authored tasks.
+Full tool reference, architecture notes, and the optional agent skill pack live here: [`libs/cua-driver/README.md`](libs/cua-driver/README.md).
-### Key Innovations
+---
-**1. Milestone-Based Evaluation**
+## Cua - Agent-Ready Sandboxes for Any OS
-Edition 2 introduces tasks divided into 2–5 independently verifiable milestones that the agent must complete in sequence. Each milestone has its own test verifiers and rubrics, yielding 3–5x more training signal per task. This enables diagnostic precision — instead of knowing the agent failed, you know it failed at milestone 3 after succeeding at milestones 1 and 2.
+Build agents that see screens, click buttons, and complete tasks autonomously. One API for any VM or container image — cloud or local.
-**2. Real-World Context Layers**
+```sh
+pip install cua
+```
-Edition 1 defines what type of problem an agent must solve. Edition 2 adds a second axis defining the conditions under which the agent must work. These cut across all task types:
+<!-- <img src="img/cua-architecture.png" alt="Cua Architecture" width="100%"> -->
-| Context Layer | What It Tests |
-|---|---|
-| Long Context | Reasoning over 50,000+ token documents that cannot be parsed programmatically |
-| API Integration | Building, debugging, and interacting with APIs where source code is in the environment |
-| DB Interaction | Solving problems through direct database querying (SQL, NoSQL, vector databases) |
-| UI Building | Creating and modifying UIs verified via Playwright and Vitest |
-| Tool-Specific | Working with niche tools that have complex SDKs but limited training data representation |
+```python
+# Requires Python 3.11 or later
+from cua import Sandbox, Image
-**3. Codebase Scale Variation**
+# Same API regardless of OS or runtime
+async with Sandbox.ephemeral(Image.linux()) as sb: # or .macos() .windows() .android()
+ result = await sb.shell.run("echo hello")
+ screenshot = await sb.screenshot()
+ await sb.mouse.click(100, 200)
+ await sb.keyboard.type("Hello from Cua!")
+ await sb.mobile.gesture((100, 500), (100, 200)) # multi-touch gestures
+```
-Edition 1 tasks operate in contained environments. Edition 2 systematically varies environment complexity with labeled codebase sizes:
+| | Linux container | Linux VM | macOS | Windows | Android | BYOI (.qcow2, .iso) |
+| ------------------ | --------------- | -------- | ----- | ------- | ------- | ------------------- |
+| **Cloud (cua.ai)** | ✅ | ✅ | ✅ | ✅ | ✅ | 🔜 soon |
+| **Local (QEMU)** | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-| Scale | Description |
-|---|---|
-| Large | 200+ files, production-scale with documentation and test suites |
-| Small | 20+ files, aligned to smaller services and projects |
-| Minimal | Little to no codebase context — pure terminal tasks |
+**[Get Started](https://cua.ai/docs/cua/guide/get-started/set-up-sandbox)** | **[Examples](https://cua.ai/docs/cua/examples)** | **[API Reference](https://cua.ai/docs/cua/reference/agent-sdk)**
-### Edition 1 vs. Edition 2 at a Glance
+---
-| Dimension | Edition 1 | Edition 2 |
-|---|---|---|
-| Task scope | Atomic (1 competency per task) | Composite (2–5 competencies chained across milestones) |
-| Evaluation | Pass/fail + rubrics | Milestone-level verifiers + rubrics per milestone |
-| Reward signal | Binary (1 signal per task) | Dense trajectory (per milestone + per rubric) |
-| Real-world context | — | Long Context, API Integration, DB Interaction, UI Building, Tool-Specific |
-| Codebase scale | Not differentiated | Large / Small / Minimal, labeled in metadata |
-| Languages | Primarily Python | Python ≤50%, 5+ languages at ≥5%, multi-language tasks |
-| Difficulty calibration | GPT-5, Claude Sonnet 4.5 | GPT-5.2, Claude Opus 4.6 |
+## CuaBot - Co-op computer-use for any agent
-## Task Type Taxonomy
+<div align="center">
+ <img src="img/cuabot-screenshot.png" alt="cuabot screenshot" width="720">
+</div>
-The tasks in this dataset span multiple categories that reflect real-world technical work:
+`cuabot` gives any coding agent a seamless sandbox for computer-use. Individual windows appear natively on your desktop with H.265, shared clipboard, and audio.
-- **Software Engineering:** Tasks focused on developing or testing features and algorithms, fixing bugs and improving/optimizing existing features, implementing tests, or maintaining software projects.
+```bash
+npx cuabot # Setup onboarding
+```
-- **Machine Learning / Model Training / Inference:** Tasks requiring training, fine-tuning, running inference, or evaluating machine learning models, including dependency setup, running training loops, and managing data pipelines for ML tasks.
+```bash
+# Run any agent in a sandbox
+cuabot claude # Claude Code
+cuabot openclaw # OpenClaw in the sandbox
-- **System/Environment Setup & Configuration:** Tasks involving OS-level configuration, user management, package management, processes, or installing, configuring, and bringing up services, networks, and environments.
+# Run any GUI workflow in a sandbox
+cuabot chromium
+cuabot --screenshot
+cuabot --type "hello"
+cuabot --click <x> <y> [button]
+```
-- **Build / Compilation / Dependency Management:** Compile code, manage dependencies, and build components.
+Built-in support for `agent-browser` and `agent-device` (iOS, Android) out of the box.
-- **Data / File Processing / ETL / Scripting:** Tasks that transform, parse, filter, and aggregate datasets or files and directories to generate derived output.
+<div align="center">
-- **Interactive / Simulation Tasks / Games:** Tasks centered on game-like or simulated environments, interactive puzzles, or simulation games that run in the terminal.
+**[Get Started](https://cua.ai/docs/cuabot/guide/getting-started/introduction)** | **[Installation](https://cua.ai/docs/cuabot/guide/getting-started/installation)** | First spotted at [ClawCon](https://www.claw-con.com/)
-- **Debugging / Repair / Fixing Tasks:** Tasks that require identifying, diagnosing, and fixing errors in scripts, codebases, or system configurations.
+<img height="64" alt="cuaXclawdbot_nbg" src="https://github.com/user-attachments/assets/8b92237d-6e9b-4b3a-ae9a-b3560622ec1d" />
-- **Security / Cryptography / Vulnerability Demonstration:** Tasks related to cryptography, authentication, permissions, penetration-style tests, exploits, validating vulnerabilities, reverse engineering, or security configuration.
+</div>
-- **Scientific Computing:** Tasks using scientific libraries or workflows, such as numerical computation, simulations, or domain-specific research code.
+---
-## Data Sample
+## Cua-Bench - Benchmarks & RL Environments
-This sample contains **50 tasks** from the Snorkel TerminalBench+ offering spanning software engineering, scientific computing, debugging, data processing, security, machine learning, system administration, system setup & configuration, and games categories.
+Evaluate computer-use agents on OSWorld, ScreenSpot, Windows Arena, and custom tasks. Export trajectories for training.
-### Difficulty Tiers
+<!-- <img src="img/cua-bench-architecture.png" alt="Cua-Bench Architecture" width="100%"> -->
-Each task is categorized based on model performance using GPT-5.2 + Terminus 2 agent and Claude Opus 4.6 + Terminus 2 agent. Difficulty tiers are defined as follows:
+```bash
+# Install and create base image
+cd cua-bench
+uv tool install -e . && cb image create linux-docker
-| Tier | Definition |
-|---|---|
-| **Frontier** | Accuracy <= 20% on best model |
-| **Advanced Plus** | Accuracy <= 20% on worst model (exclusive of Frontier) |
-| **Advanced** | 20% < Accuracy <= 60% on worst model |
-| **Core** | 60% < Accuracy <= 80% on worst model |
+# Run benchmark with agent
+cb run dataset datasets/cua-bench-basic --agent cua-agent --max-parallel 4
+```
-The difficulty breakdown of these 50 sample tasks is:
+**[Get Started](https://cua.ai/docs/cuabench/guide/getting-started/first-steps)** | **[Partner With Us](https://cuabench.ai/)** | **[Registry](https://cuabench.ai/registry)** | **[CLI Reference](https://cua.ai/docs/cuabench/reference/cli-reference)**
-- 6 Frontier
-- 24 Advanced Plus
-- 19 Advanced
-- 1 Core
+---
-## Task Structure
+## Lume - macOS Virtualization
-Each task directory contains the following required files:
+Create and manage macOS/Linux VMs with near-native performance on Apple Silicon using Apple's Virtualization.Framework.
-- **`task.toml`** - Contains the task metadata including difficulty level, category, timeout settings, and other configuration parameters
+<!-- <img src="img/lume-architecture.png" alt="Lume Architecture" width="100%"> -->
-- **`instruction.md`** - Detailed task instructions in Markdown format that describe what needs to be accomplished
+```bash
+# Install Lume
+/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/trycua/cua/main/libs/lume/scripts/install.sh)"
-- **`environment/Dockerfile`** - Defines the Docker environment and dependencies required to run the task
+# Pull & start a macOS VM
+lume run macos-sequoia-vanilla:latest
+```
-- **`solution/solve.sh`** - Reference solution created by a human expert demonstrating how to complete the task correctly
+**[Get Started](https://cua.ai/docs/lume)** | **[FAQ](https://cua.ai/docs/lume/guide/getting-started/faq)** | **[CLI Reference](https://cua.ai/docs/lume/reference/cli-reference)**
-- **`tests/test.sh`** - Shell script that executes the test suite to verify task completion
+---
-- **`tests/test_outputs.py`** - Deterministic Python tests that check whether the task was solved correctly based on the final state of the environment (or `tests/test_outputs.spec.ts` for Playwright-verified UI tasks)
+## Packages
-- **`rubrics.txt`** - Evaluation criteria for LLM-as-judge scoring of the agent's trace, covering both process and outcome (or multiple `rubrics_x.txt` files for milestone tasks)
+| Package | Description |
+| --------------------------------------------------------------------------- | ---------------------------------------------------------- |
+| [cuabot](https://docs.trycua.com/cuabot/guide/getting-started/introduction) | Multi-agent computer-use sandbox CLI |
+| [cua-agent](https://cua.ai/docs/cua/reference/agent-sdk) | AI agent framework for computer-use tasks |
+| [cua-sandbox](https://cua.ai/docs/cua/reference/sandbox-sdk) | SDK for creating and controlling sandboxes |
+| [cua-computer-server](https://cua.ai/docs/cua/reference/sandbox-sdk) | Driver for UI interactions and code execution in sandboxes |
+| [cua-bench](https://cua.ai/docs/cuabench) | Benchmarks and RL environments for computer-use |
+| [lume](https://cua.ai/docs/lume) | macOS/Linux VM management on Apple Silicon |
+| [lumier](https://cua.ai/docs/lume/guide/advanced/lumier) | Docker-compatible interface for Lume VMs |
-Note: Tasks may also include additional files such as `milestones.md`, per-milestone test files (`tests/test_m*.py`), per-milestone solution scripts (`solution/solve*.sh`), and other data or configuration files specific to the task requirements.
+## Resources
-## Task Summaries
+- [Documentation](https://cua.ai/docs) — Guides, examples, and API reference
+- [Blog](https://www.cua.ai/blog) — Tutorials, updates, and research
+- [Discord](https://discord.com/invite/mVnXXpdE85) — Community support and discussions
+- [GitHub Issues](https://github.com/trycua/cua/issues) — Bug reports and feature requests
-This sample includes the following tasks:
+## Contributing
-### Software Engineering (18 tasks)
+We welcome contributions! See our [Contributing Guidelines](CONTRIBUTING.md) for details.
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `api-incident-escalation-repair-v5` | Advanced | Repair a broken Flask on-call service — fix incident listing, risk scoring, and escalation plan generation across 3 milestones. |
-| `cartpole-ppo-reward-shaping-grid-search` | Advanced Plus | Fix a broken PPO reward-grid search script for CartPole-v1 producing JSON/CSV artifacts with Pareto flags. |
-| `ci-pipeline-monitor` | Advanced | Fix a CI pipeline monitor extracting build logs and computing health metrics across 4 milestones. |
-| `data-migration-pipeline` | Advanced Plus | Fix a warehouse migration pipeline pulling shipment CSVs into SQLite with transformation and reporting across 4 milestones. |
-| `doc-analysis-api` | Advanced Plus | Extend a Flask app with a document analysis API supporting full-text search and TF-IDF processing across 3 milestones. |
-| `expense-report` | Advanced | Fix an expense report processor that reads JSON files, converts currencies, and generates reports across 4 milestones. |
-| `fix-json-patch` | Advanced Plus | Fix a broken RFC 6902 JSON Patch and RFC 6901 JSON Pointer implementation (small codebase). |
-| `github-actions-yaml-fix_20260307_134248` | Advanced Plus | Fix multiple errors in a GitHub Actions CI YAML preventing the pipeline from running (small codebase). |
-| `go-microservice-user-api` | Advanced Plus | Build a Go microservice with user CRUD operations on SQLite, fixing a broken database connection skeleton (small codebase). |
-| `gym-gridworld-multi-objective-pareto-cli` | Core | Fix a broken multi-objective Gymnasium grid-world CLI producing Pareto-optimal configs with JSON/CSV reports. |
-| `headless-git-release-surgery` | Advanced Plus | Fix a git repository left in a broken release state after partial cherry-pick/revert work and complete the 2.4.1 release. |
-| `jwt_auth_service_evolution_task` | Frontier | Build a Flask authentication microservice evolving through JWT auth, OAuth2 server, and social login across 3 milestones. |
-| `kuhn-poker-cfr-plus-selfplay-cli` | Advanced | Fix a broken CFR+ self-play implementation for Kuhn Poker producing correct exploitability and policy JSON. |
-| `kuhn-poker-openspiel-cfr-plus-policy-export` | Frontier | Fix an OpenSpiel Kuhn Poker CFR+ trainer and export average policy meeting strict exploitability thresholds. |
-| `multi-agent-pursuit-evasion-selfplay-ppo` | Frontier | Fix a broken multi-agent pursuit-evasion RL prototype with PPO self-play training. |
-| `redis-caching-express-js` | Advanced Plus | Review and fix a Redis-based caching layer added to an Express.js API. |
-| `text-analysis-visualization-react` | Advanced Plus | Build a CLI text analysis workflow for large documents and a React UI for visualizing generated insights. |
-| `unicode-filename-normalizer-cli` | Advanced | Fix a broken Python CLI that normalizes filesystem names for a migration script. |
+## License
-### Scientific Computing (10 tasks)
+MIT License — see [LICENSE](LICENSE.md) for details.
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `battery-dispatch-opt` | Advanced Plus | Write a dispatch controller for a grid-scale battery energy storage system to optimize buy/sell decisions. |
-| `chem-kinetics-sim` | Frontier | Build a CHEMKIN-format kinetics simulator across 4 milestones: thermodynamic parsing, rate constants, ODE integration, and sensitivity analysis. |
-| `dc-workload-opt` | Advanced | Write a resource controller for a data center with an indoor battery storage system to optimize workload scheduling (small codebase). |
-| `geochem-reactive-transport` | Advanced Plus | Build a geochemical reactive transport simulator with thermodynamics, equilibrium, and reactive flow across 4 milestones. |
-| `multiref-qchem-engine` | Advanced Plus | Build a multi-configurational quantum chemistry engine implementing CI, CASSCF, NEVPT2, and FMO2 methods across 4 milestones. |
-| `phase-coupling-inference` | Advanced Plus | Infer coupling coefficients from recordings of 30 coupled phase oscillators over 5000 timesteps. |
-| `pumped-hydro-dispatch` | Advanced Plus | Write a dispatch controller for a pumped hydro storage facility to optimize grid power usage and water pumping. |
-| `qchem-output-parser` | Advanced Plus | Build a Python library parsing output files from 5 quantum chemistry programs with UV-Vis simulation and thermochemistry across 4 milestones. |
-| `sparse-preconditioner-synthesis` | Advanced | Fix a sparse linear system solver driver that loads problem specs and applies preconditioning (small codebase). |
-| `xc-functional-library-v2` | Advanced | Build a Python exchange-correlation functional library with LDA, GGA, and hybrid functionals across 3 milestones. |
+Third-party components have their own licenses:
-### Debugging / Repair (6 tasks)
+- [Kasm](libs/kasm/LICENSE) (MIT)
+- [OmniParser](https://github.com/microsoft/OmniParser/blob/master/LICENSE) (CC-BY-4.0)
+- Optional `cua-agent[omni]` includes ultralytics (AGPL-3.0)
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `c-react-stats-task` | Advanced Plus | Fix a three-layer data visualization dashboard with a C statistical engine, Express API, and React frontend across 3 milestones (small codebase). |
-| `circuit-debugger` | Advanced Plus | Debug a logic circuit server with 8 two-input gates by probing inputs/outputs to identify the faulty gate. |
-| `etl-csv-image-tool` | Advanced Plus | Fix an Express/SQLite ETL pipeline for CSV ingestion, statistical analysis, and image processing across 3 milestones. |
-| `flaky-test-suite-debugging` | Advanced | Diagnose and fix a flaky Python test suite whose failures depend on execution order and shared state. |
-| `kotlin-mlflow-task` | Advanced | Debug a Kotlin linear regression project with MLflow tracking — fix build, model training, and inference API across 3 milestones. |
-| `python-docker-c-extension-fix` | Advanced Plus | Fix a Python application with a C extension module that builds successfully but crashes at runtime (small codebase). |
+## Trademarks
-### Data Processing (5 tasks)
+Apple, macOS, Ubuntu, Canonical, and Microsoft are trademarks of their respective owners. This project is not affiliated with or endorsed by these companies.
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `csv-profiler-cli-updated` | Advanced Plus | Build a CLI tool to profile CSV datasets with statistical summaries and visualization. |
-| `data-contract-compliance-enforcer-v5` | Advanced Plus | Enforce data contract compliance in a warehouse snapshot pipeline with schema auditing and quality checks across 3 milestones. |
-| `grand-slam-upset-index-v4` | Advanced | Build a ranked Grand Slam upset index from 1990–2024 match CSV data, grouped by surface and decade, with JSON output. |
-| `panel-position-extractor-v8` | Advanced Plus | Extract per-speaker positions and agreements from a ~55k-word panel transcript and produce structured JSON output. |
-| `sql-agent` | Frontier | Explore an unknown SQLite database, decode encoded values via a codec registry, and reconstruct a secret token through foreign key chain analysis. |
+---
-### Security (5 tasks)
+<div align="center">
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `issue-x509-server-certificates` | Advanced | Build a small PKI with OpenSSL: root CA, intermediate CA, and multiple X.509 server certificates with appropriate extensions. |
-| `linux-elf-malware-correlator` | Advanced | Analyze a Linux ELF malware sample — unpack if UPX-packed, run static analysis, and correlate with threat intel across 4 milestones. |
-| `linux-persistence-hunter-remediator-cli` | Advanced Plus | Build a CLI tool to hunt stealthy Linux persistence mechanisms and generate a remediation report across 4 milestones. |
-| `opaque-wire-decode` | Frontier | Reverse-engineer an undocumented industrial monitoring protocol from a binary capture and decode 397 length-prefixed messages. |
-| `secure-flask-api` | Advanced Plus | Audit and fix security vulnerabilities in a Flask REST API handling user registration, JWT authentication, and profile management (small codebase). |
+[![Stargazers over time](https://starchart.cc/trycua/cua.svg?variant=adaptive)](https://starchart.cc/trycua/cua)
-### Machine Learning (2 tasks)
+Thank you to all our [GitHub Sponsors](https://github.com/sponsors/trycua)!
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `debug-ml-training-pipeline` | Advanced | Fix a broken binary classifier training pipeline with a custom weighted focal loss on an imbalanced dataset (small codebase). |
-| `parallel-rollout-prioritized-replay-buffer` | Advanced | Fix a broken RL utility implementing parallel rollouts with a prioritized experience replay buffer. |
+<img width="300" alt="coderabbit-cli" src="https://github.com/user-attachments/assets/23a98e38-7897-4043-8ef7-eb990520dccc" />
-### System Administration (1 task)
-
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `cis-level1-linux-compliance-cli` | Advanced | Build a CIS Level 1 Linux compliance checker that evaluates system controls and emits a remediation script across 4 milestones. |
-
-### System Setup & Configuration (1 task)
-
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `terraform-gcp-env` | Advanced | Create production-ready Terraform config for a GCP environment with VPC, Cloud SQL, GCS, and Redis with remote state. |
-
-### Games (2 tasks)
-
-| Task | Difficulty | Description |
-|------|------------|-------------|
-| `blind-auction` | Advanced | Play a blind auction game via HTTP API — maximize winnings across 8 rounds by bidding strategically on items. |
-| `process-scheduler` | Advanced | Interact with a CPU job scheduler server to produce optimal process schedules for 5 evaluation seeds. |
-
-## Milestone Tasks
-
-The following tasks use a milestone-based structure where the agent must complete sequential stages of increasing complexity:
-
-- **`api-incident-escalation-repair-v5`** (3 milestones): Repair incident listing/filtering/pagination; repair risk scoring; generate escalation plan with checksum.
-
-- **`c-react-stats-task`** (3 milestones): Build System & Server Startup; Statistical Computation; Visualization & Integration.
-
-- **`chem-kinetics-sim`** (4 milestones): Parse CHEMKIN thermodynamic properties; compute rate constants with Troe falloff; integrate reactor ODEs; perform sensitivity and rate-of-production analysis.
-
-- **`ci-pipeline-monitor`** (4 milestones): Fix data parsing and aggregation; fix flaky test detection and build health; fix stage ranking and report summary; fix trend analysis.
-
-- **`cis-level1-linux-compliance-cli`** (4 milestones): Load benchmark JSON and implement control parser; scan system settings and evaluate controls; produce compliance report JSON; emit POSIX-compliant remediation script.
-
-- **`data-contract-compliance-enforcer-v5`** (3 milestones): Build schema auditor; build quality and freshness checks; build compliance summarizer.
-
-- **`data-migration-pipeline`** (4 milestones): Fix data loading; fix processing and reporting; fix data enrichment; fix audit trail.
-
-- **`doc-analysis-api`** (3 milestones): Set up document fetching API endpoints; implement full-document analysis with TF-IDF, bigrams, and similarity; build results listing and aggregate statistics.
-
-- **`etl-csv-image-tool`** (3 milestones): CSV Ingestion Pipeline; Analysis Engine; Advanced Analytics and Dashboard.
-
-- **`expense-report`** (4 milestones): Fix data loading and parsing; correct reporting scope and limit checks; fix approver dict mutation and categorizer sorting; resolve numeric tax edge-cases and per-diem dates.
-
-- **`geochem-reactive-transport`** (4 milestones): Thermodynamic database parser and activity coefficient models; Gibbs energy minimization with Newton-Raphson; 1D reactive transport with MUSCL advection-dispersion; mineral saturation indices and benchmark validation.
-
-- **`jwt_auth_service_evolution_task`** (3 milestones): JWT Authentication Foundation; OAuth2 Authorization Server; Social Login Integration.
-
-- **`kotlin-mlflow-task`** (3 milestones): Fix Kotlin build environment; fix model training with correct normalization and MLflow integration; fix inference API route paths and prediction denormalization.
-
-- **`linux-elf-malware-correlator`** (4 milestones): Detect and unpack UPX sample; produce static imports and strings; run strace and capture syscalls; correlate and output structured report JSON.
-
-- **`linux-persistence-hunter-remediator-cli`** (4 milestones): Hunter CLI with inventory mode; classification against allowlist; safe remediation with atomic report; validate mode confirming persistence state.
-
-- **`multiref-qchem-engine`** (4 milestones): Full CI with Davidson diagonalization; CASSCF orbital optimization; NEVPT2 perturbation theory; restricted Hartree-Fock and FMO2.
-
-- **`qchem-output-parser`** (4 milestones): Gaussian/ORCA parsers with core unit conversion; NWChem/GAMESS/Psi4 parsers with auto-detection; UV-Vis simulation and thermochemistry; JSON serialization and CLI.
-
-- **`xc-functional-library-v2`** (3 milestones): LDA functionals with analytical derivatives; GGA functionals with gradient-dependent derivatives; hybrid functionals with expression parser and finite-difference verification.
+</div>