Recently updated open-source projects

VitaBench evaluates agents in real-world service scenarios such as delivery, in-store operations, and OTA, using databases, API tools, and 100 tasks per domain. Configure the models, run the vita command, save simulations under data/simulations, and re-evaluate existing runs when needed. The current release updates datasets, tools, evaluator models, and metrics, so scores should be interpreted alongside the chosen configuration.

Fable Workflow teaches coding agents to inspect the real codebase and expose unspecified decisions before implementation instead of guessing silently. During the work it records assumptions and deviations, then asks for verification evidence before declaring completion. The Claude Code plugin adds advisory hooks and an optional strict stop gate, though its heuristics can occasionally misfire.

image-prompt takes a brief idea, identifies the visual archetype, and expands it into a structured prompt covering composition zones, typography, palette, mood, and finish for posters, ads, UI mockups, game art, logos, and more. It can adapt the same concept for several image models and reuse private style presets; image generation is optional, requires a configured provider key, and may incur API charges.

DvalinCode targets regulated teams that need AI coding under explicit policy, local data controls, and reviewable run records. Chat stays read-only, Cowork previews and asks for each write, while Code can run routines; the runtime also imports scanner findings into isolated remediation worktrees and supports OpenAI-compatible or local models.

Kodo assigns a goal to coding backends such as Claude Code, Cursor, Codex, Gemini, or Kimi and coordinates repeated work cycles with independent verification. It can implement features, run improvement reviews, test software through realistic user workflows, and fix findings from earlier runs, with effort levels controlling how aggressively agents iterate. At least one supported coding-agent backend must be installed.

tower-defense-skill is a native Canvas 2D tower-defense game built around protecting a goddess core. Choose from two maps and three difficulties, build and upgrade six tower types, exploit fire, ice, and lightning matchups, and survive enemy and boss waves. Souls earned during waves unlock hero draws, while deployed heroes can receive a patrol radius; results include progression, achievements, and map-specific rankings.

Use this Claude Code skill for large, well-specified implementation work that can be divided into task packets. The main model handles decomposition and review, cheaper workers write code, and type checks or tests gate their output. Lite mode reduces top-tier quota use at relatively low total cost; max mode saves more top-tier quota but raised total dollar cost in the benchmark. The skill steps aside for small tasks and judgment-heavy debugging where delegation is not worthwhile.

prompt-injection-review maps trust boundaries in LLM applications that use RAG, external content, tools, agents, or model-rendered output, tracing untrusted sources through the model to outbound actions, execution, data access, or UI sinks. It recommends least privilege, human approval, egress controls, structured validation, and authorization enforced in code. It explicitly treats prompt injection as having no complete fix: the goal is to contain what a successful injection can reach, not to rely on filters for false certainty.

llm-cost-guard statically reviews LLM-calling code for repeated uncached prefixes, oversized model tiers, bloated context, unbounded output, unbudgeted agent loops, and repeated embeddings. It estimates the dominant tokens-times-price-times-volume term, then proposes prompt caching, model routing, context trimming, output caps, or batching while naming latency and quality trade-offs. It has no billing access, and projected savings still require current prices and evaluation results.

api-break-check reviews API specifications, schemas, protobuf definitions, or implementation changes for removed response fields, newly required inputs, narrowed types, removed enum values, reused field numbers, and other breaking effects. It reasons about request-versus-response direction, explains which consumers fail, and proposes additive fields, deprecation windows, versioning, or reserved numbers. It is a pre-merge review rather than a substitute for contract-diff tools and direct consumer validation.

terraform-blast-radius reviews Terraform or OpenTofu plans and configuration diffs for forced replacements, resource renames, count-index shifts, dangerous state operations, and shared-resource cascades, distinguishing downtime from data-loss risk. It can propose prevent_destroy, create_before_destroy, moved blocks, for_each, or snapshot-based migrations without cloud access. Its output is a pre-apply review, not a guarantee, so the real plan, provider version, and tested backups still need checking.

n-plus-one-hunter statically reviews controllers, serializers, resolvers, and query code for classic or nested N+1s, per-row aggregates, cartesian eager-load blowups, and related patterns. It selects ORM-specific preload or join strategies for Rails, Django, Prisma, SQLAlchemy, and other listed stacks, explains the scaling cost, and reports a before-and-after query count. The proposed fix still needs verification against the actual ORM version, indexes, access pattern, and realistic data.

safe-migrations reviews PostgreSQL and MySQL schema changes for locks, table scans, and downtime risks before they merge. It explains the blocking mechanism, rewrites risky operations into copy-ready staged steps in SQL or framework idioms, and sequences expand-contract deployments when application changes must remain compatible.

This Agent Skills collection covers a video subtitle workflow from audio transcription to publishing support. It converts common audio formats into SRT files, can refine wording and proper nouns against an `origin.md` transcript without changing timestamps, adds card annotations, and generates platform-specific summaries for Facebook, Threads, and YouTube. Each stage can be invoked separately or combined into the provided workflow. Users need to prepare the expected files in the same working directory and install the dependencies required by the individual skills.

ARC-AGI-2 contains 1,000 public training tasks and 120 public evaluation tasks built from small colored-integer grids. Solvers infer a transformation from demonstration pairs and must produce every test output with the correct dimensions and cells. The evaluation set is intended for previously unseen testing, so repeatedly tuning against it would compromise the result.

Yagua brings common image-processing jobs into a local desktop interface for designers, photographers, and web developers. It can batch-compress and convert files, resize or crop them, remove backgrounds, run OCR, vectorize images, and edit EXIF data without relying on a subscription web service. Windows and Linux are the main supported platforms, macOS remains experimental, and each module has its own practical batch limit.

Check an academic manuscript before submission for fabricated references, incorrect DOIs, bibliographic errors, missing text-to-list matches, and citations that do not support the associated claim. The skill accepts DOCX, TeX, and Bib inputs and uses CrossRef plus web search for cross-checking.

report-writing is a reusable method library for formal DOCX reports, especially Traditional Chinese government and tender deliverables in Taiwan. It covers requirements, content strategy, layout, equations, clickable cross-references, page-numbered tables of contents, figures, and delivery checks, with linters for wording and English AI tells. The skill works with Claude Code and Codex and focuses on verified document output.