Recently updated open-source projects

perf-optimization-loop refuses to edit code before a measured baseline exists. It isolates one experiment at a time, re-runs the same benchmark to decide KEEP, REVERT, or INCONCLUSIVE, records both wins and negatives, and generates a CI performance budget to catch regressions.

upgrade-pilot sequences major framework or dependency upgrades by dependency order, applies codemods before hand fixes, and isolates the work in a worktree. A resumable ledger records checkpoints, while install, typecheck, build, and test gates block the next step until the current one is green.

api-contract-governor governs REST, GraphQL, and protobuf/gRPC contract changes through a checked-in registry. It classifies deltas, discovers affected consumers, blocks unregistered breaking changes, and tracks versioned deprecation, migrations, and the final sunset gate before retirement.

extraction-automation-architect takes a web-extraction idea through legality and terms checks, source selection, API versus browser strategy, anti-bot constraints, typed schemas, selector healing, scheduling, validation, and monitoring. It produces durable configuration and run records instead of a single fetch. Rate limits, robots.txt, and site changes remain explicit operating constraints, so recurring jobs can be checked when data becomes stale.

medical-bill-and-claim-appeal-navigator turns a denial notice, itemized bill, and explanation of benefits into an error log, appeal materials, and a negotiation script. It can organize checks for duplicate charges, coding issues, unbundling, balance billing, and medical-necessity support, but never invents codes, statutes, or deadlines. It is not legal or medical advice; dates and citations must be confirmed for the plan and state.

agent-system-designer starts with a capability specification and routes the design among a skill, subagent, MCP server, or workflow. It defines tool contracts, context and state strategies, least-privilege permissions, safety boundaries, and a runnable test harness before implementation. The process is intended to catch overlapping schemas, context failures, and destructive access early, when architectural changes are still reversible.

learning-experience-architect applies backward design to turn source material into measurable objectives, a module map, lessons, and aligned assessments. Its completion check requires every objective to be taught and assessed without orphaned goals, then it can export the course to Markdown, SCORM, or an LMS-ready structure. The result is a learnable sequence rather than a flat outline.

This skill designs source-to-warehouse data ingestion with contracts, idempotency, duplicate protection, schema-drift checks, data-quality gates, lineage, backfills, freshness monitoring, and runbooks. It covers ETL, ELT, and CDC through ordered phases and blocks unsafe progression when a dependency or validation fails. The result is a recoverable workflow with clearer failure handling rather than a brittle one-off load.

raise-room-orchestrator runs the fundraise beyond a pitch deck: it coordinates the narrative and financial model, audits gaps in the data room, tracks investors, and prepares diligence Q&A from the company’s actual materials. The workflow is designed for rounds from pre-seed through Series B and keeps missing documents and follow-up work visible as the process moves forward.

pro-se-dispute-resolver turns a consumer or tenant problem into a dated fact record, demand letter, and route to the appropriate complaint forum. It helps check jurisdiction, citations, addresses, and deadlines, but it does not file anything, provide legal representation, or guarantee an outcome. Every fact and date must be verified before sending, with complex matters escalated to a licensed attorney.

newsletter-program-engine treats a newsletter as a recurring program rather than a one-off email. It helps define positioning and cadence, plan repeatable issues, segment the list, test subject lines, and design growth loops. Performance from each issue feeds the next planning cycle, giving the program state and a way to improve over time.

content-localization-orchestrator localizes marketing pages, docs, and product copy rather than only extracting code strings. It sets locale strategy, maintains glossaries and translation memory, supports transcreation, checks cultural and legal fit, and accounts for RTL direction and layout length. The result is prepared for coordinated publishing across multiple locales.

docs-product-orchestrator treats documentation as a product: it defines audiences and information architecture, separates tutorials, how-to guides, explanations, and reference, and uses Diataxis with docs-as-code. It can generate API reference from source, check links and examples in CI, detect drift, and verify the published site. The focus is a coherent, maintainable documentation experience rather than a single page.

authz-architect maps protected resources, actors, roles, and tenancy before choosing RBAC, ABAC, or ReBAC and an enforcement layer. A policy matrix drives consistent object-level checks and coverage tests, while adversarial probes target IDOR, privilege escalation, and cross-tenant leaks. The workflow requires previously failing cases to pass after remediation before authorization is attested.

relocation-logistics-orchestrator turns a cross-state or international move into a backward-planned checklist covering housing, healthcare and records, licenses and vehicle registration, address changes, schools, taxes, utilities, movers, and budget. It connects cross-domain dependencies and highlights jurisdiction-specific date windows. The plan still requires the household to confirm local rules and complete each task.

migration-conductor keeps a stateful ledger for schema and data migrations and gates the expand, migrate, and contract phases across deployments. It can drive tools such as pgroll, gh-ost, Atlas, or an ORM while retaining the same checks for hand-written SQL. Backfill parity, dual-write status, reader removal, and reversibility must be demonstrated before destructive cleanup, so unsafe phase skips are refused.

LLM App Productionizer guides a working LLM prototype through ordered engineering phases instead of treating production work as one prompt. It covers evaluation sets, prompt-injection defenses, output validation, retries and fallbacks, caching, token-cost limits, latency, tracing, canary releases, and rollback. The skill is aimed at teams that need evidence at each stage before expanding a prototype.

D3OXY/skills is a small pack of independently installable capabilities for coding agents. The included skills cover frontend design with an optional cinematic glass aesthetic, pull request preparation, and extraction of a domain glossary with ambiguity flags. The Skills CLI can list the pack, install one named skill, or install everything, so teams can keep their tool surface focused.