git

A library of ten read-only persona reviewer subagents (grumpy senior, paranoid SRE, fact-checker, DX advocate, threat modeller, goal auditor, outsider, contract keeper, cost hawk, noise filter) and an /assemble-panel skill that picks the right panel for a plan, spec, diff or decision, briefs and dispatches it in parallel, and filters the reports into one triage.

1 skill · 10 agents

Autonomous coding loop for long-running tasks (refactors, new features, test sweeps). A Python harness owns the state machine and runs one phase at a time as its own headless process — scout, worker, gate, render, evaluate, commit, learn, with a judge on any failure path — each with its own model tier, timeout, budget and usage record. It enforces the allow-list, screenshots the product and checks copy tasks for real similarity before anything commits, resumes or splits a timed-out worker instead of losing the tick, and settles the decisions it can settle on its own into an auditable ledger rather than waking a human. Pause and Stop land at a phase boundary; a browser dashboard shows the phase in flight, the screenshots and the decisions. On an incident it runs a budgeted /agent-loop-medic that repairs only from an allowlist and hands over to a human (NEEDS_HUMAN.md, exit 2) when it cannot. Requires the superpowers plugin; bundles the postmortem plugin and recommends the permissions plugin for its /permissions-advisor pre-flight check.

4 skills

Retrieves up-to-date technical documentation, code examples, and performs people/company research using Context7 MCP and Exa AI MCP

2 skills · 2 MCP servers

Finds unsafe, dead and contradictory permission rules across managed/user/project/local settings. SessionStart alert, /permissions-review, /permissions-advisor, perm-scan and perm-apply commands. No per-call hooks. Points to the built-in /fewer-permission-prompts for new allow rules.

2 skills · hooks

Token-frugal orchestration doctrine for an expensive top-level model: reserve the expensive tier for judgment, route token-heavy bounded work to the cheapest capable subagent tier, demand compact returns with traceable source references, and vet before acting. Model- and harness-agnostic.

1 skill

Generic structured retrospective generator. Writes searchable postmortem documents to docs/postmortems/ with an 8-section template. Callable after any significant task.

1 skill

Scaffold, validate, and publish new Claude Code plugins with proper structure

3 skills · 1 command · 1 agent

Claude Code skill for creating and presenting Slidev slidedecks

1 skill

Observability plugin for Claude Code with MLflow tracing

Skills for adding DeepEval evaluations, tracing, datasets, Confident AI reports, and iterative improvement loops to AI applications.

3 skills

Connect to your Phoenix instance to debug, evaluate, and improve LLM applications.

1 MCP server

OpenRouter for tools - 4,000+ agent-friendly tools, pay for the usage, not subscription

1 skill · 1 MCP server