Session role switches plus the specialist agent workforce they run. Roles: manager — turn the session into the lead of a subagent fleet that briefs teammates, delegates whole loops, approves their output inside the delegated scope, integrates by ground truth, and reports upward in short digests instead of executing directly. Bundled agents, spawnable as fleet teammates or callable directly as @agent-become:<name>: architecture (technical decisions and .architecture.yaml), go, python, templ (Templ and HTMX frontend), kubernetes (manifests and ArgoCD), helm (charts), containers (image builds), quality (lint, tests, security, then commit and push), hygiene (AI-artifact cleanup)
1 skill · 9 agents
Local disk cleanup utilities: stale git worktrees, dead personal forks, Docker prune (incl. non-default buildx builders and VM disk trim), Go caches, and Homebrew — each runs report → confirm → execute with interactive prompts, plus one orchestrating pipeline
6 skills
Design-of-experiments toolkit for cutting debugging and tuning iterations: taguchi orthogonal arrays for factor isolation, pairwise covering arrays for arbitrary level mixes, group-testing shrink for single-culprit hunts, S/N-ratio tuning for knob optimization
4 skills
Git workflow utilities: fork refresh and repo setup with branch protection
2 skills
Russian text editing in информационный стиль (Ilyahov, Sarycheva). Ships **infostyle** — a full edit pass over prose that starts from context, audience and the reader's useful action, then works structure, paragraph, sentence, word and presentation, with a stop-word catalogue (вводные, неопределённое, заумь, оценки, штампы), bureaucratese repairs, genre checklists (текст о себе, о компании, отчёты, пресс-релиз, слайды, промостраница) and a hard ban on inventing the facts that replace value judgements — and **infostyle-letter** — business correspondence on respect and care instead of ritual politeness: informative subject lines, substance in the first paragraph, boundaries, the do-not-send-on-emotion check, cold emails, complaints, pars, reminders and job replies. Three modes: rewrite, review-only, or re-layout without touching a single word
2 skills
Comprehensive PR review pipeline: a pre-review readiness gate (stops on merge conflicts or red code-related CI with a fix-it note, no verdict), branch-level analysis with merge-base attribution (the verdict is decided only by defects the diff introduced, worsened, or interacts with — purely pre-existing defects are reported in full in a separate outside-findings section and routed to a filed issue, parked branch, or memory entry instead of blocking), sequential Claude+Codex dual-model review with evidence-based verification, and opt-in GitHub PR review publishing under a five-frame substance framework. A value/design gate weighs whether the change should exist at all — value versus permanent maintenance cost, right layer, root cause in-repo versus upstream — and can block a PR even when the code is flawless. Remedies stay proportional: changed logic needs tests in the existing test surface, fixed prose is its own deliverable, and a new checker or guard script is never a required remedy. Findings carry git-computed origin labels (feature vs churn) that diagnose diverging review loops without ever touching the verdict. Reviews land on an LGTM or NOT LGTM verdict, compatible with non-blocking recommendations; the only no-verdict path is the readiness gate's fix-CI / fix-conflicts note. A provenance pass blocks model-advertising trailers (Assisted-by takes only LLM), session links, commit messages and comments that narrate WHAT instead of WHY, and machine-written code without an Assisted-by: LLM disclosure.
2 skills