xsefirosus
UserA 17-agent software company in a plugin: plan, build, and adversarially judge every change under hard budget caps. Route-evidence checks, a multi-surface release ledger, and a sandboxed A/B benchmark harness. Markdown + POSIX shell (Python stdlib only for the benchmark tooling).
Categories
Indexed Skills (20)
backend-design
Use when designing or building anything below the API seam -- endpoints, data models, migrations, background jobs, or service boundaries. Contract-first API design with trust-boundary validation, idempotent mutations, reversible migrations, and an explicit error taxonomy.
technical-writing
Use when writing or revising user-facing prose -- READMEs, changelogs, guides, adapter docs, or release notes. Audience-first structure with a quickstart before theory, every claim verified against the repo, and no invented benchmarks or features.
anti-hallucination
Use when producing any factual output -- paths, APIs, numbers, citations, config keys, or claims about code behavior. The canonical no-invention rule every agent points to; unknown lookups become UNKNOWN, unrun executions become PENDING, and every claim traces to a source.
frontend-design
Use when designing, building, auditing, or redesigning any user interface, or studying a design reference. The anti-AI-slop craft skill -- one committed aesthetic direction from a named lane catalog, typography-first hierarchy, a spacing system, tokens, designed states, and non-negotiable accessibility.
loop-engineering
Use when designing, running, or resuming a loop, or mapping the five moves to repo mechanics. Covers discovery, handoff via worktrees, verification, persistence, and scheduling, plus grep-countable stop conditions, deterministic tripwires, and the git-trust and cycle-count resume rules.
memory-protocol
Use when nominating, closing, recovering, searching, or indexing Sefi's local-first Memory Journalist notes.
retro-improve
Use when running the self-improvement retro over qa-engineer rejects, gate failures, and metrics to propose bounded edits. Single-writer, append-only-safe self-improvement that edits only managed-by sefi-agents files and stops when improvement is disabled.
security-review
Use when reviewing a diff, dependency change, or configuration that touches a trust boundary before it can reach a PR. The security gate covering secrets, injection surfaces, unsafe constructs, dependency risk, authorization, and data handling.
sefi-orchestration
Use when routing a request to the right agent, handing off between agents, or dispatching a subagent. The always-loaded routing brain covering roster, handoff rules, the parse ladder for structured output, model routing, and pointers to the harness map and never-auto-merge rule.
focus
Use when focus_mode.enabled is true in sefi.config.yml. Shapes orchestrator replies to the human for actionability -- lead with the next action, number multi-step work, restate progress every turn, cap lists at five, matter-of-fact errors, no preamble or closers. Governs structure, not phrasing -- pairs with, and never replaces, terse-mode's phrasing compression.
premortem
Use when a plan needs to survive contact with reality before execution starts, especially high-stakes or high-budget work. Forensic pre-execution failure analysis -- treats the plan as already failed, ranks the 7 likeliest causes of death, delivers a verdict, rebuilds the plan, plays the adversary, and sets measurable tripwires. Not for work already done (that is a postmortem).
release-tracking
Use for every version bump, release, tag, or investigation of a mismatched sefi-agents version. Reconciles six independent publication surfaces -- plugin.json, marketplace.json, CHANGELOG, git tag, GitHub release, GitHub marketplace index -- against one append-only evidence ledger, and reports partially released, never released, until every surface matches verbatim.
run-sefi-benchmark
Operate this repository's real-token-spend blinded paired A/B benchmark -- the full sefi-agents chain versus one strong model on frozen cases, needing explicit per-invocation human authorization of its dollar spend before any model call. Use only when a human explicitly asks to run the sefi benchmark, never for routine retro or efficiency review, and never from a loop.
n8n-workflow-design
Use when designing an n8n or similar automation workflow as an implementable spec. Opens with the must-haves Rule block covering triggers, idempotency, retries, secrets, webhook security, observability, and cost-per-run, and points to references/node-patterns.md for deep patterns.
strategy-gate
Use when judging a trading-strategy artifact against hard performance gates and assigning a tier. Opens with the Rule block for profit factor, drawdown, expectancy, and coefficient-of-variation thresholds at 4-6 folds plus the tier ladder; gate math and worked examples live in references/gate-math.md.
terse-mode
Use when terse_mode.enabled is true in sefi.config.yml and a chat reply exceeds roughly 1.5-2k output tokens. Narration compression with an explicit drop-list and never-touch-list; it compresses phrasing, never scope, and never touches code, paths, or safety warnings.
design-style-profiles
Use when a selected UI direction needs a constrained visual profile to tune tokens, density, depth, borders, typography treatment, or composition.
expo-native-design
Use when designing or auditing an Expo or React Native interface after a UI direction is selected.
motion-design
Use when a selected UI direction has nontrivial animation, transition interruption, focus behavior, or reduced-motion requirements.
swiftui-design
Use when designing or auditing a SwiftUI or native Apple interface after a UI direction is selected.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.