xSquad (xSquad-code)

Delegate your work to a squad of agents — and get back proof it actually works.

Give xSquad a high-level goal. One orchestrator agent plans it, writes self-contained task briefs, and fans the work out to parallel implementer subagents (Claude Code, Pi, or Codex). Nothing is reported done until it's proven:

• Claimed output earns nothing — the orchestrator re-runs project validators (build/test/lint plus feature validators that drive the real app) after every task.
• Validators never rot — behavior changes update the validator suite before a task counts as green.
• Thermo-nuclear review built in — two independent reviewers (security/correctness, code quality) audit the whole diff; their findings go back to the squad as fix briefs and are re-verified until clean.
• The squad learns — durable lessons persist across runs in .xsquad/MEMORY.md.
• Cheap triage and checks — a System One pre-pass on Jev (TypeSafe AI, ~200 ms per call) buckets validator failures and review findings, and checks subagent reports and logs against the brief's acceptance criteria, so agents read only what needs judgment instead of raw output. Use a TYPESAFE_API_KEY, Vercel AI Gateway (AI_GATEWAY_API_KEY, model typesafe-ai/jev), or the keyless classifier.dev proxy — /xSq-setup asks which.

Six installable skills: the core orchestrator plus five standalone slash commands.

Token cost: solo Opus 5.5 vs. xSquad

These are modeled estimates, not measurements. The xSquad orchestrator runs on Opus 5.5 and the implementer and reviewer subagents run on GPT-6 Luna ($0.10 / $0.50 per 1M input / output tokens, with no cache discount assumed). On a 6-task feature, the estimate is about $4.13 with xSquad vs. $9.46 solo. On a 12-task feature, it's about $7.44 vs. $25.99.

Install

npx skills add https://github.com/areai51/xsquad         # skills.sh — installs all six skills

Quickstart

1. /xSq-setup — pick runner (claude / pi / codex), models per role, and the Jev backend → .xsquad/config.json (keys go to the gitignored .env); if no validator suite exists yet, it automatically chains the validator setup in the same pass
2. /xsquad build the export-to-CSV feature end to end — plan → briefs → parallel subagents → continuous validator verification → automatic review → fix loop → report

Commands

Command	What it does
/xsquad <goal>	Full squad run: plan → briefs → parallel subagents → validator verification → automatic code review → fix loop → report
/xSq-setup	Configure runner + models per role; chains validator setup when no suite exists
/xSq-validator-setup	Generate the project validator suite
/xSq-validator-run	Run validators, triage failures (product bug / validator drift / env blocker)
/xSq-validator-update	Upkeep pass keeping validators honest (full pass + mid-run variant)
/xSq-code-review	Thermo-nuclear review; findings fixed by the squad

Layout

<this-repo>/
└── skills/
    ├── xsquad/            # core skill: workflow + command router
    │   ├── SKILL.md
    │   ├── commands/      # the five command procedures
    │   ├── agents/        # subagent contracts (implementer, two reviewers)
    │   ├── references/    # config schema, dispatch commands, brief/report templates
    │   └── scripts/       # classify.sh / classify.py — Jev triage + check pre-pass
    └── xsq-*/SKILL.md     # the five command wrappers

In a project the squad works on, everything persists under .xsquad/: config.json and MEMORY.md (committed), validators/ (committed), and runs/<run-id>/ — briefs, reports, logs, evidence, review artifacts (gitignored).