softhouselisted

Multi-agent development pipeline — plans a requirement into a task graph, executes it with parallel isolated worktree agents, gates new UI behind a human design approval, reviews and UAT-verifies every branch, then merges and records learned patterns. Use when the user runs /softhouse, asks to run the softhouse pipeline, or wants a multi-step feature planned and executed end-to-end by coordinated agents. "/softhouse resume" continues an interrupted run.
buya-v/claude-softhouse · ★ 0 · AI & Automation · score 70

Install: claude install-skill buya-v/claude-softhouse

# /softhouse — Multi-Agent Development Pipeline (v2)  **PIPELINE VERSION: `2.2`** — this is the authoritative version string. Report it in STEP 0, stamp it into every run archive, and bump it whenever the task schema or a pipeline stage changes (see the Versioning section). Orchestrates a 9-step pipeline that plans, executes, reviews, and merges development tasks using parallel worker agents. v2 changes (2026-07-02, from production build + hardening runs): `executor` routing (agent/orchestrator/user), fork-from-current-main guarantee + merge-base check, mandatory environment-safety preamble in every worker prompt, run-namespaced handoffs, a concrete review checklist + orchestrator micro-fix rule, batched integration testing, project-parameterized UAT, backlog carry-forward, and worktree/branch hygiene in cleanup. v2.1 adds the **design stage** (2026-07-02): a `designer` agent_role that produces HTML mockups from the project design system (`.softhouse/design/`) BEFORE any UI code is written; a human **design gate** (orchestrator publishes the mockup via the Artifact tool, user approves); and a best-effort post-merge **DesignSync** step that keeps the component library browsable on claude.ai/design. Rationale: rejecting a mockup costs minutes; rejecting implemented code costs an agent cycle. ## Usage - `/softhouse <requirement>` — Plan and execute a full pipeline run - `/softhouse resume` — Resume an interrupted run ## Pipeline Step