aSPARK Brings Gated Agile Delivery to Claude Code

Summary

aSPARK is a Claude Code plugin that organizes software work as a gated AI product team. Product Owner, Designer, Engineering Manager, Developer, Reviewer, QA Tester, and Release Manager roles move a feature through specification, planning, implementation, review, browser testing, and release, with the human approving each required gate. The plugin has no service, account, or runtime; it installs as Markdown and records its work in reviewable `.spark/` files. These artifacts include a constitution, specification, plan, review, QA, and release documents, while stable story and acceptance-criteria IDs trace requirements from planning through code review and testing. Its five phases are Specify, Plan, Act, Review, and Keep, and failed review or QA findings return to the Developer; after three failed rounds of one phase, the process asks whether the specification or plan is wrong. `/spark` runs or resumes the loop, while each ceremony can also run independently. Real-browser QA requires Claude in Chrome or a Playwright/Chrome DevTools MCP server unless a project declares a substitute method. The project reports 77 features, 429 role-agent runs, and 377 human gate decisions across three machines, based on committed reports. One reported 47-minute run used 42,307,448 tokens and cost $35.52 at stated Opus 5 list prices, with 96.3% cache reads; the author says this is a single data point, not a subscription estimate. Optional companions can enforce gate conditions in code or build a deterministic artifact graph, while the workflow remains usable without them.