By name: ry-start description: "Полный lifecycle задачи: init, research, plan, implement, verify, commit, sync; ревью только по явному запросу. Используй для: реализуй задачу, доработай, исправь качественно, сделай фичу, end-to-end, доведи до конца. EN triggers: ry-start, full SDLC, implement task, build feature, complete lifecycle, explicit review only."
ry-start
Purpose
Implement a task to a high-quality, scalable, synchronized state. Speed is secondary to correctness, consistency, maintainability, and clean git history.
Workflow
- If context is missing, run a scoped
ry-initautomatically. - Understand the prompt. For ambiguity, ask concise Russian questions with options.
- Research current code through Serena memories and semantic tools (
serena_find_symbol,serena_get_symbols_overview,serena_find_referencing_symbols). - Research current docs, patterns, and alternatives through
ry-explore. Prefer Context7 MCP for official versioned docs, DeepWiki MCP for public repo architecture, and Grep MCP for real production usage patterns. - Read
references/context-sufficiency-gate.mdand pass the gate before editing code. - Write a detailed plan. Verify each plan item against code using Serena before editing.
- Create or use a feature branch/worktree. Use stacked PRs only when the task naturally splits into independent logical PRs.
- Implement strictly by plan, adapting only after code evidence. Make frequent atomic Conventional Commits.
- Provide progress checkpoints after meaningful milestones or every 2-3 completed plan groups.
- Fix all issues in touched scope plus affected integration path. If wider technical debt is found, ask whether to expand scope.
- Run quality gates using project scripts, OpenCode LSP (auto-starts for detected file extensions), and detected stack checks. Use
verification-quality-gatesskill. - Trigger browser validation for UI/browser-visible work unless auth blocks it; if auth blocks, report the limitation and use available evidence. Use
browser-tool-routingandbrowser-validationskills. - Apply security implementation guidance for security-sensitive changes; run
ry-sec-reviewonly when the owner explicitly asks for review/audit/security review. - Run reviewer subagents only when the owner explicitly asks for review, audit, security review, rules review, or
ry-review; otherwise skip the expensive review phase. Reviewer output transport followsreferences/reviewer-protocol.md: file-first reports plus compact summaries. Explicit review may invoke up to six parallel reviewer subagents in a single Task fan-out with self-contained read-only prompts. - Run
flow-post-task-syncbefore final response.
Deploy Intent Routing
If the owner request includes deploy, production, server rollout,
sync-and-deploy, or a named deployment target, do not finish after
implementation. After code validation and Serena/docs sync, route into
/ry-deploy with the same scope and target. If the deploy contract is
incomplete, ask for the missing server, branch, environment, health-check,
rollback, or credential decision before deploying. Never invent server access
or deployment targets.
Automatic Helper Routing
The owner normally writes prompts in Russian. ry-start must route helper skills automatically instead of waiting for explicit helper skill names:
| Russian intent pattern | Helper skills |
|---|---|
| изучи код, посмотри проект, реализуй, доработай, исправь, рефакторинг, ревью, архитектура, файлы, директории, symbols, implementation scope | serena-code-workflow, lsp-routing, quality-first-engineering, implementation-discipline |
| исследуй интернет, изучи в интернете, посмотри документацию, best practices, migration, API behavior, framework/library setup, MCP/tool sources | tech-research (Context7, DeepWiki, Grep MCP), then web-research for broader sources |
| проверь в браузере, визуально, UI, адаптив, скриншот, pixel-perfect, user flow, business-logic checks | browser-tool-routing, browser-validation |
| консоль, сеть, runtime, layout, hydration, Lighthouse, performance, browser-only failures | browser-debug |
| Figma, дизайн, UI, верстка, дизайн-система, shadcn/ui, ReactBits, FSD, tokens, pixel-perfect design | ry-design, figma-to-code, design-system-implementation, fsd-frontend-architecture, design-validation |
| auth/authz/API/input/file/dependency/config/secrets/payment/admin/external-integration | owasp-top-10-implementation |
explicit review, audit, ry-review, ry-sec-review, ry-rules-review, reviewer subagents, security/rules review |
matching reviewer tracks; security track only for explicit security-review request or security-sensitive explicit review |
| завершение, финализация, durable code/config/docs/memory changes produced | verification-quality-gates, serena-memory-sync, flow-post-task-sync |
Context Sufficiency
Do not implement from a shallow prompt. Before editing, the model must know the relevant architecture, files, symbols, DB/schema/API/config contracts, tests, integration paths, current project patterns, and current external API/framework guidance needed for the task.
If the model cannot answer the gate questions in references/context-sufficiency-gate.md, it must gather more evidence through Serena, LSP, ry-explore, browser/security/design workflows, or ask the owner with options. This is a quality guard, not a hard blocker: the correct response is to enrich context until implementation is safe.
Reviewer Opt-In
Invoking ry-start alone is not permission to use parallel reviewer subagents.
Reviewer tracks are orchestrated only by explicit review intents such as
ry-review, ry-sec-review, ry-rules-review, "сделай ревью", "security
review", or "аудит". Prompts must be self-contained and read-only for
reviewers. Reviewer subagents use mode: subagent, hidden: true,
permission: { edit: "deny" } as defined in .opencode/agents/.
Non-Negotiables
- No hacks, temporary workarounds, or untracked debt in touched scope.
- No fake green checks. If a check cannot run, say why.
- No silent destructive git actions. Branch/worktree cleanup requires verified merged state.
- No secrets in commits, logs, docs, memories, or prompts.