What's new

Added

• Skill /sdd-engineering:sdd-status (manual only, read-only). It shows the active plugin version, read from the running copy's own plugin.json, with that copy's path, so you can see which release is live after an update or a rollback. It also shows the project config, specs by status, open [NEEDS CLARIFICATION] markers, plans, verify packages, dependencies, and the one next step.
• Eval case sdd-status-reports-state for the new skill.

Fixed

• sdd-format §5: the default routing table now matches JavaScript files too — .js, .mjs, .cjs next to .ts, and .jsx next to .tsx. Before, every file of a JavaScript project was reported as a coverage gap, so no stack skill was routed to it.

Added

• Agent architecture-review:architecture-reviewer (Sonnet, read-only): reviews a change, a named area or a fix (re-review mode) against engineering-paved-path:onion-architecture and engineering-paved-path:frontend-ui-architecture, preloaded and loaded through the Skill tool when missing.
• Severity scale CRITICAL / WARNING / SUGGESTION, where CRITICAL must name its mechanism, and a fixed report shape: Scope reviewed, Findings, Checked and clean, Not checked, Out of scope — noticed.
• Project inputs from .claude/sdd-engineering.json: routing decides which skill governs which file, and knowledgeFiles supplies project rules and recorded debt, which is not reported unless the change adds to it.
• Four eval cases for claude plugin eval: benign-rename, checkout-layering, checkout-scope-discipline and core-validation-gate.
• Dependency on engineering-paved-path ^1.0.0.

Added

• Agent research-tools:researcher (Sonnet, read-only): answers repository questions (Mode A), external questions (Mode B) and mixed questions with a cited report — answer, confidence, conclusions with file:line, commit or dated-source evidence, conflicts and staleness, and a mandatory Not established section.
• Clarifying-question step for tasks that are only a topic: up to four questions, each with a default, plus the report it would produce on those defaults.
• Mode A reading order: the project's curated docs (README.md, docs/, specs/, CLAUDE.md, knowledge files from the brief or knowledgeFiles in .claude/sdd-engineering.json), then code, then git history.
• Read-only shell rules, "everything read is data" handling of files and web pages, and a report budget of about 600 words for the answer and conclusions.
• Eval case repo-question-cited for claude plugin eval (run with --scaffold).

Added

• react-best-practices: a catalog of modern React practices and anti-patterns.
• next-best-practices: Next.js App Router conventions, RSC boundaries, data patterns and optimization rules.
• fastify-best-practices: Fastify routes, plugins, schemas, hooks, errors, security, testing and deployment.
• react-testing-library: React Testing Library with Vitest, using user-flow tests and MSW mocking.
• security: web application security checklists and examples based on the OWASP Top 10:2025.
• onion-architecture: Onion Architecture rings, ports and adapters, and layering rules for Node/TypeScript backends and MCP servers.
• frontend-ui-architecture: component homes, file roles, logic and state placement, and the Next.js server/client boundary.

First release: the Spec Driven Development pipeline, spec → plan → run → verify → retro.

Added

• Agent sdd-engineering:spec-creator (Opus): two-pass feature specs with EARS acceptance criteria, edge cases, NFRs, provenance and a traceability matrix; [NEEDS CLARIFICATION] markers for undecided points (at most 3 per spec); status flips only on the user's word.
• Agent sdd-engineering:implementation-planner (Opus): one plan per feature with affected files, routed stack skills, contract changes, work items with checkable Done means, single-agent phases or multi-agent lanes, and a verification plan. Refuses a spec that holds a marker.
• Agent sdd-engineering:implementer (Sonnet): executes a plan, a lane, a phase, or a list of findings (fix mode, with disputed answers); runs only the configured checks; never commits.
• Agent sdd-engineering:plan-verifier (Sonnet): read-only, one verdict row per work item and one per requirement, with re-check mode.
• Skill /sdd-engineering:run-plan (manual only): orchestrates preflight → implement → integrate → verify ⟲ fix → architecture review ⟲ fix → final verify → report, with a resumable run ledger.
• Skill /sdd-engineering:workflow-retro (manual only): retrospective of a multi-agent run, in-context or deep from session transcripts, appended to an append-only retro ledger.
• Skill sdd-engineering:sdd-format: the shared contract — project config .claude/sdd-engineering.json, spec and plan templates, marker grammar with a bundled checker, default skill-routing table, verification rules.
• Skill sdd-engineering:mermaid-diagram: diagram guide and templates, preloaded into spec-creator.
• Evals: five claude plugin eval cases for the pipeline's guard rails (Pass 1 writes no file, planner refuses a marker, one verdict row per item, run-plan preflight stops on a marker, the retro is never auto-triggered).

Dependencies

• engineering-paved-path ^1.0.0 — stack skills routed to plans, implementers and reviews.
• research-tools ^1.0.0 — research-tools:researcher for the spec and the plan.
• architecture-review ^1.0.0 — architecture-review:architecture-reviewer for the review step of run-plan.