The sdcd plugin

A Claude Code plugin that operationalises SDCD as namespaced slash-commands and specialised subagents. You can use the methodology without the plugin — but the plugin automates what would otherwise be manual discipline.

What you get

• 11 skills covering the full project lifecycle.
• 8 subagents for independent review, test design, post-commit checks, and plan-drift detection.
• 4 project archetypes for common project shapes.

All skills are namespaced: /sdcd:new-project, /sdcd:audit, etc. That keeps them distinct from any other plugin you have installed.

---

Skills

Planning pipeline

Skill	Purpose	Writes
/sdcd:new-project	Drive the Ur-Plan phase of a brand-new project	design/UR_PLAN.md, design/CURRENT_STATE.md, scaffolding
/sdcd:data-plan	Design the data model from the Ur-Plan	design/DATA_PLAN.md
/sdcd:backend-plan	Design the backend API contract	design/BACKEND_PLAN.md
/sdcd:design-system-plan	Establish visual direction + tokens	design/DESIGN_SYSTEM.md
/sdcd:frontend-plan	Design routes + components + state	design/FRONTEND_PLAN.md

Each planning skill dispatches a subset of the challenger pool tuned for its concerns. For example, design-system-plan dispatches UX + accessibility + maintainability instead of the default security/performance trio — because visual work is where user-facing concerns dominate.

Audits

Skill	When	Writes
/sdcd:audit	Pre-implementation, after all plans are drafted	design/AUDIT.md (GO / NO-GO verdict)
/sdcd:milestone-audit	Mid-implementation, after a feature / milestone completes	Appends to design/AUDIT_LOG.md, updates CURRENT_STATE.md

The two audits do not overlap: audit reads plans only, milestone-audit reads code + state + the milestone’s plan section.

Retrofit (for running projects)

Skill	Purpose
/sdcd:adopt	Scan an existing repo, scaffold design/, reverse-engineer an Ur-Plan stub from README + git-log + code-evidence

adopt is additive only — it never modifies existing files. If your project already has a CLAUDE.md, the skill proposes a diff for you to merge manually rather than overwriting. See the retrofit guide.

Session management

Skill	When	What
/sdcd:session-start	Start of each work session	Loads CURRENT_STATE.md, active plan section, brain files of recently-changed files
/sdcd:session-end	End of each work session	Updates CURRENT_STATE.md, appends to SESSION_HISTORY.md (if present), refreshes brain files

Brain files

Skill	Purpose
/sdcd:auto-brain	Write or refresh a <file>.brain summary for one source file

auto-brain is typically invoked automatically by session-end or as part of §1.6. You can call it explicitly when you’ve just hand-edited a file and want the brain updated immediately.

---

Subagents

Challenger pool — five lenses

Challenger subagents run independently of the author. They read a planning document (or several, for cross-cutting audits) and return pushback reports. They never see the author’s reasoning — which is the point; independent review is cheaper when the reviewer literally cannot inherit the author’s blind spots.

Agent	Lens	Dispatched by
challenger-security	Trust boundaries, auth, data sensitivity, injection surfaces	new-project, data-plan, backend-plan, audit (cross-cut), frontend-plan
challenger-performance	Latency vs. targets, N+1, index coverage, bundle size	new-project, data-plan, backend-plan, audit (cross-cut), frontend-plan
challenger-maintainability	Abstractions that rot, ownership, scope sprawl	All planning skills + audit
challenger-ux	Flow dead-ends, feedback loops, empty / first-time states	design-system-plan, frontend-plan, audit (if UI)
challenger-accessibility	Contrast vs. mandate, keyboard coverage, screen-reader gaps, colour-as-signal	design-system-plan, frontend-plan, audit (if UI)

Each returns max 7 findings with severity (block / high / medium / note). The skill synthesises the findings into a ## Challenger pushback section in the plan document itself — not the chat. That way the feedback survives across sessions.

Implementation-phase subagents

Agent	Purpose
test-designer	Writes the test plan before implementation starts. Outside-in TDD — every test anchors to a spec bullet.
sdcd-reviewer	Post-commit / pre-push review. Focuses on §1 compliance + code quality. Solo-dev stand-in reviewer.
plan-drift-detector	Enforces §1.4 mechanically. Compares code state to plan state and halts on contradictions.

sdcd-reviewer and plan-drift-detector are deliberately separate — one is opinion-bearing (is this good code?), the other is mechanical (does it match the plan?). Skills that need both call both.

---

Project archetypes

Each archetype ships a CLAUDE.md, UR_PLAN.md, CURRENT_STATE.md (and ARCHITECTURE.md where relevant) with sensible defaults for that shape of project.

Archetype	Tier	Typical applicable plans
cli	Serious	UR_PLAN, BACKEND_PLAN (as public-API plan)
web-service	Larger	UR_PLAN, DATA_PLAN, BACKEND_PLAN, optional FE
library	Serious	UR_PLAN, BACKEND_PLAN (as public-API contract)
full-stack	Larger	All five plans

/sdcd:new-project picks the closest match based on user intent and stack hints. You can override or adapt the template after scaffolding — the archetype just saves typing for the common case.

See the archetype guide for detail on when to pick which.

---

Anatomy of a skill dispatch

When you invoke /sdcd:backend-plan, the skill:

1. Reads design/UR_PLAN.md. If missing, stops and points you at /sdcd:new-project.
2. Optionally reads design/DATA_PLAN.md if it exists — recommends /sdcd:data-plan first if persistent state is implied but no data plan was written.
3. Drafts design/BACKEND_PLAN.md with structured sections (Domain model / API contract / Storage / Auth / Error handling / Observability / Open questions). Section headers are exact and consistent across skills — that consistency is what lets audit trace-link decisions later.
4. Dispatches the three core challengers (security / performance / maintainability) in parallel against the draft. Each returns a short findings report.
5. Synthesises the pushback into a ## Challenger pushback section at the bottom of the draft.
6. Presents a top-three summary of the pushback to you in chat. Waits.
7. You decide: address now (update the plan in place) or defer (move to ## Open questions).

No skill proceeds to the next phase automatically. You drive the transitions.

---

Customisation

The plugin is opinionated but not locked.

• Add an archetype: drop a new directory under plugin/sdcd/templates/<name>/. The new-project skill picks it up automatically if the user’s stated project matches its intent.
• Add a challenger lens: drop a new challenger-<lens>.md under plugin/sdcd/agents/. Update the skill procedures to dispatch it where appropriate.
• Override rules: project-level CLAUDE.md can override user-level rules with a stated reason (per §1.5).

The plugin and the rules are intentionally thin. If you want something heavier (CI integration, multi-repo coordination, team-specific workflows), fork it.