rpikit

Architecture

rpikit is a Claude Code plugin built on a layered component model. Skills are user-facing entry points that auto-register as slash commands from skills/*/SKILL.md. Skills contain methodology instructions and orchestrate agents. Agents are autonomous task performers. Hooks enforce quality automatically.

Component Types

Type	Location	Count	Role
Skills	skills/*/SKILL.md	17	Methodology and entry points
Agents	agents/*.md	7	Autonomous task performers

Skills

Skills are self-contained methodology documents that define how to perform each activity and which agents to use.

Core RPI Workflow

Skill	Phase	Agents Used
research-plan-implement	All	Orchestrates subagents for each phase
researching-codebase	Research	file-finder, web-researcher
synthesizing-research	Research	(none — reads and consolidates research files)
writing-plans	Plan	file-finder, web-researcher
implementing-plans	Implement	file-finder, web-researcher, code-reviewer, security-reviewer

Design and Review

Skill	Purpose
brainstorming	Collaborative design exploration
documenting-decisions	Record decisions as ADRs
reviewing-code	Code review with Conventional Comments
security-review	Security-focused vulnerability analysis

Development Discipline

Skill	Purpose
test-driven-development	RED-GREEN-REFACTOR cycle enforcement
systematic-debugging	Root cause investigation before fixes
verification-before-completion	Evidence-based completion claims

Workflow Support

Skill	Purpose
finishing-work	Structured completion (merge, PR, push)
receiving-code-review	Verification-first response to feedback
git-worktrees	Isolated workspaces for parallel work
parallel-agents	Concurrent dispatch for independent tasks

Agents

Agents are specialized task performers invoked by skills via the Task tool. Each has a model assignment and color for terminal display.

Agent	Model	Color	Purpose
file-finder	haiku	cyan	Locate relevant files in the codebase
web-researcher	sonnet	magenta	Internet research with citations
code-reviewer	sonnet	blue	Quality/design review (soft-gating)
security-reviewer	sonnet	red	Vulnerability analysis (hard-gating)
test-runner	haiku	green	Test execution and diagnostics
debugger	sonnet	orange	Root cause investigation
verifier	haiku	yellow	Verification checks before completion

Model selection follows a pattern: haiku for fast, mechanical tasks (file search, test running, verification) and sonnet for tasks requiring judgment (research, review, debugging).

How Components Connect

graph TD
    User --> Skills["Skills (methodology + entry points)"]
    Skills --> Agents["Agents (autonomous task performers)"]
    Skills --> Output["Output artifacts (docs/plans/*.md, docs/decisions/*.md)"]

RPI Pipeline Flow

The /rpikit:research-plan-implement skill orchestrates the full workflow in a single session using subagents. The orchestrator spawns subagents for each phase.

graph TD
    RPI["/rpikit:research-plan-implement"] --> RTIS["research-plan-implement skill"]
    RTIS --> T1["subagent: codebase-researcher"]
    RTIS --> T2["subagent: web-researcher"]
    T1 --> SYN["subagent: synthesizer"]
    T2 --> SYN
    SYN --> RD["docs/plans/YYYY-MM-DD-*-research.md"]
    RD --> G1user approval gate
    G1 --> T3["subagent: planner"]
    T3 --> PD["docs/plans/YYYY-MM-DD-*-plan.md"]
    PD --> G2user approval gate
    G2 --> T4["subagent: implementer"]
    T4 --> CODE["implementation code"]

Core RPI Flow (Individual Skills)

graph TD
    R["/rpikit:researching-codebase"] --> RS["researching-codebase skill"]
    RS --> RF["file-finder (locate files)"]
    RS --> RW["web-researcher (external context)"]
    RS --> RO["docs/plans/YYYY-MM-DD-*-research.md"]
    RO --> G1user approval gate

    G1 --> P["/rpikit:writing-plans"]
    P --> PS["writing-plans skill"]
    PS --> PF["file-finder (locate files)"]
    PS --> PW["web-researcher (external dependencies)"]
    PS --> PO["docs/plans/YYYY-MM-DD-*-plan.md"]
    PO --> G2user approval gate

    G2 --> I["/rpikit:implementing-plans"]
    I --> IS["implementing-plans skill"]
    IS --> IF["file-finder (locate target files)"]
    IS --> IW["web-researcher (unfamiliar issues)"]
    IS --> IC["code-reviewer (quality gate)"]
    IS --> ISR["security-reviewer (security gate)"]
    IS --> IO["implementation code"]

Decision Flow

Decisions can be recorded after planning or design work:

graph LR
    D["/rpikit:documenting-decisions"] --> DS["documenting-decisions skill"]
    DS --> DR["reads: docs/plans/*-design.md (or user input)"]
    DS --> DW["writes: docs/decisions/NNNN-*.md"]

Review Flow

Reviews can be used at any point, independent of the RPI phases:

graph LR
    RC["/rpikit:reviewing-code"] --> RCS["reviewing-code skill"] --> CRA["code-reviewer agent"]
    RSec["/rpikit:security-review"] --> RSS["security-review skill"] --> SRA["security-reviewer agent"]

Infrastructure

CI: .github/workflows/ci.yml provides automated validation on push and pull requests.

Git hooks (via Husky, activated by npm install):

• Pre-commit: markdownlint, shellcheck (fast checks)
• Pre-push: full test suite, plugin validation (full gate)

Design Principles

1. Skills are entry points - Skills auto-register as slash commands from skills/*/SKILL.md. No separate command wrappers needed.
2. Skills own methodology - All decision logic and instructions live in skills, not agents.
3. Agents are reusable - file-finder and web-researcher are used across multiple skills rather than duplicated.
4. Model selection by task type - haiku for mechanical tasks, sonnet for judgment tasks.
5. Output to docs/ - Research and plans written to docs/plans/, decisions written to docs/decisions/ as numbered ADRs.
6. Human approval gates - Users review output between each RPI phase before the next phase begins.

This site is open source. Improve this page.