Contents

Interview-driven, test-first, human-in-the-loop workflow that builds a reviewed, tested, committed Python feature via a conductor + isolated, model/effort-pinned subagent gates.

Features

Implement Feature

/implement-feature runs as a conductor (the interactive session that holds the through-line and talks to you) walking a fixed sequence of gates. The bias-sensitive gates — write tests, review tests, implement, verify, review code — run as isolated subagents. A guard hook enforces the isolation; a deterministic analyzer proves after the fact what actually happened. There is no hand-written orchestration code — the whole workflow is expressed in Markdown, and the agent is the runtime.

What that buys you:

• Tests are written by an algorithm-blind subagent, reviewed by an independent critic before any code exists, and the implementer is barred from editing them.
• Design and every review gate run on a higher model than implementation
• A separate verifier drives the real feature against every acceptance criterion and exercises every external boundary un-mocked.
• Nothing is committed until you review the real artifacts and approve.
• Every gate is isolated (it reads only the files curated for its role) and runs at a pinned model/effort.
• After each run a deterministic audit produces a receipt that verifies both from the ground-truth session transcript — turning "we isolate and we bound the reasoning budget" from a claim into a per-run, checkable fact. The model, effort, and isolation are all verified by the receipt from the transcript, with best-effort real-time prevention by the guard (isolation).

---

User Guide

Read the User Guide for full install, prerequisites, and FAQ. Quick start:

1. Install the plugin:

   /plugin marketplace add Sdaas/claude-plugins
   /plugin install sdlc-lite@sdaas
   

2. Make sure your target project is a git repo.
3. Install the required toolchain into that repo's environment (see User Guide).
4. Run /implement-feature and point it at a GitHub issue, a file, or a 1-2 line description of the feature.

Developer Guide

Read the Developer Guide - Architecture (conductor + isolated gates), the guard hook, the analyzer, the design decisions (ADRs), and the container testing methodology.

Also there is a Tutorial that demonstrates how to build a basic plugin. The concepts (plugin vs command vs skill vs subagent), subagent isolation, and a runnable toy-greet example to build intuition before reading the real product.

---

What's in this repo

README.md                      # this router
CLAUDE.md                      # guidance for Claude Code working in this repo
docs/
  user-guide.md                # run it on your own repo
  developer-guide.md           # understand / extend it
  tutorial.md                  # learn the underlying concepts
DEVCONTAINER.md                # the dev-container test harness (referenced by the Developer Guide)
design/                        # standalone design-investigation records referenced by the ADRs
sdlc-lite-plugin/      # ← the product
toy-greet-plugin/              # a minimal 2-gate example plugin (used by the Tutorial)
.claude-plugin/marketplace.json  # the DEV catalog (name: sdlc-lite-dev)
.devcontainer/                 # the dev container definition
REVIEW-PROMPT.md               # read-only review methodology (findings now tracked as GitHub issues)

This repo's root .claude-plugin/marketplace.json is the dev catalog (name: sdlc-lite-dev, a live directory source used by the maintainer + dev container). It carries:

• sdlc-lite — the product this repo exists to ship (command /implement-feature).
• toy-greet — a two-file, two-gate /greet workflow kept as the Tutorial's runnable example (tutorial-only; never published to customers).

---

Requirements at a glance

• Claude Code (the CLI, desktop, or IDE extension).
• Python 3.12+ on the target repo.
• The pinned dev toolchain (ruff, mypy, pytest, pytest-cov, mutmut, hypothesis, pytest-asyncio) installed into the target repo's environment — the User Guide walks through this. Gate 0 hard-fails if any tool is missing, so nothing runs on a broken environment.

Status

sdlc-lite is v1: it has been run end-to-end against real Python features (a duration parser, a slugifier, and an async cached JSON fetcher), including a fault-injection pass, inside the dev container. See the Developer Guide for the testing methodology and the recorded design decisions.

---

Built by Soumendra Daas. Licensed MIT (see sdlc-lite-plugin/.claude-plugin/plugin.json).