AI Architect Feedback Loop

Feed it findings, get back pull requests. A 14-stage autonomous pipeline for Claude Code that takes issues, specs, or research findings and turns them into production-ready PRs — with impact analysis, verified PRDs, tested implementations, and quality gates.

/plugin marketplace add cdeust/ai-architect-feedback-loop
/plugin install ai-architect-feedback-loop

Then run /run-pipeline from your project. Free, open source, works with any tech stack.

---

What You Get

A summary like this at the end of every run:

=== Pipeline Complete: 20260221_020000 ===
Findings analyzed: 3
  [PASS] tv-042: PR #187 — improve core, api
  [PASS] tv-051: PR #188 — improve worker
  [FAIL] tv-063: Stage 7 build failure (3 attempts)
PRs created: 2

Each PR includes: impact analysis, integration plan, PRD excerpt, quality enforcement results, semantic verification, and retry history.

How It Works

Phase	Stages	What happens
Discovery	1	Parse findings, filter by relevance, prioritize by multi-module impact
Analysis	2-6	Impact scoring, integration design, PRD generation + review (64 quality rules enforced)
Implementation	7-11	Feature branch, code changes, build + test, quality gates, semantic verification
Delivery	12-14	Benchmark, deployment simulation, PR creation per finding

Each finding retries up to 3 times. Failed findings are skipped so the pipeline keeps moving. Full stage details: docs/configuration.md

Why This Exists

Most AI coding tools generate code from a prompt. This one:

• Starts from findings, not prompts — feed it output from code analysis tools, design reviews, bug reports, or research papers
• Generates verified specs before code — 9-file PRDs with claim-by-claim verification, JIRA tickets, test cases
• Implements and tests autonomously — feature branches, build + test cycles, quality gates, up to 3 retries
• Opens real PRs — structured pull requests with full context, one per finding

Quick Start

Via Claude Code Marketplace

/plugin marketplace add cdeust/ai-architect-feedback-loop
/plugin install ai-architect-feedback-loop

Manual Setup

git clone https://github.com/cdeust/ai-architect-feedback-loop.git
cd ai-architect-feedback-loop
export PIPELINE_BUILDER="/absolute/path/to/your-product"
make setup            # wizard auto-detects your stack
make pipeline-health  # verify everything is ready
claude                # then type: /run-pipeline

The setup wizard detects your project language, module structure, build/test commands, and git conventions. It generates config/pipeline.yml — the single source of truth for all settings.

Docker (zero dependencies)

make docker-build                                    # one-time
TARGET_REPO=/path/to/your-product make docker-setup  # first time
TARGET_REPO=/path/to/your-product make docker-run   # run pipeline

The container clones your repo (original mounted read-only), installs quality gate hooks, and runs Claude Code in sandbox mode.

Requires: Docker, CLAUDE_CODE_OAUTH_TOKEN, GH_TOKEN. Details: docs/docker.md

Feeding Findings

Findings are structured JSON — issues, specs, or improvements for the pipeline to analyze and implement:

{
  "source": "your-tool-name",
  "findings": [
    {
      "id": "spec-001",
      "title": "Add token refresh to auth flow",
      "description": "Silent refresh using refresh_token grant type",
      "relevance_category": "api_change",
      "relevance_score": 0.9
    }
  ]
}

Full schema: docs/findings-format.md

Configuration

The setup wizard handles everything. To customize later:

vim config/pipeline.yml

What you can customize: build/test commands, branch naming, commit format, PR labels, quality thresholds, max findings per run, notification sound, nightly schedule.

Full reference: docs/configuration.md

The PRD Generator

Stage 5 uses the AI PRD Generator — a standalone plugin you can also use independently. It produces 9 verified files per PRD: overview, requirements, user stories, technical spec, acceptance criteria, roadmap, JIRA tickets, test cases, and a verification report.

If you only need PRD generation (no pipeline), install that plugin directly:

/plugin marketplace add cdeust/ai-prd-generator-plugin
/plugin install ai-prd-generator

Troubleshooting

Issue	Fix
PIPELINE_BUILDER not set	export PIPELINE_BUILDER="/path/to/your-product"
pipeline.yml validation errors	make validate-config for details
Stage 5 fails with "skill not found"	Install the ai-prd-generator plugin
Build failures in Stage 7	Verify build_command in pipeline.yml works in your target repo
gh errors in Stage 14	gh auth login to authenticate
Docker clone fails	Set TARGET_REPO to a valid git repo path

System Requirements

• Python 3.10+, gh CLI (authenticated), git, jq, PyYAML
• Claude Code (Anthropic)
• Docker (optional, for containerized runs)

---

Built by Clement Deust