PRD Task Loop

Autonomous PRD-driven development loop. Reads a PRD/task list, loops through tasks using AI coding agents or Hermes subagents, auto-commits, and marks tasks complete.

Inspired by Ralphy, adapted for the Hermes Agent skill ecosystem.

Install

# Any agent (Claude Code, OpenCode, Codex, Cursor, etc.)
npx skills add BarryYin/prd-task-loop

# Global install (all projects)
npx skills add BarryYin/prd-task-loop -g

What It Does

1. Read a PRD file (Markdown or YAML)
2. Resolve task dependencies and execution order
3. Run AI agents/subagents (OpenCode, Claude Code, Codex, Hermes subagents)
4. Auto-commit changes
5. Mark task complete in the PRD
6. Repeat until all tasks done

What's New in v2.0.0

• Subagent execution — use Hermes delegate_task for full tool access per task
• 3 parallel strategies — subagent parallel, git worktree isolation, hybrid
• Dependency graph — topological sort with depends field and cycle detection
• Dry-run mode — preview execution plan before running
• Config system — .prd-loop/config.yaml for project/execution/verification settings
• Structured progress — JSON state + human-readable Markdown log
• PRD auto-generation — analyze project structure and generate tasks automatically
• Skill composition — pipelines with kiro-spec-tasks, writing-plans
• Enhanced parsing — [priority], [depends], [tags], [engine] metadata in Markdown

Usage

Single Task

ralphy --opencode "add login button"

PRD Mode (sequential)

Create a PRD.md:

## Features
- [ ] Create user model [priority:high]
- [ ] Add auth endpoints [depends:user-model]
- [ ] Build dashboard [depends:auth-endpoints]

Then run:

ralphy --opencode --prd PRD.md

YAML Tasks with Dependencies

tasks:
  - id: setup-db
    title: Setup database schema
    priority: high

  - id: user-model
    title: Create user model
    depends: [setup-db]
    parallel_group: 1

  - id: post-model
    title: Create post model
    depends: [setup-db]
    parallel_group: 1  # runs in parallel with user-model

  - id: relationships
    title: Add model relationships
    depends: [user-model, post-model]  # runs after both

ralphy --opencode --yaml tasks.yaml --parallel

Hermes Subagent Mode (Recommended)

In Hermes, this skill uses delegate_task for each task:

# Sequential
result = delegate_task(
    goal="Create user model with email, password fields using Prisma",
    context="Project: Next.js + Prisma at ~/my-project",
    toolsets=["terminal", "file"]
)

# Parallel (multiple tasks at once)
results = delegate_task(tasks=[
    {"goal": "Create utils.js", "context": "...", "toolsets": ["terminal", "file"]},
    {"goal": "Create types.ts", "context": "...", "toolsets": ["terminal", "file"]},
])

Execution Modes

Mode	Best For	Tools
Subagent	Complex multi-file tasks, needs reasoning	Full (file, terminal, web, browser)
Engine	Simple, known patterns	Terminal only
Hybrid	Worktree isolation + subagent reasoning	Both

Parallel Strategies

Strategy	When to Use
A: Subagent Parallel	Complex tasks, need full reasoning
B: Worktree Parallel	Tasks modifying overlapping files
C: Hybrid	Complex tasks + file conflicts

Configuration

Create .prd-loop/config.yaml:

project:
  name: "my-app"
  language: "TypeScript"
  framework: "Next.js"

execution:
  mode: "subagent"        # subagent | engine | hybrid
  parallel: true
  max_concurrent: 3
  auto_commit: true

verification:
  test_command: "npm test"
  run_tests: true

rules:
  - "use strict TypeScript"
  - "follow existing code patterns"

boundaries:
  never_touch:
    - "*.lock"
    - "src/legacy/**"

Supported Engines

Engine	Command
OpenCode	opencode run '<prompt>' --format json
Claude Code	claude --dangerously-skip-permissions -p '<prompt>'
Codex	codex exec --full-auto --json '<prompt>'
Subagent	delegate_task(goal=...)

With Hermes Agent

This skill integrates with Hermes Agent's tool ecosystem:

• terminal() — run AI coding agents
• delegate_task() — parallel task execution
• patch() — update PRD task status
• session_search() — recall past runs
• memory — persist project context

Skill Composition

kiro-spec-tasks → prd-task-loop     # spec → auto-execute
writing-plans → prd-task-loop       # plan → auto-execute

Comparison with Ralphy

	Ralphy (bash)	Engine Mode	Subagent Mode
Setup	npm i -g ralphy-cli	No install	No install
Parallel	Git worktrees	Manual	Built-in
Tools	Terminal only	Terminal only	Full access
Reasoning	None	CLI agent	Full LLM
Cost	Lowest	Low	Higher

License

MIT