3.1 KiB

Raw Blame History

Super-Dev-Pipeline v2.0 - GSDMAD Architecture

Multi-agent pipeline with independent validation and adversarial code review

Quick Start

# Use v2.0 for a story
/story-full-pipeline mode=multi_agent story_key=17-10

# Use v1.x (fallback)
/story-full-pipeline mode=single_agent story_key=17-10

What's New in v2.0

Multi-Agent Validation

4 independent agents instead of 1
Builder → Inspector → Reviewer → Fixer
Each agent has fresh context
No conflict of interest

Honest Reporting

Inspector verifies Builder's work (doesn't trust claims)
Reviewer is adversarial (wants to find issues)
Main orchestrator does final verification
Can't fake completion

Wave-Based Execution

Independent stories run in parallel
Dependencies respected via waves
57% faster than sequential

Architecture

See workflow.md for complete architecture details.

Agent Prompts:

agents/builder.md - Implementation agent
agents/inspector.md - Validation agent
agents/reviewer.md - Adversarial review agent
agents/fixer.md - Issue resolution agent

Workflow Config:

workflow.yaml - Main configuration
workflow.md - Complete documentation

Why v2.0?

The Problem with v1.x

Single agent does ALL steps:

Implement code
Validate own work ← Conflict of interest
Review own code ← Even worse
Commit changes

Result: Agent can lie, skip steps, fake completion

The Solution in v2.0

Separate agents for each phase:

Builder implements (no validation)
Inspector validates (fresh context, no knowledge of Builder)
Reviewer reviews (adversarial, wants to find issues)
Fixer fixes (addresses review findings)
Main orchestrator verifies (final quality gate)

Result: Honest reporting, real validation, quality enforcement

Comparison

Metric	v1.x	v2.0
Agents	1	4
Context Fresh	No	Yes (each phase)
Validation	Self	Independent
Review	Self	Adversarial
Honesty	60%	95%
Completion Accuracy	Low	High

Migration Guide

For new stories: Use v2.0 by default For existing workflows: Keep v1.x until tested

Testing v2.0:

Run on 3-5 stories
Compare results with v1.x
Measure time and quality
Make v2.0 default after validation

Files in This Directory

story-full-pipeline/
├── README.md (this file)
├── workflow.yaml (configuration)
├── workflow.md (complete documentation)
├── agents/
│   ├── builder.md (implementation agent prompt)
│   ├── inspector.md (validation agent prompt)
│   ├── reviewer.md (review agent prompt)
│   └── fixer.md (fix agent prompt)
└── steps/
    └── (step files from v1.x, adapted for multi-agent)

Next Steps

Test v2.0 on Epic 18 stories
Measure improvements (time, quality, honesty)
Refine agent prompts based on results
Make v2.0 default after validation
Deprecate v1.x in 6 months

Philosophy: Trust but verify. Every agent's work is independently validated by a fresh agent with no conflict of interest.

3.1 KiB Raw Blame History