Agentic SDLC Control Layer

Deterministic validation for agentic software-development workflows.

This is not an agent runtime. It does not run Codex, Claude, GitHub Copilot, LangGraph, CrewAI, AutoGen, or any other agent. It validates the files around an agentic workflow: task state, events, handoffs, resume packets, review gates, evidence, ownership claims, privacy boundaries, and memory candidates.

The goal is simple: make agent work auditable enough that a human or another agent can safely resume it, review it, or reject it without trusting stale chat context.

Why It Exists

Agent tools can write code quickly, but real SDLC work fails around the seams:

• stale resume context
• unclear handoffs
• skipped tests reported as success
• missing review evidence
• parallel agents editing the same files
• raw secrets or private paths leaking into durable memory

This repo treats those as validation failures.

V1 Promise

Given a bounded workflow directory, agentic-sdlc validates whether the workflow artifacts are structurally safe to continue.

It checks:

• task graph references and cycles
• stale resume packets
• least-privilege handoff shape
• unresolved backward-pass gates
• required evidence for review gates
• required tests skipped without waiver
• overlapping ownership claims
• privacy violations such as secrets, raw transcripts, or private absolute paths

Quick Start

From the repo root:

$env:PYTHONPATH = "src"
python -m agentic_sdlc validate fixtures/valid_workflow
python -m agentic_sdlc validate fixtures/privacy_violation
python -m agentic_sdlc status fixtures/ownership_conflict
python -m agentic_sdlc review-plan fixtures/missing_review_evidence

Machine-readable output:

$env:PYTHONPATH = "src"
python -m agentic_sdlc validate fixtures/stale_resume --json

Run tests:

$env:PYTHONPATH = "src"
python -m unittest discover -s tests -v

Fixture Demo

The fixtures intentionally include both valid and invalid workflows:

• fixtures/valid_workflow
• fixtures/stale_resume
• fixtures/bad_handoff
• fixtures/failed_backward_pass
• fixtures/ownership_conflict
• fixtures/privacy_violation
• fixtures/skipped_required_test
• fixtures/missing_review_evidence

Each invalid fixture is expected to fail with stable rule IDs.

Non-Goals

V1 is not:

• an agent runtime
• a scheduler or watcher
• a prompt policy engine
• a memory system that writes durable facts automatically
• a broad filesystem scanner
• a live GitHub or MCP integration
• a replacement for existing orchestration frameworks

Adapters can come later. First the workflow records need to be trustworthy.

Documentation

• V1 spec
• Rule catalog
• JSON Schemas