Agentic Sdlc Orchestrator

Agentic SDLC Orchestrator

**The deliverable being evaluated here is the orchestration engine** - a governed, gated, dependency-graph execution system that drives a real Claude agent through a full software delivery lifecycle: understanding a requirement, decomposing it into stages, executing them sequentially and in parallel, gating high-impact decisions on a human, retrying and rolling back on real failure, re-planning around it, and producing an audited, metric-backed trail of everything that happened.

A URL shortener service is the **worked example** the engine is proven against - the "customer" whose SDLC the orchestrator actually ran, live, against the real Anthropic API, three times over (greenfield, brownfield, ambiguous). Every claim below is backed by a real run: a `runId`, a persisted event log, and (for the two code-producing scenarios) a real `mvn test` execution — not a simulated or scripted demo.

Contents

• What was asked for, and where it lives
• Architecture
• Repository layout
• Setup and running it yourself
• Test results
• The three scenarios
• Reliability in practice: what the engine caught, live
• Scope boundaries (deliberate)

---

What was asked for, and where it lives

Requirement	Implementation	Evidence
Requirement understanding	Requirements-Analysis stage — no tools, normalizes the input requirement and states assumptions explicitly rather than guessing	Scenario 3 below: one vague sentence → an 8-feature-family ambiguity inventory, then a 7-entry assumption register with blast radii once resolved
Task decomposition	WorkflowGraph — a fixed 7-stage DAG with explicit dependsOn per stage	[WorkflowGraph.java](orch