The team — the pipeline as a relay race

Six executors and five reviewers, split between two standalone product roots and one engineering orchestrator.

SKRAFT is not a monolithic agent. DISCOVER and DISCUSS are two product workflows selected directly by the developer. skraft-orchestrator then takes a refined story and sequences engineering agents. No global orchestrator drives all six steps.

The relay at a glance

graph LR
    H((human)) --> D[backlog-discoverer]
    D -->|triage| H
    H --> DI[backlog-planner]
    DI -->|refined story| H
    H --> O[skraft-orchestrator]
    O -->|when needed| R[solution-researcher]
    R -->|research| DE[solution-architect]
    O -.sequences.-> DE
    DE -->|architecture| DIS[acceptance-designer]
    DIS -->|scenarios| DEL[software-engineer]
    style H fill:#102016,stroke:#6f8478
    style O fill:#102016,stroke:#6f8478
    style D fill:#1a3a2a,stroke:#4ed58a
    style DI fill:#1a3a2a,stroke:#4ed58a
    style DE fill:#1a3a2a,stroke:#4ed58a
    style DIS fill:#1a3a2a,stroke:#4ed58a
    style DEL fill:#1a3a2a,stroke:#4ed58a

The orchestrator — skraft-orchestrator

The engineering conductor only. It produces no business artifact: it sequences RESEARCH → DESIGN → DISTILL → DELIVER, checks pre-conditions, invokes declared reviewers, and applies their verdicts. It invokes neither backlog-discoverer nor backlog-planner. It owns the state.json (see the HVE-Core substrate).

The two standalone product roots

1. backlog-discoverer — the triager (DISCOVER)

2. backlog-planner — the refiner (DISCUSS)

The four engineering executors

3. solution-researcher — the researcher (RESEARCH)

4. solution-architect — the architect (DESIGN)

5. acceptance-designer — the specifier (DISTILL)

6. software-engineer — the craftsperson (DELIVER)

Why five reviewers for six executors

« Peer reviews are the single most effective quality practice a software organization can employ. » — Wiegers, K., Peer Reviews in Software, 2002.

DISCOVER, DISCUSS, DESIGN, DISTILL, and DELIVER each have a dedicated reviewer that never modifies the judged work. RESEARCH is the exception: no phase reviewer is declared, so the orchestrator closes it after checking its output contract. No one validates their own artifact in the five reviewed loops.

Going further