# Engineering model

`sdd-agentic-flow` is a **repository-native engineering control plane for
Spec-Driven coding-agent workflows**. It is a control layer, not a scheduler,
worker runtime, session store, or agent framework.

> Intent is durable. Work is bounded. Context is scoped. State is external.
> Evidence is observable. Verification is independent. Autonomy is conditional.
> Humans retain authority.

## Product identity

`sdd-agentic-flow` is a **Spec-Driven Agentic Workflow Harness** for coding agents.
It acts as a repository-native engineering control plane while the coding-agent host
owns runtime execution.

| Level | SAF term |
| --- | --- |
| Public category | Agentic Workflow Harness |
| Primary full category | Spec-Driven Agentic Workflow Harness |
| Explicit target form | Spec-Driven Coding-Agent Workflow Harness |
| Architecture role | repository-native engineering control plane |
| Methodology | Spec-Driven Development |
| Runtime owner | coding-agent host |

The host runtime harness owns models, sessions, tools, workers, hooks, and sandboxing.
SAF is the outer workflow harness: intent, specifications, capabilities, policy, artifacts,
evidence, verification, and admissible transitions. Skills are public capabilities; they are
not runtime execution mechanics. See the [canonical vocabulary](../shared/references/canonical-vocabulary.md)
and the [developer journey](developer-journey.md).

## Principles

1. **Durable intent** — specifications preserve the behavior and decision context.
2. **Bounded work** — each task has observable completion criteria (`GOAL-001`).
3. **Minimum sufficient context** — workers load a task context package, not a whole history.
4. **Capabilities over persistent personas** — a Skill names an engineering capability; a worker or subagent is host execution context.
5. **Externalized control state** — artifacts and loop state make transitions reconstructable.
6. **Evidence before verdict** — sensors observe; evidence records; verification evaluates against requirements.
7. **Conditional autonomy** — guardrails permit a next admissible action only while humans retain consequential authority.

## Authority boundaries

The effective contract remains the current canonical specification, applicable repository
contracts, and configured gates. An agent may propose a consequential contract change when the
current contract cannot safely be followed, but the proposal is not authority or evidence. The
existing human Decision Gate and canonical authoring path decide whether the effective contract
changes. This keeps proposal, implementation, verification, and human authority distinct without
adding a second lifecycle or runtime admission system.

## Control boundary

```text
human intent → specification → bounded work
                              ↓
                 scoped context + capabilities
                              ↓
                       policy / gates
                              ↓
                      admissible action
                              ↓
               host/human chooses execution
                              ↓
                  host execution → tools → sensors
                                             ↓
                                      current evidence
                                    ↙                 ↘
                           verification       evidence graph
                                    ↓            (projection)
                             transition state
                          ↙        ↓         ↘
                    admissible  blocked  human decision
```

The model proposes. The host executes. SAF defines constraints and admissible
transitions. Sensors observe. Evidence records. Verification evaluates. Humans
govern. A blocked transition is not permission for the host to continue; a
human decision is not a host decision.

## Graphs and correctness

| Graph | Owner | Question |
| --- | --- | --- |
| Dependency graph | SAF tasks DAG | What may run in parallel? |
| Execution graph | Host | Who executes the next turn or worker? |
| Evidence graph | SAF projection | Why is REQ-X believed satisfied? |

Correctness dimensions are explanatory, not a mandatory five-gate pipeline:
structural validity, behavioral evidence, domain/contract validity, and
requirement coverage. Domain semantics belong to the oracle; sensors observe
whether the oracle holds. Passing these dimensions never resolves a
consequential human-judgment boundary.

See the [mental model](sdd-agentic-flow-model.md), [architecture](architecture.md),
and [bounded execution](../shared/references/bounded-execution.md).