# Agent Personas

Thirteen agent personas live in `.claude/agents/` for use with Agent Teams.

## Review Agents (spawned by `/speckit.review`)

| Agent | Model | Focus | Gates |
|---|---|---|---|
| `product-strategist` | Sonnet | User value, requirements completeness, stakeholder coverage | Spec |
| `systems-architect` | Sonnet | Scalability, modularity, technical design quality | Plan, Tasks |
| `security-reviewer` | Sonnet | Vulnerabilities, threat vectors, data protection, data & privacy | Spec, Plan |
| `delivery-reviewer` | Sonnet | Dependencies, risk, test coverage, execution readiness | Tasks, Pre-impl |
| `operational-reviewer` | Sonnet | Observability, failure modes, rollback, runbooks, on-call burden | Plan, Tasks |
| `devils-advocate` | Opus | Assumption challenges, hidden risks, anti-convergence | All gates |
| `synthesis-judge` | Opus | Integrates findings, preserves dissent, produces consensus | All gates |

## Code Review Agents (spawned by `/speckit.codereview`)

| Agent | Model | Focus | Gates |
|---|---|---|---|
| `code-reviewer` | Sonnet | Correctness, test quality, ADR compliance, maintainability, conventions | Code |
| `security-reviewer` | Sonnet | In-code vulnerabilities: injection, secrets, missing auth, unsafe deserialization | Code (STANDARD/FULL) |

## Audit Agent (spawned by `/speckit.audit`)

| Agent | Model | Focus |
|---|---|---|
| `consistency-auditor` | Opus | Bidirectional doc-code drift, decision record discovery |

## Brainstorming Agents (spawned by `/speckit.brainstorm`)

| Agent | Model | Focus |
|---|---|---|
| `visionary` | Sonnet | Unconstrained creative ideation, boundary-pushing features |
| `user-advocate` | Sonnet | Daily user needs, pain points, practical value |
| `technologist` | Sonnet | Elegant solutions, platform capabilities, enabling infrastructure |
| `provocateur` | Sonnet | Lateral thinking, assumption inversion, contrarian ideas |

Brainstorming agents operate in **divergent mode** — no judgment, no feasibility checks.
Ideas are evaluated only after all rounds complete.

## Calibration

Agents receive injected project context from the orchestrator and calibrate review intensity accordingly (FULL/STANDARD/LIGHTWEIGHT).

The devil's advocate and synthesis judge always run at FULL rigor regardless of project context.

The `operational-reviewer` defaults to LIGHTWEIGHT — it scales up only when the project context signals a production service, on-call rotation, or 24/7 availability requirement.

The `security-reviewer` data & privacy section is skipped entirely for projects with no personal data, and applied in full only for PII/regulated data.