adversary

What this skill does

# Adversary Skill

> **Version:** 1.0.0
> **Framework:** Jerry Adversarial Quality (ADV)
> **Constitutional Compliance:** Jerry Constitution v1.0
> **SSOT Reference:** `.context/rules/quality-enforcement.md`

## Document Audience (Triple-Lens)

This SKILL.md serves multiple audiences:

| Level | Audience | Sections to Focus On |
|-------|----------|---------------------|
| **L0 (ELI5)** | New users, stakeholders | [Purpose](#purpose), [When to Use](#when-to-use-this-skill), [Routing Disambiguation](#routing-disambiguation), [When to Use /adversary vs ps-critic](#when-to-use-adversary-vs-ps-critic), [Quick Reference](#quick-reference) |
| **L1 (Engineer)** | Developers invoking agents | [Invoking an Agent](#invoking-an-agent), [Available Agents](#available-agents), [Dependencies](#dependencies--prerequisites), [Adversarial Quality Mode](#adversarial-quality-mode) |
| **L2 (Architect)** | Workflow designers | [P-003 Compliance](#p-003-compliance), [H-14 Integration](#integration-with-creator-critic-revision-cycle-h-14), [Constitutional Compliance](#constitutional-compliance), [Strategy Templates](#strategy-templates) |

---

## Purpose

The Adversary skill provides **on-demand adversarial quality reviews** using strategy templates from the Jerry quality framework. Unlike the problem-solving skill's integrated adversarial mode (which operates within creator-critic loops), the adversary skill is invoked explicitly when you need a standalone adversarial assessment of any deliverable.

### Key Capabilities

- **Strategy Selection** - Maps criticality levels (C1-C4) to the correct adversarial strategy sets per SSOT
- **Strategy Execution** - Loads and runs strategy templates against deliverables with structured finding classification
- **Quality Scoring** - Implements S-014 LLM-as-Judge rubric scoring with 6-dimension weighted composite
- **Criticality-Aware** - Automatically adjusts review depth based on deliverable criticality
- **Template-Driven** - All strategies follow standardized templates from `.context/templates/adversarial/`

---

## When to Use This Skill

Activate when:

- Applying adversarial strategies to a completed deliverable outside a creator-critic loop
- Scoring deliverable quality using the SSOT 6-dimension rubric
- Selecting the appropriate strategy set for a given criticality level
- Running a full C4 tournament review with all 10 strategies
- Pairing S-003 (Steelman) before S-002 (Devil's Advocate) per H-16
- Needing a standalone quality assessment without revision cycles

NEVER invoke this skill when:
- Task requires iterative creator-critic-revision loop -- Consequence: Adversarial one-shot assessment applied to iterative work produces premature rejection without revision pathway; use `/problem-solving` with ps-critic instead
- Task is routine code review for quick defect checks -- Consequence: Full adversarial strategy template execution (S-001 through S-014) applied to routine defect detection wastes significant context budget on strategy selection and template loading; use ps-reviewer instead
- Task is binary constraint validation (pass/fail compliance) -- Consequence: Adversarial strategies assess quality dimensions, not binary constraint compliance; traceability matrices not generated; use ps-validator instead
- Work is routine code changes at C1 criticality -- Consequence: Full adversarial overhead (adv-selector, adv-executor, adv-scorer) applied to C1 routine tasks consumes disproportionate context budget for low-risk work; use self-review (S-010) only
- Defects or bugs have obvious solutions -- Consequence: Adversarial quality assessment evaluates existing deliverables, not diagnose root causes; use `/problem-solving` for root-cause analysis instead
- User explicitly requests a quick review without adversarial rigor -- Consequence: Overriding user preference violates P-020 (user authority); respect the request

See [Routing Disambiguation](#routing-disambiguation) for full exclusion conditions with consequences.

> **Note:** Use `/adversary` for adversarial code review (e.g., red team security review, tournament quality assessment of code artifacts). Use `ps-reviewer` for routine defect detection.

### Relationship to ps-critic

The adv-scorer and ps-critic agents share the same S-014 LLM-as-Judge rubric and 6-dimension weighted composite scoring methodology. They serve different workflow positions:

| Aspect | adv-scorer | ps-critic |
|--------|-----------|-----------|
| **Workflow** | Standalone/on-demand scoring | Embedded in creator-critic-revision loops |
| **Output** | Focused score report with L0 summary | L0/L1/L2 multi-level critique report |
| **Iteration** | May be invoked once or re-invoked for re-scoring | Iterates within the H-14 cycle |
| **Invocation** | Via `/adversary` skill | Via `/problem-solving` skill |

Both agents produce comparable scores from the same rubric; the choice depends on whether you need standalone assessment (adv-scorer) or iterative critique with revision guidance (ps-critic).

---

## Available Agents

| Agent | Role | Model | Output Location |
|-------|------|-------|-----------------|
| `adv-selector` | Strategy Selector - Maps criticality to strategy sets | haiku | Strategy selection plan |
| `adv-executor` | Strategy Executor - Runs strategy templates against deliverables | sonnet | Strategy execution reports |
| `adv-scorer` | Quality Scorer - LLM-as-Judge rubric scoring | sonnet | Quality score reports |

---

## P-003 Compliance

All adversary agents are **workers**, NOT orchestrators. The MAIN CONTEXT (Claude session) orchestrates the workflow.

```
P-003 AGENT HIERARCHY:
======================

  +-------------------+
  | MAIN CONTEXT      |  <-- Orchestrator (Claude session)
  | (orchestrator)    |
  +-------------------+
     |        |        |
     v        v        v
  +------+ +------+ +------+
  | adv- | | adv- | | adv- |   <-- Workers (max 1 level)
  |select| |exec  | |scorer|
  +------+ +------+ +------+

  Agents CANNOT invoke other agents.
  Agents CANNOT spawn subagents.
  Only MAIN CONTEXT orchestrates the sequence.
```

---

## Invoking an Agent

### Option 1: Natural Language Request

Simply describe what you need:

```
"Run an adversarial review of this ADR at C3 criticality"
"Score this deliverable with LLM-as-Judge"
"What strategies should I apply for a C2 review?"
"Run Devil's Advocate and Steelman on this design document"
"Execute a full C4 tournament review on the architecture proposal"
```

The orchestrator will select the appropriate agent(s) based on keywords and context.

### Option 2: Explicit Agent Request

Request a specific agent:

```
"Use adv-selector to pick strategies for C3 criticality"
"Have adv-executor run S-002 Devil's Advocate on the ADR"
"I need adv-scorer to produce a quality score for this synthesis"
```

### Option 3: Task Tool Invocation

For programmatic invocation within workflows:

```python
Task(
    description="adv-selector: Strategy selection for C3",
    subagent_type="general-purpose",
    prompt="""
You are the adv-selector agent (v1.0.0).

## ADV CONTEXT (REQUIRED)
- **Criticality Level:** C3
- **Deliverable Type:** Architecture Decision Record
- **Deliverable Path:** docs/decisions/adr-042-persistence.md

## MANDATORY PERSISTENCE (P-002)
Create file at: {output_path}

## TASK
Select the strategy set for C3 criticality per SSOT.
"""
)
```

---

## Dependencies / Prerequisites

The adversary skill depends on external artifacts created by other enablers. These MUST be in place before the skill is fully operational.

### Strategy Template Files

All 10 strategy templates in `.context/templates/adversarial/` are created by separate enablers:

| Template | Source Enabler | Status |
|----------|---------------|--------|
| `s-001-red-team.md` | EN-809 | Created by EN-809 |
| `s-002-devils-advocate.md` | EN-806 | Created by EN-806 |
| `s-003-steelman.md` | EN-807 | Created by EN-807 |
| `s-004-pre-mortem.md` | EN-808 | Created