EDD (Eval-Driven Development) Framework v2.64

Eval-Driven Development is a quality-first development pattern that enforces define-before-implement workflow with structured evaluations.

v2.88 Key Changes (MODEL-AGNOSTIC)

• Model-agnostic: Uses model configured in ~/.claude/settings.json or CLI/env vars
• No flags required: Works with the configured default model
• Flexible: Works with GLM-5, Claude, Minimax, or any configured model
• Settings-driven: Model selection via ANTHROPIC_DEFAULT_*_MODEL env vars

What is EDD?

EDD provides a systematic approach to software development with three phases:

1. DEFINE - Create structured eval specifications using TEMPLATE.md
2. IMPLEMENT - Build features according to eval definitions
3. VERIFY - Validate implementation against eval criteria

Check Types

Prefix	Type	Purpose
CC-	Capability Checks	Feature capabilities and functionality
BC-	Behavior Checks	Expected behaviors and responses
NFC-	Non-Functional Checks	Performance, security, maintainability

Usage

# Invoke EDD workflow
/edd "Define memory-search feature"

# CLI script (if available)
ralph edd define memory-search
ralph edd check memory-search

Components

• TEMPLATE.md: Template for creating eval definitions
• edd.sh: CLI script for eval management
• /edd skill: Skill invocation from Claude Code
• ~/.claude/evals/: Directory for eval definitions

Template Structure

Each eval definition includes:

1. Capability Checks (CC-) - What the feature can do
2. Behavior Checks (BC-) - How the feature behaves
3. Non-Functional Checks (NFC-) - Performance, security, etc.
4. Implementation Notes - Technical guidance
5. Verification Evidence - Test results

Example: memory-search.md

# Memory Search Eval

**Status**: DRAFT
**Created**: 2026-01-30

## Capability Checks
- [ ] CC-1: Search across semantic memory
- [ ] CC-2: Support filtering by type

## Behavior Checks
- [ ] BC-1: Returns ranked results
- [ ] BC-2: Handles empty queries gracefully

## Non-Functional Checks
- [ ] NFC-1: Search completes in <2s
- [ ] NFC-2: Memory usage <100MB

## Implementation Notes
- Use parallel search for performance
- Cache frequent queries

## Verification Evidence
- Test results attached

Integration with Orchestrator

EDD integrates with the orchestrator workflow to ensure quality-first development:

1. Clarify phase - Define evals
2. Plan phase - Review eval requirements
3. Implement phase - Build to eval specs
4. Validate phase - Verify against evals

---

Swarm Mode Integration (v2.81.1)

EDD framework now supports swarm mode for parallel evaluation across multiple check types.

Auto-Spawn Configuration

When invoked via /edd, the framework automatically spawns a specialized evaluation team:

Task:
  subagent_type: "general-purpose"
  model: "sonnet"
  team_name: "edd-evaluation-team"
  name: "edd-coordinator"
  mode: "delegate"
  run_in_background: true
  prompt: |
    Execute Eval-Driven Development workflow for: $ARGUMENTS

    EDD Pattern:
    1. DEFINE - Create structured eval specifications
    2. DISTRIBUTE - Assign check types to specialists
    3. VERIFY - Validate against eval criteria
    4. CONSOLIDATE - Merge findings from all evaluators

Team Composition

Role	Purpose	Specialization
Coordinator	EDD workflow orchestration	Manages eval lifecycle, consolidates findings
Teammate 1	Capability Checks specialist	CC- prefix: feature capabilities and functionality
Teammate 2	Behavior Checks specialist	BC- prefix: expected behaviors and responses
Teammate 3	Non-Functional Checks specialist	NFC- prefix: performance, security, maintainability

Swarm Mode Workflow

User invokes: /edd "Define memory-search feature"

1. Team "edd-evaluation-team" created
2. Coordinator (edd-coordinator) receives task
3. 3 Teammates spawned with check-type specializations
4. Eval definition distributed:
   - Teammate 1 → Capability Checks (CC-)
   - Teammate 2 → Behavior Checks (BC-)
   - Teammate 3 → Non-Functional Checks (NFC-)
5. Teammates work in parallel (background execution)
6. Coordinator monitors progress and gathers results
7. Findings consolidated into single eval specification
8. Final eval document returned

Parallel Evaluation Pattern

Each teammate focuses on their check type:

# Teammate 1: Capability Checks
CC-1: Feature can perform X
CC-2: Feature supports Y configuration
CC-3: Feature integrates with Z system

# Teammate 2: Behavior Checks
BC-1: Feature handles error case A gracefully
BC-2: Feature returns expected response for B
BC-3: Feature maintains state across C

# Teammate 3: Non-Functional Checks
NFC-1: Response time < 100ms
NFC-2: Memory usage < 50MB
NFC-3: Security vulnerability scan passes

Communication Between Teammates

Teammates use the built-in mailbox system:

# Teammate sends finding to coordinator
SendMessage:
  type: "message"
  recipient: "edd-coordinator"
  content: "CC-3 defined: Feature integrates with auth system via OAuth2"

Task List Coordination

All teammates share a unified task list:

# Location: ~/.claude/tasks/edd-evaluation-team/tasks.json

# Example tasks:
[
  {"id": "1", "subject": "Define Capability Checks", "owner": "teammate-1"},
  {"id": "2", "subject": "Define Behavior Checks", "owner": "teammate-2"},
  {"id": "3", "subject": "Define Non-Functional Checks", "owner": "teammate-3"},
  {"id": "4", "subject": "Consolidate eval specification", "owner": "edd-coordinator"}
]

Manual Override

To disable swarm mode:

/edd "Define feature X" --no-swarm

Output Location

# Evals saved to ~/.claude/evals/
ls ~/.claude/evals/

# View last eval
cat ~/.claude/evals/latest.md

---

Testing

Test suite: tests/test_v264_edd_framework.bats (33 tests)

Run tests:

bats tests/test_v264_edd_framework.bats

Swarm Mode Tests

Additional tests for swarm mode integration:

# Test swarm team creation
tests/edd/test-swarm-team-creation.sh

# Test parallel evaluation
tests/edd/test-parallel-evaluation.sh

Status

Current: Framework defined with swarm mode integration (v2.81.1) Note: TEMPLATE.md and evals directory structure ready for use

---

Version: v2.64 | Status: DRAFT | Tests: 33 passing

Action Reporting (v2.93.0)

Esta skill genera reportes automáticos completos para trazabilidad:

Reporte Automático

Cuando esta skill completa, se genera automáticamente:

1. En la conversación de Claude: Resultados visibles
2. En el repositorio: docs/actions/edd/{timestamp}.md
3. Metadatos JSON: .claude/metadata/actions/edd/{timestamp}.json

Contenido del Reporte

Cada reporte incluye:

• ✅ Summary: Descripción de la tarea ejecutada
• ✅ Execution Details: Duración, iteraciones, archivos modificados
• ✅ Results: Errores encontrados, recomendaciones
• ✅ Next Steps: Próximas acciones sugeridas

Ver Reportes Anteriores

# Listar todos los reportes de esta skill
ls -lt docs/actions/edd/

# Ver el reporte más reciente
cat $(ls -t docs/actions/edd/*.md | head -1)

# Buscar reportes fallidos
grep -l "Status: FAILED" docs/actions/edd/*.md

Generación Manual (Opcional)

source .claude/lib/action-report-lib.sh
start_action_report "edd" "Task description"
# ... ejecución ...
complete_action_report "success" "Summary" "Recommendations"

Referencias del Sistema

• Action Reports System - Documentación completa
• action-report-lib.sh - Librería helper
• action-report-generator.sh - Generador