How niravrvaghasiya/dependency-first-architect works
Dependency-First Architect is a planning skill for coding agents: users request a plan, and the skill structures it around dependencies, early tradeoffs, a production walking skeleton, and validation gates. The repository also contains an evaluation harness that generates, blinds, judges, and aggregates plans, plus an outcomes benchmark. The README documents the skill’s capabilities; the sampled harness excerpts and source index support its main internal stage relationships. Some documented skill references and evaluation subsystems were not sampled, so those links are documented rather than source-verified.
Main entry points: SKILL.md Planning Skill · cli.py Evaluation CLI · records.py Run Records · generate.py Plan Generation
Components 21
Plan Method
- Planning Skill
SKILL.md - Tradeoff Gates
SKILL.md - Dependency Order
SKILL.md - Walking Skeleton
SKILL.md - Validation Gates
SKILL.md - Cross-cutting Risks
SKILL.md
Plan Evaluation
- Evaluation CLI
eval/dfa_eval/cli.py - Run Configuration
eval/dfa_eval/config.py - Plan Generation
eval/dfa_eval/generate.py - Model Providers
eval/dfa_eval/providers.py - Session Workspace
eval/dfa_eval/workspace.py - Run Records
eval/dfa_eval/records.py - Plan Blinding
eval/dfa_eval/blinding.py - Plan Judging
eval/dfa_eval/judge.py - Evidence Aggregation
eval/dfa_eval/aggregate.py - Record Validation
eval/dfa_eval/schema.py - Record Schemas
eval/dfa_eval/schemas.py - Statistical Analysis
eval/dfa_eval/stats.py
Outcome Benchmark
- Outcome Runner
eval/outcomes/runner.py - Webhook Ledger Task
eval/outcomes/tasks/webhook-ledger
Other components
- Plan Requester
Connections 23
- Plan Requester → Planning Skill: requests plan
README.md - Planning Skill → Tradeoff Gates: frames decisions
README.md - Planning Skill → Dependency Order: orders by dependencies
README.md - Planning Skill → Walking Skeleton: defines early slice
README.md - Planning Skill → Validation Gates: sets validation gates
README.md - Planning Skill → Cross-cutting Risks: threads concerns
README.md - Evaluation CLI → Run Configuration: loads configuration
eval/dfa_eval/cli.py - Evaluation CLI → Plan Generation: runs generation
eval/dfa_eval/cli.py - Evaluation CLI → Plan Blinding: runs blinding
eval/dfa_eval/cli.py - Evaluation CLI → Plan Judging: runs judging
eval/dfa_eval/cli.py - Evaluation CLI → Run Records: validates run files
eval/dfa_eval/cli.py - Plan Generation → Model Providers: makes model calls
eval/dfa_eval/generate.py - Plan Generation → Session Workspace: prepares sessions
eval/dfa_eval/generate.py - Plan Generation → Run Records: writes generation records
eval/dfa_eval/generate.py - Plan Judging → Model Providers: makes judge calls
eval/dfa_eval/judge.py - Plan Judging → Run Records: reads run records
eval/dfa_eval/judge.py - Plan Blinding → Run Records: reads run records
eval/dfa_eval/blinding.py - Evidence Aggregation → Run Records: reads run data
eval/dfa_eval/aggregate.py - Evidence Aggregation → Statistical Analysis: computes statistics
eval/dfa_eval/aggregate.py - Evidence Aggregation → Run Configuration: uses run configuration
eval/dfa_eval/aggregate.py - Record Validation → Record Schemas: validates against
eval/dfa_eval/schemas.py - Outcome Runner → Webhook Ledger Task: runs task attempts
- Outcome Runner → Planning Skill: evaluates plan conditions
README.md