SDLC Workflow Plugin

A multi-agent SDLC orchestration system: 10 agents × 5 phases → plan, build, verify, ship, operate. Tracer bullet development with Ralph Loop self-correction and Linear issue integration.

Project

• Type: Claude Code plugin
• Version: 2.0.0
• Repo: github.com/saitarrun/devforge-ai

Repository Map

• agents/ — 10 role-specific agents (product-manager, ux-designer, fullstack-engineer, qa-engineer, security-engineer, performance-engineer, devops-engineer, sre-engineer, data-engineer, technical-writer)
• skills/ — 34 knowledge skills (architecture, testing, code-review, ralph-loop, grill-me, etc.)
• commands/ — 10 orchestrator commands (/sdlc, /sdlc-plan, /sdlc-build, /sdlc-verify, /sdlc-ship, /sdlc-operate, /sdlc-review, /sdlc-implement, /to-prd, /to-issues)
• scripts/ — CLI, install, uninstall, validation helpers
• README.md — User documentation and feature overview

Commands

• Install: npm run install-local or make install-local
• Validate: make validate
• Uninstall: make uninstall

Agent Conventions

Frontmatter required: name, description, tools (comma-separated), model (haiku/sonnet/opus)

Model tiers: Haiku (templated), Sonnet (development), Opus (architecture/security decisions)

Tools: Include only what the agent uses — Read, Write, Edit, Bash, Glob, Grep, WebFetch

• Note: When querying codebase, prefer code-review-graph (bundled integration) over grep for structural analysis. See "Codebase Querying Strategy" below.

Embed book principles (QUANTS/INVEST/Testing Pyramid, Architecture: ADR/fitness functions, Pragmatic Programmer DRY/ETC, Clean Code naming/SOLID) in agent bodies.

Skill Conventions

Structure: skills/skill-name/SKILL.md (no tools, no model — pure knowledge)

Frontmatter: name, description, version

Content: Methodology, checklists, reference patterns. Load on-demand, not always-on.

Agent-Skill Auto-Wiring

All 10 agents are auto-wired to phase-specific skills. When an agent is invoked:

1. Agent frontmatter declares skills: [skill-a, skill-b, ...]
2. Command loads all phase skills into agent context (pure methodology, not tools)
3. Agent system prompt says: "You have access to these skills: X, Y, Z. Apply them."
4. Agent applies skill principles → polished industry-standard outputs

Codebase Querying Strategy

All agents and skills should query the codebase using this priority order:

1. code-review-graph (preferred)
   • Provides structural understanding, dependency graphs, coupling analysis
   • More accurate than text search (understands relationships, not just text)
   • Usage: code_review_graph analyze <path> or code_review_graph query <symbol>
   • Best for: "What depends on this function?", "Show me the call chain", "What's the coupling?"
2. Grep (fallback when code-review-graph unavailable)
   • Fast text search when structural analysis not needed
   • Used for simple pattern matching, file location, quick searches
   • Usage: grep -r "pattern" <path> or Grep tool
   • Acceptable for: "Find all usages of X", "Locate error message", "Find TODO comments"

Why this order?

• code-review-graph understands codebase structure and dependencies
• grep is fast but misses semantic relationships and context
• Fallback ensures agents work even if code-review-graph is unavailable
• Layered approach gives best results with graceful degradation

Extending

New agent:

1. Create agent file with frontmatter skills: [...]
2. Wire into command or /sdlc orchestrator

New skill:

1. Create skills/skill-name/SKILL.md with frontmatter + methodology
2. Add to agent skills: field where relevant

New command:

1. Edit /commands/sdlc.md or create /commands/sdlc-phase-name.md
2. Command injects all phase skills before spawning agents

New phase:

1. Add agents with skills: field
2. Create phase skill with output standards
3. Create command /sdlc-phase-name.md
4. Wire into /sdlc orchestrator

Troubleshooting

• Plugin not loading: Validate frontmatter (name/description/tools/model required), run make validate, restart session
• Agent not invoked: Check command references, verify description matches trigger, review tools field
• Permission prompts: Add patterns to ~/.claude/settings.local.json under permissions.allow

References

Read only when relevant:

• README.md — Feature overview, usage examples
• docs/INTEGRATIONS.md — code-review-graph integration guide and codebase querying best practices
• INSTALLATION.md — Setup and distribution

Agent skills

Issue tracker

Issues are tracked in Linear; agents interact via the Linear API or CLI. See docs/agents/issue-tracker.md.

Triage labels

Default label strings: needs-triage, needs-info, ready-for-agent, ready-for-human, wontfix. See docs/agents/triage-labels.md.

Domain docs

Single-context layout — one CONTEXT.md and docs/adr/ at the repo root. See docs/agents/domain.md.