| name | sdlc |
| description | SDLC orchestrator — new projects (/sdlc init), codebase onboarding (/sdlc onboard), feature addition (/sdlc feature), audit & improvement (/sdlc improve). Runs Discovery Interview before producing any documents. Route here whenever the user asks to review, audit, evaluate, or improve a codebase — never freelance the analysis. |
SDLC Workflow
Load and follow the instructions in the sdlc-lead agent.
Four operating modes:
-
/sdlc init <name> "<desc>" — New project: phases 0-5 with proper engineering artifacts (SRS, SAD, C4 diagrams, sequence diagrams, ERD).
Starts with a Discovery Interview — asks 7 targeted questions about problem, users, constraints, and tech before writing any documents.
-
/sdlc onboard [--quick | --deep] — Existing codebase: reverse engineer, produce architecture docs, C4 diagrams, onboarding guide.
-
/sdlc feature "<description>" — Add feature: starts with a Feature Discovery Interview (scope, success criteria, constraints), then impact analysis, modular design, backward compatibility, implementation, verification.
Asks Design Clarification Questions before architecture work begins.
-
/sdlc improve ["<focus>"] — Audit & improve an existing system. Runs specialist audits (UX, code quality, performance, security, database), synthesizes a prioritized backlog, gets user approval, then routes fixes through coding-agent (S/M items) or spawns Mode 3 sub-workflows (L items). Use this whenever the user asks to review, evaluate, audit, or improve a codebase — including phrases like "review for gaps", "what could we improve", "audit this", "the UI looks bad", "make it better", "this is slow".
Other commands:
/sdlc status — Current phase/milestone progress/sdlc resume — Continue after clearing context: rehydrates from docs/work/STATE.md (the after-each-step checkpoint) instead of chat history, so you can /clear on a large loop and pick up exactly where you left off/sdlc gate — Check exit criteria before advancing
The lead delegates to expert agents via HANDOFF blocks (and the task tool for git-expert only) — it coordinates, it doesn't do technical work itself.
Natural-language routing (MANDATORY). When the user asks for any of the following without naming a mode, route into Mode 4 (/sdlc improve) — do not start scanning files yourself:
- "review the product / codebase"
- "find gaps / look for improvements / what should we fix"
- "audit (the code / UX / performance / security)"
- "make it better / this needs work / the UI feels off"
- "is there anything wrong with X"
- "evaluate / give me an assessment / health check"
If the project has not been onboarded yet (docs/LANDSCAPE.md missing), Mode 4 runs a lightweight landscape scan first; you do not need to run /sdlc onboard separately.
Interactive questioning phases:
- Mode 1 (init): Discovery Interview before Phase 0, Design Clarification before Phase 3
- Mode 3 (feature): Feature Discovery Interview before impact analysis
- Mode 4 (improve): Improvement Discovery Interview before any audit fans out
Confidence loops: Each phase gate runs a score loop (1-10 per deliverable, min 7 to advance, up to 3 revision iterations) — not a one-shot pass/fail.
Architecture principles enforced:
- Feature-sliced directory structure (not layered)
- Interface-driven design with dependency injection
- Mermaid diagrams for all architecture documentation
- Modular code with clear module boundaries
