sbeeredd04/software-factory

Use to discover agent capabilities, choose risk-aware worker routes, project approved principles into selected review-bot rules, or triage pull-request reviews. Keeps native fallback, preserves user files, and renders commands without executing agents or GitHub writes.

Maps a codebase read-only into CODEBASE.md: module map, call edges and entry points from the Python AST, documented best effort for other languages. Use when the factory needs to learn or refresh its picture of a codebase.

Routes design work to sf-smell-audit, sf-refactor, sf-patterns and the architecture references. Use for dirty code, code smells, refactoring, technical debt, design patterns, a design report card, or architecture calls: service split, CQRS, event-driven, hexagonal, decision records.

Makes evals a precondition of work: per-surface suites signed off before stories queue, per-story fail_to_pass/pass_to_pass cases with stable IDs, deterministic-first grading, cost-tiered runs, sha256 receipts. Use for eval setup, story evaluation or failure triage.

Use to register a skill or MCP capability, preserve a reviewed factory customization through reinstall, customize pinned BMad, or turn observed rework into a bounded independently reviewed improvement with canary and rollback evidence.

Set up or re-onboard Software Factory in a project. Read existing context, prefill twenty grouped decisions, propose configuration and document drafts, and apply only the approved setup. Use for first setup, missing configuration, or reviewing an existing setup.

Judges whether a Gang of Four design pattern belongs, or whether it would be speculative generality. Use when the user asks about design patterns, Strategy, Factory, Singleton, Facade, or whether to add a pattern during a refactor.

Authors the project's core principles with the user and keeps the committed .software-factory/context/ home honest. Use when drafting or changing principles (product, conventions, risk, evals), judging work against them, or recording a decision.

Chooses refactoring techniques and a safe order of moves for a smell that was already measured. Use when the user asks how to refactor, which technique to apply, or how to pay down technical debt without a rewrite.

Audits a codebase against the 23 Fowler code smells and writes a scored report card. Use when the user asks to rate dirty code, score code smells, find bloat, or produce a design report card for a repository.

Turns a plain request (feature, bug, refactor) into shipped, proven code. The main agent orchestrates: grounds work in project principles, plans with BMad epics and stories, staffs best-fit agents, demands independent proof, merges to the integration branch and checkpoints so work survives resets.