architecture

Lead architecture decisions. Produce architectural decisions (components, data flow, APIs, dependencies). Hands off via GitHub issue body under ## Implementation plan.

Process

1. Research

Spawn in parallel:

• Agent("agents/workflow-codebase-scanner.md") — pass cwd and scope (feature area or module list).
• Agent("agents/workflow-patterns-researcher.md") — pass problem and tech_stack. Gate: skip if codebase-scanner returns >= 3 internal patterns (unless security/payments/privacy domain).

2. Analyze

Spawn Agent("agents/workflow-constraint-analyzer.md") — pass problem, cwd, codebase_findings and patterns_findings from step 1 outputs.

3. Decide

Invoke Skill("grill-me") with devil's advocate. For each major point, evaluate 2-3 approaches with trade-offs, diagrams, and code structure previews. User selects the recommendation.

4. Deepen

Scan for vague language or thin sections → spawn Agent("agents/workflow-deepening-agent.md") per gap — pass gap, cwd, and context (summary of prior research). Max 2 rounds; user approves before each dispatch.

5. Output

Invoke Skill("preflight"). Read _shared/handoff-artifact.md. Write decisions to issue body under ## Implementation plan:

• Component diagram (Mermaid).
• Key interfaces and data flow.
• Sub-issues with GitHub relationships.
• Dependency graph for parallelization.
• Research summary (informed patterns).

Rules

• Code-First: Never propose architecture without reading existing code.
• Pattern Adherence: Respect existing patterns unless justifying deviation.
• Concrete Only: No vague placeholders; every section must be decisive.
• Stay interactive: Never skip user-facing deliberation — the research phase is pre-work, not a replacement for discussion.