DEC-0005 — Expert team & first skills

Context

With the mission, knowledge model, and audit loop in place, we need the "development team" — and per DEC-0004, it should be Claude primitives (subagents = team, skills = procedures), not a bespoke framework.

Options considered

1. One generalist agent doing everything.
2. A bespoke agent framework.
3. A roster of specialist subagents + Agent Skills, one per architecture layer.

Decision

Create the roster in .claude/agents/ and the first skills in .claude/skills/:

Subagents — principal-architect, knowledge-architect (principals); ingestion-engineer, extraction-engineer (the moat), mcp-engineer, platform-engineer, qa-reviewer; plus the existing log-auditor.

Skills — author-okf, verify-kb, scaffold-vertical, ingest-client.

Rationale

• Maps every layer (ingest → extract/OKF → serve → agentic foundation) to a primitive; mirrors the deep-research roster.
• The team is the fastest way to make the platform real and build everything else.
• The docs-site / frontend agent is intentionally deferred — output rendering is one downstream surface, not the product.

Consequences

• Subagents and skills activate on the next Claude Code session start (they are not picked up mid-session — same as hooks).
• The roster and skills will evolve as the Agent SDK runtime is built (extraction, MCP, provisioning).

Review

Revisit after the extraction runtime exists and the first DXA vertical is scaffolded — confirm the roster boundaries still hold.