Product Documentation

docs/architecture/, quickstart.md and current product decisions own self-contained Skill use and maintenance truth. Author semantic composition diagrams as .puml under docs/architecture/diagrams/ and regenerate tracked .svg projections with bash scripts/render-diagrams.sh; do not edit generated SVG by hand.

Reader-facing decision maps use .dot sources under docs/architecture/diagrams/decisions/ and SVG projections under docs/architecture/generated/decisions/; render them with bash scripts/render-decision-diagrams.sh. Decision Guidance maps each diagram to its owning Skill or method. These maps explain conditional choices, not a required agent workflow or executable acceptance contract; do not make their labels or edges CI assertions.

Shared boundaries and all new designs/plans belong to $AGENT_ARCHITECTURE_DIR/docs/; use its plans/skills/ and evaluations/skills/ domains for this product. Historical originals live in its archived/agent-skills/ tree. Product checks must not require these external records or local docs/plans files.

Default search here covers product truth. Honor the architecture owner's search boundary when reading stage artifacts, using explicit paths and rg --no-ignore. Git tracking and search exclusion are distinct. Preserve product rationale, exceptions and compatibility obligations; do not copy the shared architecture explanation back into this tree.