@bazidiy/ontology

BaziDIY ontology plugin: deterministic wuxing/bead/style rules exposed as model-facing dsh tools. No LLM calls inside execute — the rules derive 旺衰/喜用神/珠子匹配/款式方案 from data only.

Tools

Tool	Input	Output type
calculate_bazi	birth_date, birth_hour, gender	bazi_result
propose_designs	bead_ids?, wrist_size?, day_master_element, month_branch_wuxing	design_proposal
generate_design	style_name, beads, wrist_size?, summary, rationale	design_result
save_design	style_name, slots, wrist_size, summary, rationale	design_saved
load_design	—	design_loaded

Data

src/data/*.ts is generated from backend/ontology/data (schema v3, hardness removed). Regenerate with node scripts/gen-data.mjs.

Model Experience

What the model sees

Five tool schemas. calculate_bazi returns day_master_element / month_branch_wuxing which the model passes straight into propose_designs. propose_designs returns favorable / unfavorable / designs that the model quotes verbatim. generate_design validates bead names are full names (no abbreviations). save_design persists the settled design per session id; load_design restores it in the same session.

Token effect

Schema cost per request where the tools are visible; result tokens are data-dependent.

KV Cache effect

Prefix-stable while the tool definitions and visibility are unchanged.

Known Limitations and Deferred Work

• calculate_bazi ignores gender; 节气 uses a simplified month boundary (same as the v8 reference).
• generate_design validates bead names but not that the style_name/beads sequence came from propose_designs — the prompt constrains this, not the tool.
• save_design/load_design key by the calling agent's session id; a design is only reachable from the session that saved it.