What problem does it solve?
Single-shot top-k search misses the full picture on recall tasks like "where was X mentioned" or "have we ever discussed Y". This Skill teaches agents how to run multi-query, multi-scope searches with union dedup, read original sources instead of summaries, and know exactly when to stop searching.
Core Features & Use Cases
- 8 Question-Type Recipes: Maps recall task types (definition lookup, relationship traversal, decision archaeology, cold-start onboarding, full coverage, source-map provenance, absence check, delta comparison) to concrete search sequences using search_evidence, graph_resolve, get_thread_context, and list_recent.
- Agent-Driven Query Expansion: The agent expands queries with domain knowledge, synonyms, abbreviations, and cross-language variants instead of relying on black-box system expansion.
- Stopping Criteria Calibration: Per-question-type stop conditions plus bidirectional calibration for model families that stop too early or too late.
- Use Case: When asked "where have we mentioned AUDHD across all docs and threads", run hybrid doc search plus semantic thread search, expand to related terms in Chinese and English, follow source thread links from canonical docs, and output a coverage matrix.
Quick Start
Ask the agent to find every place a topic has been mentioned across docs and threads and produce a coverage matrix with sources.
