graph-traversal

Graph Traversal

The deep-dive rung of memory search. Once memory-search (rungs 1–2) has surfaced the relevant entities, walk their relationships to answer questions a summary can't: how things connect, what depends on what, and whether something is still true. You enter already holding entry results — so you start at PICK, never at a fresh search.

Hard rules (non-negotiable — read first)

1. Never the first call — and try rung 2 before rung 3. Enter from memory-search's rung 1–2 results already in context. Crucially, rung 2 — search_memory_facts(center_node_id=…, edge_type=…) — is valid-only by default, so it often settles "is X still …?" without the unbounded nav tools. Drop to rung-3 traversal only for what rung 2 can't give (see "When to go deeper" below). If invoked with no entry node in hand, run memory-search rung 1 first.

2. Never trust an edge's currency without checking dates — and one date isn't enough. The nav tools return superseded edges with no warning. Before presenting any edge as current:

   • drop edges whose invalid_at is in the past — expired_at only records when the end was noticed, and a future invalid_at is a scheduled end, still current until it arrives;
   • when two live edges of the same kind disagree (duplicate HAS_STATUS, etc. — it happens), prefer the newest valid_at;
   • for "is X still blocked / true?", also look for a replacement — an old edge can expire while a narrower one takes its place the same day.

   Never trust a node summary for currency: summaries fold old and current facts together with no date.

3. find_path is not for typed chains. It is undirected and edge-type- blind, so it routes through shared hubs — two tickets both "To Do" or sharing a creator get a bogus 2-hop "path" through the Status or Person node, often over a stale edge. Use it only for "are these connected at all", and distrust any path running through a status/person hub. For a dependency or impact chain, walk get_node_edges(edge_type=BLOCKS|DEPENDS_ON) hop by hop.