v0.13.1

Request and response shapes checked against a live v0.12.0 server (the same contracts as
v0.11.108). Every fix below was a call that failed or a field that was wrong on both servers.

Changed (signatures)

• memory_link() (sync and async) takes a required keyword agent_id and an optional
  label. The server's POST /v1/memories/{id}/links requires agent_id; without it every
  call failed with 422, so no working code breaks. edge_type is still accepted but not sent:
  the server records every explicit link as linked_by.
• DakeraDelegationHelper.link_delegation() (TealTiger integration) takes a required keyword
  agent_id. Links are linked_by edges labelled delegated_from (the server cannot store a
  custom edge type, so the old delegated_from edges were never created) and
  get_delegation_chain() follows linked_by edges.
• summarize(dry_run=True) raises ValueError before sending anything: the server has no dry
  run and always stores the summary memory (dry_run was silently ignored). summarize() also
  raises ValueError with fewer than two memory_ids, and knowledge_graph() without
  memory_id; the server answered both with 422.

Added

• tags on recall() and search_memories() (sync and async): memories carrying at least one
  of the tags. The server accepts it; the SDK had no way to send it.
• EdgeType.SUPERSEDES (edges the server creates when a newer memory supersedes an older one);
  EdgeType is now lenient, so a type this SDK does not know parses instead of raising.
• GraphNode.edges, MemoryGraph.node_count, GraphLinkResponse.from_id / to_id /
  edge_type, GraphExport.namespace / edges, MemoryEntitiesResponse.count.

Fixed

• update_memory() (sync and async) sends agent_id as a query parameter. PUT /v1/memory/update/{id} reads it from the query string, so every update failed with
  400 INVALID_REQUEST (missing field agent_id).
• AsyncDakeraClient.get_memory() sends agent_id as a query parameter too (the sync client
  already did); GET /v1/memory/get/{id} needs it.
• Graph edges: the server sends {from_id, to_id, edge_type, weight, created_at} with no edge
  id. GraphEdge.from_dict() required id, source_id and target_id, so memory_graph(),
  knowledge_query(), knowledge_export() and agent_graph_export() raised KeyError on any
  graph with edges. source_id / target_id now carry from_id / to_id; id is "".
• memory_graph(): the server lists edges per node; MemoryGraph.edges is now their union
  (it was always empty). types is applied client-side (the server ignores it and returns
  every type).
• memory_path() sent the target as target; the server reads to, so every call failed with
  400. The answer {from_id, to_id, path, hop_count} is parsed into GraphPath.
• memory_link() parsed {"edge": ...}; the server answers {from_id, to_id, edge_type}
  (KeyError after the request). GraphLinkResponse.edge is built from it (weight 1.0).
• agent_graph_export() read format and data; the server answers {agent_id, namespace, node_count, edge_count, edges} and ignores format (KeyError). GraphExport.data holds the
  answer as a JSON string.
• memory_entities(): the server answers {entities, count} without memory_id (KeyError);
  memory_id is now the requested id.
• Examples: advanced.py and playground/quickstart.py link with agent_id;
  knowledge_graph.py is rewritten against the real API (it called methods with arguments the
  SDK does not have).
• Integration tests cover update/get (sync and async), recall with tags, link, graph, path,
  agent export, knowledge query and memory entities against the real server.

What's Changed

Bug Fixes

• fix: request and response shapes the server really uses (0.13.1) @ferhimedamine (#200)

Full Changelog: v0.13.0...v0.13.1