graph.*

graph.* is the gateway's read/write surface over the agent's knowledge graph (the same one the Panel renders as a 2D / 3D galaxy). The data is NoteStore + notes_graph_cache on the in-core MemoryBackend, not a per-call LLM synthesis — every method here is a deterministic read or a direct indexer write.

The three read methods (graph.query, graph.search, graph.node_detail) are unconditional. The three write methods (graph.update_note, graph.rename_note, graph.delete_note) require a live NoteIndexer; in simulated mode the writes return Method not found.

The 26.7.21 changelog added a loop-graph governance layer (the harness's persistent store for goal / audit / team nodes); that lives on top of teams.task.trace + the governance.* event topic and is reachable through events.subscribe, not through graph.*.

Methods

graph.query

Read the note index for an agent — nodes (with community id from the Louvain cache) and edges (with kind and confidence). Similarity edges materialized by GraphRecompute are deduped against real wikilink edges.

Parameter	Type	Required	Description
`limit`	number	no	Defaults to 100
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: GraphQueryResponse: { "nodes": [ ...NoteNodeDto... ], "edges": [ ...NoteLinkDto... ], "total": <number> | null, "bridge_nodes": [ ...path... ], "surprising_edges": [ [from, to]... ] }. bridge_nodes and surprising_edges are filtered to nodes / edges visible in the response; empty until the first GraphRecompute run.

NoteNodeDto carries id (path category/title), name (filename), path, category, tags, link_count, community_id?, updated_at?. NoteLinkDto carries from, to, label?, kind (wikilink for plain body wikilinks, related_similarity for the materialized similarity edges), confidence?.

graph.search

Full-text search over note content via the NoteStore FTS index.

Parameter	Type	Required	Description
`query`	string	yes	FTS query string
`limit`	number	no	Defaults to 20
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: { "results": [ ...SearchResultDto... ] }. Each row carries id, name, category, match_field (title when the query hit the filename, content otherwise), agent_id, created_at, updated_at, tags, link_count.

graph.node_detail

Fetch one node's full row, markdown content (read from disk via notes_dir()), incoming backlinks, and outgoing links with full lifecycle provenance.

Parameter	Type	Required	Description
`node_id`	string	yes	Note path `category/title`
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: NoteDetailResponse: { "node": { ...NoteNodeDto... }, "content", "backlinks": [ ...path... ], "outgoing": [ ...OutgoingLinkDto... ] }. outgoing carries to (resolved or raw), raw, relation?, label?, confidence, resolved_by?, status (active / dangling / tombstone).

graph.update_note

Persist the edited markdown for a single note verbatim — the body is written byte-for-byte (not the lossy KnowledgeNote::to_markdown reconstruction), so hand-edited prose, headings, and code blocks survive. The incoming content is scanned for exfiltration payloads before persisting (exfiltration-only scope so a user's own security-research notes are not rejected).

Parameter	Type	Required	Description
`node_id`	string	yes	Note path `category/title`
`content`	string	yes	Full raw markdown (frontmatter + body)
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: { "node_id", "saved": true }. INVALID_PARAMS on a node_id without category/, on a category or agent_id containing path-traversal components (.., \, / in agent_id), or when the exfiltration scan refuses the body.

graph.rename_note

Rename the file, rewrite every inbound [[old]] wikilink, re-index, and backfill. The new path is re-derived from find_by_filename so a stale client category does not break the rename.

Parameter	Type	Required	Description
`node_id`	string	yes	Current note path `category/title`
`new_title`	string	yes	New filename (without `.md`); sanitized server-side
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: { "node_id", "new_id", "renamed": true }. new_id is the canonical category/new_title (re-derived from the file system, not just the client-supplied category). INVALID_PARAMS on the same path-traversal cases as graph.update_note.

graph.delete_note

Remove the file + index. Inbound links become tombstones (the source bodies are untouched, so the rename-revival path is preserved).

Parameter	Type	Required	Description
`node_id`	string	yes	Note path `category/title`
`agent_id`	string	no	Defaults to `routing::DEFAULT_AGENT_ID`

Response: { "node_id", "deleted": true }. INVALID_PARAMS on path-traversal cases.

Methods that are not RPC

The following methods are described in earlier versions of this page but are not registered on the current aleph-server:

Method	Why it is not an RPC
`graph.entities` / `graph.add_entity` / `graph.add_relation` / `graph.delete_entity`	There is no free-standing `Entity` table; the graph is built from the `notes_index` (`KnowledgeNote` + wikilinks + similarity edges). The Panel calls `graph.query` for a render and never had a single-entity insert path.
`graph.neighbors`	Surfaces the same data as the `edges` of `graph.query`; the Panel iterates the edges client-side.

graph.neighbors is occasionally referenced in older design docs; in the runtime it is just a property of the edges array returned by graph.query.