use serde_json::json;
use crate::protocol::ToolDef;
pub fn all_tools(allow_labels: bool) -> Vec<ToolDef> {
let _ = allow_labels;
let search_schema = json!({
"type": "object",
"properties": {
"label": { "type": "string", "description": "Node label (e.g. 'Person'). Honoured by default. Set MNEM_LABELS=0 (or legacy MNEM_BENCH=0) at server launch to force every label to Node::DEFAULT_NTYPE." },
"where": { "type": "object", "description": "Optional prop-equality filter, e.g. {\"name\": \"Alice\"}. Single property only in this version." },
"with_outgoing": { "type": "array", "items": { "type": "string" }, "description": "Edge labels to include on each hit." },
"limit": { "type": "integer", "minimum": 1, "maximum": 500, "default": 10 }
},
"additionalProperties": false
});
let commit_nodes_item_schema = json!({
"type": "object",
"properties": {
"ntype": { "type": "string", "description": "Node type / label. Honoured by default. Set MNEM_LABELS=0 (or legacy MNEM_BENCH=0) at server launch to force the handler to substitute Node::DEFAULT_NTYPE." },
"summary": { "type": "string", "description": "Short LLM-facing summary. Indexed by text + retrieve." },
"props": { "type": "object" },
"content": { "type": "string", "description": "Optional text/markdown body (UTF-8)." }
},
"additionalProperties": false
});
let list_nodes_schema = json!({
"type": "object",
"properties": {
"label": { "type": "string", "description": "Optional label (ntype) filter. Honoured by default. Set MNEM_LABELS=0 (or legacy MNEM_BENCH=0) at server launch to force the filter to be silently dropped." },
"limit": { "type": "integer", "minimum": 1, "maximum": 1000, "default": 50 },
"offset": { "type": "integer", "minimum": 0, "default": 0 }
},
"additionalProperties": false
});
let resolve_or_create_schema = json!({
"type": "object",
"properties": {
"label": { "type": "string", "description": "Node label / kind. Honoured by default. Set MNEM_LABELS=0 (or legacy MNEM_BENCH=0) at server launch to force the handler to substitute Node::DEFAULT_NTYPE." },
"kind": { "type": "string", "description": "Alias for `label`. Pick one." },
"prop_name": { "type": "string", "description": "Property to anchor the find-or-create on. Defaults to `name` when the `name` alias is used." },
"name": { "type": "string", "description": "Alias for the natural-language entity name. When set, `prop_name` defaults to \"name\" and `value` defaults to this string." },
"value": { "description": "String, number, bool, or JSON object/array. Canonicalised before indexing." },
"agent_id": { "type": "string", "description": "Commit author. Defaults to 'mnem mcp' when absent." },
"task_id": { "type": "string" },
"extra_props": { "type": "object", "description": "Additional properties to set if the node has to be created." },
"global": { "type": "boolean", "description": "When true, also resolve-or-create the same entity in the global graph (~/.mnemglobal/.mnem/) and stamp its UUID as `_global_anchor` on the local node. Best-effort: silently skipped if the global graph has not been initialised." }
},
"additionalProperties": false
});
let retrieve_schema = json!({
"type": "object",
"properties": {
"label": { "type": "string", "description": "Label filter. Honoured by default. Set MNEM_LABELS=0 (or legacy MNEM_BENCH=0) at server launch to force the filter to be silently dropped." },
"where": { "type": "object", "description": "Optional single-property equality gate, e.g. {\"team\": \"eng\"}." },
"text": { "type": "string", "description": "Query text. Retained so a reranker can read (query, candidate) pairs jointly. For retrieval proper, pass a `vector` in the matching embed model or configure the sparse lane separately." },
"vector": {
"type": "object",
"properties": {
"model": { "type": "string", "minLength": 1 },
"values": { "type": "array", "items": { "type": "number" }, "minItems": 1 }
},
"required": ["model", "values"],
"additionalProperties": false
},
"token_budget": { "type": "integer", "minimum": 0, "description": "Max rendered-text tokens to return. Default: unlimited." },
"limit": { "type": "integer", "minimum": 1, "description": "Max items to return, independent of the token budget. No hard ceiling; callers own back-pressure." },
"vector_cap": { "type": "integer", "minimum": 1, "description": "Override the per-lane cap on vector candidates (default: retriever-built-in). Raising it lets rerank / graph-expand see more of the long tail." },
"rerank_top_k": { "type": "integer", "minimum": 1, "description": "If a reranker is wired in via the host config, how many fused candidates to rerank. Has no effect without a reranker." },
"fusion": { "type": "string", "enum": ["convex_min_max", "rrf"], "description": "Rank-fusion strategy over the lane outputs. `convex_min_max` (default) per Bruch 2023; `rrf` for the classic Reciprocal Rank Fusion baseline." },
"graph_expand": { "type": "integer", "minimum": 1, "description": "Enable graph-expand: after hybrid fusion produces a top-K, traverse authored edges up to this many frontier nodes. Disables when absent." },
"graph_decay": { "type": "number", "minimum": 0.0, "maximum": 1.0, "description": "Score decay applied per hop during graph-expand. Default preserves retriever built-in." },
"graph_depth": { "type": "integer", "minimum": 1, "maximum": 4, "description": "Multi-hop traversal depth. 1 = single-hop; 2+ for MuSiQue-style compositional queries. Clamped to [1, 4]." },
"graph_etype": { "type": "array", "items": { "type": "string" }, "description": "Edge-type allowlist for graph-expand. Empty / absent means all edge types." },
"graph_max_per_seed": { "type": "integer", "minimum": 1, "description": "Per-seed outgoing-edge cap: prevents a hot-seed node from starving siblings in the global graph_expand budget." },
"graph_mode": { "type": "string", "enum": ["decay", "ppr"], "description": "Graph-expand strategy. `decay` (default) = historical BFS with decay^depth scoring; `ppr` = personalised PageRank over the hybrid adjacency index (E2+). PPR falls through to decay when no adjacency index is wired." },
"ppr_damping": { "type": "number", "minimum": 0.0, "maximum": 0.999, "description": "PPR damping factor. Default 0.85. Ignored unless graph_mode = \"ppr\"." },
"ppr_iter": { "type": "integer", "minimum": 1, "description": "PPR power-iteration cap. Default 15. Ignored unless graph_mode = \"ppr\"." }
},
"additionalProperties": false
});
#[cfg_attr(not(feature = "summarize"), allow(unused_mut))]
let mut tools: Vec<ToolDef> = vec![
ToolDef {
name: "mnem_stats",
description: "Repository overview: op-head, head commit, ref summary, known labels. \
Cheap; call this first to discover what a repo contains.",
input_schema: json!({
"type": "object",
"properties": {},
"required": [],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_schema",
description: "List every node label and edge label present in the current commit, \
along with the property names the IndexSet has built for each label. \
Agents use this to write well-scoped queries.",
input_schema: json!({
"type": "object",
"properties": {},
"required": [],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_search",
description: "Search for nodes. Uses the indexed path when a label + exact property \
match is specified; falls back to label-scoped scan or full scan \
otherwise. Optionally include each hit's outgoing edges of named \
labels.",
input_schema: search_schema,
},
ToolDef {
name: "mnem_get_node",
description: "Fetch a single node by UUID (as returned by mnem_search / mnem_commit). \
Returns full props + content size + outgoing edge count.",
input_schema: json!({
"type": "object",
"properties": {
"id": { "type": "string", "description": "Node UUID (hyphenated form)." }
},
"required": ["id"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_traverse",
description: "From a start node, list outgoing neighbours reachable via specified \
edge labels. One-hop only in this version; deeper traversal lands in a future version.",
input_schema: json!({
"type": "object",
"properties": {
"start": { "type": "string", "description": "Start node UUID." },
"edge_labels": { "type": "array", "items": { "type": "string" }, "description": "Edge labels to follow." },
"limit": { "type": "integer", "minimum": 1, "maximum": 200, "default": 25 }
},
"required": ["start"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_commit",
description: "Add nodes and/or edges as a single commit. `agent_id` (required) is \
stored as the Commit author. `task_id` is accepted and reserved for \
future Operation.task_id plumbing (tracked in ); today it is \
not persisted. Returns the new op-id, commit CID, and created node UUIDs.",
input_schema: json!({
"type": "object",
"properties": {
"agent_id": { "type": "string", "description": "Required. Stored as the Commit author." },
"task_id": { "type": "string", "description": "Reserved. Accepted but not yet persisted ." },
"message": { "type": "string", "default": "" },
"nodes": {
"type": "array",
"items": commit_nodes_item_schema
},
"edges": {
"type": "array",
"items": {
"type": "object",
"properties": {
"etype": { "type": "string" },
"src": { "type": "string", "description": "Source node UUID." },
"dst": { "type": "string", "description": "Destination node UUID." },
"props": { "type": "object" }
},
"required": ["etype", "src", "dst"],
"additionalProperties": false
}
}
},
"required": ["agent_id"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_commit_relation",
description: "Compound write: resolve-or-create a subject node, resolve-or-create an \
object node, and connect them with a typed edge - all in one commit. \
Audit fix G6 (2026-04-25): collapses the 3-tool dance \
(resolve_or_create + resolve_or_create + commit-edge) that an LLM under \
no specific instruction was unlikely to perform fully, leaving the graph \
flat. Anchor property defaults to `name`; pass `anchor` to switch to \
`email` / `slug` / `id`. Typical call: \
{\"subject\": \"Alice\", \"subject_kind\": \"Entity:Person\", \
\"predicate\": \"works_at\", \"object\": \"Globex\", \
\"object_kind\": \"Entity:Organization\"}.",
input_schema: json!({
"type": "object",
"properties": {
"subject": { "type": "string", "description": "Subject entity natural-language name (the value of the anchor property)." },
"subject_kind": { "type": "string", "description": "Subject ntype (e.g. 'Entity:Person'). Honoured when labels are enabled (default); otherwise the handler substitutes Node::DEFAULT_NTYPE." },
"predicate": { "type": "string", "description": "Edge type (e.g. 'works_at', 'lives_in', 'has_preference')." },
"object": { "type": "string", "description": "Object entity natural-language name (the value of the anchor property)." },
"object_kind": { "type": "string", "description": "Object ntype (e.g. 'Entity:Organization'). Honoured when labels are enabled (default); otherwise the handler substitutes Node::DEFAULT_NTYPE." },
"anchor": { "type": "string", "default": "name", "description": "Property name to anchor the resolve_or_create on. Defaults to `name`." },
"subject_props":{ "type": "object", "description": "Optional extra props to set on the subject node." },
"object_props": { "type": "object", "description": "Optional extra props to set on the object node." },
"edge_props": { "type": "object", "description": "Optional props to set on the edge." },
"agent_id": { "type": "string", "description": "Commit author. Defaults to 'mnem mcp' when absent." },
"message": { "type": "string", "default": "mnem_mcp commit_relation" }
},
"required": ["subject", "predicate", "object"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_delete_node",
description: "Remove a node from the current head. Commits a new op with the removal. \
The node is no longer reachable from the new commit's node tree, but its \
prior CID and any prior commits that referenced it remain addressable \
(mnem's history is append-only). Edges incident to the node are NOT \
auto-removed; delete them explicitly or via a future cascade flag.",
input_schema: json!({
"type": "object",
"properties": {
"id": { "type": "string", "description": "Node UUID to remove." },
"agent_id": { "type": "string", "description": "Required. Stored as the Commit author." },
"message": { "type": "string", "default": "mnem_mcp delete" }
},
"required": ["id", "agent_id"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_tombstone_node",
description: "Logically \"forget\" a node without deleting its content. Unlike \
mnem_delete_node this does NOT remove the node from the node tree - the \
node's CID stays stable and any prior edges / commits that reference \
it remain intact. What changes is that subsequent retrieves filter the \
node out by default (agent can no longer see the memory). Use this when \
a user says \"forget X\" or revokes consent; use mnem_delete_node only \
when the goal is to free storage, not memory hygiene. Errors if the \
node does not exist or has already been tombstoned.",
input_schema: json!({
"type": "object",
"properties": {
"id": { "type": "string", "description": "Node UUID to tombstone." },
"reason": { "type": "string", "description": "Free-form reason recorded on the tombstone (e.g. the user's own phrasing)." },
"agent_id": { "type": "string", "description": "Required. Stored as the Commit author." },
"message": { "type": "string", "default": "mnem_mcp tombstone" }
},
"required": ["id", "agent_id"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_list_nodes",
description: "Enumerate nodes at the current head, optionally filtered by label. \
Returns UUID + label + optional summary per node. Cheap discovery tool \
an agent can call before composing a retrieval: lets it see what's in \
the repo without a text-search guess.",
input_schema: list_nodes_schema,
},
ToolDef {
name: "mnem_resolve_or_create",
description: "Find-or-create a node by a primary-key property. Accepts EITHER the \
friendly `{name: \"Alice\", kind: \"Person\"}` shape (anchors on the \
`name` property) OR the canonical \
`{prop_name: \"email\", value: \"a@x\", label: \"Person\"}` shape \
(anchors on whatever property you choose). If a node with the same \
(label, anchor-property) == value already exists, its UUID is \
returned; otherwise a new node is committed. Prevents the duplicate-\
entity problem agents hit when the same fact is re-asserted across \
tool calls. audit-2026-04-25 C3-10: `name`/`kind` aliases added.",
input_schema: resolve_or_create_schema,
},
ToolDef {
name: "mnem_recent",
description: "Walk the op-log from the current head backwards. Returns the last N \
operations with time, author, agent_id, task_id, and one-line message.",
input_schema: json!({
"type": "object",
"properties": {
"limit": { "type": "integer", "minimum": 1, "maximum": 100, "default": 10 }
},
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_vector_search",
description: "Cosine-similarity nearest-neighbour search over stored node embeddings. \
Pass the embedding-model identifier and a query vector; receive the \
top-k matches. Nodes whose embedding.model differs from the query are \
silently skipped - each index binds to one (model, dim).",
input_schema: json!({
"type": "object",
"properties": {
"model": { "type": "string", "minLength": 1 },
"vector": { "type": "array", "items": { "type": "number" }, "minItems": 1 },
"k": { "type": "integer", "minimum": 1, "maximum": 500, "default": 10 }
},
"required": ["model", "vector"],
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_retrieve",
description: "Composite retrieval: combines label + prop-eq filters with cosine \
vector search and (optionally) the learned-sparse lane, fuses ranked \
lists via min-max convex combination (Bruch 2023) or RRF, optionally \
runs multi-hop graph expansion over the authored edges, and greedily \
packs rendered nodes under a token budget. Use this as the default \
tool when assembling LLM context: it returns nodes pre-rendered to \
text plus tokens_used / dropped / candidates_seen metadata so you \
know whether the budget was tight. All retrieval knobs exposed by \
POST /v1/retrieve are available here so MCP callers reach parity \
with the HTTP surface.",
input_schema: retrieve_schema,
},
ToolDef {
name: "mnem_ingest",
description: "Ingest a source as a Doc + Chunk + Entity subgraph. Accepts EITHER \
{path: \"<file>\"} (server reads the file from disk) OR \
{text: \"...\", source?: \"label\"} (caller has already buffered the \
document). Runs parse + chunk + rule-based-NER and commits in one \
transaction. Chunker choice: 'auto' (picks per source kind), \
'paragraph' (blank-line split, best for markdown), 'recursive' \
(token-budgeted sliding window, best for PDFs), 'session' (groups \
conversation messages). Typical calls: \
{\"path\": \"notes.md\"}, \
{\"path\": \"book.pdf\", \"chunker\": \"recursive\", \"max_tokens\": 1024}, \
{\"text\": \"Alice met Bob.\", \"source\": \"convo-2026-04-25\"}. \
File / text size is capped at 32 MiB and max_tokens at 8192 for DoS \
resistance. Returns commit_cid plus per-run node / chunk / entity / \
relation counts. audit-2026-04-25 C3-8: schema accepts both shapes.",
input_schema: json!({
"type": "object",
"properties": {
"path": { "type": "string", "description": "Absolute or relative path to the source file on the MCP server's filesystem. Mutually exclusive with `text`." },
"text": { "type": "string", "description": "Inline document body. Use this when the caller already has the bytes; mutually exclusive with `path`." },
"source": { "type": "string", "description": "Cosmetic label rendered as the `path:` field in the output when ingesting via `text`. Defaults to 'inline-text'." },
"ntype": { "type": "string", "description": "Root Doc node label (default 'Doc').", "default": "Doc" },
"chunker": { "type": "string", "enum": ["auto", "paragraph", "recursive", "session"], "default": "auto" },
"max_tokens": { "type": "integer", "minimum": 1, "maximum": 8192, "default": 512 },
"overlap": { "type": "integer", "minimum": 0, "maximum": 8192, "default": 32 },
"agent_id": { "type": "string", "description": "Commit author. Defaults to 'mnem mcp' when absent." },
"message": { "type": "string", "default": "mnem_mcp ingest" }
},
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_global_retrieve",
description: "Semantic search on the global anchor graph (~/.mnemglobal/.mnem/) only. \
Always targets the global graph regardless of which repo the MCP server \
is pointed at. Use this when you explicitly want to read from the shared \
cross-session memory store. Use mnem_retrieve for the current local repo.",
input_schema: json!({
"type": "object",
"properties": {
"text": { "type": "string", "description": "Query text. Passed as BM25/rerank input and optionally auto-embedded when an embedder is configured." },
"vector": {
"type": "object",
"description": "Pre-computed query vector.",
"properties": {
"model": { "type": "string" },
"values": { "type": "array", "items": { "type": "number" } }
},
"required": ["model", "values"],
"additionalProperties": false
},
"limit": { "type": "integer", "minimum": 1, "maximum": 1000, "default": 10, "description": "Max results to return." },
"token_budget": { "type": "integer", "minimum": 1, "description": "Soft token cap on total rendered output." }
},
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_global_ingest",
description: "Ingest a source as a Doc + Chunk + Entity subgraph directly into the \
global anchor graph (~/.mnemglobal/.mnem/). Always targets the global \
graph regardless of which repo the MCP server is pointed at. Accepts \
EITHER {path: \"<file>\"} (server reads the file from disk) OR \
{text: \"...\", source?: \"label\"} (caller has already buffered the \
document). Same chunker options as mnem_ingest. Use this for documents \
that should be queryable across all sessions and projects.",
input_schema: json!({
"type": "object",
"properties": {
"path": { "type": "string", "description": "Absolute or relative path to the source file on the MCP server's filesystem. Mutually exclusive with `text`." },
"text": { "type": "string", "description": "Inline document body. Mutually exclusive with `path`." },
"source": { "type": "string", "description": "Cosmetic label for the `path:` field when ingesting via `text`. Defaults to 'inline-text'." },
"ntype": { "type": "string", "description": "Root Doc node label (default 'Doc').", "default": "Doc" },
"chunker": { "type": "string", "enum": ["auto", "paragraph", "recursive", "session"], "default": "auto" },
"max_tokens": { "type": "integer", "minimum": 1, "maximum": 8192, "default": 512 },
"overlap": { "type": "integer", "minimum": 0, "maximum": 8192, "default": 32 },
"agent_id": { "type": "string", "description": "Commit author. Defaults to 'mnem mcp' when absent." },
"message": { "type": "string", "default": "mnem_mcp global_ingest" }
},
"additionalProperties": false
}),
},
ToolDef {
name: "mnem_global_add",
description: "Write nodes and/or edges directly to the global graph \
(~/.mnemglobal/.mnem/). Use this when an entity or fact should \
belong to the shared cross-repo graph rather than (or in addition \
to) the current local repo. Typical use: named entities \
(people, orgs, places) that appear across multiple projects.",
input_schema: json!({
"type": "object",
"properties": {
"nodes": {
"type": "array",
"items": {
"type": "object",
"properties": {
"ntype": { "type": "string", "description": "Node label (e.g. 'Entity:Person'). Defaults to Node::DEFAULT_NTYPE." },
"summary": { "type": "string", "description": "Human-readable summary sentence." },
"props": { "type": "object", "description": "Arbitrary key/value props." }
},
"additionalProperties": false
}
},
"edges": {
"type": "array",
"items": {
"type": "object",
"properties": {
"src": { "type": "string", "description": "Source node UUID." },
"predicate": { "type": "string", "description": "Edge label (e.g. 'works_at')." },
"dst": { "type": "string", "description": "Destination node UUID." }
},
"required": ["src", "predicate", "dst"],
"additionalProperties": false
}
},
"agent_id": { "type": "string", "description": "Commit author. Defaults to 'mnem mcp' when absent." },
"message": { "type": "string", "default": "mnem_mcp global_add" }
},
"additionalProperties": false
}),
},
];
#[cfg(feature = "summarize")]
{
tools.push(ToolDef {
name: "mnem_community_summarize",
description: "Extractive Centroid + MMR summarizer over a caller-supplied set of node \
UUIDs. Looks up each node's `summary` field, embeds the collected \
sentences through the server's configured embedder (MNEM_EMBED_* env \
vars or `[embed]` in <repo>/config.toml), and picks `k` sentences \
balancing proximity to the community centroid against MMR diversity. \
No LLM call, no rewrite: the returned sentences are verbatim slices \
from the input summaries. Optional `query` biases selection toward \
query-relevant sentences. This is the MCP mirror of POST /v1/retrieve \
with `summarize: true`, except you choose the node set directly \
(typical callers: a Leiden-community node list, or a hand-curated \
subgraph). Degree-centrality fallback is uniform today; PPR slots in \
unchanged once E2 lands.",
input_schema: json!({
"type": "object",
"properties": {
"node_ids": {
"type": "array",
"items": { "type": "string" },
"minItems": 1,
"maxItems": 10000,
"description": "Node UUIDs (as produced by other tool outputs)."
},
"query": {
"type": "string",
"description": "Optional query text. When set, biases sentence selection toward query-relevance (beta=0.3 in the Centroid+MMR weighting)."
},
"k": {
"type": "integer",
"minimum": 0,
"maximum": 1000,
"default": 3,
"description": "Max number of sentences to return. Clamped to min(k, sentences)."
},
"mmr_lambda": {
"type": "number",
"minimum": 0.0,
"maximum": 1.0,
"default": 0.5,
"description": "MMR diversity weight. 0 = pure relevance, 1 = pure diversity."
}
},
"required": ["node_ids"],
"additionalProperties": false
}),
});
}
tools
}