feat: mirror tag-vocabulary summary and note-type validation

Ports two changes from the knowledge-base vault this skeleton is extracted
from, keeping the shared scripts byte-identical across the pair:

- `graph_summary` reports `by_tag` and `tag_count`, sized by `topTags`, and
  drops the `by_type`/`by_status` breakdowns. An MCP-only agent could not
  otherwise discover the tag spellings that `graph_search_nodes` matches
  verbatim.
- `generate-graph.py --validate` now enforces a controlled vocabulary for a
  note's `type`, which is where the drift signal `by_type` used to give now
  lives. `AGENT_QUERY_GUIDE.md`, both skills and the-knowing-one quote the
  enforced enum.

`tests/test_generate_graph.py` is new; `make test` is 50 tests. Regenerating
the graph also drops a stale `GLOBAL_AGENTS` node for a file this skeleton
does not carry.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

AGENT_QUERY_GUIDE.md

@@ -23,7 +23,9 @@ Schema:
   - `title` (string): note title
   - `path` (string): vault-relative markdown path
   - `repo` (string): repo scope (`tenant-controller`, `controller-util`, `global`, ...)
-  - `type` (string): note classification (`guideline`, `policy`, `project`, etc.)
+  - `type` (string): note classification from a controlled vocabulary —
+    `hub|project|area|guideline|policy|feature|roadmap|decision|note|meta|reference`;
+    absent frontmatter means `note`, and `--validate` fails on anything else
   - `status` (string): lifecycle state (`active`, `draft`, `deprecated`)
   - `tags` (string array): topic tags
   - `last_reviewed` (string): `YYYY-MM-DD`
@@ -64,6 +66,10 @@ Examples with `jq`:
 # list repos
 jq -r '.nodes[].repo' index/graph.json | sort -u
 
+# list the tag vocabulary, most frequent first (tags are matched verbatim)
+jq -r '[.nodes[].tags[]?] | group_by(.) | map({tag: .[0], count: length})
+  | sort_by(-.count)[] | "\(.count)\t\(.tag)"' index/graph.json
+
 # list top linked notes
 jq -r '.nodes | sort_by(-(.inbound_count // 0))[] | "\(.inbound_count // 0)\t\(.id)\t\(.path)"' index/graph.json
 
@@ -96,8 +102,8 @@ For `jg`, use the same filters if your local `jg` is jq-compatible.
 Validation:
 
 ```bash
-# validate: non-zero exit on unresolved links, edge types outside the
+# validate: non-zero exit on unresolved links, edge types or note types
-# vocabulary, OR any remaining legacy [[wiki links]]
+# outside their vocabularies, OR any remaining legacy [[wiki links]]
 python3 scripts/generate-graph.py --validate
 
 # regenerate graph and validate

agents/the-knowing-one.md

@@ -26,7 +26,8 @@ through MCP.
 
 ## MCP tools
 
-- `graph_summary` — overview of the graph: counts, types, tags, central nodes.
+- `graph_summary` — overview of the graph: counts, repo scopes, the tag
+  vocabulary (`by_tag`, most frequent first) and central nodes.
 - `graph_search_nodes` — find nodes by tag / type / repo / id / text.
 - `graph_neighbors` — outbound + inbound links for a node id, each with its edge
   `type` + `annotation` (traverse the graph).
@@ -56,15 +57,18 @@ README.md                   vault hub
 ```
 
 Every note carries taxonomy frontmatter: `id`, `title`, `type`
-(`hub|project|guideline|policy|feature|roadmap|decision|note|meta`), `status`
+(`hub|project|area|guideline|policy|feature|roadmap|decision|note|meta|reference`, enforced by
+`generate-graph.py --validate`), `status`
 (`active|draft|deprecated`), `tags`, `last_reviewed`, optional `repo`. Notes link
 with typed markdown links — `[text](note-id){"type":"…"}` — resolved by `id`.
 Legacy `[[wiki links]]` are rejected by the graph validator and must not be written.
 
 ## 1. Orient
 
-Start broad with `graph_summary` to see node counts, types, tags, and the
+Start broad with `graph_summary` to see node counts, repo scopes, the tag
-most-central notes. Then `read_note` the entry points:
+vocabulary and the most-central notes. Its `by_tag` list is the only place the
+tag spellings are visible — read it before filtering by `tag`, which matches
+verbatim. Then `read_note` the entry points:
 
 - vault hub (`README.md`) — entry points.
 - `03_resources/repos/repo-landscape.md` — index of documented repos, grouped.
@@ -80,8 +84,9 @@ Use `graph_search_nodes` to locate notes:
 
 - by **repo** — every note tied to a repo (e.g. `tenant-controller`).
 - by **type** — `policy`, `decision`, `feature`, `guideline`, `project`, ...
-- by **tag** — e.g. `kcp`.
+- by **tag** — e.g. `kcp`. Exact match; take the spelling from `by_tag`.
-- by **text** — when tag/type/id unknown, full-text search the notes.
+- by **text** — `query` matches a node's id, title, path and tags, not the note
+  body. Widen with a tag or repo filter rather than a longer phrase.
 
 Use `graph_summary` to spot central (most-linked) notes worth reading early.
 

agents/the-scribe.md

@@ -25,7 +25,7 @@ already captured into `raw/` (a PDF, an image) at the absolute path
 ## MCP tools
 
 Read:
-- `graph_summary` — counts, types, tags, most-linked notes.
+- `graph_summary` — counts, repo scopes, the tag vocabulary (`by_tag`), most-linked notes.
 - `graph_search_nodes` — find notes by tag / type / repo / id / free text.
 - `graph_neighbors` — inbound + outbound links of a node, with edge `type` and `annotation`.
 - `read_note` — the Markdown of a note (by `id` or vault-relative `path`). On a

index/graph.json

@@ -1,6 +1,6 @@
 {
   "version": 1,
-  "generated_at": "2026-09-11T07:34:35.807809+00:00",
+  "generated_at": "2026-09-14T09:33:34.687166+00:00",
   "nodes": [
     {
       "id": ".claude/agents/the-knowing-one",
@@ -50,18 +50,6 @@
       "outbound_links": [],
       "inbound_count": 0
     },
-    {
-      "id": "GLOBAL_AGENTS",
-      "title": "GLOBAL_AGENTS.md",
-      "path": "GLOBAL_AGENTS.md",
-      "repo": "global",
-      "type": "note",
-      "status": "active",
-      "tags": [],
-      "last_reviewed": "",
-      "outbound_links": [],
-      "inbound_count": 0
-    },
     {
       "id": "agents/README",
       "title": "Agents",

scripts/generate-graph.py

@@ -1,4 +1,13 @@
 #!/usr/bin/env python3
+"""Build `index/graph.json` from the vault's Markdown, and validate it.
+
+The graph is the derived navigation layer agents query instead of walking the
+tree: one node per note (frontmatter plus link counts) and one edge per typed
+markdown link.  `--validate` is the gate that keeps it trustworthy — it fails on
+links that resolve to nothing, on edge relationship types or note `type` values
+outside their controlled vocabularies, and on legacy `[[wiki links]]`.
+"""
+
 import argparse
 import json
 import re
@@ -36,6 +45,24 @@ RELATION_TYPES = {
 }
 
 
+# Strict controlled vocabulary for a note's frontmatter `type`. A note without
+# one is a "note"; --validate fails on any value outside this set, which is what
+# keeps the facet meaningful as the vault grows.
+NOTE_TYPES = {
+    "hub",
+    "project",
+    "area",
+    "guideline",
+    "policy",
+    "feature",
+    "roadmap",
+    "decision",
+    "note",
+    "meta",
+    "reference",
+}
+
+
 @dataclass
 class LinkRef:
     target: str  # raw target text (resolution key: id / rel-path / title / stem, or a URL)
@@ -402,7 +429,8 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--validate",
         action="store_true",
-        help="validate links: unresolved targets, edge relationship types, "
+        help="validate the graph: unresolved targets, edge relationship types, "
+        "note types, "
         "and legacy [[wiki links]] (all rejected)",
     )
     parser.add_argument(
@@ -453,6 +481,25 @@ def print_invalid_types(invalid: List[Dict[str, object]]) -> None:
         print(f"- {e.get('from')} -> {e.get('to')}: type={e.get('type')!r}")
 
 
+def find_invalid_node_types(graph: Dict[str, object]) -> List[Dict[str, object]]:
+    return [
+        n
+        for n in graph.get("nodes", [])  # type: ignore[union-attr]
+        if n.get("type") not in NOTE_TYPES
+    ]
+
+
+def print_invalid_node_types(invalid: List[Dict[str, object]]) -> None:
+    if not invalid:
+        print("Validation passed: all node types are in the controlled vocabulary")
+        return
+
+    print(f"Validation failed: {len(invalid)} note(s) with unknown type")
+    print(f"  allowed: {', '.join(sorted(NOTE_TYPES))}")
+    for n in invalid:
+        print(f"- {n.get('id')} ({n.get('path')}): type={n.get('type')!r}")
+
+
 def main() -> None:
     args = parse_args()
     root = Path(__file__).resolve().parents[1]
@@ -490,8 +537,10 @@ def main() -> None:
         print_unresolved(unresolved)
         invalid = find_invalid_types(graph)
         print_invalid_types(invalid)
+        invalid_nodes = find_invalid_node_types(graph)
+        print_invalid_node_types(invalid_nodes)
         print_wiki_links(wiki_uses)
-        if unresolved or invalid or wiki_uses:
+        if unresolved or invalid or invalid_nodes or wiki_uses:
             sys.exit(1)
 
 

scripts/mcp-server.py

@@ -25,6 +25,7 @@ import subprocess
 import sys
 import urllib.error
 import urllib.request
+from collections import Counter
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Dict, List
@@ -196,6 +197,11 @@ _WILDCARD_PATHS = {".", "./", "-A", "--all", "*", "-a", ":/"}
 _DEFAULT_MAX_CAPTURE_BYTES = 100 * 1024 * 1024
 _DOWNLOAD_TIMEOUT_SECONDS = 60
 
+# graph_summary defaults: enough most-linked notes to spot the hubs, and enough
+# tags to cover the vocabulary an agent actually filters on.
+_DEFAULT_TOP_LINKED = 10
+_DEFAULT_TOP_TAGS = 50
+
 
 def _resolve_in_root(state: AppState, raw_path: Any, *, label: str = "path") -> Path:
     """Turn a vault-relative path into an absolute path inside the repo root.
@@ -245,12 +251,31 @@ def _tool_definitions() -> List[Dict[str, Any]]:
     return [
         {
             "name": "graph_summary",
-            "description": "Return aggregate counts from index/graph.json.",
+            "description": (
+                "Orientation for a cold start: node and edge counts, the repo "
+                "scopes, the most-linked notes, and the tag vocabulary as "
+                "`by_tag` (most frequent first) plus `tag_count`, the number of "
+                "distinct tags. Read `by_tag` before filtering "
+                "`graph_search_nodes` by `tag` \u2014 that filter matches a tag "
+                "verbatim, so a guessed spelling returns nothing."
+            ),
             "inputSchema": {
                 "type": "object",
                 "properties": {
-                    "topLinked": {"type": "integer", "minimum": 1, "maximum": 50}
+                    "topLinked": {
+                        "type": "integer",
+                        "minimum": 1,
+                        "maximum": 50,
+                        "description": "Most-linked notes to return (default 10).",
                     },
+                    "topTags": {
+                        "type": "integer",
+                        "minimum": 1,
+                        "maximum": 500,
+                        "description": "Tags to return, most frequent first (default 50).",
+                    },
+                },
+                "additionalProperties": False,
             },
         },
         {
@@ -453,21 +478,16 @@ def _tool_graph_summary(state: AppState, args: Dict[str, Any]) -> Dict[str, Any]
     graph = _load_graph(state)
     nodes = graph.get("nodes", [])
     edges = graph.get("edges", [])
-    top_n = int(args.get("topLinked", 10))
+    top_n = int(args.get("topLinked", _DEFAULT_TOP_LINKED))
+    top_tags = int(args.get("topTags", _DEFAULT_TOP_TAGS))
 
-    by_repo: Dict[str, int] = {}
+    by_repo: Counter[str] = Counter(str(node.get("repo", "unknown")) for node in nodes)
-    by_type: Dict[str, int] = {}
+    tag_counts: Counter[str] = Counter(
-    by_status: Dict[str, int] = {}
+        str(tag) for node in nodes for tag in (node.get("tags") or [])
-    for node in nodes:
-        by_repo[str(node.get("repo", "unknown"))] = (
-            by_repo.get(str(node.get("repo", "unknown")), 0) + 1
-        )
-        by_type[str(node.get("type", "unknown"))] = (
-            by_type.get(str(node.get("type", "unknown")), 0) + 1
-        )
-        by_status[str(node.get("status", "unknown"))] = (
-            by_status.get(str(node.get("status", "unknown")), 0) + 1
     )
+    # Frequent tags first, alphabetical within a tie, so a truncated list is
+    # both the most useful slice and stable across calls.
+    ranked_tags = sorted(tag_counts.items(), key=lambda item: (-item[1], item[0]))
 
     sorted_nodes = sorted(
         nodes,
@@ -479,9 +499,11 @@ def _tool_graph_summary(state: AppState, args: Dict[str, Any]) -> Dict[str, Any]
         "generated_at": graph.get("generated_at"),
         "node_count": len(nodes),
         "edge_count": len(edges),
-        "by_repo": by_repo,
+        "by_repo": dict(by_repo),
-        "by_type": by_type,
+        "tag_count": len(tag_counts),
-        "by_status": by_status,
+        "by_tag": [
+            {"tag": tag, "count": count} for tag, count in ranked_tags[:top_tags]
+        ],
         "top_linked": [
             {
                 "id": node.get("id"),

skills/knowledge-base-ingest-resource/SKILL.md

@@ -229,7 +229,8 @@ last_reviewed: <today, YYYY-MM-DD>
 
 - `ingested` and `last_reviewed` = today (the `currentDate` from session context).
 - `type` must stay within the vault taxonomy enum
-  (`hub|project|guideline|policy|feature|roadmap|decision|note|meta`) — use `note`;
+  (`hub|project|area|guideline|policy|feature|roadmap|decision|note|meta|reference`),
+  which `generate-graph.py --validate` enforces — use `note`;
   the `resource` tag is what marks it.
 - Markdown link targets resolve by `id`, rel-path, title, or unique stem — always
   use the `id` as the target, e.g. `[Human title](resource/foo/bar)`.

skills/knowledge-base-read/SKILL.md

@@ -29,7 +29,8 @@ README.md                   vault hub; AGENT_QUERY_GUIDE.md   graph query refere
 ```
 
 Every note carries taxonomy frontmatter: `id`, `title`, `type`
-(`hub|project|guideline|policy|feature|roadmap|decision|note|meta`), `status`
+(`hub|project|area|guideline|policy|feature|roadmap|decision|note|meta|reference`, enforced by
+`generate-graph.py --validate`), `status`
 (`active|draft|deprecated`), `tags`, `last_reviewed`, optional `repo`. Notes are
 linked with typed markdown links — `[text](note-id){"type":"…"}` — that resolve
 by `id`. Legacy `[[wiki links]]` are rejected by `make validate-graph`.

tests/test_generate_graph.py

@@ -0,0 +1,106 @@
+#!/usr/bin/env python3
+"""Tests for the graph generator's validation gate.
+
+`--validate` is the only thing standing between a typo in a note's frontmatter
+and a vault whose facets quietly stop meaning anything, so these tests pin the
+controlled vocabularies it enforces: the edge relationship types and the node
+`type` enum. Each filesystem test builds a throwaway vault, never the real one.
+"""
+
+from __future__ import annotations
+
+import importlib
+import subprocess
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+SCRIPTS = Path(__file__).resolve().parents[1] / "scripts"
+sys.path.insert(0, str(SCRIPTS))
+
+gen = importlib.import_module("generate-graph")
+
+
+def _node(node_id: str, note_type: str) -> dict:
+    return {"id": node_id, "path": f"{node_id}.md", "type": note_type}
+
+
+class TestNoteTypeVocabulary(unittest.TestCase):
+    """The node `type` enum is the drift signal; it has to be enforced."""
+
+    def test_covers_every_type_the_taxonomy_documents(self) -> None:
+        self.assertLessEqual(
+            {
+                "hub",
+                "project",
+                "area",
+                "guideline",
+                "policy",
+                "feature",
+                "roadmap",
+                "decision",
+                "note",
+                "meta",
+                "reference",
+            },
+            gen.NOTE_TYPES,
+        )
+
+    def test_accepts_a_graph_whose_nodes_are_all_in_the_vocabulary(self) -> None:
+        graph = {"nodes": [_node("a", "note"), _node("b", "area")], "edges": []}
+        self.assertEqual(gen.find_invalid_node_types(graph), [])
+
+    def test_reports_a_node_whose_type_is_outside_the_vocabulary(self) -> None:
+        graph = {"nodes": [_node("a", "note"), _node("b", "refrence")], "edges": []}
+        invalid = gen.find_invalid_node_types(graph)
+        self.assertEqual([n["id"] for n in invalid], ["b"])
+
+
+class TestValidateExitCode(unittest.TestCase):
+    """--validate must fail the run, not just print, on an unknown node type."""
+
+    def setUp(self) -> None:
+        self._tmp = tempfile.TemporaryDirectory()
+        self.root = Path(self._tmp.name).resolve() / "vault"
+        (self.root / "scripts").mkdir(parents=True)
+        (self.root / "03_resources").mkdir()
+        (self.root / "scripts" / "generate-graph.py").write_text(
+            (SCRIPTS / "generate-graph.py").read_text(encoding="utf-8"),
+            encoding="utf-8",
+        )
+
+    def tearDown(self) -> None:
+        self._tmp.cleanup()
+
+    def write_note(self, name: str, note_type: str) -> None:
+        (self.root / "03_resources" / name).write_text(
+            f"---\nid: {name[:-3]}\ntitle: {name}\ntype: {note_type}\n---\n\n# {name}\n",
+            encoding="utf-8",
+        )
+
+    def validate(self) -> subprocess.CompletedProcess:
+        return subprocess.run(
+            [sys.executable, "scripts/generate-graph.py", "--validate"],
+            cwd=self.root,
+            capture_output=True,
+            text=True,
+        )
+
+    def test_passes_on_a_vault_of_known_types(self) -> None:
+        self.write_note("good.md", "decision")
+        proc = self.validate()
+        self.assertEqual(proc.returncode, 0, proc.stdout + proc.stderr)
+        self.assertIn("all node types are in the controlled vocabulary", proc.stdout)
+
+    def test_fails_and_names_the_offending_note(self) -> None:
+        self.write_note("good.md", "decision")
+        self.write_note("bad.md", "descision")
+        proc = self.validate()
+        self.assertEqual(proc.returncode, 1)
+        self.assertIn("descision", proc.stdout)
+        self.assertIn("bad", proc.stdout)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_mcp_server.py

@@ -1,11 +1,11 @@
 #!/usr/bin/env python3
-"""Tests for the knowledge-base MCP server's write and git tools.
+"""Tests for the knowledge-base MCP server's query, write and git tools.
 
-Covers the tools an ingesting agent needs when the MCP server is its only way
+Covers the tools an agent needs when the MCP server is its only way into the
-into the vault: creating notes, editing notes, capturing raw bytes (local files
+vault: orienting with graph_summary, creating notes, editing notes, capturing
-and downloads), committing staged paths, and pushing. Each test drives the
+raw bytes (local files and downloads), committing staged paths, and pushing.
-server through `_call_tool` against a throwaway repository so no test touches
+Each test drives the server through `_call_tool` against a throwaway repository
-the real vault.
+so no test touches the real vault.
 """
 
 from __future__ import annotations
@@ -430,3 +430,89 @@ class TestReadNoteBinary(ToolTestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class TestGraphSummary(ToolTestCase):
+    """graph_summary is the agent's vocabulary lookup, not a stats dump.
+
+    Tags are the vault's only high-cardinality facet and the `tag` filter of
+    graph_search_nodes matches them verbatim, so the summary has to name them.
+    """
+
+    def write_graph(self, nodes: list) -> None:
+        self.state.graph_path.parent.mkdir(parents=True, exist_ok=True)
+        self.state.graph_path.write_text(
+            json.dumps(
+                {
+                    "version": 1,
+                    "generated_at": "2026-09-14T00:00:00Z",
+                    "nodes": nodes,
+                    "edges": [],
+                }
+            ),
+            encoding="utf-8",
+        )
+
+    def node(self, node_id: str, **fields) -> dict:
+        base = {
+            "id": node_id,
+            "title": node_id,
+            "path": f"03_resources/{node_id}.md",
+            "repo": "global",
+            "type": "note",
+            "status": "active",
+            "tags": [],
+            "inbound_count": 0,
+        }
+        base.update(fields)
+        return base
+
+    def test_counts_tags_and_orders_them_by_frequency(self) -> None:
+        self.write_graph(
+            [
+                self.node("a", tags=["kcp", "flux"]),
+                self.node("b", tags=["kcp"]),
+                self.node("c", tags=["kcp", "helm"]),
+            ]
+        )
+        payload = self.structured("graph_summary")
+        self.assertEqual(
+            payload["by_tag"],
+            [
+                {"tag": "kcp", "count": 3},
+                {"tag": "flux", "count": 1},
+                {"tag": "helm", "count": 1},
+            ],
+        )
+
+    def test_reports_the_distinct_tag_total_even_when_truncated(self) -> None:
+        self.write_graph(
+            [self.node(f"n{i}", tags=[f"tag-{i}"]) for i in range(10)]
+        )
+        payload = self.structured("graph_summary", topTags=3)
+        self.assertEqual(len(payload["by_tag"]), 3)
+        self.assertEqual(payload["tag_count"], 10)
+
+    def test_drops_the_type_and_status_facets(self) -> None:
+        self.write_graph([self.node("a"), self.node("b", status="deprecated")])
+        payload = self.structured("graph_summary")
+        self.assertNotIn("by_type", payload)
+        self.assertNotIn("by_status", payload)
+
+    def test_keeps_repo_counts_and_most_linked_notes(self) -> None:
+        self.write_graph(
+            [
+                self.node("hub", repo="global", inbound_count=9),
+                self.node("leaf", repo="kunnus", inbound_count=1),
+            ]
+        )
+        payload = self.structured("graph_summary", topLinked=1)
+        self.assertEqual(payload["by_repo"], {"global": 1, "kunnus": 1})
+        self.assertEqual([n["id"] for n in payload["top_linked"]], ["hub"])
+        self.assertEqual(payload["node_count"], 2)
+
+    def test_advertises_the_top_tags_argument(self) -> None:
+        tool = next(
+            t for t in mcp._tool_definitions() if t["name"] == "graph_summary"
+        )
+        self.assertIn("topTags", tool["inputSchema"]["properties"])