fix(search): compare frontmatter tags ignoring letter case and store them by Obsidian's tag rule

.devin/wiki.json

@@ -4,7 +4,7 @@
       "content": "Vault Cortex is a remote MCP (Model Context Protocol) server that exposes an Obsidian vault over HTTPS — vault CRUD, hybrid search, task management, structured memory, link-graph queries, and non-markdown file reading (images, PDFs, canvases, data files), plus user-initiated prompts. It runs headless in Docker, working directly with .md files on disk using its own Obsidian-parity parsers and search index. README.md is the authoritative feature description and ARCHITECTURE.md the authoritative design document — derive capability details, defaults, and security specifics from those and the code rather than restating them. The repo also ships a CLI (cli/, published as the vault-cortex npm package) that scaffolds a deployment: npx vault-cortex@latest init."
     },
     {
-      "content": "The codebase is organized into four internal layers under src/vault-mcp/: (1) obsidian-markdown/ — pure Obsidian/Markdown parsers with no I/O (links, lines/fences, frontmatter, callouts, headings, tasks, memory entries, canvas, plaintext). (2) vault-operations/ — filesystem I/O for vault content (vault-filesystem, vault-patcher, note-mover, memory-store, daily-notes, vault-folder-config, task-mutations, task-format-config, trash-sweeper, asset-operations). (3) search/ — the SQLite FTS5 + sqlite-vec index, embedding pipeline, and chokidar file watcher. (4) mcp-core/ — the MCP protocol surface: streamable-http transport, tool group modules under mcp-core/tools/, and prompt group modules under mcp-core/prompts/. Authentication lives in src/vault-mcp/oauth/ (OAuth 2.1) and src/functions/authorizer.ts (AWS Lambda, path-aware). The :remote image's setup mode — the browser sign-in served before an Obsidian Sync token exists — lives in src/vault-mcp/setup/. Infrastructure-as-code (sst.config.ts) is specific to the AWS reference deployment and should not be documented as the primary deployment path."
+      "content": "The codebase is organized into four internal layers under src/vault-mcp/: (1) obsidian-markdown/ — pure Obsidian/Markdown parsers with no I/O (links, lines/fences, frontmatter, tags, callouts, headings, tasks, memory entries, canvas, plaintext, PDF). (2) vault-operations/ — filesystem I/O for vault content (vault-filesystem, vault-patcher, note-mover, memory-store, daily-notes, vault-folder-config, task-mutations, task-format-config, trash-sweeper, asset-operations). (3) search/ — the SQLite FTS5 + sqlite-vec index, embedding pipeline, and chokidar file watcher. (4) mcp-core/ — the MCP protocol surface: streamable-http transport, tool group modules under mcp-core/tools/, and prompt group modules under mcp-core/prompts/. Authentication lives in src/vault-mcp/oauth/ (OAuth 2.1) and src/functions/authorizer.ts (AWS Lambda, path-aware). The :remote image's setup mode — the browser sign-in served before an Obsidian Sync token exists — lives in src/vault-mcp/setup/. Infrastructure-as-code (sst.config.ts) is specific to the AWS reference deployment and should not be documented as the primary deployment path."
     }
   ],
   "pages": [
@@ -40,7 +40,7 @@
     },
     {
       "title": "Obsidian Markdown Parsing",
-      "purpose": "The pure parser layer in obsidian-markdown/ — no filesystem I/O: links.ts, lines.ts, frontmatter.ts, callouts.ts, headings.ts, tasks.ts, recurrence.ts, memory-entries.ts, canvas.ts, plaintext.ts, moment-format.ts. This layer encodes the Obsidian-parity behavior (every link form and 'New link format' mode Obsidian recognizes; both Tasks-plugin formats) that distinguishes Vault Cortex from simpler file-based integrations.",
+      "purpose": "The pure parser layer in obsidian-markdown/ — no filesystem I/O: links.ts, lines.ts, frontmatter.ts, tags.ts, callouts.ts, headings.ts, tasks.ts, recurrence.ts, memory-entries.ts, canvas.ts, plaintext.ts, pdf.ts, pdf-engine.ts, moment-format.ts. This layer encodes the Obsidian-parity behavior (every link form and 'New link format' mode Obsidian recognizes; both Tasks-plugin formats) that distinguishes Vault Cortex from simpler file-based integrations.",
       "parent": "Architecture"
     },
     {

AGENTS.md

@@ -85,7 +85,7 @@ src/
   __tests__/docker/ # :remote image boot tests (npm run test:remote-boot): docker-harness, and fixtures/ob, a stub of the obsidian-headless CLI
   vault-mcp/
     server.ts, config.ts # Entry point and env-var loader, the only loose files here
-    obsidian-markdown/ # Pure parsers and transforms with no I/O (pdf-engine excepted): lines, frontmatter, callouts, headings, links, tasks, recurrence, memory-entries, canvas, plaintext, pdf + pdf-engine, moment-format
+    obsidian-markdown/ # Pure parsers and transforms with no I/O (pdf-engine excepted): lines, frontmatter, callouts, headings, links, tags, tasks, recurrence, memory-entries, canvas, plaintext, pdf + pdf-engine, moment-format
     vault-operations/ # Vault read, write, and patch: vault-filesystem (base I/O), vault-patcher, note-mover, memory-store, daily-notes, vault-folder-config, task-mutations, task-format-config, trash-config, trash-sweeper, asset-operations
     mcp-core/ # MCP protocol surface: mcp-router, tool-registry, tool-availability, tool-definitions, prompt-definitions
       tools/ # One <domain>-tools module per data-layer domain (vault-crud, search, task, memory, daily-note, asset), plus tool-helpers

ARCHITECTURE.md

@@ -221,6 +221,10 @@ Both `vault_delete_note` and `vault_move_note` support `prune_empty_folders` to
 
 `vault_search_by_tag` matches a frontmatter tag and every tag nested under it (`project` matches `project/a/b` but not `projects`), or the tag alone with `exact: true`. It returns the most recently modified notes first, up to `limit` (default 20).
 
+Every tag input (`tag` on `vault_search_by_tag` and `vault_list_tasks`, the `tags` filter on `vault_search`) ignores letter case and, unless `exact: true` is set on `vault_search_by_tag`, matches nested tags the same way. Both sides of a comparison are folded with `toLowerCase()`, the rule Obsidian's Tags view groups spellings by. A leading `#` on an input is dropped, and an input that is only `#` is rejected.
+
+`vault_list_tags` merges spellings that differ only in letter case into one entry. The entry shows the spelling with the most occurrences, where a repeat within one note counts again, and its count is the number of notes carrying any of the spellings.
+
 `vault_recent_notes` sorts by `sort_by` — `"created"` or `"modified"` (default `"modified"`).
 
 ### Property discovery + daily notes
@@ -232,8 +236,10 @@ Both `vault_delete_note` and `vault_move_note` support `prune_empty_folders` to
 | `vault_list_property_values` | `key, folder?, limit?`        | readOnlyHint |
 | `vault_search_by_property`   | `key, value, folder?, limit?` | readOnlyHint |
 
-**Promoted properties:** Five frontmatter keys — `title`, `tags`, `type`, `created`, `related` — get dedicated columns in the `notes` table for direct `WHERE`-clause filtering (no `json_extract` needed). In tool responses, these appear as top-level fields; remaining frontmatter keys are returned under `additional_properties` (via `formatNoteMetadata` in `tool-helpers.ts`). All other properties live in a JSON `properties` column — functional for any schema, but without dedicated columns. The property queries match a key as data through `json_each` over that column, never as a JSON path, so a property named `a.b` or `k[0]` is matched like any other. Array values are unpacked via `json_each`, so scalar and list properties both match.
+**Promoted properties:** Five frontmatter keys — `title`, `tags` (the first key so spelled in any letter case), `type`, `created`, `related` — get dedicated columns in the `notes` table for direct `WHERE`-clause filtering (no `json_extract` needed). In tool responses, these appear as top-level fields; remaining frontmatter keys are returned under `additional_properties` (via `formatNoteMetadata` in `tool-helpers.ts`).
 
+- **Tags column:** holds the tags Obsidian's Tags view shows for the note, read by `obsidian-markdown/tags.ts` from the first key spelled `tags` in any letter case. A text value is one tag, so `tags: a, b` gives no tags, because a tag name cannot contain `,`. A list keeps its text entries. Each entry loses a leading `#` and a trailing `/`, and only names Obsidian accepts as tags are kept (no spaces, none of the punctuation Obsidian rejects such as `,`, `:` or `+`, and not digits only).
+- **Other properties:** every unpromoted key lives in a JSON `properties` column — functional for any schema, but without dedicated columns. The property queries match a key as data through `json_each` over that column, never as a JSON path, so a property named `a.b` or `k[0]` is matched like any other. Array values are unpacked via `json_each`, so scalar and list properties both match.
 - **Listing values:** `vault_list_property_values`, and the `sample_values` in `vault_list_property_keys`, count each value by the text the tool returns. The number `1` and the text `"1"` both return `"1"`, so they share one count; the text `"1.0"` keeps its own. Counting happens before `limit` and sample selection apply, and each item of a list property counts on its own.
 - **Matching a value:** `vault_search_by_property` matches text exactly. A value written as a whole, finite number in [YAML 1.2 core schema](https://yaml.org/spec/1.2.2/#1032-tag-resolution) form (`4`, `-0.5`, `1e3`, `0x1F`, `0o17`) also matches stored numbers equal to it as [JavaScript numbers](https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Number#number_encoding). So `"04"` matches the number `4` and the text `"04"`, but not the text `"4"`. Checkbox values match only `"1"` and `"0"`.
 
@@ -576,7 +582,7 @@ The search query and indexing layer is split into seven modules (the embedding p
 | `search-index.ts`   | Factory/closure (`createSearchIndex`), schema, migrations, write operations (`upsertNote`, `removeNote`, `rebuildFromVault`, `upsertFileContent`, `removeFileContent`), embedder + reranker wiring                                                                                                                         |
 | `search-queries.ts` | `SearchQueryContext` + 16 query methods — `fullTextSearch`, `memoryRecall`, `searchByTag`, `searchByFolder`, `listTasks`, `recentNotes`, `listAllTags`, `listPropertyKeys`, `listPropertyValues`, `searchByProperty`, `getBacklinks`, `getOutgoingLinks`, `findOrphans`, `brokenLinkCount`, `modifiedOnDate`, `vaultStats` |
 | `hybrid-search.ts`  | `hybridSearch` — runs the note FTS, note vector, file-content FTS, and file-content vector legs, fuses them with RRF, applies cross-encoder reranking; falls back to FTS-only when no embeddings are available                                                                                                             |
-| `search-helpers.ts` | Pure data transforms — row mappers (`rowToMetadata`, `rowToTaskEntry`, `noteRowToSearchResult`, `fileContentRowToSearchResult`), filters (`noteMatchesSearchFilters`), snippet construction                                                                                                                                |
+| `search-helpers.ts` | Pure data transforms — row mappers (`rowToMetadata`, `rowToTaskEntry`, `noteRowToSearchResult`, `fileContentRowToSearchResult`), filters (`noteMatchesSearchFilters`), tag comparison (`foldTag`, `NESTED_TAG_PREDICATE`), snippet construction                                                                            |
 | `fts-query.ts`      | FTS5 query sanitization — compound-term handling, reserved-word stripping, phrase extraction                                                                                                                                                                                                                               |
 | `rrf.ts`            | Reciprocal Rank Fusion scoring (`computeRrfScores`) — rank accumulation, k=60, top-rank bonuses                                                                                                                                                                                                                            |
 | `reranker.ts`       | Cross-encoder factory (`createReranker`, ms-marco-MiniLM-L-6-v2) + position-aware score blending (`blendScores`, `normalizeScores`)                                                                                                                                                                                        |

src/__tests__/integration/server-error-contracts.test.ts

@@ -2501,3 +2501,37 @@ describe("cannot read daily notes config", () => {
     })
   }, 30_000)
 })
+
+// ── Empty tag inputs ─────────────────────────────────────────
+
+describe("tag must not be empty", () => {
+  const EMPTY_TAG_MESSAGE = '[Error]: tag must not be empty after its leading "#"'
+
+  it("vault_search_by_tag rejects a tag that is only a #", async () => {
+    const result = await callTool({ client, name: "vault_search_by_tag", args: { tag: "#" } })
+    expect(result).toEqual({
+      content: [{ type: "text", text: EMPTY_TAG_MESSAGE }],
+      isError: true,
+    })
+  })
+
+  it("vault_search rejects a tags filter entry that is only a #", async () => {
+    const result = await callTool({
+      client,
+      name: "vault_search",
+      args: { query: "note", filters: { tags: ["#"] } },
+    })
+    expect(result).toEqual({
+      content: [{ type: "text", text: EMPTY_TAG_MESSAGE }],
+      isError: true,
+    })
+  })
+
+  it("vault_list_tasks rejects a tag that is only a #", async () => {
+    const result = await callTool({ client, name: "vault_list_tasks", args: { tag: "#" } })
+    expect(result).toEqual({
+      content: [{ type: "text", text: EMPTY_TAG_MESSAGE }],
+      isError: true,
+    })
+  })
+})