@gmickel/gno 2.8.0 → 2.8.1

package/README.md

@@ -139,7 +139,7 @@ See the [guide](docs/COMPILED-CONTEXT.md).
 
 <!-- public-truth:current-version -->
 
-> Current source version: **v2.8.0**. See [CHANGELOG.md](./CHANGELOG.md).
+> Current source version: **v2.8.1**. See [CHANGELOG.md](./CHANGELOG.md).
 
 <!-- /public-truth -->
 

package/assets/skill/SKILL.md

@@ -355,8 +355,11 @@ rewrite, persist findings, judge factual truth, or replace retrieval.
 `--max-findings` accepts `all` (MCP `maxFindings: "all"`) to export every
 finding. Link findings carry `referenceKind`, `resolutionStatus`, and
 `resolvedScope` in their evidence detail; ambiguous vault links list the tied
-`candidates`. Report `truncation.snapshotTruncated` as "totals cover the
-bounded snapshot, not the whole index".
+`candidates`. `links.outside-index` `info` findings are link targets that
+exist in the vault but are not indexed (attachments, excluded or unindexed
+folders); Obsidian resolves them, so do not report them as broken links.
+Report `truncation.snapshotTruncated` as "totals cover the bounded snapshot,
+not the whole index".
 
 Treat exit `4` as a complete report with findings. Exit `5` or report status
 `partial`/`changed_during_audit` means evidence is unavailable, inconclusive,

package/assets/skill/recipes/memory-scoped-recall.md

@@ -32,11 +32,16 @@ the flag.
      Superseded records never appear.
    - `budget.omitted` > 0: facts matched but did not fit. Narrow the query
      or raise the budget.
-   - Empty `facts` plus a `hint`: nothing is stored in scope yet. Say so;
-     the hint names `gno remember` for when the user wants to store one.
-   - `retrieval.mode: "lexical"`: the collection has no cached embeddings, so
-     the query matched every term. Rephrase to the fact's own words, or embed
-     the collection (`gno embed <collection>`) for question-shaped queries.
+   - Empty `facts` plus a `hint`: the hint says why. "No memories in scope
+     yet" means nothing is stored there; "No memories in scope matched this
+     query" means facts exist but none shares a word with the question, so
+     retry with other words before concluding memory is silent. The hints
+     name `gno remember` for when the user wants to store one.
+   - `retrieval.mode: "lexical"`: the collection has no cached embeddings.
+     Questions still work (question words are dropped and facts sharing a
+     content word are returned), but a paraphrase with no word in common
+     misses; embed the collection (`gno embed <collection>`) to match by
+     meaning.
 
 3. Answer from the facts and cite each by its `gno://` URI. When memory is
    silent or stale, fall through to `gno search` / `gno query` on the

package/browser-extension/artifacts/gno-browser-clipper-v2.8.1.zip.sha256

@@ -0,0 +1 @@
+3f23832d1ed8b19e24eef29d6ea888a8c84009e39c61981d0b665c95cd7d792e  gno-browser-clipper-v2.8.1.zip

package/browser-extension/dist/manifest.json

@@ -21,5 +21,5 @@
   "content_security_policy": {
     "extension_pages": "script-src 'self'; object-src 'none'; connect-src http://127.0.0.1:*"
   },
-  "version": "2.8.0"
+  "version": "2.8.1"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gmickel/gno",
-  "version": "2.8.0",
+  "version": "2.8.1",
   "description": "Local semantic search for your documents. Index Markdown, PDF, and Office files with hybrid BM25 + vector search.",
   "keywords": [
     "embeddings",

package/spec/cli.md

@@ -240,7 +240,12 @@ the model with `id`, `model`, `dimensions`, `state` (`active`|`shadow`),
 `provenance` (building runtime, e.g. `CUDA, Bun 1.4.2`),
 `compatibleRuntimes` (runtimes that read it) and `incompatibleRuntimes`.
 Terminal output prints a `Vector partitions:` block
-unless there is exactly one healthy partition. Per-collection chunk totals remain deduplicated by canonical chunk;
+unless there is exactly one healthy partition. `collections`, `totalDocuments`
+and `totalChunks` cover configured collections only: a collection removed from
+config (for example by `gno collection remove`) is not reported even while its
+rows remain in the index until the next `gno update`. `totalChunks` counts the
+distinct chunks of active documents; chunks left by deleted documents do not
+count. Per-collection chunk totals remain deduplicated by canonical chunk;
 embedded counts require matching current inputs for every active owner within
 that collection. Status reads persisted identity and coverage without loading
 models. Legacy storage remains the fallback before variant authority; ambiguous
@@ -949,9 +954,9 @@ healthy.
 Link findings resolve with the workspace-aware resolver (see
 docs/ARCHITECTURE.md "Resolution"). Their evidence `detail` JSON carries
 separate `referenceKind` (`wiki-name`, `wiki-path`, `explicit-collection`,
-`markdown`), `resolutionStatus` (`unresolved`, `ambiguous`) and
-`resolvedScope` (`same-collection`, `cross-collection`, `explicit-collection`,
-or null when unresolved). Ambiguous workspace links add `candidateCount` and
+`markdown`), `resolutionStatus` (`unresolved`, `ambiguous`, `outside-index`)
+and `resolvedScope` (`same-collection`, `cross-collection`,
+`explicit-collection`, or null when unresolved). Ambiguous workspace links add `candidateCount` and
 `candidates` (tied candidate URIs in canonical path order); in a
 collection-scoped audit, candidates outside the requested collections are
 counted in `candidatesWithheld` and never named, and a list longer than the
@@ -960,6 +965,27 @@ Orphans stay "no incoming or outgoing resolved links", with connectivity drawn
 from the whole index even when the audited documents are scoped; a tied link
 connects nothing.
 
+A plain wiki link or embed from a document in a link workspace whose target
+is not an indexed document but exists as a file inside the workspace root (a
+non-Markdown attachment, a note in an unindexed folder, a note excluded by a
+collection pattern) is not reported by `links.local-targets`. It is an `info`
+finding of `links.outside-index` (evidence kind `outside-index-target`,
+`resolutionStatus: "outside-index"`). That rule stays `pass`, so these
+findings never change the exit code. The check is existence-only: each
+involved workspace root is listed once per run (hidden files and folders
+skipped, symlinked folders not descended, a symlink counted only when it
+resolves to a regular file inside the workspace, at most 200,000 files), no
+file is opened, and no graph edge is created. Targets match files by the workspace
+link rules; a non-Markdown target needs its extension, `.md` is optional. A
+target missing from the listing stays unresolved; an unreadable folder or the
+file bound makes the listing incomplete, which the `links.outside-index` rule
+message states once.
+
+The wiki link parser reads Obsidian's table-escaped alias `[[Note\|Alias]]`
+as target `Note` with alias `Alias`. Markdown link text may contain balanced
+square brackets (`[see [1]](note.md)`); a destination containing a square
+bracket is not a link.
+
 The JSON contract is versioned as `gno://schemas/audit-report@1.0`. Finding IDs
 are stable SHA-256 identities derived from rule, normalized subject/location,
 and load-bearing evidence. Wall-clock timing and traversal order do not affect
@@ -2014,6 +2040,12 @@ gno recall <query> --scope <scope> [--scope <scope>...] [--collection <name>] [-
   `retrieval.semanticUnavailable` explaining why). Query expansion, graph
   expansion, and reranking are disabled. Scope and supersession filtering run
   inside the retrieval query; superseded facts are never returned.
+- The lexical leg searches the query's content words (question and function
+  words dropped; quoted phrases and `-term` exclusions kept). Facts containing
+  every content word are returned first; when none does, facts sharing any
+  content word are returned in BM25 order, dropping those scoring below 0.1
+  of the best match.
+- The budget is filled in retrieval-rank order.
 - Budget: at most `--max-facts` facts (default 8) under `--max-tokens`
   (default 512). Both must be positive integers. Recall never downloads a
   model.
@@ -2023,8 +2055,11 @@ gno recall <query> --scope <scope> [--scope <scope>...] [--collection <name>] [-
   (`caller`, `session`, `issuedAt`, `memoryIds`, `spanHashes`, `digest`) plus
   `budget` and `retrieval`. Derived output inherits the strictest source
   egress policy (`egressLineage`).
-- Empty recall prints the self-teaching line naming `gno remember`
-  (`hint` in JSON) and exits 0.
+- Empty recall prints a hint (`hint` in JSON) and exits 0. The hint
+  distinguishes an empty scope (`No memories in scope yet. Store one with:
+gno remember ...`), a populated scope with no match (`No memories in scope
+matched this query. ...`, also naming `gno remember`), and matches that did
+  not fit the token budget.
 
 **Output:** `--json` prints the shared `RecallResult`. Terminal output lists
 numbered facts with URI, text, scopes, hash, and identity, then `Budget:`,

package/spec/mcp.md

@@ -361,7 +361,10 @@ modifies source files, config, index rows, graph edges, daemon state, or a
 persisted audit baseline.
 
 Rules report `pass`, `fail`, `skip`, `unavailable`, or `inconclusive`; reports
-are `complete`, `partial`, `changed_during_audit`, or `failed`. Unavailable or
+are `complete`, `partial`, `changed_during_audit`, or `failed`. A passing rule
+may carry `info` findings (`links.outside-index`: link targets that exist as
+files in the link workspace but are not indexed); only `warning` and `error`
+findings fail a rule. Link semantics match `gno audit` (spec/cli.md). Unavailable or
 changing evidence never appears healthy. Stable finding IDs are derived from
 rule, normalized subject/location, and evidence fingerprint. Responses are
 canonically ordered and bounded while retaining exact pre-truncation totals.
@@ -1459,8 +1462,9 @@ Budgeted, cited, current-state recall from a memory-managed collection
   the vector leg did not run
 - `egressLineage` — strictest source policy across returned facts (absent when
   empty)
-- `hint` — self-teaching line naming `gno remember`, present only when no fact
-  was returned
+- `hint` — present only when no fact was returned; says why: the scope holds
+  no current fact (names `gno remember`), nothing in scope matched the query,
+  or the matches did not fit `maxTokens`
 
 **Identity:** `caller` is the MCP client implementation name from the
 `initialize` handshake (`mcp` when absent); `session` is the Streamable HTTP
@@ -2185,7 +2189,7 @@ Find semantically similar documents using vector embeddings.
 **Algorithm:**
 
 1. Get all chunks for the source document
-2. Retrieve embeddings for each chunk from content_vectors
+2. Retrieve each chunk's stored embedding from the active vector partition (legacy `content_vectors` only before any partition activates); no model is loaded
 3. Compute average embedding across all chunks
 4. Search for nearest neighbors using sqlite-vec
 5. Exclude self and filter by collection if not crossCollection

package/spec/output-schemas/memory-recall.schema.json

@@ -38,7 +38,7 @@
     },
     "hint": {
       "type": "string",
-      "description": "Self-teaching line naming `gno remember`; present only when no fact was returned."
+      "description": "Why nothing was returned (empty scope, no match in scope, or matches over the token budget); present only when no fact was returned."
     }
   },
   "$defs": {

package/spec/output-schemas/status.schema.json

@@ -98,7 +98,7 @@
     },
     "collections": {
       "type": "array",
-      "description": "Collection statistics",
+      "description": "Statistics for configured collections; a collection removed from config is omitted even before its index rows are pruned",
       "items": {
         "type": "object",
         "required": ["name", "documentCount", "chunkCount", "embeddedCount"],
@@ -146,12 +146,12 @@
     },
     "totalDocuments": {
       "type": "integer",
-      "description": "Total documents across all collections",
+      "description": "Active documents across configured collections",
       "minimum": 0
     },
     "totalChunks": {
       "type": "integer",
-      "description": "Total chunks across all collections",
+      "description": "Distinct chunks of active documents across configured collections",
       "minimum": 0
     },
     "embeddingBacklog": {

package/src/cli/commands/graph.ts

@@ -14,6 +14,7 @@ import type {
 
 import { normalizeContentTypes } from "../../config";
 import { diagnoseGraphQuery } from "../../core/graph-query";
+import { getActivePreset } from "../../llm/registry";
 import { initStore } from "./shared";
 
 // ─────────────────────────────────────────────────────────────────────────────
@@ -125,7 +126,7 @@ export async function graph(
   if (!initResult.ok) {
     return { success: false, error: initResult.error };
   }
-  const { store } = initResult;
+  const { store, config } = initResult;
 
   try {
     const storeOptions: GetGraphOptions = {
@@ -136,6 +137,7 @@ export async function graph(
       threshold: options.threshold,
       linkedOnly: !options.includeIsolated,
       similarTopK: options.similarTopK,
+      embedModel: getActivePreset(config).embed,
     };
 
     const result = await store.getGraph(storeOptions);

package/src/cli/commands/links.ts

@@ -5,12 +5,7 @@
  * @module src/cli/commands/links
  */
 
-import type {
-  DocEdgeRow,
-  DocLinkRow,
-  DocumentRow,
-  StorePort,
-} from "../../store/types";
+import type { DocEdgeRow, DocLinkRow, StorePort } from "../../store/types";
 
 import { resolveDocRef } from "../../core/ref-parser";
 import { initStore } from "./shared";
@@ -592,27 +587,23 @@ export async function similar(
     }
     const db = store.getRawDb();
 
-    // Get document embedding from content_vectors (prefer seq=0)
-    interface VectorRow {
-      embedding: Uint8Array;
-    }
-
-    const embedModel = modelPreset.embed;
-    const vectorRow = db
-      .query<VectorRow, [string, string]>(
-        "SELECT embedding FROM content_vectors WHERE mirror_hash = ? AND model = ? AND seq = 0 LIMIT 1"
-      )
-      .get(doc.mirrorHash, embedModel);
-
-    const fallbackRow =
-      vectorRow ??
-      db
-        .query<VectorRow, [string, string]>(
-          "SELECT embedding FROM content_vectors WHERE mirror_hash = ? AND model = ? ORDER BY seq LIMIT 1"
-        )
-        .get(doc.mirrorHash, embedModel);
-
-    if (!fallbackRow) {
+    // Stored vector of the document's first chunk, from the active partition
+    const {
+      readStoredDocumentVectors,
+      resolveStoredVectorSource,
+      similarityHitDocuments,
+      storedVectorSearchOptions,
+    } = await import("../../store/vector/stored-vectors.js");
+    const source = resolveStoredVectorSource(db, modelPreset.embed);
+    const [embedding] =
+      readStoredDocumentVectors(
+        db,
+        source,
+        [{ id: doc.id, mirrorHash: doc.mirrorHash }],
+        { firstChunkOnly: true }
+      ).get(doc.id) ?? [];
+
+    if (!embedding) {
       return {
         success: false,
         error: "Document has no embeddings. Run: gno embed",
@@ -621,9 +612,6 @@ export async function similar(
     }
 
     // Normalize embedding for cosine similarity
-    const { decodeEmbedding } =
-      await import("../../store/vector/sqlite-vec.js");
-    const embedding = decodeEmbedding(fallbackRow.embedding);
     const dimensions = embedding.length;
     let norm = 0;
     for (let i = 0; i < dimensions; i++) {
@@ -641,7 +629,7 @@ export async function similar(
     const { createVectorIndexPort } =
       await import("../../store/vector/sqlite-vec.js");
     const vecResult = await createVectorIndexPort(db, {
-      model: embedModel,
+      model: modelPreset.embed,
       dimensions,
     });
     if (!vecResult.ok) {
@@ -663,13 +651,13 @@ export async function similar(
     const searchResult = await vectorIndex.searchNearest(
       embedding,
       candidateLimit,
-      {}
+      storedVectorSearchOptions(source)
     );
     if (!searchResult.ok) {
       return { success: false, error: searchResult.error.message };
     }
 
-    // Build mirrorHash -> doc map from a single listDocuments call
+    // Candidate documents from a single listDocuments call
     const docsResult = crossCollection
       ? await store.listDocuments()
       : await store.listDocuments(doc.collection);
@@ -678,28 +666,18 @@ export async function similar(
       return { success: false, error: docsResult.error.message };
     }
 
-    const docsByHash = new Map<string, DocumentRow>();
-    for (const d of docsResult.value) {
-      if (d.active && d.mirrorHash) {
-        // Only keep first doc per hash (they have same content)
-        if (!docsByHash.has(d.mirrorHash)) {
-          docsByHash.set(d.mirrorHash, d);
-        }
-      }
-    }
-
-    // Map results to documents, excluding self
+    // Map hits to their owning documents, excluding self
     const similarItems: SimilarItem[] = [];
     const seenDocids = new Set<string>();
 
-    for (const vec of searchResult.value) {
+    for (const { document: d, distance } of similarityHitDocuments(
+      searchResult.value,
+      docsResult.value.filter((d) => d.active && d.mirrorHash)
+    )) {
       if (similarItems.length >= limit) {
         break;
       }
 
-      const d = docsByHash.get(vec.mirrorHash);
-      if (!d) continue;
-
       // Exclude self
       if (d.docid === doc.docid) continue;
 
@@ -708,7 +686,7 @@ export async function similar(
 
       // Compute similarity score from cosine distance
       // sqlite-vec with cosine metric returns distance where similarity = 1 - distance
-      const score = Math.max(0, Math.min(1, 1 - vec.distance));
+      const score = Math.max(0, Math.min(1, 1 - distance));
 
       if (score < threshold) continue;
 

package/src/cli/commands/status.ts

@@ -339,6 +339,7 @@ export async function status(
     const statusResult = await store.getStatus({
       embedModel: resolveModelUri(config, "embed"),
       chunking: config.chunking ?? {},
+      configuredCollections: config.collections.map(({ name }) => name),
     });
     if (!statusResult.ok) {
       return { success: false, error: statusResult.error.message };

package/src/core/audit-links.ts

@@ -13,8 +13,11 @@ import {
   compareAuditFindingDrafts,
 } from "./audit";
 
-/** 1.1: workspace-wide resolution and reference/resolution/scope evidence. */
-export const LINK_AUDIT_RULE_VERSION = "1.1" as const;
+/**
+ * 1.1: workspace-wide resolution and reference/resolution/scope evidence.
+ * 1.2: `links.outside-index` for workspace files that are not indexed.
+ */
+export const LINK_AUDIT_RULE_VERSION = "1.2" as const;
 /** Default per-rule cap; runs pass the effective `maxFindings` instead. */
 export const LINK_AUDIT_MAX_FINDINGS_PER_RULE = 1000;
 
@@ -70,6 +73,7 @@ const linkFinding = (
   onEvidenceTruncated: () => void
 ): AuditFindingDraft => {
   const ambiguous = (link.resolved?.matchCount ?? 0) > 1;
+  const outsideIndex = !ambiguous && link.outsideIndex === true;
   const target = `${link.targetCollection}:${link.targetRef}`;
   const kind = referenceKind(link);
   const tied = (link.resolved?.candidates ?? []).map((candidate) =>
@@ -101,7 +105,11 @@ const linkFinding = (
     matchRank: link.resolved?.matchRank ?? null,
     normalizedTarget: link.targetRefNorm,
     referenceKind: kind,
-    resolutionStatus: ambiguous ? "ambiguous" : "unresolved",
+    resolutionStatus: ambiguous
+      ? "ambiguous"
+      : outsideIndex
+        ? "outside-index"
+        : "unresolved",
     resolvedScope,
   };
   // Candidate URIs are added while the detail stays within the evidence
@@ -140,6 +148,26 @@ const linkFinding = (
     candidates.uris.pop();
   }
   if (candidates.truncated) onEvidenceTruncated();
+  if (outsideIndex) {
+    return {
+      subject: link.sourceUri,
+      location: lineLocation(link.startLine, link.startCol),
+      severity: "info",
+      message: `Link target exists outside the index: ${target}`,
+      evidence: [
+        {
+          kind: "outside-index-target",
+          summary: target,
+          uri: link.sourceUri,
+          path: link.sourceRelPath,
+          detail: render(),
+        },
+      ],
+      guidance: [
+        "No action needed; index the target's folder to add it to the link graph",
+      ],
+    };
+  }
   return {
     subject: link.sourceUri,
     location: lineLocation(link.startLine, link.startCol),
@@ -210,6 +238,7 @@ export const evaluateLinkAudit = (
   policy: AuditOrphanPolicy
 ): AuditRuleContribution[] => {
   const unresolved: AuditFindingDraft[] = [];
+  const outsideIndex: AuditFindingDraft[] = [];
   const ambiguous: AuditFindingDraft[] = [];
   const connected = new Set<number>();
   const auditedDocumentIds = new Set(
@@ -229,7 +258,11 @@ export const evaluateLinkAudit = (
   for (const link of snapshot.links) {
     const sourceAudited = auditedDocumentIds.has(link.sourceId);
     if (!link.resolved) {
-      if (sourceAudited) unresolved.push(finding(link));
+      if (sourceAudited) {
+        (link.outsideIndex === true ? outsideIndex : unresolved).push(
+          finding(link)
+        );
+      }
       continue;
     }
     // A tied workspace link is not a resolved link: it connects nothing.
@@ -327,6 +360,25 @@ export const evaluateLinkAudit = (
       findingCount: orphanFindings.length,
       skipReason: partial ? "snapshot_truncated" : null,
     },
+    {
+      // Informational: Obsidian resolves these links, so they never fail
+      // the audit. Info findings keep the rule passing.
+      ...common,
+      ruleId: "links.outside-index",
+      category: "links",
+      status: partial ? ("inconclusive" as const) : ("pass" as const),
+      message: [
+        partial
+          ? "Outside-index scan was truncated"
+          : `${outsideIndex.length} links point to existing files outside the index`,
+        ...(snapshot.outsideIndexDiagnostic
+          ? [snapshot.outsideIndexDiagnostic]
+          : []),
+      ].join("; "),
+      findings: boundedFindings(outsideIndex, maxFindingsPerRule),
+      findingCount: outsideIndex.length,
+      skipReason: partial ? "snapshot_truncated" : null,
+    },
     {
       ...common,
       ruleId: "links.parser-boundary",