@theglitchking/semantic-memory 1.2.0 → 1.2.1

package/.claude-plugin/marketplace.json

@@ -6,18 +6,18 @@
   },
   "metadata": {
     "description": "Single-plugin marketplace for semantic-memory — the unified memory-layer plugin. v1.1.0 absorbs brain's strengths: AGENTS.md contract artifact, hard-gated verification sessions, multi-agent skill bundler (codex/copilot/pi), and SessionStart drift detection — all backwards-compatible with v1.0.x via deprecation shims.",
-    "version": "1.2.0"
+    "version": "1.2.1"
   },
   "plugins": [
     {
       "name": "semantic-memory",
       "description": "Unified memory-layer plugin for Claude Code. One MCP server, 40 tools across multi-corpus search/read/write/graph/lint/sessions/contract. v1.1.0 adds: AGENTS.md contract artifact, hard-gated verification sessions (session_start/run/finish/status), proposal-mode synthesize_note + synthesize_promote, multi-agent skill bundler (codex/copilot/pi via the new `skills` CLI), and SessionStart drift detection. Backwards-compatible with v1.0.x — every previous tool name remains callable through deprecation shims (removed in v2.0.0).",
-      "version": "1.2.0",
+      "version": "1.2.1",
       "author": {
         "name": "TheGlitchKing"
       },
-      "homepage": "https://github.com/TheGlitchKing/semantic-sidekick",
-      "repository": "https://github.com/TheGlitchKing/semantic-sidekick.git",
+      "homepage": "https://github.com/TheGlitchKing/semantic-memory",
+      "repository": "https://github.com/TheGlitchKing/semantic-memory.git",
       "license": "MIT",
       "keywords": [
         "mcp",

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 All notable changes to semantic-memory (formerly semantic-sidekick) will be documented here.
 
+## [1.2.1] - 2026-06-11 — Capture-cue false-positive fixes
+
+Fixes two classes of false positives in the Stop-hook capture nudge (`hooks/vault-context.js`), where the cue scanner over-fired and asked to synthesize moments that carried no new knowledge.
+
+### Fixed
+
+- **Conversational "gotcha" no longer fires the cue.** The `gotcha` cue was a bare-word match (`\bgotcha\b`), so filler acknowledgments like "Gotcha, ok…" tripped it. It now requires noun-context (`a/the gotcha`, `gotcha:`, `gotchas`, `gotcha is/was/here/with`). `workaround`/`hack` are unchanged.
+- **Quoted machinery no longer re-primes the cue.** Cue detection ran over the entire raw prompt, so pasting tool output back — especially this hook's own `<vault-capture-prompt>` block, which contains the words "gotcha"/"workaround"/"hack" and the cue regexes themselves — re-primed `capture-pending` in a self-referential loop. New `stripQuotedMachinery()` removes self-emitted `<vault-*>` blocks, fenced code, and inline-code spans before matching, so only the user's own prose is scanned.
+
+### Tests
+
+- Extended `test/phase4/mode-hook.test.ts` with four regression cases: conversational "gotcha" does not prime; noun-context "the gotcha is…" does; pasting a `<vault-capture-prompt>` block does not re-prime; and a real cue alongside fenced machinery still primes (the strip is surgical, not total).
+
 ## [1.2.0] - 2026-05-09 — State consolidation under .claude/.semantic-memory/
 
 Consolidates every transient state file under one namespace. Renames the legacy `.sidekick-*` files to drop the obsolete prefix. All legacy paths remain readable through v1.x via fallback; v2.0 will remove the fallback.

package/hooks/vault-context.js

@@ -267,13 +267,29 @@ const CAPTURE_CUES = [
   /\bbecause\b/i,
   /\bdecided\b|\bwe chose\b|\bthe decision\b/i,
   /\bturned out to be\b|\bthe bug was\b|\bthe fix was\b/i,
-  /\bgotcha\b|\bworkaround\b|\bhack\b/i,
+  // "gotcha" is also a common acknowledgment ("Gotcha, ok…") — require noun-context
+  // (a/the gotcha, gotcha:, gotchas, gotcha is/was/here/with) so bare filler doesn't misfire.
+  /\b(?:a|the|one|another)\s+gotcha\b|\bgotcha[:s]\b|\bgotcha (?:is|was|here|with)\b|\bworkaround\b|\bhack\b/i,
   /\bnew convention\b|\bfrom now on\b|\bgoing forward\b/i,
 ];
 
+// Cue detection must scan the user's own prose, NOT quoted machinery. Pasting
+// tool output back into the prompt — especially this hook's own <vault-*> blocks,
+// which literally contain the words "gotcha"/"workaround"/"hack" and the cue
+// regexes themselves — would otherwise re-prime capture-pending in a self-
+// referential loop. Strip self-emitted blocks, fenced code, and inline-code spans
+// before matching. (Unterminated blocks/fences from truncated pastes strip to EOL.)
+function stripQuotedMachinery(prompt) {
+  return prompt
+    .replace(/<vault-[a-z-]+[\s\S]*?(?:<\/vault-[a-z-]+>|$)/gi, " ")
+    .replace(/```[\s\S]*?(?:```|$)/g, " ")
+    .replace(/`[^`]*`/g, " ");
+}
+
 function detectCaptureCue(prompt) {
+  const prose = stripQuotedMachinery(prompt);
   for (const re of CAPTURE_CUES) {
-    if (re.test(prompt)) return re.source;
+    if (re.test(prose)) return re.source;
   }
   return null;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@theglitchking/semantic-memory",
-  "version": "1.2.0",
+  "version": "1.2.1",
   "description": "Unified memory-layer plugin for Claude Code. One MCP server, multiple typed corpora (vault, code, plans, docs, research, project-map), unified cross-corpus knowledge graph, conditional tool registration, in-process drift detection, and human↔code translation. Rebrand of @theglitchking/semantic-sidekick at v1.0 — see CHANGELOG for migration.",
   "type": "module",
   "main": "./dist/core/index.js",