get-claudia 1.55.19 → 1.55.20

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 All notable changes to Claudia will be documented in this file.
 
+## 1.55.20 (2026-03-19)
+
+### Community Fixes
+
+Four fixes from GitHub issues #24, #26, #28, #31.
+
+- **Fixed MCP double-spawn crash (#24)** -- Removed `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` from the template settings. This env var caused Claude Code to spawn the memory daemon twice on Linux, crashing both with BrokenPipeError. The installer now strips it from existing user settings on upgrade.
+- **Alias overlap false positives (#26)** -- Single-token aliases (common first names like "Joel") no longer flag unrelated entities as 95% similar. The filter checks if full entity names diverge beyond the shared alias. Multi-token aliases work normally.
+- **Stale dedupe predictions (#28)** -- After merging or deleting entities, dedupe predictions referencing them are now expired immediately instead of lingering for up to 14 days in briefings.
+- **Memory-health skill schema reference (#31)** -- Added complete column-name reference to the skill file. Documents `sacred_reason` (not `sacred`), `invalid_at` (not `invalidated_at` on relationships), and that embeddings live in separate tables.
+- 717 tests pass, 0 regressions, 11 new tests across 2 new test files.
+
 ## 1.55.19 (2026-03-19)
 
 ### The Self-Healer

package/bin/index.js

@@ -667,6 +667,20 @@ async function main() {
     console.log(` ${colors.cyan}✓${colors.reset} Framework updated (data preserved)`);
   }
 
+  // Self-heal: strip CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS from settings (#24)
+  // This env var causes double-spawn crashes on Linux and some macOS setups
+  try {
+    const settingsPath = join(targetPath, '.claude', 'settings.local.json');
+    if (existsSync(settingsPath)) {
+      const raw = readFileSync(settingsPath, 'utf8');
+      const settings = JSON.parse(raw);
+      if (settings.env && settings.env.CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS) {
+        delete settings.env.CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS;
+        writeFileSync(settingsPath, JSON.stringify(settings, null, 2) + '\n');
+      }
+    }
+  } catch { /* non-fatal */ }
+
   // Restore MCP servers that earlier versions incorrectly disabled.
   restoreMcpServers(targetPath);
 

package/memory-daemon/claudia_memory/services/consolidate.py

@@ -2496,6 +2496,22 @@ class ConsolidateService:
                     e1 = self.db.get_one("entities", where="id = ?", where_params=(row["eid1"],))
                     e2 = self.db.get_one("entities", where="id = ?", where_params=(row["eid2"],))
                     if e1 and e2:
+                        shared = row["alias"].strip()
+
+                        # Single-token alias filter (#26): a shared first name
+                        # like "joel" is weak evidence when full names diverge.
+                        if " " not in shared:
+                            tokens1 = set(e1["name"].lower().split())
+                            tokens2 = set(e2["name"].lower().split())
+                            overlap = tokens1 & tokens2
+                            non_shared = (tokens1 | tokens2) - overlap
+                            if len(non_shared) >= 2 and len(overlap) <= 1:
+                                logger.debug(
+                                    f"Alias overlap skipped: '{e1['name']}' / '{e2['name']}' "
+                                    f"share only single-token alias '{shared}'"
+                                )
+                                continue
+
                         candidates.append({
                             "entity_1": {"id": e1["id"], "name": e1["name"], "type": e1["type"]},
                             "entity_2": {"id": e2["id"], "name": e2["name"], "type": e2["type"]},

package/memory-daemon/claudia_memory/services/remember.py

@@ -996,6 +996,9 @@ class RememberService:
         result["success"] = True
         logger.info(f"Merged entity {source_id} ({source['name']}) into {target_id} ({target['name']})")
 
+        # Expire dedupe predictions referencing the merged entity (#28)
+        self._expire_dedupe_predictions(source_id)
+
         # Audit log
         _audit_log(
             "entity_merge",
@@ -1048,6 +1051,9 @@ class RememberService:
 
         logger.info(f"Soft-deleted entity {entity_id} ({entity['name']}): {reason}")
 
+        # Expire dedupe predictions referencing the deleted entity (#28)
+        self._expire_dedupe_predictions(entity_id)
+
         # Audit log
         _audit_log(
             "entity_delete",
@@ -1063,6 +1069,28 @@ class RememberService:
             "deleted_at": now,
         }
 
+    def _expire_dedupe_predictions(self, entity_id: int) -> None:
+        """Expire dedupe predictions that reference the given entity ID.
+
+        After merging or deleting an entity, stale dedupe suggestions should
+        not keep appearing in briefings. This expires (not deletes) any
+        prediction where the entity_id appears in the dedupe_pair metadata.
+        """
+        now = datetime.utcnow().isoformat()
+        try:
+            self.db.execute(
+                """
+                UPDATE predictions
+                SET expires_at = ?
+                WHERE prediction_type = 'suggestion'
+                  AND expires_at > ?
+                  AND (metadata LIKE ? OR metadata LIKE ?)
+                """,
+                (now, now, f'%"dedupe_pair": [{entity_id},%', f'%, {entity_id}]%'),
+            )
+        except Exception as e:
+            logger.debug(f"Failed to expire dedupe predictions for entity {entity_id}: {e}")
+
     def correct_memory(
         self,
         memory_id: int,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "get-claudia",
-  "version": "1.55.19",
+  "version": "1.55.20",
   "description": "An AI assistant who learns how you work.",
   "keywords": [
     "claudia",

package/template-v2/.claude/settings.local.json

@@ -1,7 +1,5 @@
 {
-  "env": {
-    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
-  },
+  "env": {},
   "hooks": {
     "SessionStart": [
       {

package/template-v2/.claude/skills/memory-health/SKILL.md

@@ -15,11 +15,32 @@ Provide a dashboard view of the memory system's health, including entity counts,
 - User says "how much do you remember?", "what's in your brain?"
 - Periodic self-check (weekly review, morning brief)
 
+## Schema Reference
+
+**Use these exact column names in all SQLite queries. Do NOT guess column names.**
+
+**entities** table: `id`, `name`, `type`, `canonical_name`, `description`, `importance` (REAL), `created_at`, `updated_at`, `metadata`, `last_contact_at`, `contact_frequency_days`, `contact_trend`, `attention_tier`, `close_circle` (BOOLEAN), `close_circle_reason`, `deleted_at`, `deleted_reason`
+
+**memories** table: `id`, `content`, `content_hash`, `type`, `importance` (REAL), `confidence` (REAL), `source`, `source_id`, `source_context`, `created_at`, `updated_at`, `last_accessed_at`, `access_count`, `verified_at`, `verification_status`, `metadata`, `source_channel`, `deadline_at`, `temporal_markers`, `lifecycle_tier`, `sacred_reason` (NOT `sacred`), `archived_at`, `fact_id`, `hash`, `prev_hash`, `workspace_id`, `corrected_at`, `corrected_from`, `invalidated_at`, `invalidated_reason`, `origin_type`
+
+**relationships** table: `id`, `source_entity_id`, `target_entity_id`, `relationship_type`, `strength` (REAL), `origin_type`, `direction`, `valid_at`, `invalid_at` (NOT `invalidated_at`), `created_at`, `updated_at`, `metadata`, `lifecycle_tier`
+
+**predictions** table: `id`, `content`, `prediction_type`, `priority` (REAL), `expires_at`, `is_shown`, `is_acted_on`, `created_at`, `shown_at`, `prediction_pattern_name`, `metadata`
+
+**Important distinctions:**
+- Embeddings are in SEPARATE tables (`entity_embeddings`, `memory_embeddings`), NOT columns on the main tables
+- `memories.sacred_reason` exists, but there is no column called `sacred` or `critical`
+- `relationships.invalid_at` (not `invalidated_at`) marks invalid relationships
+- `memories.invalidated_at` marks invalidated memories (different column name than relationships)
+- Always filter with `deleted_at IS NULL` on entities and `invalidated_at IS NULL` on memories
+
 ## Workflow
 
 ### Step 1: Gather Statistics
 
-Use the Claudia CLI to get current system state:
+Use the `memory.system_health` MCP tool or direct SQLite queries with the schema above.
+
+Alternatively, use the Claudia CLI to get current system state:
 
 ```bash
 claudia memory session context --scope full --project-dir "$PWD"