@unblocklabs/unblock-memory 0.3.25 → 0.3.26

package/dist/src/training-retrieval.js

@@ -2,6 +2,7 @@ import { createHash } from "node:crypto";
 import { existsSync } from "node:fs";
 import { join } from "node:path";
 import { createRequire } from "node:module";
+import { setImmediate as yieldToEventLoop } from "node:timers/promises";
 import { readSessionManifest } from "./session-sync.js";
 import { resolveSessionSource } from "./sources.js";
 import { trainingCandidates } from "./training-candidates.js";
@@ -95,7 +96,15 @@ export async function historicalTrainingSearch(stateDir, chatTypes, cutoff, open
         FROM content_vectors cv JOIN vectors_vec v ON v.hash_seq=cv.hash||'_'||cv.seq
         WHERE cv.hash=? ORDER BY cv.seq`);
             let dimensions;
+            // Copy cooperatively: concurrent snapshots must not starve lease timers,
+            // provider responses or native-model disposal. Keep the source transaction
+            // open across yields so every row still comes from the same read snapshot.
+            let yieldAt = performance.now() + 25;
             for (const session of Object.values(manifest.sessions).sort((a, b) => a.documentPath.localeCompare(b.documentPath))) {
+                if (performance.now() >= yieldAt) {
+                    await yieldToEventLoop();
+                    yieldAt = performance.now() + 25;
+                }
                 // Loggie projections can retroactively annotate old text using later revisions.
                 // Workspace files and meetings lack immutable historical content proof here.
                 if (!chatTypes.includes(session.chatType) || session.provider === "loggie" || session.startedAt >= cutoff) {
@@ -120,7 +129,11 @@ export async function historicalTrainingSearch(stateDir, chatTypes, cutoff, open
                 report.sessions++;
                 if (prefix.spans.length < (session.messages?.length ?? 0))
                     report.truncated++;
-                for (const chunk of chunks.all(row.hash)) {
+                for (const chunk of chunks.iterate(row.hash)) {
+                    if (performance.now() >= yieldAt) {
+                        await yieldToEventLoop();
+                        yieldAt = performance.now() + 25;
+                    }
                     if (!Number.isSafeInteger(chunk.pos) || !Number.isSafeInteger(chunk.chunk_len) || chunk.pos < 0 || chunk.chunk_len <= 0 ||
                         chunk.pos + chunk.chunk_len > prefix.body.length) {
                         report.excludedChunks++;

package/dist/src/training-store.js

@@ -43,6 +43,8 @@ export class TrainingStore {
           status TEXT NOT NULL CHECK(status IN ('pending','attempted','complete','failed','ambiguous')),
           result_json TEXT, error TEXT, completed_at INTEGER
         ) STRICT;
+        CREATE INDEX IF NOT EXISTS training_judgment_exclusions ON training_steps(json_extract(request_json,'$.identity'))
+          WHERE stage='judge' AND status='complete' AND json_extract(result_json,'$.excluded')=1;
         CREATE TABLE IF NOT EXISTS training_step_attempts (
           id INTEGER PRIMARY KEY, step_id TEXT NOT NULL REFERENCES training_steps(id), started_at INTEGER NOT NULL,
           finished_at INTEGER, status TEXT NOT NULL, error TEXT, result_json TEXT

package/docs/memory-training.md

@@ -128,8 +128,13 @@ No blanket failure skipping.
 
 The host must support isolated completion and grant
 `plugins.entries.unblock-memory.llm.allowModelOverride: true` with
-`allowedModels: ["openai/gpt-6-luna"]`. Preserve other grants. The plugin does not
-change its own permissions. Credentials stay with the host, never in provenance.
+`allowedModels: ["openai/gpt-6-luna"]`. The operator CLI explicitly selects an
+agent, so an unbound CLI runtime also requires
+`plugins.entries.unblock-memory.llm.allowAgentIdOverride: true`. Preserve other
+grants. The host's model catalog and selected runtime must support that exact
+model with usable host-managed authentication; a model permission grant alone
+does not establish model availability. The plugin does not change its own
+permissions. Credentials stay with the host, never in provenance.
 
 ## Export and consolidation
 
@@ -145,3 +150,16 @@ identical-input groups together across nodes for train/validation. Use the actua
 **LFM2.5-230M-Base** tokenizer and an explicit causal-LM input/target format before
 fine-tuning; the interim byte limit is not a token count. Runtime abstention remains
 a separate TypeSafe gate; positive-only query training does not teach abstention.
+
+The repository helper uses a pinned official tokenizer revision and chat template,
+with assistant-only loss labels, a conservative 32,768-token training ceiling
+(the model config supports 128,000 positions), and deterministic grouped splits:
+
+```sh
+python scripts/prepare-query-training.py /private/node1-queries.jsonl /private/node2-queries.jsonl \
+  --output /private/new-prepared-directory --cache-dir /private/tokenizer-cache
+```
+
+Install `transformers` and `jinja2` in a separate environment first. It downloads
+tokenizer files only, never model weights. Regex secret screening is not an
+exhaustive privacy audit; inspect quarantine references before including those rows.

package/openclaw.plugin.json

@@ -1,7 +1,7 @@
 {
   "id": "unblock-memory",
   "name": "Unblock Memory",
-  "version": "0.3.25",
+  "version": "0.3.26",
   "description": "Indexes, retrieves, and analyzes configured workspace memory with existing QMD vectors.",
   "kind": "memory",
   "activation": { "onStartup": true },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@unblocklabs/unblock-memory",
-  "version": "0.3.25",
+  "version": "0.3.26",
   "description": "Workspace-native memory for OpenClaw, powered by QMD",
   "type": "module",
   "license": "MIT",