@yemi33/minions 0.1.2368 → 0.1.2370

package/engine/playbook.js

@@ -11,6 +11,9 @@ const shared = require('./shared');
 const queries = require('./queries');
 const { wrapUntrusted, buildSource } = require('./untrusted-fence');
 const { MINIONS_BRAND_URL } = require('./comment-format');
+const features = require('./features');
+const memoryRetrieval = require('./memory-retrieval');
+const memoryStore = require('./memory-store');
 
 const { safeJsonArr, safeRead, getProjects, log, ts, dateStamp, truncateTextBytes, ENGINE_DEFAULTS, WI_STATUS, WORK_TYPE, PR_STATUS, DISPATCH_RESULT, getProjectOrg } = shared;
 const { getConfig, getDispatch, getNotes, getPrs, getKnowledgeBaseIndex, AGENTS_DIR } = queries;
@@ -716,10 +719,61 @@ function renderPlaybook(type, vars) {
     inertAppendices.push('\n\n---\n\n## Pinned Context (CRITICAL — READ FIRST)\n\n' + (fenced || pinnedContent));
   }
 
+  // Structured memory rollout is deliberately two-stage. The feature flag
+  // turns retrieval on; shadow mode (default) records selection telemetry but
+  // leaves the legacy notes/personal-memory prompt unchanged. Once shadow mode
+  // is disabled, a non-empty bounded retrieval pack replaces those broad
+  // appendices. Pinned operator context always remains fully visible.
+  let relevantMemoryInjected = false;
+  if (features.isFeatureOn('memoryRetrieval', configEarly)) {
+    const engineConfig = configEarly.engine || {};
+    const shadow = engineConfig.memoryRetrievalShadowMode ?? ENGINE_DEFAULTS.memoryRetrievalShadowMode;
+    try {
+      const retrieved = memoryRetrieval.retrieveRelevantMemory({
+        workItemId: vars.item_id || vars.task_id || '',
+        title: vars.item_name || vars.pr_title || '',
+        description: vars.item_description || vars.task_description || vars.pr_description || '',
+        type: vars.work_type || type,
+        project: projectName || '',
+        agent: vars.agent_id || '',
+        failureClass: vars.failure_class || '',
+        sourcePlan: vars.source_plan || '',
+        references: vars.references || '',
+        prTitle: vars.pr_title || '',
+        prBranch: vars.pr_branch || '',
+      }, {
+        topK: engineConfig.memoryRetrievalTopK ?? ENGINE_DEFAULTS.memoryRetrievalTopK,
+        maxBytes: engineConfig.memoryRetrievalMaxBytes ?? ENGINE_DEFAULTS.memoryRetrievalMaxBytes,
+        candidateLimit: engineConfig.memoryRetrievalCandidateLimit ?? ENGINE_DEFAULTS.memoryRetrievalCandidateLimit,
+      });
+      memoryStore.recordRetrievalRun({
+        workItemId: vars.item_id || vars.task_id || null,
+        project: projectName,
+        agent: vars.agent_id || null,
+        query: retrieved.query,
+        candidateCount: retrieved.candidates.length,
+        selectedCount: retrieved.selected.length,
+        selectedBytes: retrieved.bytes,
+        durationMs: retrieved.durationMs,
+        shadow,
+      });
+      if (!shadow && retrieved.text) {
+        const fenced = wrapUntrusted(
+          retrieved.text,
+          buildSource('memory-retrieval', { item: vars.item_id || vars.task_id || 'unknown' })
+        );
+        inertAppendices.push('\n\n---\n\n## Relevant Memory (task-selected; verify against live code)\n\n' + (fenced || retrieved.text));
+        relevantMemoryInjected = true;
+      }
+    } catch (e) {
+      log('warn', `Memory retrieval failed for ${vars.item_id || vars.task_id || type}; using legacy memory: ${e.message}`);
+    }
+  }
+
   // Inject team notes (single injection point — not in buildAgentContext) — capped via ENGINE_DEFAULTS.
   // F5: wrap in <UNTRUSTED-INPUT> fence — notes.md is an LLM-consolidated mix
   // of agent inbox notes (semi-trusted) and human edits.
-  let notes = getNotes();
+  let notes = relevantMemoryInjected ? '' : getNotes();
   if (notes) {
     if (Buffer.byteLength(notes, 'utf8') > ENGINE_DEFAULTS.maxNotesPromptBytes) {
       const sections = notes.split(/(?=^### )/m);
@@ -738,7 +792,7 @@ function renderPlaybook(type, vars) {
   // notes budget; missing file degrades gracefully (silent skip).
   // F5: fence — agent-authored inbox notes routed into this file; any agent
   // could include attacker-controlled quoted material.
-  const agentIdForMemory = vars.agent_id;
+  const agentIdForMemory = relevantMemoryInjected ? '' : vars.agent_id;
   if (agentIdForMemory && /^[a-z][a-z0-9-]{0,40}$/i.test(agentIdForMemory) && !String(agentIdForMemory).toLowerCase().startsWith('temp-')) {
     const agentMemRel = `knowledge/agents/${String(agentIdForMemory).toLowerCase()}.md`;
     const agentMemPath = path.join(MINIONS_DIR, agentMemRel);

package/engine/shared.js

@@ -3683,6 +3683,11 @@ const ENGINE_DEFAULTS = {
   agentMemorySummaryEnabled: false, // opt-in: when true, eviction batches go through an LLM-compressed summary before being dropped. Default off to mirror the conservative gating on the existing reconcile pass (LLM cost + test stability). Operators flip via engine.agentMemorySummaryEnabled.
   agentMemorySummaryThreshold: 30, // batch window: when summary is enabled and a prune evicts entries, fold at least this many oldest sections into one summary. Means "summary every ~30 entries" in steady state (the original PRD intent).
   agentMemorySummaryDays: 30, // age trigger: when the oldest section is older than this and >= agentMemorySummaryThreshold entries exist, summarize the oldest window even if the file is under the entry cap.
+  memoryRetrievalShadowMode: true, // compute + measure FTS5 retrieval but preserve legacy prompt injection until explicitly disabled
+  memoryRetrievalTopK: 8, // maximum structured memory records injected into one dispatch
+  memoryRetrievalMaxBytes: 8 * 1024, // hard byte budget for the Relevant Memory prompt appendix
+  memoryRetrievalCandidateLimit: 50, // bounded FTS5 candidate pool before deterministic re-ranking
+  memoryEpisodicCapture: false, // opt-in capture of compact task outcomes; never stores transcripts or chain-of-thought
   untrustedFenceMaxBytes: 64 * 1024, // F5 (W-mpeklod3000we69c): per-block cap for `<UNTRUSTED-INPUT>` fences in engine/untrusted-fence.js. 64KB is long enough for realistic PR comments / pinned notes / agent memory sections, short enough that a megabyte-bomb comment cannot blow up the prompt. Content above the cap is truncated INSIDE the fence with a `[truncated N more bytes]` marker so the agent still sees the provenance attribute.
   maxMeetingPromptBytes: 16 * 1024, // cap meeting findings/debate context injected into prompts
   maxMeetingHumanNotesBytes: 2 * 1024, // cap human note bullet lists injected into meeting prompts

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yemi33/minions",
-  "version": "0.1.2368",
+  "version": "0.1.2370",
   "description": "Multi-agent AI dev team that runs from ~/.minions/ — five autonomous agents share a single engine, dashboard, and knowledge base",
   "bin": {
     "minions": "bin/minions.js"