@loreai/core 0.20.2 → 0.22.0

package/dist/node/index.js

@@ -816,6 +816,17 @@ var MIGRATIONS = [
   );
   CREATE INDEX IF NOT EXISTS idx_dedup_feedback_project
     ON dedup_feedback(project_id);
+  `,
+  `
+  -- Version 26: Persist sub-agent parent\u2013child session relationships.
+  -- parent_session_id: Lore internal session ID of the parent session
+  -- (resolved from the x-parent-session-id header at detection time).
+  -- NULL for root (non-sub-agent) sessions.
+  -- is_subagent: boolean flag (1 = sub-agent) so the flag survives
+  -- gateway restarts without requiring the header again.
+  ALTER TABLE session_state ADD COLUMN parent_session_id TEXT;
+  ALTER TABLE session_state ADD COLUMN is_subagent INTEGER NOT NULL DEFAULT 0;
+  CREATE INDEX IF NOT EXISTS idx_session_state_parent ON session_state(parent_session_id);
   `
 ];
 function dbPath() {
@@ -1039,13 +1050,11 @@ function loadForceMinLayer(sessionID) {
   return row?.force_min_layer ?? 0;
 }
 function saveForceMinLayer(sessionID, layer) {
-  if (layer === 0) {
-    db().query("DELETE FROM session_state WHERE session_id = ?").run(sessionID);
-  } else {
-    db().query(
-      "INSERT OR REPLACE INTO session_state (session_id, force_min_layer, updated_at) VALUES (?, ?, ?)"
-    ).run(sessionID, layer, Date.now());
-  }
+  const now = Date.now();
+  db().query(
+    "INSERT OR IGNORE INTO session_state (session_id, force_min_layer, updated_at) VALUES (?, 0, ?)"
+  ).run(sessionID, now);
+  db().query("UPDATE session_state SET force_min_layer = ?, updated_at = ? WHERE session_id = ?").run(layer, now, sessionID);
 }
 function saveSessionCosts(sessionID, costs) {
   db().query(
@@ -1226,6 +1235,14 @@ function saveSessionTracking(sessionID, state) {
     sets.push("last_bust_at = ?");
     vals.push(state.lastBustAt);
   }
+  if (state.parentSessionId !== void 0) {
+    sets.push("parent_session_id = ?");
+    vals.push(state.parentSessionId);
+  }
+  if (state.isSubagent !== void 0) {
+    sets.push("is_subagent = ?");
+    vals.push(state.isSubagent ? 1 : 0);
+  }
   db().query(
     "UPDATE session_state SET " + sets.join(", ") + " WHERE session_id = ?"
   ).run(...vals, sessionID);
@@ -1238,7 +1255,8 @@ function loadSessionTracking(sessionID) {
               fingerprint, header_session_id, header_name,
               resolved_conversation_ttl, warmup_state,
               dynamic_context_cap, bust_rate_ema, inter_bust_interval_ema,
-              last_layer, last_known_input, last_turn_at, last_bust_at
+              last_layer, last_known_input, last_turn_at, last_bust_at,
+              parent_session_id, is_subagent
        FROM session_state WHERE session_id = ?`
   ).get(sessionID);
   if (!row) return null;
@@ -1262,7 +1280,9 @@ function loadSessionTracking(sessionID) {
     lastLayer: row.last_layer,
     lastKnownInput: row.last_known_input,
     lastTurnAt: row.last_turn_at,
-    lastBustAt: row.last_bust_at
+    lastBustAt: row.last_bust_at,
+    parentSessionId: row.parent_session_id,
+    isSubagent: row.is_subagent === 1
   };
 }
 function loadHeaderSessionIndex() {
@@ -1277,6 +1297,18 @@ function loadHeaderSessionIndex() {
     headerName: row.header_name
   }));
 }
+function loadParentChildMap() {
+  const rows = db().query(
+    `SELECT session_id, parent_session_id
+       FROM session_state
+       WHERE parent_session_id IS NOT NULL`
+  ).all();
+  const map5 = /* @__PURE__ */ new Map();
+  for (const row of rows) {
+    map5.set(row.session_id, row.parent_session_id);
+  }
+  return map5;
+}
 function getKV(key) {
   const row = db().query("SELECT value FROM kv_meta WHERE key = ?").get(key);
   return row?.value ?? null;
@@ -12245,7 +12277,9 @@ var CURATOR_SYSTEM = `You are a long-term memory curator. Your job is to extract
 Focus ONLY on knowledge that helps a coding agent work effectively on THIS codebase:
 - Architectural decisions and their rationale (why something was built a certain way)
 - Non-obvious implementation patterns and conventions specific to the project
-- Recurring gotchas, constraints, or traps in the codebase
+- Recurring gotchas, constraints, or traps in the codebase \u2014 always include WHY the
+  wrong approach seems right, not just the trap and fix. Without this, a future session
+  will re-propose the broken approach because it looks like a reasonable improvement.
 - Environment/tooling setup details that affect development
 - Important relationships between components that aren't obvious from reading the code
 - User preferences and working style specific to how they use this project
@@ -12264,10 +12298,19 @@ Do NOT extract:
 - Knowledge about unrelated projects or repositories unless explicitly cross-project
 - Restatements of what the code obviously does (e.g. "the auth module handles authentication")
 
+INCLUDE THE "WHY" \u2014 decisions and gotchas without rationale get undone:
+- Every "decision" MUST include the rejected alternative and why it was rejected.
+  Format: "Chose X over Y because Z." Without the rejected option, a future session
+  will re-propose Y because it looks like a reasonable improvement.
+- Every "gotcha" MUST explain why the wrong approach seems correct, not just the trap
+  and its fix. Format: "Trap: X looks right because [reason]. Fix: Y, because [reason]."
+- Any standard or rule without its rationale is vulnerable to being optimized away by
+  a session that doesn't know what problem it was solving.
+
 BREVITY IS CRITICAL \u2014 each entry must be concise:
 - content MUST be under 150 words (~600 characters). Capture ONE specific actionable
   insight in 2-3 sentences. Prefer terse technical language.
-- Each "gotcha": one specific trap + its fix in 1-2 sentences
+- Each "gotcha": one specific trap + WHY it looks right + its fix in 2-3 sentences
 - Each "architecture": one design decision and its key constraint
 - Focus on the actionable insight, not the full story behind it
 - If a pattern requires more detail, split into multiple focused entries (each under 150 words)
@@ -12293,6 +12336,20 @@ crossProject flag:
 - Default is true \u2014 most useful knowledge is worth sharing across projects
 - Set crossProject to false for things that are meaningless outside this specific repo (e.g. a config path, a project-local naming convention that conflicts with your usual style)
 
+Confidence values (0.0\u20131.0) \u2014 determines injection priority when budget is tight:
+- 1.0: Unconditional directive \u2014 user used "NEVER", "ALWAYS", "from now on", or similarly
+  absolute language. These must always be respected regardless of context.
+- 0.9: Strong preference \u2014 explicit user preference ("I prefer", "I want", "make sure to",
+  "don't forget to"). Clear intent but not absolute.
+- 0.8: Moderate preference \u2014 inferred from repeated user behavior or gentle correction across
+  sessions. Not explicitly stated as a rule.
+- 0.6: Mild/contextual preference \u2014 may not apply universally. Observed once or context-dependent.
+- For non-preference categories (gotcha, pattern, architecture, decision), confidence reflects
+  how well-established the knowledge is: 1.0 = verified/confirmed, 0.8 = high confidence,
+  0.6 = probable but unverified.
+- Default to 1.0 for preferences with strong directive language, 0.8 for other preferences.
+- Always set confidence on create ops \u2014 it determines injection priority.
+
 Produce a JSON array of operations:
 [
   {
@@ -12301,7 +12358,8 @@ Produce a JSON array of operations:
     "title": "Short descriptive title",
     "content": "Concise knowledge entry \u2014 under 150 words",
     "scope": "project" | "global",
-    "crossProject": true
+    "crossProject": true,
+    "confidence": 1.0
   },
   {
     "op": "update",
@@ -12339,7 +12397,8 @@ IMPORTANT:
 4. Only create a new entry for genuinely distinct knowledge with no existing home.
 5. Keep all entries under 150 words. If an existing entry is too long, use an update op to trim it.
 6. Pay special attention to user instructions ("always do X", "never do Y", "make sure to X").
-   These are strong signals for "preference" entries with high confidence.`;
+   These are strong signals for "preference" entries with high confidence (1.0 for absolute
+   directives like "never"/"always", 0.9 for explicit preferences like "I prefer").`;
 }
 var CONSOLIDATION_SYSTEM = `You are a long-term memory curator performing a consolidation pass. The knowledge base has grown too large and needs to be trimmed.
 
@@ -26617,8 +26676,10 @@ var LoreConfig = external_exports.object({
     distilled: external_exports.number().min(0.05).max(0.5).default(0.25),
     raw: external_exports.number().min(0.1).max(0.7).default(0.4),
     output: external_exports.number().min(0.1).max(0.5).default(0.25),
-    /** Max fraction of usable context reserved for LTM system-prompt injection. Default: 0.05 (5%). */
+    /** Max fraction of usable context reserved for context-bound LTM system-prompt injection. Default: 0.05 (5%). */
     ltm: external_exports.number().min(0.02).max(0.3).default(0.05),
+    /** Fraction of usable context for stable LTM (preferences). Independent of `ltm`. Default: 0.02 (2%). */
+    preferenceLtm: external_exports.number().min(0.01).max(0.1).default(0.02),
     /** Per-turn cache-read cost target in dollars. Controls when layer 0 (full
      *  passthrough) escalates to layer 1 (compressed). The cap is derived as:
      *  maxLayer0Tokens = max(target / model.cost.cache.read, 40K).
@@ -26633,7 +26694,7 @@ var LoreConfig = external_exports.object({
     targetBustCost: external_exports.number().min(0).default(1).optional(),
     /** @deprecated Ignored. Tier-based bust-vs-continue replaces static cap. */
     maxContextTokens: external_exports.number().min(0).optional()
-  }).default({ distilled: 0.25, raw: 0.4, output: 0.25, ltm: 0.05, targetCacheReadCostPerTurn: 0.1 }),
+  }).default({ distilled: 0.25, raw: 0.4, output: 0.25, ltm: 0.05, preferenceLtm: 0.02, targetCacheReadCostPerTurn: 0.1 }),
   /**
    * Cold-cache idle-resume handling.
    *
@@ -27779,6 +27840,7 @@ __export(ltm_exports, {
   recordDedupFeedback: () => recordDedupFeedback,
   recordDedupResultFeedback: () => recordDedupResultFeedback,
   remove: () => remove,
+  rerankPreferences: () => rerankPreferences,
   resolveRef: () => resolveRef2,
   saveCalibratedThreshold: () => saveCalibratedThreshold,
   search: () => search3,
@@ -28363,28 +28425,33 @@ function create(input) {
     ).get(pid, input.title) : db().query(
       "SELECT id FROM knowledge WHERE project_id IS NULL AND LOWER(title) = LOWER(?) AND confidence > 0 LIMIT 1"
     ).get(input.title);
+    const dedupUpdate = {
+      content: input.content,
+      ...input.confidence != null ? { confidence: input.confidence } : {}
+    };
     if (existing) {
-      update(existing.id, { content: input.content });
+      update(existing.id, dedupUpdate);
       return existing.id;
     }
     const crossExisting = db().query(
       "SELECT id FROM knowledge WHERE cross_project = 1 AND LOWER(title) = LOWER(?) AND confidence > 0 LIMIT 1"
     ).get(input.title);
     if (crossExisting) {
-      update(crossExisting.id, { content: input.content });
+      update(crossExisting.id, dedupUpdate);
       return crossExisting.id;
     }
     const fuzzyMatch = findFuzzyDuplicate({ title: input.title, projectId: pid });
     if (fuzzyMatch) {
-      update(fuzzyMatch.id, { content: input.content });
+      update(fuzzyMatch.id, dedupUpdate);
       return fuzzyMatch.id;
     }
   }
   const id = input.id ?? uuidv72();
   const now = Date.now();
+  const confidence = input.confidence != null ? Math.max(0, Math.min(1, input.confidence)) : 1;
   db().query(
     `INSERT INTO knowledge (id, project_id, category, title, content, source_session, cross_project, confidence, created_at, updated_at)
-       VALUES (?, ?, ?, ?, ?, ?, ?, 1.0, ?, ?)`
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`
   ).run(
     id,
     pid,
@@ -28393,6 +28460,7 @@ function create(input) {
     input.content,
     input.session ?? null,
     crossProject2 ? 1 : 0,
+    confidence,
     now,
     now
   );
@@ -28543,6 +28611,24 @@ async function forSession(projectPath2, sessionID, maxTokens, options) {
        ORDER BY confidence DESC, updated_at DESC`
   ).all(...categoryParams);
   if (!crossEntries.length && !projectEntries.length) return [];
+  const isPreferenceOnly = categoryFilter?.length === 1 && categoryFilter[0] === "preference";
+  if (isPreferenceOnly) {
+    const allPrefs = [...projectEntries, ...crossEntries];
+    allPrefs.sort(
+      (a, b) => a.confidence !== b.confidence ? b.confidence - a.confidence : b.updated_at - a.updated_at
+    );
+    const HEADER_OVERHEAD_TOKENS2 = 15;
+    let used2 = HEADER_OVERHEAD_TOKENS2;
+    const result2 = [];
+    for (const entry of allPrefs) {
+      if (used2 >= maxTokens) break;
+      const cost = estimateTokens3(entry.title + entry.content) + 10;
+      if (used2 + cost > maxTokens) continue;
+      result2.push(entry);
+      used2 += cost;
+    }
+    return result2;
+  }
   let sessionContext = "";
   if (sessionID) {
     const distRow = db().query(
@@ -28615,10 +28701,24 @@ async function forSession(projectPath2, sessionID, maxTokens, options) {
   const allScored = [...scoredProject, ...scoredCross];
   allScored.sort((a, b) => b.score - a.score);
   const HEADER_OVERHEAD_TOKENS = 15;
+  const ARCH_BUDGET_FRACTION = 0.2;
   let used = HEADER_OVERHEAD_TOKENS;
   const result = [];
+  const packedIds = /* @__PURE__ */ new Set();
+  const archBudget = Math.floor(maxTokens * ARCH_BUDGET_FRACTION);
+  const archEntries = allScored.filter((s) => s.entry.category === "architecture");
+  archEntries.sort((a, b) => b.score - a.score);
+  for (const { entry } of archEntries) {
+    if (used >= archBudget + HEADER_OVERHEAD_TOKENS) break;
+    const cost = estimateTokens3(entry.title + entry.content) + 10;
+    if (used + cost > maxTokens) continue;
+    result.push(entry);
+    packedIds.add(entry.id);
+    used += cost;
+  }
   for (const { entry } of allScored) {
     if (used >= maxTokens) break;
+    if (packedIds.has(entry.id)) continue;
     const cost = estimateTokens3(entry.title + entry.content) + 10;
     if (used + cost > maxTokens) continue;
     result.push(entry);
@@ -28680,6 +28780,28 @@ function crossProject() {
        ORDER BY confidence DESC, updated_at DESC`
   ).all();
 }
+function rerankPreferences() {
+  const prefs = db().query(`SELECT ${KNOWLEDGE_COLS} FROM knowledge WHERE category = 'preference' AND confidence = 1.0`).all();
+  const STRONG_DIRECTIVE_RE = /\b(never|always|must not|must)\b/i;
+  const EXPLICIT_PREF_RE = /\b(I (?:want|need|prefer|expect)|make sure to|don'?t forget)\b/i;
+  let updated = 0;
+  for (const entry of prefs) {
+    const text4 = entry.title + " " + entry.content;
+    let newConfidence;
+    if (STRONG_DIRECTIVE_RE.test(text4)) {
+      newConfidence = 1;
+    } else if (EXPLICIT_PREF_RE.test(text4)) {
+      newConfidence = 0.9;
+    } else {
+      newConfidence = 0.8;
+    }
+    if (newConfidence !== entry.confidence) {
+      update(entry.id, { confidence: newConfidence });
+      updated++;
+    }
+  }
+  return updated;
+}
 function searchLike2(input) {
   const terms = input.query.toLowerCase().split(/\s+/).filter((t2) => t2.length > 2);
   if (!terms.length) return [];
@@ -29972,13 +30094,20 @@ function getTier(tokens) {
 function recordCacheUsage(cacheWrite, cacheRead, inputTokens, sessionID) {
   if (!sessionID) return;
   const state = getSessionState(sessionID);
-  const total = inputTokens > 0 ? inputTokens : cacheWrite + cacheRead;
+  const total = cacheWrite + cacheRead + inputTokens;
   if (total > 0) {
-    if (cacheWrite / total > 0.5) {
+    const bustRatio = cacheWrite / total;
+    const prev = state.consecutiveBusts;
+    if (bustRatio > 0.5) {
       state.consecutiveBusts++;
     } else {
       state.consecutiveBusts = 0;
     }
+    if (state.consecutiveBusts !== prev) {
+      info(
+        `bust-tracker: session=${sessionID.slice(0, 16)} ratio=${bustRatio.toFixed(3)} (write=${cacheWrite} read=${cacheRead} uncached=${inputTokens}) busts=${prev}\u2192${state.consecutiveBusts}`
+      );
+    }
   }
 }
 var FIRST_TURN_OVERHEAD = 15e3;
@@ -30015,7 +30144,6 @@ function getSessionState(sessionID) {
       state.lastLayer = persisted.lastLayer;
       state.lastKnownInput = persisted.lastKnownInput;
       state.lastTurnAt = persisted.lastTurnAt;
-      state.consecutiveBusts = persisted.dynamicContextCap;
     }
     sessionStates.set(sessionID, state);
   }
@@ -30074,6 +30202,7 @@ function getLtmBudget(ltmFraction) {
   const usable = Math.max(0, contextLimit - outputReserved - overhead);
   return Math.floor(usable * ltmFraction);
 }
+var getPreferenceLtmBudget = getLtmBudget;
 function calibrate(actualInput, sessionID, messageCount) {
   const messageEstimate = sessionID ? getSessionState(sessionID).lastTransformEstimate : 0;
   if (messageEstimate > 0 || actualInput === 0) {
@@ -31089,6 +31218,7 @@ function sliceTokens(messages, start, end) {
 function splitSegments(messages, maxTokens) {
   const totalTokens = messages.reduce((s, m) => s + m.tokens, 0);
   if (totalTokens <= maxTokens) return [messages];
+  if (messages.length <= 1) return [messages];
   const splitIdx = findSplitIndex(messages, maxTokens);
   const left = messages.slice(0, splitIdx);
   const right = messages.slice(splitIdx);
@@ -31727,7 +31857,8 @@ function applyOps(ops, input) {
         content: content3,
         session: input.sessionID,
         scope: op.scope,
-        crossProject: op.crossProject ?? true
+        crossProject: op.crossProject ?? true,
+        confidence: op.confidence
       });
       idsToSync.push(id);
       created++;
@@ -33968,6 +34099,7 @@ export {
   getLtmBudget,
   getLtmTokens,
   getMeta,
+  getPreferenceLtmBudget,
   getTier,
   h,
   importFromFile,
@@ -33988,6 +34120,7 @@ export {
   loadAllSessionCosts,
   loadForceMinLayer,
   loadHeaderSessionIndex,
+  loadParentChildMap,
   loadSessionCosts,
   loadSessionTracking,
   log_exports as log,