claude-mem-lite 3.64.0 → 3.66.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "3.64.0",
+      "version": "3.66.0",
       "source": "./",
       "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark)."
     }

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "3.64.0",
+  "version": "3.66.0",
   "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark).",
   "author": {
     "name": "sdsrss"

package/deep-search.mjs

@@ -36,6 +36,7 @@ import { sanitizeFtsQuery } from './utils.mjs';
 import { RRF_K } from './tfidf.mjs';
 import { rrfAccumulate } from './lib/rrf.mjs';
 import { llmRerankOrder, defaultRerankLLM } from './rerank.mjs';
+import { liveObsFilterSql } from './lib/inject-search-core.mjs';
 
 // original + up to 3 rewrites (keyword / concept-expansion / HyDE).
 export const MAX_VARIANTS = 4;
@@ -68,7 +69,7 @@ export const AUTO_DEEP_MIN_CORPUS = 10;
  */
 export function hasEscalatableCorpus(db, project, min = AUTO_DEEP_MIN_CORPUS) {
   try {
-    const where = ['superseded_at IS NULL', 'COALESCE(compressed_into, 0) = 0'];
+    const where = [liveObsFilterSql('')];
     const params = [];
     if (project) { where.push('project = ?'); params.push(project); }
     const row = db.prepare(`SELECT COUNT(*) AS c FROM observations WHERE ${where.join(' AND ')}`).get(...params);

package/format-utils.mjs

@@ -1,3 +1,4 @@
+import { DAY_MS } from './lib/time-constants.mjs';
 // claude-mem-lite: String formatting and display utilities
 // Extracted from utils.mjs for focused responsibility
 
@@ -231,7 +232,7 @@ export function isoWeekKey(epochMs) {
   const tmp = new Date(Date.UTC(d.getUTCFullYear(), d.getUTCMonth(), d.getUTCDate()));
   tmp.setUTCDate(tmp.getUTCDate() + 4 - (tmp.getUTCDay() || 7));
   const yearStart = new Date(Date.UTC(tmp.getUTCFullYear(), 0, 1));
-  const weekNum = Math.ceil(((tmp - yearStart) / 86400000 + 1) / 7);
+  const weekNum = Math.ceil(((tmp - yearStart) / DAY_MS + 1) / 7);
   const isoYear = tmp.getUTCFullYear();
   return `${isoYear}-W${String(weekNum).padStart(2, '0')}`;
 }

package/haiku-client.mjs

@@ -100,6 +100,28 @@ const MODEL_MAP = {
 // A call that genuinely needs sampling can pass opts.temperature to override.
 const DEFAULT_LLM_TEMPERATURE = 0;
 
+/**
+ * Timeout budget for BACKGROUND LLM work (detached enrich/optimize/summary
+ * workers, registry indexing) — the calls with no latency budget at all.
+ *
+ * Every dispatcher below degrades to `claude -p` when the keyed provider fails,
+ * and the CLI leg pays a full Claude Code boot before inference: measured
+ * 8.1s / 9.2s / 11.7s / 13.4s on an idle machine for a 400-token JSON reply
+ * (2026-08-16), against API-leg latencies under 2s. Callers that sized their
+ * timeout for the API leg (15–20s) were therefore killing the fallback
+ * mid-flight — save-enrich's 15s budget left 1.6s of headroom over the worst
+ * sample, which is how 6/57 (10.5%) of instrumented runs landed on
+ * reason:'llm-null' and why manual saves stopped getting search_aliases.
+ *
+ * Deliberately NOT applied as a floor inside callModelCLI / callHaikuCLI /
+ * callModelCLIAsync: those are also reached from latency-bound callers — the
+ * lesson bridge's 2.5s fail-open budget on the PreToolUse hook, and deep-search
+ * rerank on the MCP request path — where failing fast beats blocking a user for
+ * 45s. The allowance is caller-side policy, not a clamp.
+ * Pinned both ways by `tests/llm-timeout-budget.test.mjs`.
+ */
+export const BG_LLM_TIMEOUT_MS = 45000;
+
 /**
  * Resolve the LLM model to use for background calls.
  * Reads CLAUDE_MEM_MODEL env var, defaults to 'haiku'.
@@ -409,11 +431,20 @@ async function callModelAPI(prompt, model, { timeout, maxTokens, temperature = D
 function callModelCLI(prompt, model, { timeout }) {
   const modelName = MODEL_MAP[model] ? model : 'haiku';
   try {
-    const result = execFileSync(getClaudePath(), ['-p', '--model', modelName], {
+    // --no-session-persistence + DISABLE_CLAUDEMD_HOOKS (2026-08-16): these
+    // headless calls were paying the full interactive-session tax — 1,004
+    // transcripts piled up in ~/.claude/projects/-tmp/, and every spawn ran
+    // the claudemd plugin's whole hook fan-out (its SessionStart banner alone
+    // logged 682 rows in 3 days, drowning that project's telemetry). The
+    // persistence flag is OAuth-safe (probed); `--bare`/CLAUDE_CODE_SIMPLE are
+    // NOT (they hard-require ANTHROPIC_API_KEY — "Not logged in" on OAuth
+    // machines). The user's global CLAUDE.md injection has no OAuth-safe
+    // opt-out; accepted (haiku + prompt caching keeps it cheap).
+    const result = execFileSync(getClaudePath(), ['-p', '--model', modelName, '--no-session-persistence'], {
       input: flattenForCLI(prompt),
       timeout,
       encoding: 'utf8',
-      env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1' },
+      env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1', DISABLE_CLAUDEMD_HOOKS: '1' },
       stdio: ['pipe', 'pipe', 'pipe'],
       cwd: '/tmp',
     });
@@ -453,8 +484,9 @@ export function callModelCLIAsync(prompt, model, { timeout }) {
     const modelName = MODEL_MAP[model] ? model : 'haiku';
     let child;
     try {
-      child = spawn(getClaudePath(), ['-p', '--model', modelName], {
-        env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1' },
+      // Same headless-tax flags as callModelCLI (rationale there).
+      child = spawn(getClaudePath(), ['-p', '--model', modelName, '--no-session-persistence'], {
+        env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1', DISABLE_CLAUDEMD_HOOKS: '1' },
         cwd: '/tmp',
         stdio: ['pipe', 'pipe', 'pipe'],
       });
@@ -603,11 +635,12 @@ async function callOpenRouterAPI(prompt, tier, { timeout, maxTokens, temperature
 function callHaikuCLI(prompt, { timeout }) {
   const { cli: modelName } = resolveModel();
   try {
-    const result = execFileSync(getClaudePath(), ['-p', '--model', modelName], {
+    // Same headless-tax flags as callModelCLI (rationale there).
+    const result = execFileSync(getClaudePath(), ['-p', '--model', modelName, '--no-session-persistence'], {
       input: flattenForCLI(prompt),
       timeout,
       encoding: 'utf8',
-      env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1' },
+      env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1', DISABLE_CLAUDEMD_HOOKS: '1' },
       stdio: ['pipe', 'pipe', 'pipe'],
       cwd: '/tmp', // Prevent ghost sessions in user's /resume list
     });

package/hook-context.mjs

@@ -12,10 +12,12 @@ import {
   debugLog, debugCatch, neutralizeContextDelimiters,
   DECAY_HALF_LIFE_BY_TYPE, DEFAULT_DECAY_HALF_LIFE_MS, notLowSignalTitleClause,
 } from './utils.mjs';
-import { STALE_SESSION_MS, FALLBACK_OBS_WINDOW_MS, RUNTIME_DIR, effectiveQuiet, isQuietHooks } from './hook-shared.mjs';
+import { STALE_SESSION_MS, FALLBACK_OBS_WINDOW_MS, RUNTIME_DIR, effectiveQuiet, isQuietHooks, KEY_CONTEXT_LIMIT } from './hook-shared.mjs';
 import { extractUnfinishedSummary } from './hook-handoff.mjs';
 import { recentInjectableEvents, renderInjectableEvent } from './lib/events-injection.mjs';
+import { liveObsFilterSql } from './lib/inject-search-core.mjs';
 
+import { DAY_MS } from './lib/time-constants.mjs';
 /**
  * Infer the project directory from environment variables or cwd.
  * @returns {string} Absolute path to the project directory
@@ -42,7 +44,7 @@ function mdCell(s) {
 }
 
 export function computeAdaptiveWindows(db, project) {
-  const sevenDaysAgo = Date.now() - 7 * 86400000;
+  const sevenDaysAgo = Date.now() - 7 * DAY_MS;
   const row = db.prepare(`
     SELECT COUNT(*) as c FROM observations
     WHERE project = ? AND created_at_epoch > ? AND COALESCE(compressed_into, 0) = 0
@@ -51,13 +53,13 @@ export function computeAdaptiveWindows(db, project) {
 
   if (velocity > 10) {
     // High velocity: tighter windows, focus on very recent
-    return { tier1: 12 * 3600000, tier2: 3 * 86400000, tier3: 14 * 86400000, sessWindow: 3 * 86400000 };
+    return { tier1: 12 * 3600000, tier2: 3 * DAY_MS, tier3: 14 * DAY_MS, sessWindow: 3 * DAY_MS };
   } else if (velocity >= 3) {
     // Medium velocity: default windows
-    return { tier1: 24 * 3600000, tier2: 7 * 86400000, tier3: 30 * 86400000, sessWindow: 7 * 86400000 };
+    return { tier1: 24 * 3600000, tier2: 7 * DAY_MS, tier3: 30 * DAY_MS, sessWindow: 7 * DAY_MS };
   } else {
     // Low velocity: wider windows, older data still relevant
-    return { tier1: 48 * 3600000, tier2: 14 * 86400000, tier3: 60 * 86400000, sessWindow: 14 * 86400000 };
+    return { tier1: 48 * 3600000, tier2: 14 * DAY_MS, tier3: 60 * DAY_MS, sessWindow: 14 * DAY_MS };
   }
 }
 
@@ -83,12 +85,7 @@ export function selectWithTokenBudget(db, project, budget = 2000) {
   const obsPool = db.prepare(`
     SELECT id, type, title, narrative, importance, created_at_epoch, files_modified, lesson_learned
     FROM observations
-    WHERE project = ? AND COALESCE(compressed_into, 0) = 0
-      -- superseded invisibility: auto-dedup (hook.mjs) sets superseded_at but leaves
-      -- compressed_into=0, so the compressed filter alone lets the hidden near-duplicate
-      -- resurface in the most-visible surface (injected every SessionStart). Sibling
-      -- keyObs already filters this; obsPool + fallbackObs must match.
-      AND superseded_at IS NULL
+    WHERE project = ? AND ${liveObsFilterSql('')}
       AND ${notLowSignalTitleClause('')}
       AND (
         (created_at_epoch > ? AND importance >= 1)
@@ -293,9 +290,15 @@ export function cleanupClaudeMdLegacyBlock() {
  * @param {string|null} [currentCcSessionId=null] Claude Code session id — when provided,
  *   the "Working State (from /clear)" block is filtered to handoffs owned by this
  *   session, preventing parallel-session bleed (see docs/bug.txt).
+ * @param {object|null} [collector=null] Optional out-param: when given, its
+ *   `keyContextIds` property is set to the obs ids ACTUALLY rendered into the
+ *   File Lessons / Key Context sections ([] under quiet/adopted or when the
+ *   sections are empty). handleUserPrompt persists this as its exclude-set
+ *   (D#123: the exclude-set must mirror real injections, not the keyObs query).
  * @returns {string} Joined markdown lines (without <claude-mem-context> wrappers)
  */
-export function buildSessionContextLines(db, project, now = new Date(), currentCcSessionId = null) {
+export function buildSessionContextLines(db, project, now = new Date(), currentCcSessionId = null, collector = null) {
+  if (collector) collector.keyContextIds = [];
   // 1. Token-budgeted observation selection
   const selected = selectWithTokenBudget(db, project, 2000);
   const observations = selected.observations;
@@ -308,8 +311,7 @@ export function buildSessionContextLines(db, project, now = new Date(), currentC
     fallbackObs = db.prepare(`
       SELECT id, type, title, project, created_at
       FROM observations
-      WHERE COALESCE(compressed_into, 0) = 0
-        AND superseded_at IS NULL
+      WHERE ${liveObsFilterSql('')}
         AND ${notLowSignalTitleClause('')}
         AND (
           (created_at_epoch > ? AND importance >= 1)
@@ -335,10 +337,9 @@ export function buildSessionContextLines(db, project, now = new Date(), currentC
   //    and Key Context (informational). Pushed into summaryLines.
   const keyObs = db.prepare(`
     SELECT o.id, o.type, o.title, o.lesson_learned, o.files_modified FROM observations o
-    WHERE o.project = ? AND COALESCE(o.compressed_into, 0) = 0
-      AND o.superseded_at IS NULL
+    WHERE o.project = ? AND ${liveObsFilterSql('o')}
       AND COALESCE(o.importance, 1) >= 2
-    ORDER BY o.created_at_epoch DESC LIMIT 10
+    ORDER BY o.created_at_epoch DESC LIMIT ${KEY_CONTEXT_LIMIT}
   `).all(project);
 
   if (keyObs.length > 0) {
@@ -357,30 +358,35 @@ export function buildSessionContextLines(db, project, now = new Date(), currentC
           const files = JSON.parse(o.files_modified);
           const fname = basename(Array.isArray(files) && files.length > 0 ? files[0] : '');
           if (fname) {
-            fileLessons.push(`- ${fname}: ${truncate(o.lesson_learned, 100)} (#${o.id})`);
+            fileLessons.push({ id: o.id, line: `- ${fname}: ${truncate(o.lesson_learned, 100)} (#${o.id})` });
             continue;
           }
         } catch { /* fall through to keyContext */ }
       }
       const lesson = hasLesson ? ` — ${truncate(o.lesson_learned, 60)}` : '';
-      keyContext.push(`- [${o.type || 'discovery'}] ${truncate(clean, 80)} (#${o.id})${lesson}`);
+      keyContext.push({ id: o.id, line: `- [${o.type || 'discovery'}] ${truncate(clean, 80)} (#${o.id})${lesson}` });
     }
 
     // Phase A (QUIET_HOOKS) + Phase D (adopted sentinel): drop descriptive
     // File Lessons / Key Context sections when the user has opted into low-noise
     // hooks OR adopted invited-memory (MEMORY.md sentinel carries the triggers
     // at higher system-prompt authority). The Recent table still fires so #IDs
-    // remain reachable via mem_get.
+    // remain reachable via mem_get. The collector sees only rows that survive
+    // BOTH the quiet gate and the per-section slice — rendered rows, nothing else.
     const quiet = effectiveQuiet();
     if (fileLessons.length > 0 && !quiet) {
+      const shown = fileLessons.slice(0, 5);
       summaryLines.push('### File Lessons');
-      summaryLines.push(...fileLessons.slice(0, 5));
+      summaryLines.push(...shown.map((e) => e.line));
       summaryLines.push('');
+      if (collector) collector.keyContextIds.push(...shown.map((e) => e.id));
     }
     if (keyContext.length > 0 && !quiet) {
+      const shown = keyContext.slice(0, 5);
       summaryLines.push('### Key Context');
-      summaryLines.push(...keyContext.slice(0, 5));
+      summaryLines.push(...shown.map((e) => e.line));
       summaryLines.push('');
+      if (collector) collector.keyContextIds.push(...shown.map((e) => e.id));
     }
   } else if (!latestSummary && !effectiveQuiet()) {
     // Fallback: no summary AND no key observations — show recent activity.

package/hook-handoff.mjs

@@ -13,6 +13,7 @@ import {
 // immutable in ESM and cannot be mocked after the fact.
 import * as gitStateModule from './lib/git-state.mjs';
 import * as taskReaderModule from './lib/task-reader.mjs';
+import { liveObsFilterSql } from './lib/inject-search-core.mjs';
 
 /**
  * Build and save a handoff snapshot to session_handoffs table.
@@ -76,8 +77,7 @@ export function buildAndSaveHandoff(db, sessionId, project, type, episodeSnapsho
   if (subjectPrompts.length === 0) {
     const fallback = db.prepare(`
       SELECT title FROM observations
-      WHERE project = ? AND COALESCE(compressed_into, 0) = 0
-        AND superseded_at IS NULL
+      WHERE project = ? AND ${liveObsFilterSql('')}
         AND COALESCE(importance, 1) >= 3
         AND ${notLowSignalTitleClause('')}
       ORDER BY created_at_epoch DESC LIMIT 1
@@ -185,8 +185,7 @@ export function buildAndSaveHandoff(db, sessionId, project, type, episodeSnapsho
   const decisions = db.prepare(`
     SELECT title FROM observations
     WHERE memory_session_id = ? AND COALESCE(importance, 1) >= 2
-      AND COALESCE(compressed_into, 0) = 0
-      AND superseded_at IS NULL ${obsWindowClause}
+      AND ${liveObsFilterSql('')} ${obsWindowClause}
     ORDER BY created_at_epoch DESC LIMIT 10
   `).all(sessionId, ...obsWindowParams).filter(d => d.title && !LOW_SIGNAL_TITLE.test(d.title)).slice(0, 5);
 

package/hook-llm.mjs

@@ -10,6 +10,7 @@ import {
   getCurrentBranch, notLowSignalTitleClause,
 } from './utils.mjs';
 import { acquireLLMSlot, releaseLLMSlot } from './hook-semaphore.mjs';
+import { BG_LLM_TIMEOUT_MS } from './haiku-client.mjs';
 import { scrubRecord } from './lib/scrub-record.mjs';
 import { getVocabulary, computeVector, vecTextForRow } from './tfidf.mjs';
 import { insertObservationRow, insertObservationFiles, insertObservationVector, normalizeScope } from './lib/observation-write.mjs';
@@ -23,6 +24,7 @@ import { isNoiseObservation, capNoiseImportance, isLowYieldChangeObs } from './l
 import { episodeHasSignificantContent } from './hook-episode.mjs';
 import { OBS_TYPE_SET } from './lib/obs-types.mjs';
 
+import { DAY_MS } from './lib/time-constants.mjs';
 // T9: memdir-incompatible types live in the `events` table, not `observations`.
 // Set lookup is O(1) — authoritative source is lib/activity.mjs::EVENT_TYPES.
 const EVENT_TYPE_SET = new Set(EVENT_TYPES);
@@ -84,7 +86,7 @@ export function recordRetryAttempt(db, recovered, bucket = dateBucketUtc()) {
  * YYYY-MM-DD lexicographic order).
  */
 export function readRetryStats(db, days = 30) {
-  const cutoff = new Date(Date.now() - days * 86400000);
+  const cutoff = new Date(Date.now() - days * DAY_MS);
   return db.prepare(
     `SELECT date_bucket, attempts, recovered FROM lesson_retry_stats
      WHERE date_bucket >= ? ORDER BY date_bucket DESC`
@@ -205,8 +207,8 @@ export function saveObservation(obs, projectOverride, sessionIdOverride, externa
     // 3-day Jaccard catches near-duplicates without blocking legitimately new observations
     const LOW_SIGNAL = LOW_SIGNAL_TITLE;
     if (obs.title && LOW_SIGNAL.test(obs.title)) {
-      const sevenDaysAgo = now.getTime() - 7 * 86400000;
-      const threeDaysAgo = now.getTime() - 3 * 86400000;
+      const sevenDaysAgo = now.getTime() - 7 * DAY_MS;
+      const threeDaysAgo = now.getTime() - 3 * DAY_MS;
       // Phase 1: exact title match within 7 days
       const exactDup = db.prepare(`
         SELECT 1 FROM observations
@@ -856,7 +858,7 @@ ${actionList}`;
         const retrySlot = await acquireLLMSlot();
         try {
           const retryPrompt = buildLessonRetryPrompt(episode, parsed);
-          const retryRaw = retrySlot ? await callLLM(retryPrompt, 10000) : null;
+          const retryRaw = retrySlot ? await callLLM(retryPrompt, BG_LLM_TIMEOUT_MS) : null;
           if (retryRaw) {
             const retry = parseJsonFromLLM(retryRaw);
             const retryLesson = typeof retry?.lesson === 'string' ? retry.lesson.trim() : '';
@@ -1135,7 +1137,7 @@ ${obsList}`;
 
     let raw, llmParsed;
     try {
-      raw = await callLLM(prompt, 20000);
+      raw = await callLLM(prompt, BG_LLM_TIMEOUT_MS);
       llmParsed = parseJsonFromLLM(raw);
     } finally {
       releaseLLMSlot();

package/hook-memory.mjs

@@ -9,8 +9,9 @@ import { DB_DIR } from './schema.mjs';
 import { extractIdents } from './lib/lesson-idents.mjs';
 import { formatSubagentContext } from './lib/task-imperative.mjs';
 
+import { DAY_MS } from './lib/time-constants.mjs';
 const MAX_MEMORY_INJECTIONS = 3;
-const MEMORY_LOOKBACK_MS = 60 * 86400000; // 60 days
+const MEMORY_LOOKBACK_MS = 60 * DAY_MS; // 60 days
 // Aligned with TYPE_QUALITY_CASE in scoring-sql.mjs (R2 rebalance).
 // Weights calibrated to empirical avg access_count:
 //   decision 6.05, discovery 3.32, bugfix 2.24, feature 2.04, change 0.93, refactor 0.54.
@@ -98,7 +99,7 @@ function candidateCoverage(row, queryTerms) {
   return hits / queryTerms.length;
 }
 
-const FILE_RECALL_LOOKBACK_MS = 60 * 86400000; // 60 days
+const FILE_RECALL_LOOKBACK_MS = 60 * DAY_MS; // 60 days
 const MAX_FILE_RECALL = 2;
 
 // P1: stale-obs verify-before-use threshold. An injected obs older than this
@@ -107,7 +108,7 @@ const MAX_FILE_RECALL = 2;
 // renamed since capture. Pure-decision/architecture obs (no file_paths)
 // don't get the hint: their drift is text-only and Claude already verifies
 // at consumption time per the project mem-usage contract.
-const STALE_OBS_THRESHOLD_MS = 30 * 86400000;
+const STALE_OBS_THRESHOLD_MS = 30 * DAY_MS;
 
 /**
  * Format a single line for the <memory-context> block emitted by

package/hook-optimize.mjs

@@ -13,14 +13,16 @@ import {
   computeMinHash, estimateJaccardFromMinHash, jaccardSimilarity, clampImportance, cjkBigrams,
   notLowSignalTitleClause, scrubSecrets,
 } from './utils.mjs';
-import { callModelJSON } from './haiku-client.mjs';
+import { callModelJSON, BG_LLM_TIMEOUT_MS } from './haiku-client.mjs';
 import { acquireLLMSlot, releaseLLMSlot } from './hook-semaphore.mjs';
 import { scrubRecord } from './lib/scrub-record.mjs';
 import { getVocabulary, computeVector, cosineSimilarity, vecTextForRow } from './tfidf.mjs';
 import { MERGE_JACCARD_LOW, AUTO_MERGE_THRESHOLD } from './lib/dedup-constants.mjs';
 import { DB_DIR } from './schema.mjs';
 import { OBS_TYPE_SET } from './lib/obs-types.mjs';
+import { liveObsFilterSql } from './lib/inject-search-core.mjs';
 
+import { DAY_MS } from './lib/time-constants.mjs';
 const RUNTIME_DIR = join(DB_DIR, 'runtime');
 
 // ─── Budget ─────────────────────────────────────────────────────────────────
@@ -104,8 +106,7 @@ export function findReenrichCandidates(db, limit = 10, { scope = 'narrow', proje
     const stmt = db.prepare(`
       SELECT id, title, narrative, type, subtitle, concepts, facts, text, search_aliases, importance, project
       FROM observations
-      WHERE COALESCE(compressed_into, 0) = 0
-        AND superseded_at IS NULL
+      WHERE ${liveObsFilterSql('')}
         AND (search_aliases IS NULL OR search_aliases = '')
         AND LENGTH(COALESCE(narrative, '')) > 100
         AND ${notLowSignalTitleClause('')}
@@ -119,8 +120,7 @@ export function findReenrichCandidates(db, limit = 10, { scope = 'narrow', proje
     const stmt = db.prepare(`
       SELECT id, title, narrative, type, subtitle, concepts, facts, search_aliases, importance, project
       FROM observations
-      WHERE COALESCE(compressed_into, 0) = 0
-        AND superseded_at IS NULL
+      WHERE ${liveObsFilterSql('')}
         AND optimized_at IS NULL
         AND type IN ('bugfix','refactor','feature','decision')
         AND (lesson_learned IS NULL OR lesson_learned = '')
@@ -175,7 +175,7 @@ Narrative: ${truncate(cand.narrative || '(no narrative)', 500)}
 
 JSON: {"search_aliases":["alt phrasing","synonym","spelled-out jargon","CJK term if the domain word has one"]}
 Give 3-6 aliases: words a user might search for the SAME concept but that are NOT already in the title (synonyms, the spelled-out form of an acronym, the jargon term for a described symptom, a CJK translation of a key domain term).`;
-        const parsed = await callModelJSON(aliasPrompt, 'haiku', { timeout: 15000, maxTokens: 300 });
+        const parsed = await callModelJSON(aliasPrompt, 'haiku', { timeout: BG_LLM_TIMEOUT_MS, maxTokens: 300 });
         const aliasArr = parsed && Array.isArray(parsed.search_aliases)
           ? parsed.search_aliases.filter((a) => typeof a === 'string' && a.trim().length > 0)
           : [];
@@ -204,7 +204,7 @@ importance: 0=no value, 1=routine, 2=notable non-obvious insight, 3=critical. De
 lesson_learned: State what was learned. If routine, write "none".
 search_aliases: 2-6 alternative search terms (include CJK if applicable).`;
 
-      const parsed = await callModelJSON(prompt, 'haiku', { timeout: 15000, maxTokens: 500 });
+      const parsed = await callModelJSON(prompt, 'haiku', { timeout: BG_LLM_TIMEOUT_MS, maxTokens: 500 });
       if (!parsed || !parsed.title) { skipped++; continue; }
 
       // Auto-hide on importance:0 targets fully-degraded NARROW rows (this branch predates
@@ -289,7 +289,7 @@ search_aliases: 2-6 alternative search terms (include CJK if applicable).`;
 // ─── Task 2: Normalize ─────────────────────────────────────────────────────
 
 const NORMALIZE_GATE_FILE = join(RUNTIME_DIR, 'last-normalize.json');
-const NORMALIZE_INTERVAL_MS = 7 * 86400000; // 7 days
+const NORMALIZE_INTERVAL_MS = 7 * DAY_MS; // 7 days
 
 // Pure gate decision (no IO) — exported for testing. Fail-OPEN on a
 // malformed-but-valid-JSON gate: a missing/non-numeric `epoch` makes
@@ -356,7 +356,7 @@ Rules:
 - Include CJK ↔ English equivalents if present
 - Skip terms that have no synonyms in the list`;
 
-    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: 20000, maxTokens: 1000 });
+    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: BG_LLM_TIMEOUT_MS, maxTokens: 1000 });
     if (!parsed?.groups || !Array.isArray(parsed.groups)) return [];
     return parsed.groups.filter(g => g.canonical && Array.isArray(g.aliases) && g.aliases.length > 0);
   } catch (e) {
@@ -448,7 +448,7 @@ export async function executeNormalize(db, force = false, { project } = {}) {
 
 // ─── Task 3: Cluster-merge ─────────────────────────────────────────────────
 
-const MERGE_TIME_WINDOW_MS = 30 * 86400000;
+const MERGE_TIME_WINDOW_MS = 30 * DAY_MS;
 // Merge-review band [MERGE_JACCARD_LOW, AUTO_MERGE_THRESHOLD): titles in this
 // Jaccard range are LLM-reviewed for merge; at/above AUTO_MERGE_THRESHOLD they'd
 // already auto-merge elsewhere, below MERGE_JACCARD_LOW they're too dissimilar.
@@ -459,8 +459,7 @@ export function findMergeCandidates(db, maxClusters = 5, { project } = {}) {
   const stmt = db.prepare(`
     SELECT id, title, narrative, project, type, access_count, importance, created_at_epoch, minhash_sig, lesson_learned, concepts, facts
     FROM observations
-    WHERE COALESCE(compressed_into, 0) = 0
-      AND superseded_at IS NULL
+    WHERE ${liveObsFilterSql('')}
       AND optimized_at IS NULL
       AND title IS NOT NULL AND title != ''
       AND created_at_epoch > ?
@@ -525,7 +524,7 @@ Return ONLY valid JSON:
 - If they should NOT be merged: {"should_merge":false}
 - If they SHOULD be merged: {"should_merge":true,"merged_title":"≤120 char comprehensive title","merged_narrative":"comprehensive ≤800 char summary preserving all key details","merged_concepts":["kw1","kw2"],"merged_facts":["specific fact 1"],"merged_lesson":"synthesized non-obvious lesson or null","importance":2}`;
 
-    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: 20000, maxTokens: 1000 });
+    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: BG_LLM_TIMEOUT_MS, maxTokens: 1000 });
     if (!parsed || !parsed.should_merge) return { merged: false };
 
     // Keeper = highest importance, then highest access_count. Previously access_count
@@ -643,11 +642,11 @@ export async function executeClusterMerge(db, maxClusters = 5, { project } = {})
 
 // ─── Task 4: Smart-compress ────────────────────────────────────────────────
 
-const COMPRESS_TIME_SPLIT_MS = 14 * 86400000;
+const COMPRESS_TIME_SPLIT_MS = 14 * DAY_MS;
 const COMPRESS_COSINE_THRESHOLD = 0.3;
 
 export function findSmartCompressCandidates(db, ageDays = 30, { project } = {}) {
-  const cutoff = Date.now() - ageDays * 86400000;
+  const cutoff = Date.now() - ageDays * DAY_MS;
   const projectClause = project ? 'AND project = ?' : '';
   const stmt = db.prepare(`
     SELECT id, title, narrative, lesson_learned, project, type, created_at_epoch
@@ -759,7 +758,7 @@ ${obsDescriptions}
 
 JSON: {"title":"descriptive summary ≤120 chars","narrative":"comprehensive summary ≤800 chars preserving key decisions and lessons","concepts":["kw1","kw2"],"facts":["all specific facts preserved"],"lesson_learned":"most important synthesized lesson or 'none'","search_aliases":["alt search 1","alt search 2"]}`;
 
-    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: 20000, maxTokens: 1000 });
+    const parsed = await callModelJSON(prompt, 'sonnet', { timeout: BG_LLM_TIMEOUT_MS, maxTokens: 1000 });
     if (!parsed || !parsed.title) return { compressed: false };
 
     // Scrub BEFORE truncate (see re-enrich note): boundary cut on scrubbed text.

package/hook-precompact.mjs

@@ -7,9 +7,13 @@
 
 import { buildSessionContextLines } from './hook-context.mjs';
 import { inferProject, debugCatch, debugLog } from './utils.mjs';
+import { RUNTIME_DIR } from './hook-shared.mjs';
+import { recordKeyContextInjection } from './lib/keyctx-marker.mjs';
 
 /**
- * Build + emit the memory context block on stdout. Pure read; no DB writes.
+ * Build + emit the memory context block on stdout. Writes the Key Context ids
+ * the re-emitted block renders (refreshing handleUserPrompt's exclude-set — see
+ * D#123 in hook.mjs) and bumps injection_count on those rows (D#124).
  *
  * @param {object} ctx
  * @param {import('better-sqlite3').Database} ctx.db
@@ -19,9 +23,18 @@ import { inferProject, debugCatch, debugLog } from './utils.mjs';
  */
 export function handlePreCompact({ db, project, sessionId }) {
   try {
-    const body = buildSessionContextLines(db, project, new Date(), sessionId || null);
+    const collector = {};
+    const body = buildSessionContextLines(db, project, new Date(), sessionId || null, collector);
     if (!body || String(body).trim() === '') return;
     process.stdout.write(`<claude-mem-context>\n${body}\n</claude-mem-context>\n`);
+    // Same recorder as handleSessionStart: marker + injection_count bump (D#124).
+    // A re-render into a compacted context is a fresh injection of those rows.
+    recordKeyContextInjection(db, {
+      runtimeDir: RUNTIME_DIR,
+      project,
+      sessionId: sessionId || null,
+      ids: collector.keyContextIds || [],
+    });
   } catch (e) {
     debugCatch(e, 'handlePreCompact');
   }