claude-mem-lite 3.65.0 → 3.66.1

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "3.65.0",
+      "version": "3.66.1",
       "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": "3.65.0",
+  "version": "3.66.1",
   "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/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

@@ -17,6 +17,7 @@ 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
@@ -43,7 +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
@@ -52,13 +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 };
   }
 }
 

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

@@ -13,7 +13,7 @@ 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';
@@ -22,6 +22,7 @@ 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 ─────────────────────────────────────────────────────────────────
@@ -174,7 +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)
           : [];
@@ -203,7 +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
@@ -288,7 +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
@@ -355,7 +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) {
@@ -447,7 +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.
@@ -523,7 +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
@@ -641,11 +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
@@ -757,7 +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

@@ -5,17 +5,15 @@
 // Differs from SessionStart-on-compact (which fires AFTER compaction):
 // PreCompact ensures memory survives the compaction step itself.
 
-import { writeFileSync } from 'fs';
-import { join } from 'path';
 import { buildSessionContextLines } from './hook-context.mjs';
 import { inferProject, debugCatch, debugLog } from './utils.mjs';
 import { RUNTIME_DIR } from './hook-shared.mjs';
-import { keyContextIdsFileName } from './lib/injected-ids.mjs';
+import { recordKeyContextInjection } from './lib/keyctx-marker.mjs';
 
 /**
- * Build + emit the memory context block on stdout. Pure read; no DB writes
- * (one runtime marker file: the Key Context ids the re-emitted block renders,
- * refreshing handleUserPrompt's exclude-set — see D#123 in hook.mjs).
+ * 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
@@ -29,12 +27,14 @@ export function handlePreCompact({ db, project, sessionId }) {
     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`);
-    try {
-      writeFileSync(
-        join(RUNTIME_DIR, keyContextIdsFileName(project, sessionId || null)),
-        JSON.stringify({ ids: collector.keyContextIds || [], ts: Date.now(), session: sessionId || null }),
-      );
-    } catch (e) { debugCatch(e, 'pre-compact-keyctx-marker'); }
+    // 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');
   }

package/hook-shared.mjs

@@ -10,7 +10,7 @@ import { ensureDbWithWalRecovery, DB_DIR } from './schema.mjs';
 // Pure-`node:`/local module (it imports only binding-probe + native-binding-hint, and
 // neither imports this file) — no cycle.
 import { recordHookError } from './lib/hook-telemetry.mjs';
-import { getClaudePath as getClaudePathShared, resolveModel as resolveModelShared, flattenForCLI as _flattenForCLI, detectMode as detectLLMMode, callHaiku } from './haiku-client.mjs';
+import { getClaudePath as getClaudePathShared, resolveModel as resolveModelShared, flattenForCLI as _flattenForCLI, detectMode as detectLLMMode, callHaiku, BG_LLM_TIMEOUT_MS } from './haiku-client.mjs';
 // Phase D: invited-memory sentinel detection. memdir.mjs/claudemd.mjs only pull in
 // fs/path/os/crypto; adopt-content.mjs is pure strings. No circular deps —
 // neither imports hook-shared.
@@ -18,6 +18,7 @@ import { memdirPath as _memdirPath, isAdopted as _isAdoptedMemdir } from './memd
 import { isAdopted as _isAdoptedClaudeMd } from './claudemd.mjs';
 import { PLUGIN_SLUG as _PLUGIN_SLUG } from './adopt-content.mjs';
 
+import { DAY_MS } from './lib/time-constants.mjs';
 // ─── Constants ────────────────────────────────────────────────────────────────
 
 export const RUNTIME_DIR = join(DB_DIR, 'runtime');
@@ -30,7 +31,7 @@ export const SESSION_EXPIRY_MS = 12 * 60 * 60 * 1000;    // 12h
 export const STALE_SESSION_MS = 24 * 60 * 60 * 1000;     // 24h
 export const STALE_LOCK_MS = 30000;                       // 30s
 export const DEDUP_WINDOW_MS = 5 * 60 * 1000;            // 5 min (title dedup)
-export const RELATED_OBS_WINDOW_MS = 7 * 86400000;       // 7 days
+export const RELATED_OBS_WINDOW_MS = 7 * DAY_MS;       // 7 days
 export const FALLBACK_OBS_WINDOW_MS = RELATED_OBS_WINDOW_MS; // same window
 // Candidate rows the SessionStart Key Context surface considers (hook-context.mjs
 // keyObs; each of the two sections then renders at most 5). The user-prompt
@@ -123,6 +124,120 @@ export function sweepOrphanEpisodeFiles(runtimeDir, { ageMs = ORPHAN_EPISODE_AGE
   return count;
 }
 
+// ─── Per-project marker GC (P2-15) ───────────────────────────────────────────
+// RUNTIME_DIR had three sweeps and a hole. Per-SESSION files age out at 24h
+// (hook.mjs) and orphaned episode/read trackers at 1h/24h (above), but the
+// per-PROJECT markers — one file per project, written once, never revisited —
+// had no reclamation path at all. A live install on 2026-08-16 held 253 files,
+// 152 of them past 30 days, including entire families for test sandboxes
+// deleted months earlier (session-tmp--sdscc-e2e-*, cite-recall-scratchpad--
+// fixture-*) and a .skill-reco-cooldown-* family that nothing had ever swept.
+//
+// Deliberately a NAMED list rather than a wildcard: these markers share a shape
+// but not a meaning. The GC-able ones are caches — delete them and the next
+// session re-derives the state (or, for a cooldown, merely allows a suggestion
+// sooner). The preserved ones are records of a side effect already performed;
+// removing them re-arms it (.auto-adopt-* re-attempts a write into the user's
+// project CLAUDE.md, the migration sentinels re-run their one-time work), which
+// is a bad trade for the 13-45 bytes each occupies.
+export const STALE_PROJECT_MARKER_AGE_MS = 30 * 24 * 60 * 60 * 1000;
+
+// Regenerated on demand; safe to lose at any time.
+export const GC_PROJECT_MARKER_PREFIXES = Object.freeze([
+  'session-',                 // project → memory-session-id pointer
+  'cite-recall-',             // last session's cite-recall snapshot (nudge input)
+  '.skill-cooldown-',         // suggestion throttle timestamp
+  '.skill-reco-cooldown-',    // recommendation throttle timestamp
+]);
+
+// Records of a completed side effect — never age out. `ep-`/`ep-flush-`/
+// `pending-`/`reads-` are absent from BOTH lists on purpose: the first holds
+// unflushed observations (data, not cache) and the rest already belong to
+// sweepOrphanEpisodeFiles on tighter cutoffs.
+export const GC_PRESERVED_MARKER_PREFIXES = Object.freeze([
+  '.auto-adopt-',
+  '.deferred-block-migrated-',
+  '.legacy-claude-md-cleaned-',
+  // v3.66.1: these two shipped in the GC list for one release and had to come
+  // out. Both are version-keyed one-shot migration sentinels written by
+  // scripts/setup.sh, and their gate is `! -f <marker>` — deleting one re-runs
+  // its migration. `.mcp-dedup-v2.78` gates a block that removes
+  // mcpServers.mem / mcpServers["mem-lite"] from the user's ~/.claude.json with
+  // a raw writeFileSync (no tmp+rename, no backup), which the repo's own test
+  // documents as intentionally one-shot: "If a user later runs `claude mcp add
+  // mem ...` themselves, the gate intentionally lets it stand." A 30-day sweep
+  // turned that into a recurring purge of a config file we do not own.
+  // The mtime never refreshes (the gate skips the block once the file exists),
+  // so every install older than 30 days would have lost it on the first
+  // SessionStart after upgrading.
+  //
+  // Why it was missed: the search for writers used `grep --include=*.mjs
+  // --include=*.js`, and the writer is a SHELL script. `sentinelPrefixesFromShell`
+  // below now derives this class from scripts/*.sh instead of from memory.
+  '.mcp-dedup-',
+  '.residue-warned-',
+]);
+
+/**
+ * Marker-name prefixes that scripts/*.sh treats as one-shot sentinels, derived
+ * from the shell source rather than restated here. `tests/runtime-marker-gc`
+ * asserts none of them is GC-able: a shell-written sentinel is invisible to a
+ * JS-only grep, which is exactly how `.mcp-dedup-` reached the GC list.
+ *
+ * @param {string} shellSource concatenated contents of scripts/*.sh
+ * @returns {string[]} prefixes like `.mcp-dedup-`
+ */
+export function sentinelPrefixesFromShell(shellSource) {
+  const out = new Set();
+  // Matches `"$DATA_DIR/runtime/.mcp-dedup-v2.78"` and friends: a dotfile under
+  // runtime/ whose name carries a version-ish suffix.
+  for (const m of String(shellSource || '').matchAll(/runtime\/(\.[a-z0-9-]*?-)v?[0-9][0-9.]*/gi)) {
+    out.add(m[1]);
+  }
+  return [...out];
+}
+
+/**
+ * Sweep per-project runtime markers older than `ageMs`. fs-only, best-effort,
+ * never throws. Returns the number of files removed.
+ *
+ * The two prefix lists are injectable ONLY so the precedence rule below can be
+ * exercised: with the shipped lists they are disjoint, which makes the
+ * preserved check redundant today and load-bearing the moment a future family
+ * nests inside a GC-able one. Production callers pass neither.
+ *
+ * @param {string} runtimeDir
+ * @param {{ageMs?: number, now?: number, gcPrefixes?: string[], preservedPrefixes?: string[]}} [opts]
+ * @returns {number}
+ */
+export function sweepStaleProjectMarkers(runtimeDir, {
+  ageMs = STALE_PROJECT_MARKER_AGE_MS,
+  now = Date.now(),
+  gcPrefixes = GC_PROJECT_MARKER_PREFIXES,
+  preservedPrefixes = GC_PRESERVED_MARKER_PREFIXES,
+  env = process.env,
+} = {}) {
+  // Kill switch (naming mirrors SKIP_COMPRESS / SKIP_OPTIMIZE / SKIP_SAVE_ENRICH):
+  // this is the only sweep that deletes files a user might want to inspect, so a
+  // released default that reclaims state needs a documented way back out.
+  if (env.CLAUDE_MEM_SKIP_MARKER_GC === '1') return 0;
+  let entries;
+  try { entries = readdirSync(runtimeDir); } catch { return 0; }
+  const cutoff = now - ageMs;
+  let count = 0;
+  for (const f of entries) {
+    // Preserved wins on any overlap, so a future prefix added to both lists
+    // fails safe (kept) instead of deleting a side-effect record.
+    if (preservedPrefixes.some((p) => f.startsWith(p))) continue;
+    if (!gcPrefixes.some((p) => f.startsWith(p))) continue;
+    const full = join(runtimeDir, f);
+    try {
+      if (statSync(full).mtimeMs < cutoff) { unlinkSync(full); count++; }
+    } catch { /* concurrent unlink / permission / directory — ignore */ }
+  }
+  return count;
+}
+
 // Ensure runtime directory exists AND is owner-only (0700), matching the DB dir
 // (schema.mjs). Runtime aux files carry captured file paths + scrubbed activity; on a
 // shared host a 0755 dir would let another local user read them. hardenRuntimeFiles()
@@ -195,7 +310,7 @@ export function openDb() {
 // response string (callers run parseJsonFromLLM themselves) or null.
 // maxTokens is sized for session-summary / episode JSON (larger than the
 // registry/optimize callers' budgets).
-export async function callLLM(prompt, timeoutMs = 15000) {
+export async function callLLM(prompt, timeoutMs = BG_LLM_TIMEOUT_MS) {
   if (detectLLMMode() !== 'cli') {
     const result = await callHaiku(prompt, { timeout: timeoutMs, maxTokens: 2000 });
     return result?.text ?? null;
@@ -203,11 +318,13 @@ export async function callLLM(prompt, timeoutMs = 15000) {
 
   const { cli: modelName } = resolveModelShared();
   try {
-    const result = execFileSync(getClaudePathShared(), ['-p', '--model', modelName], {
+    // Same headless-tax flags as haiku-client.mjs#callModelCLI (rationale
+    // there): no transcript persistence, no claudemd hook fan-out.
+    const result = execFileSync(getClaudePathShared(), ['-p', '--model', modelName, '--no-session-persistence'], {
       input: _flattenForCLI(prompt),
       timeout: timeoutMs,
       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
     });