clementine-agent 1.13.3 → 1.15.0

package/README.md

@@ -307,9 +307,6 @@ WORKSPACE_DIRS=~/projects,~/work
 
 # Security
 ALLOW_ALL_USERS=false      # true = skip owner checks
-
-# Beta Features
-ENABLE_1M_CONTEXT=false    # Enable 1M token context for Sonnet (toggle in dashboard)
 ```
 
 Secrets can also be stored in macOS Keychain (`security find-generic-password`) — Clementine checks Keychain as a fallback for any missing `.env` value.
@@ -336,7 +333,6 @@ Your overrides live in `~/.clementine/.env` — **they survive every `npm update
 | `BUDGET_CRON_T2_USD` | `5.00` | Max spend per tier-2 cron job |
 | `BUDGET_HEARTBEAT_USD` | `0.50` | Max spend per heartbeat tick |
 | `DEFAULT_MODEL_TIER` | `sonnet` | Default model: `haiku` / `sonnet` / `opus` |
-| `ENABLE_1M_CONTEXT` | `false` | Enable Sonnet 1M-token context (beta) |
 | `HEARTBEAT_INTERVAL_MINUTES` | `30` | How often the agent auto-checks in |
 | `HEARTBEAT_ACTIVE_START` | `8` | First hour of the active window (0–23) |
 | `HEARTBEAT_ACTIVE_END` | `22` | Last hour of the active window |

package/dist/agent/assistant.d.ts

@@ -247,7 +247,9 @@ export declare class PersonalAssistant {
         };
         delegateProfile?: AgentProfile;
     }): Promise<string>;
-    runCronJob(jobName: string, jobPrompt: string, tier?: number, maxTurns?: number, model?: string, workDir?: string, timeoutMs?: number, successCriteria?: string[], agentSlug?: string): Promise<string>;
+    runCronJob(jobName: string, jobPrompt: string, tier?: number, maxTurns?: number, model?: string, workDir?: string, timeoutMs?: number, successCriteria?: string[], agentSlug?: string, opts?: {
+        disableAllTools?: boolean;
+    }): Promise<string>;
     /**
      * Goal-backward verification pass using Haiku after cron job execution.
      * Instead of vague quality ratings, verifies actual outcomes:

package/dist/agent/assistant.js

@@ -13,7 +13,7 @@ import fs from 'node:fs';
 import path from 'node:path';
 import { query as rawQuery, listSubagents, getSubagentMessages, SYSTEM_PROMPT_DYNAMIC_BOUNDARY, } from '@anthropic-ai/claude-agent-sdk';
 import pino from 'pino';
-import { BASE_DIR, PKG_DIR, VAULT_DIR, DAILY_NOTES_DIR, SOUL_FILE, AGENTS_FILE, MEMORY_FILE, AGENTS_DIR, ASSISTANT_NAME, OWNER_NAME, MODEL, MODELS, HEARTBEAT_MAX_TURNS, SEARCH_CONTEXT_LIMIT, SEARCH_RECENCY_LIMIT, SYSTEM_PROMPT_MAX_CONTEXT_CHARS, SESSION_EXCHANGE_HISTORY_SIZE, SESSION_EXCHANGE_MAX_CHARS, INJECTED_CONTEXT_MAX_CHARS, UNLEASHED_PHASE_TURNS, UNLEASHED_DEFAULT_MAX_HOURS, UNLEASHED_MAX_PHASES, PROJECTS_META_FILE, CRON_PROGRESS_DIR, CRON_REFLECTIONS_DIR, HANDOFFS_DIR, BUDGET, TASK_BUDGET_TOKENS, ENABLE_1M_CONTEXT, IDENTITY_FILE, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_API_KEY as CONFIG_ANTHROPIC_API_KEY, } from '../config.js';
+import { BASE_DIR, PKG_DIR, VAULT_DIR, DAILY_NOTES_DIR, SOUL_FILE, AGENTS_FILE, MEMORY_FILE, AGENTS_DIR, ASSISTANT_NAME, OWNER_NAME, MODEL, MODELS, HEARTBEAT_MAX_TURNS, SEARCH_CONTEXT_LIMIT, SEARCH_RECENCY_LIMIT, SYSTEM_PROMPT_MAX_CONTEXT_CHARS, SESSION_EXCHANGE_HISTORY_SIZE, SESSION_EXCHANGE_MAX_CHARS, INJECTED_CONTEXT_MAX_CHARS, UNLEASHED_PHASE_TURNS, UNLEASHED_DEFAULT_MAX_HOURS, UNLEASHED_MAX_PHASES, PROJECTS_META_FILE, CRON_PROGRESS_DIR, CRON_REFLECTIONS_DIR, HANDOFFS_DIR, BUDGET, TASK_BUDGET_TOKENS, IDENTITY_FILE, CLAUDE_CODE_OAUTH_TOKEN, ANTHROPIC_API_KEY as CONFIG_ANTHROPIC_API_KEY, } from '../config.js';
 import { DEFAULT_CHANNEL_CAPABILITIES } from '../types.js';
 import { enforceToolPermissions, getSecurityPrompt, getHeartbeatSecurityPrompt, getCronSecurityPrompt, getHeartbeatDisallowedTools, logToolUse, setProfileTier, setProfileAllowedTools, setAgentDir, setSendPolicy, setInteractionSource, logAuditJsonl, } from './hooks.js';
 import { scanner } from '../security/scanner.js';
@@ -1958,11 +1958,6 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
         // path that wants to log "soft target" values, but it is intentionally
         // never passed into sdkOptions.
         const supportsTaskBudget = false;
-        // 1M context beta: enable for Sonnet when toggled and context-heavy work benefits
-        const isSonnet = resolvedModel.includes('sonnet');
-        const computedBetas = ENABLE_1M_CONTEXT && isSonnet
-            ? ['context-1m-2025-08-07']
-            : undefined;
         // Merge external MCP servers (Claude Desktop, Claude Code, user-managed).
         // Skip when tools are disabled (no point connecting to servers we won't use)
         // or for internal plan steps that only need Clementine's own tools.
@@ -2036,7 +2031,6 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             ...(computedEffort ? { effort: computedEffort } : {}),
             // maxBudgetUsd intentionally omitted — see comment above.
             ...(computedThinking ? { thinking: computedThinking } : {}),
-            ...(computedBetas ? { betas: computedBetas } : {}),
             ...(outputFormat ? { outputFormat } : {}),
             canUseTool: async (toolName, toolInput, _options) => {
                 // Per-query stall guard (no global state — scoped to this query)
@@ -3897,9 +3891,14 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
     // ── Heartbeat / Cron ──────────────────────────────────────────────
     async heartbeat(standingInstructions, changesSummary = '', timeContext = '', dedupContext = '', profile) {
         setInteractionSource('autonomous');
+        // Heartbeat speaks text only — the prompt below explicitly forbids tool
+        // calls. Skipping MCP server load + tool inventory cuts the prompt by
+        // hundreds of thousands of tokens on installs with many integrations,
+        // which is what kept Haiku exceeding its 200K context window.
         const sdkOptions = await this.buildOptions({
             isHeartbeat: true,
             enableTeams: false,
+            disableAllTools: true,
             model: MODELS.haiku,
             profile: profile ?? undefined,
         });
@@ -4006,7 +4005,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
         return extractDeliverable(trace) ||
             trace.filter(t => t.type === 'text').map(t => t.content).join('').trim();
     }
-    async runCronJob(jobName, jobPrompt, tier = 1, maxTurns, model, workDir, timeoutMs, successCriteria, agentSlug) {
+    async runCronJob(jobName, jobPrompt, tier = 1, maxTurns, model, workDir, timeoutMs, successCriteria, agentSlug, opts) {
         setInteractionSource('autonomous');
         // Tag every tool_use audit event with the cron job name + agent so
         // analytics tool-usage can show "Bash×893 driven by market-leader-followup"
@@ -4040,6 +4039,7 @@ You have a cost budget per message — not a hard turn limit. Work until the tas
             enableTeams: true,
             stallGuard: cronGuard,
             profile: cronProfile,
+            disableAllTools: opts?.disableAllTools ?? false,
         });
         // Override cwd if a project workDir is specified
         if (workDir) {

package/dist/agent/route-classifier.js

@@ -318,7 +318,14 @@ export async function classifyRoute(userMessage, agents, gateway) {
     try {
         raw = await gateway.handleCronJob('route-classify', prompt, 1, // tier 1
         3, // maxTurns — classifier doesn't need tools
-        'haiku');
+        'haiku', // cheap
+        undefined, // workDir
+        'standard', // mode
+        undefined, // maxHours
+        undefined, // timeoutMs
+        undefined, // successCriteria
+        undefined, // agentSlug
+        { disableAllTools: true });
     }
     catch (err) {
         logger.warn({ err }, 'Route classifier call failed');

package/dist/channels/discord.js

@@ -11,7 +11,7 @@ import { existsSync, readFileSync, readdirSync, statSync } from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
 import { chunkText, sendChunked, DiscordStreamingMessage, friendlyToolName, formatCronEmbed, rehydrateStatusEmbed, setSavedStatusEmbed, } from './discord-utils.js';
-import { DISCORD_TOKEN, DISCORD_OWNER_ID, DISCORD_WATCHED_CHANNELS, MODELS, ASSISTANT_NAME, OWNER_NAME, PKG_DIR, VAULT_DIR, BASE_DIR, DEFAULT_MODEL_TIER, ENABLE_1M_CONTEXT, } from '../config.js';
+import { DISCORD_TOKEN, DISCORD_OWNER_ID, DISCORD_WATCHED_CHANNELS, MODELS, ASSISTANT_NAME, OWNER_NAME, PKG_DIR, VAULT_DIR, BASE_DIR, DEFAULT_MODEL_TIER, } from '../config.js';
 import { findProjectByName, getLinkedProjects } from '../agent/assistant.js';
 import * as cronParser from 'cron-parser';
 const logger = pino({ name: 'clementine.discord' });
@@ -535,8 +535,7 @@ export async function startDiscord(gateway, heartbeat, cronScheduler, dispatcher
         embed.addFields({ name: '\u{1F4CB} Scheduled', value: schedSummary, inline: true });
         // ── System info ──────────────────────────────────────────────
         const modelLabel = DEFAULT_MODEL_TIER.charAt(0).toUpperCase() + DEFAULT_MODEL_TIER.slice(1);
-        const contextTag = ENABLE_1M_CONTEXT ? ' \u00b7 1M context' : '';
-        embed.addFields({ name: '\u{2699}\u{FE0F} System', value: `${modelLabel}${contextTag}`, inline: true });
+        embed.addFields({ name: '\u{2699}\u{FE0F} System', value: modelLabel, inline: true });
         return embed;
     }
     /** Format a duration in minutes to a compact human string. */

package/dist/cli/dashboard.js

@@ -234,8 +234,7 @@ function startDaemonWatcher(broadcastFn) {
         lastKnownDaemonPid = currentPid;
     }, 5000);
 }
-// ── Memory search (direct DB access, read-only) ─────────────────────
-async function searchMemory(query, limit = 20) {
+async function searchMemory(query, limit = 20, filters = {}) {
     if (!existsSync(MEMORY_DB_PATH)) {
         return { results: [], dbExists: false, error: `Memory DB not found at ${MEMORY_DB_PATH}` };
     }
@@ -243,20 +242,39 @@ async function searchMemory(query, limit = 20) {
     const db = new Database(MEMORY_DB_PATH, { readonly: true });
     try {
         const words = query.split(/\s+/).filter((w) => w.length > 0);
-        if (words.length === 0) {
+        const where = ['sd.chunk_id IS NULL'];
+        const params = [];
+        let orderBy = 'c.updated_at DESC';
+        let fromClause = 'chunks c LEFT JOIN chunk_soft_deletes sd ON sd.chunk_id = c.id';
+        if (words.length > 0) {
+            const ftsQuery = words.map((w) => `"${w.replace(/"/g, '')}"`).join(' OR ');
+            fromClause = 'chunks_fts f JOIN chunks c ON c.id = f.rowid LEFT JOIN chunk_soft_deletes sd ON sd.chunk_id = c.id';
+            where.unshift('chunks_fts MATCH ?');
+            params.push(ftsQuery);
+            orderBy = 'bm25(chunks_fts)';
+        }
+        if (filters.chunkType) {
+            where.push('c.chunk_type = ?');
+            params.push(filters.chunkType);
+        }
+        if (filters.sinceDays && filters.sinceDays > 0) {
+            where.push("c.updated_at >= datetime('now', ?)");
+            params.push(`-${filters.sinceDays} days`);
+        }
+        if (filters.pinnedOnly) {
+            where.push('c.pinned = 1');
+        }
+        if (words.length === 0 && !filters.chunkType && !filters.sinceDays && !filters.pinnedOnly) {
             db.close();
             return { results: [], dbExists: true };
         }
-        const ftsQuery = words.map((w) => `"${w.replace(/"/g, '')}"`).join(' OR ');
-        const rows = db.prepare(`SELECT c.id, c.source_file, c.section, c.content, c.chunk_type,
-              c.updated_at, c.salience, c.pinned, bm25(chunks_fts) as score
-       FROM chunks_fts f
-       JOIN chunks c ON c.id = f.rowid
-       LEFT JOIN chunk_soft_deletes sd ON sd.chunk_id = c.id
-       WHERE chunks_fts MATCH ?
-         AND sd.chunk_id IS NULL
-       ORDER BY bm25(chunks_fts)
-       LIMIT ?`).all(ftsQuery, limit);
+        const sql = `SELECT c.id, c.source_file, c.section, c.content, c.chunk_type,
+              c.updated_at, c.salience, c.pinned${words.length > 0 ? ', bm25(chunks_fts) as score' : ', 0 as score'}
+       FROM ${fromClause}
+       WHERE ${where.join(' AND ')}
+       ORDER BY ${orderBy}
+       LIMIT ?`;
+        const rows = db.prepare(sql).all(...params, limit);
         return { results: rows, dbExists: true };
     }
     catch (err) {
@@ -5142,7 +5160,6 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
             label: 'Model',
             keys: [
                 { key: 'DEFAULT_MODEL_TIER', label: 'Default Tier', hint: 'haiku, sonnet, or opus', type: 'select:haiku,sonnet,opus' },
-                { key: 'ENABLE_1M_CONTEXT', label: '1M Context', hint: 'Enable 1M token context window for Sonnet (beta)', type: 'toggle' },
             ],
         },
         {
@@ -5340,10 +5357,6 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
             }
             writeEnvValue(key, value);
             // Apply runtime-hot settings immediately (no restart needed)
-            if (key === 'ENABLE_1M_CONTEXT') {
-                const { setEnable1MContext } = await import('../config.js');
-                setEnable1MContext(value.toLowerCase() === 'true');
-            }
             // Composio: mutate process.env in-place + drop singleton so the very
             // next /api/composio/* call picks up the new key without a daemon
             // restart. Without this, "Save key → Connect Gmail" would 503 until
@@ -5747,6 +5760,57 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
             res.status(500).json({ error: String(err) });
         }
     });
+    // Quick-add: append a sentence to today's daily note from the dashboard.
+    // Mirrors the agent's memory_write({action:'append_daily'}) path so the
+    // note gets indexed identically.
+    app.post('/api/memory/quick-add', async (req, res) => {
+        try {
+            const content = String(req.body?.content ?? '').trim();
+            const section = (req.body?.section ? String(req.body.section) : 'Interactions').trim() || 'Interactions';
+            const salienceHint = req.body?.salience_hint != null ? Number(req.body.salience_hint) : undefined;
+            if (!content) {
+                res.status(400).json({ error: 'content required' });
+                return;
+            }
+            if (content.length > 4000) {
+                res.status(400).json({ error: 'content too long (max 4000 chars)' });
+                return;
+            }
+            const sharedMod = await import('../tools/shared.js');
+            const dailyPath = sharedMod.ensureDailyNote();
+            const timestamp = sharedMod.nowTime();
+            let body = readFileSync(dailyPath, 'utf-8');
+            const entry = `\n- **${timestamp}** — ${content}`;
+            const pattern = new RegExp(`(## ${section.replace(/[.*+?^${}()|[\\]\\\\]/g, '\\$&')}.*?)(\\n## |$)`, 's');
+            const match = pattern.exec(body);
+            if (match) {
+                body = body.slice(0, match.index + match[1].length) + entry + body.slice(match.index + match[1].length);
+            }
+            else {
+                body += `\n\n## ${section}${entry}`;
+            }
+            writeFileSync(dailyPath, body, 'utf-8');
+            const rel = path.relative(VAULT_DIR, dailyPath);
+            await sharedMod.incrementalSync(rel);
+            try {
+                const store = await sharedMod.getStore();
+                if (typeof salienceHint === 'number' && Number.isFinite(salienceHint)) {
+                    store.applyWriteSalience(rel, section, salienceHint);
+                }
+                store.logExtraction({
+                    sessionKey: 'dashboard:quick-add', userMessage: content.slice(0, 200),
+                    toolName: 'memory_write',
+                    toolInput: JSON.stringify({ action: 'append_daily', section, salience_hint: salienceHint, source: 'dashboard' }),
+                    extractedAt: new Date().toISOString(), status: 'active',
+                });
+            }
+            catch { /* observability best-effort */ }
+            res.json({ ok: true, file: rel, section });
+        }
+        catch (err) {
+            res.status(500).json({ error: String(err) });
+        }
+    });
     app.post('/api/memory/health/action', async (req, res) => {
         try {
             const action = (req.body?.action ?? '');
@@ -6250,12 +6314,15 @@ If the tool returns nothing or errors, return an empty array \`[]\`.`,
     // ── Memory search route ───────────────────────────────────────────
     app.get('/api/memory/search', async (req, res) => {
         const q = String(req.query.q ?? '');
-        if (!q.trim()) {
+        const chunkType = req.query.type ? String(req.query.type) : undefined;
+        const sinceDays = req.query.since ? Number(req.query.since) : undefined;
+        const pinnedOnly = String(req.query.pinned ?? '') === 'true';
+        if (!q.trim() && !chunkType && !sinceDays && !pinnedOnly) {
             res.json({ results: [] });
             return;
         }
         try {
-            const data = await searchMemory(q, 20);
+            const data = await searchMemory(q, 20, { chunkType, sinceDays, pinnedOnly });
             // Enrich with graph relationships for entities found in results
             let graphContext = [];
             try {
@@ -12918,9 +12985,38 @@ if('serviceWorker' in navigator){navigator.serviceWorker.getRegistrations().then
           <h1>Brain</h1>
           <p class="desc">Query what you know, feed new knowledge in, and watch the system learn.</p>
         </div>
-        <div class="actions" style="flex:1;max-width:480px;display:flex;gap:8px">
+        <div class="actions" style="flex:1;max-width:560px;display:flex;gap:8px">
           <input type="text" id="memory-search-input" placeholder="Search vault, notes, memory..." style="flex:1;padding:6px 10px;border:1px solid var(--border);border-radius:6px;background:var(--bg-input);color:var(--text-primary);font-size:13px" onkeydown="if(event.key==='Enter')runMemorySearch()">
           <button class="btn-primary btn-sm" onclick="runMemorySearch()">Search</button>
+          <button class="btn-sm" onclick="openQuickAddMemory()" title="Append a quick note to today's daily log">+ Add memory</button>
+        </div>
+      </div>
+
+      <!-- Quick-add memory modal: type a sentence, hit save. Posts to memory_write
+           append_daily so the note lands in today's vault log and gets indexed. -->
+      <div id="quick-add-memory-modal" style="display:none;position:fixed;inset:0;background:rgba(0,0,0,0.5);z-index:1000;align-items:center;justify-content:center" onclick="if(event.target===this)closeQuickAddMemory()">
+        <div style="background:var(--bg-primary);border:1px solid var(--border);border-radius:10px;padding:20px;width:520px;max-width:90vw">
+          <div style="display:flex;align-items:center;justify-content:space-between;margin-bottom:14px">
+            <h3 style="margin:0;font-size:16px">Add a memory</h3>
+            <button class="btn-icon btn-sm" onclick="closeQuickAddMemory()" title="Close">×</button>
+          </div>
+          <div style="font-size:12px;color:var(--text-muted);margin-bottom:10px">
+            Appends to today's daily note. The agent will see this on its next search.
+          </div>
+          <textarea id="quick-add-memory-text" placeholder="What should Clementine remember?" rows="5" style="width:100%;padding:10px;border:1px solid var(--border);border-radius:6px;background:var(--bg-input);color:var(--text-primary);font-size:13px;font-family:inherit;resize:vertical" autofocus></textarea>
+          <div style="display:flex;gap:8px;align-items:center;margin-top:10px">
+            <label style="font-size:12px;color:var(--text-muted)">Salience:</label>
+            <select id="quick-add-memory-salience" style="padding:5px 8px;border:1px solid var(--border);border-radius:6px;background:var(--bg-input);color:var(--text-primary);font-size:12px">
+              <option value="0.5">0.5 — tentative</option>
+              <option value="1.0" selected>1.0 — normal</option>
+              <option value="1.5">1.5 — durable</option>
+              <option value="2.0">2.0 — identity-level</option>
+            </select>
+            <span style="flex:1"></span>
+            <span id="quick-add-memory-status" style="font-size:12px;color:var(--text-muted)"></span>
+            <button class="btn-sm" onclick="closeQuickAddMemory()">Cancel</button>
+            <button class="btn-primary btn-sm" onclick="submitQuickAddMemory()">Save</button>
+          </div>
         </div>
       </div>
       <div class="tab-bar" id="intelligence-tabs" style="margin:0 0 0 18px">
@@ -12931,13 +13027,41 @@ if('serviceWorker' in navigator){navigator.serviceWorker.getRegistrations().then
         <button data-icon="zap" onclick="switchTab('intelligence','health')"><span class="icon-slot"></span> Health <span class="tab-badge" id="brain-health-badge" style="display:none;background:#ef4444;color:#fff">0</span></button>
         <button data-icon="users" onclick="switchTab('intelligence','user-model')"><span class="icon-slot"></span> User Model</button>
         <button data-icon="brain" onclick="switchTab('intelligence','learning')"><span class="icon-slot"></span> Learning <span class="tab-badge" id="brain-learning-badge" style="display:none;background:#f59e0b;color:#000">0</span></button>
-        <button onclick="switchTab('intelligence','memory')">Stats</button>
         <button onclick="switchTab('intelligence','seed')">Seed</button>
         <button onclick="switchTab('intelligence','runs')">Runs</button>
       </div>
       <div id="intelligence-tab-content">
         <div class="tab-pane active" id="tab-intelligence-search">
+          <div id="memory-coverage-strip" style="margin-bottom:14px"></div>
           <div id="memory-search-results"></div>
+          <div id="memory-overview" style="margin-top:18px">
+            <div class="grid-2" id="memory-stats" style="margin-bottom:14px"></div>
+            <div class="card" style="margin-bottom:14px">
+              <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
+                <span>MEMORY.md</span>
+                <span style="font-size:11px;color:var(--text-muted)">Curated facts loaded into every session</span>
+              </div>
+              <div class="card-body" id="panel-memory"><div class="empty-state">Loading...</div></div>
+            </div>
+            <div class="card" style="margin-bottom:14px">
+              <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
+                <span>Recent writes</span>
+                <span style="font-size:11px;color:var(--text-muted)">What the agent captured, with reason &amp; salience</span>
+              </div>
+              <div class="card-body" id="panel-recent-writes" style="padding:0">
+                <div class="skel-block" style="padding:14px"><div class="skel-row med"></div><div class="skel-row short"></div></div>
+              </div>
+            </div>
+            <div class="card">
+              <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
+                <span>Self-correction (supersedes)</span>
+                <span style="font-size:11px;color:var(--text-muted)">Old facts the agent has explicitly replaced</span>
+              </div>
+              <div class="card-body" id="panel-supersedes" style="padding:0">
+                <div class="skel-block" style="padding:14px"><div class="skel-row med"></div><div class="skel-row short"></div></div>
+              </div>
+            </div>
+          </div>
         </div>
         <div class="tab-pane" id="tab-intelligence-graph">
           <div style="display:flex;gap:8px;margin-bottom:12px;flex-wrap:wrap">
@@ -12956,18 +13080,6 @@ if('serviceWorker' in navigator){navigator.serviceWorker.getRegistrations().then
           <div id="graph-legend" style="display:flex;gap:16px;margin-top:8px;flex-wrap:wrap"></div>
           <div id="graph-detail-panel" style="margin-top:12px"></div>
         </div>
-        <div class="tab-pane" id="tab-intelligence-memory">
-          <div style="margin-bottom:12px;font-size:13px;color:var(--text-muted)">
-            Stats and content browsing. For janitor, integrity, write queue, and staleness diagnostics see
-            <a href="#" onclick="navigateTo('memory-health');return false" style="color:var(--accent)">Memory Health &rarr;</a>
-          </div>
-          <div class="grid-2" id="memory-stats"></div>
-          <div class="card">
-            <div class="card-header">MEMORY.md</div>
-            <div class="card-body" id="panel-memory"><div class="empty-state">Loading...</div></div>
-          </div>
-        </div>
-
         <!-- User Model — MemGPT-style core memory blocks always loaded into context -->
         <div class="tab-pane" id="tab-intelligence-user-model">
           <div style="color:var(--muted,#888);margin-bottom:12px;font-size:13px;max-width:760px">
@@ -13165,24 +13277,6 @@ if('serviceWorker' in navigator){navigator.serviceWorker.getRegistrations().then
           <div id="memory-health-content">
             <div class="skel-block"><div class="skel-row med"></div><div class="skel-row"></div><div class="skel-row short"></div></div>
           </div>
-          <div class="card" style="margin-top:18px">
-            <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
-              <span>Recent writes</span>
-              <span style="font-size:11px;color:var(--text-muted)">What the agent captured, with reason &amp; salience</span>
-            </div>
-            <div class="card-body" id="panel-recent-writes" style="padding:0">
-              <div class="skel-block" style="padding:14px"><div class="skel-row med"></div><div class="skel-row short"></div></div>
-            </div>
-          </div>
-          <div class="card" style="margin-top:18px">
-            <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
-              <span>Self-correction (supersedes)</span>
-              <span style="font-size:11px;color:var(--text-muted)">Old facts the agent has explicitly replaced</span>
-            </div>
-            <div class="card-body" id="panel-supersedes" style="padding:0">
-              <div class="skel-block" style="padding:14px"><div class="skel-row med"></div><div class="skel-row short"></div></div>
-            </div>
-          </div>
           <div class="card" style="margin-top:18px">
             <div class="card-header" style="display:flex;align-items:center;justify-content:space-between">
               <span>Knowledge graph signals</span>
@@ -15505,7 +15599,6 @@ function navigateTo(page, opts) {
       }
       break;
     case 'brain':
-      if (typeof refreshMemory === 'function') refreshMemory();
       var bt = opts.tab || 'memory';
       // Spec tab names → internal intelligence-tab ids
       var intelTab = bt === 'memory' ? 'search'
@@ -15516,16 +15609,7 @@ function navigateTo(page, opts) {
         : bt === 'learning' ? 'learning'
         : bt;
       try { switchTab('intelligence', intelTab); } catch (e) { /* */ }
-      if (bt === 'health') {
-        if (typeof refreshMemoryHealth === 'function') refreshMemoryHealth();
-        if (typeof refreshRecentWrites === 'function') refreshRecentWrites();
-        if (typeof refreshSupersedes === 'function') refreshSupersedes();
-        if (typeof refreshGraphStats === 'function') refreshGraphStats();
-        if (typeof refreshSessionBridge === 'function') refreshSessionBridge();
-        if (typeof refreshClaims === 'function') refreshClaims();
-        if (typeof refreshRoutingAudit === 'function') refreshRoutingAudit();
-      }
-      if (bt === 'learning' && typeof refreshSelfImprove === 'function') refreshSelfImprove();
+      // switchTab() above already fires the per-tab refreshers; nothing to do here.
       break;
     case 'settings':
       // Settings tabs use the switchTab() system (id="tab-settings-<tab>"),
@@ -15923,12 +16007,16 @@ function switchTab(group, tab) {
   // Tab-specific refresh
   if (group === 'intelligence') {
     if (tab === 'graph') refreshGraph();
-    if (tab === 'memory') refreshMemory();
+    if (tab === 'search') {
+      // Consolidated Memory tab: search results + stats + MEMORY.md + recent writes + supersedes + coverage strip.
+      refreshMemory();
+      if (typeof refreshRecentWrites === 'function') refreshRecentWrites();
+      if (typeof refreshSupersedes === 'function') refreshSupersedes();
+      if (typeof refreshCoverageStrip === 'function') refreshCoverageStrip();
+    }
     if (tab === 'files' && typeof refreshVaultFiles === 'function') refreshVaultFiles();
     if (tab === 'health') {
       if (typeof refreshMemoryHealth === 'function') refreshMemoryHealth();
-      if (typeof refreshRecentWrites === 'function') refreshRecentWrites();
-      if (typeof refreshSupersedes === 'function') refreshSupersedes();
       if (typeof refreshGraphStats === 'function') refreshGraphStats();
       if (typeof refreshSessionBridge === 'function') refreshSessionBridge();
       if (typeof refreshClaims === 'function') refreshClaims();
@@ -21105,16 +21193,39 @@ async function loadBuilderAttachments(jobName) {
 }
 
 // ── Memory Search ─────────────────────────
+// Parse inline filter syntax from a query string. Supported:
+//   type:procedure          → ?type=procedure
+//   since:7d / since:30d    → ?since=7
+//   pinned:true             → ?pinned=true
+// Returns the cleaned query (filters stripped) plus the filter params.
+function parseSearchFilters(raw) {
+  var filters = {};
+  var cleaned = raw.replace(/\b(type|since|pinned):(\S+)/g, function(_m, key, val) {
+    if (key === 'type') filters.type = val;
+    else if (key === 'since') {
+      var m = /^(\d+)d?$/.exec(val);
+      if (m) filters.since = m[1];
+    }
+    else if (key === 'pinned' && val === 'true') filters.pinned = 'true';
+    return '';
+  }).replace(/\s+/g, ' ').trim();
+  return { q: cleaned, filters: filters };
+}
+
 async function runMemorySearch() {
   const input = document.getElementById('memory-search-input');
-  const q = input.value.trim();
-  if (!q) return;
+  const raw = input.value.trim();
+  if (!raw) return;
 
+  const parsed = parseSearchFilters(raw);
   const container = document.getElementById('memory-search-results');
   container.innerHTML = '<div class="empty-state">Searching...</div>';
 
   try {
-    const r = await apiFetch('/api/memory/search?q=' + encodeURIComponent(q));
+    var qs = 'q=' + encodeURIComponent(parsed.q);
+    var fkeys = Object.keys(parsed.filters);
+    for (var i = 0; i < fkeys.length; i++) qs += '&' + fkeys[i] + '=' + encodeURIComponent(parsed.filters[fkeys[i]]);
+    const r = await apiFetch('/api/memory/search?' + qs);
     const d = await r.json();
 
     if (d.error) {
@@ -21126,11 +21237,16 @@ async function runMemorySearch() {
     }
 
     if (!d.results || d.results.length === 0) {
-      container.innerHTML = '<div class="empty-state">No results found for "' + esc(q) + '"</div>';
+      container.innerHTML = '<div class="empty-state">No results found for "' + esc(raw) + '"</div>';
       return;
     }
 
-    let html = '<div style="font-size:12px;color:var(--text-muted);margin-bottom:12px">' + d.results.length + ' result(s)</div>';
+    var filterChips = '';
+    var fkeys2 = Object.keys(parsed.filters);
+    for (var fi = 0; fi < fkeys2.length; fi++) {
+      filterChips += '<span style="display:inline-block;padding:2px 8px;margin-right:4px;background:var(--bg-secondary);border:1px solid var(--border);border-radius:10px;font-size:11px">' + esc(fkeys2[fi]) + ': ' + esc(parsed.filters[fkeys2[fi]]) + '</span>';
+    }
+    let html = '<div style="font-size:12px;color:var(--text-muted);margin-bottom:12px">' + d.results.length + ' result(s)' + (filterChips ? ' &middot; ' + filterChips : '') + '</div>';
 
     // Show graph relationships if any
     if (d.graphContext && d.graphContext.length > 0) {
@@ -21149,20 +21265,24 @@ async function runMemorySearch() {
       const score = Math.abs(r.score || 0).toFixed(2);
       const pinned = r.pinned ? ' 📌' : '';
       const idAttr = r.id ? String(r.id) : '';
+      const previewSnippet = (r.content || '').slice(0, 200).replace(/\s+/g, ' ').trim();
       html += '<div class="search-result" data-chunk-id="' + esc(idAttr) + '" id="chunk-row-' + esc(idAttr) + '">'
         + '<div class="search-result-header" style="display:flex;justify-content:space-between;align-items:center">'
         + '<span class="search-result-file">' + esc(r.source_file) + pinned + '</span>'
-        + '<div style="display:flex;gap:6px;align-items:center">'
+        + '<div style="display:flex;gap:6px;align-items:center;flex-wrap:wrap">'
         + '<span class="search-result-score" style="font-size:11px;color:var(--text-muted)">score ' + score + '</span>'
         + (idAttr ? (
           '<button class="btn" style="font-size:11px;padding:2px 8px" onclick="editChunk(' + idAttr + ')">Edit</button>'
           + '<button class="btn" style="font-size:11px;padding:2px 8px" onclick="togglePinChunk(' + idAttr + ',' + (r.pinned ? 'false' : 'true') + ')">' + (r.pinned ? 'Unpin' : 'Pin') + '</button>'
           + '<button class="btn" style="font-size:11px;padding:2px 8px;color:var(--red,#ef4444)" onclick="deleteChunk(' + idAttr + ')">Delete</button>'
+          + '<button class="btn" style="font-size:11px;padding:2px 8px" data-snippet="' + esc(previewSnippet) + '" onclick="findSimilarFromButton(this)" title="Search using this chunk\\'s content">Find similar</button>'
+          + '<button class="btn" style="font-size:11px;padding:2px 8px" onclick="toggleTrace(' + idAttr + ')" id="trace-toggle-' + idAttr + '">Trace ▾</button>'
         ) : '')
         + '</div>'
         + '</div>'
         + '<div class="search-result-section">' + esc(r.section || '') + ' &middot; ' + esc(r.chunk_type || '') + '</div>'
         + '<div class="search-result-content" id="chunk-content-' + esc(idAttr) + '">' + esc((r.content || '').slice(0, 500)) + '</div>'
+        + (idAttr ? '<div id="trace-' + idAttr + '" style="display:none;margin-top:8px;padding:10px;background:var(--bg-secondary);border:1px solid var(--border);border-radius:6px;font-size:12px"></div>' : '')
         + '</div>';
     }
     container.innerHTML = html;
@@ -21328,6 +21448,156 @@ async function refreshSessionBridge() {
   }
 }
 
+// Re-run search using the selected chunk's content as the query. With dense
+// embeddings populated, this becomes a true semantic "more like this" lookup.
+// Reads the snippet from the button's data-snippet attribute to avoid HTML/JS
+// escaping issues with quotes inside chunk content.
+function findSimilarFromButton(btn) {
+  var snippet = (btn && btn.dataset && btn.dataset.snippet) ? btn.dataset.snippet : '';
+  if (!snippet) return;
+  var input = document.getElementById('memory-search-input');
+  if (!input) return;
+  input.value = snippet;
+  runMemorySearch();
+}
+
+// Toggle the inline Trace disclosure on a search result. Loads chunk metadata
+// and edit/supersede history from /api/memory/chunks/:id and /history.
+async function toggleTrace(id) {
+  var box = document.getElementById('trace-' + id);
+  var btn = document.getElementById('trace-toggle-' + id);
+  if (!box) return;
+  if (box.style.display !== 'none') {
+    box.style.display = 'none';
+    if (btn) btn.textContent = 'Trace ▾';
+    return;
+  }
+  box.style.display = 'block';
+  if (btn) btn.textContent = 'Trace ▴';
+  box.innerHTML = '<div style="color:var(--text-muted)">Loading…</div>';
+  try {
+    var chunkResp = await apiFetch('/api/memory/chunks/' + id);
+    var chunkData = await chunkResp.json();
+    var historyResp = await apiFetch('/api/memory/chunks/' + id + '/history');
+    var historyData = await historyResp.json();
+    if (!chunkData.ok || !chunkData.chunk) {
+      box.innerHTML = '<div style="color:#ef4444">' + esc(chunkData.error || 'Failed to load') + '</div>';
+      return;
+    }
+    var c = chunkData.chunk;
+    var meta = '<div style="display:grid;grid-template-columns:auto 1fr;gap:4px 12px;margin-bottom:8px">'
+      + '<span style="color:var(--text-muted)">ID:</span><span>' + esc(String(c.id || id)) + '</span>'
+      + '<span style="color:var(--text-muted)">Source:</span><span>' + esc(c.sourceFile || c.source_file || '—') + '</span>'
+      + '<span style="color:var(--text-muted)">Section:</span><span>' + esc(c.section || '—') + '</span>'
+      + '<span style="color:var(--text-muted)">Type:</span><span>' + esc(c.chunkType || c.chunk_type || '—') + '</span>'
+      + '<span style="color:var(--text-muted)">Salience:</span><span>' + esc(String(c.salience != null ? Number(c.salience).toFixed(2) : '—')) + (c.pinned ? ' (pinned)' : '') + '</span>'
+      + '<span style="color:var(--text-muted)">Confidence:</span><span>' + esc(String(c.confidence != null ? Number(c.confidence).toFixed(2) : '—')) + '</span>'
+      + '<span style="color:var(--text-muted)">Created:</span><span>' + esc(c.createdAt || c.created_at || '—') + '</span>'
+      + '<span style="color:var(--text-muted)">Updated:</span><span>' + esc(c.lastUpdated || c.updated_at || '—') + '</span>'
+      + '<span style="color:var(--text-muted)">Agent:</span><span>' + esc(c.agentSlug || c.agent_slug || 'global') + '</span>'
+      + '</div>';
+    var history = '';
+    if (historyData.ok && Array.isArray(historyData.history) && historyData.history.length > 0) {
+      history += '<div style="margin-top:8px"><b style="font-size:11px;color:var(--text-muted)">HISTORY</b>';
+      for (var i = 0; i < historyData.history.length; i++) {
+        var h = historyData.history[i];
+        history += '<div style="padding:4px 0;border-top:1px solid var(--border);font-size:11px">'
+          + esc(h.timestamp || h.at || '') + ' &middot; ' + esc(h.kind || h.action || 'edit')
+          + (h.reason ? ' &middot; ' + esc(h.reason) : '')
+          + '</div>';
+      }
+      history += '</div>';
+    }
+    box.innerHTML = meta + history;
+  } catch (err) {
+    box.innerHTML = '<div style="color:#ef4444">' + esc(String(err)) + '</div>';
+  }
+}
+
+// Quick-add memory modal — type a sentence, hit save, lands in today's daily note.
+function openQuickAddMemory() {
+  var m = document.getElementById('quick-add-memory-modal');
+  if (!m) return;
+  m.style.display = 'flex';
+  var ta = document.getElementById('quick-add-memory-text');
+  if (ta) { ta.value = ''; setTimeout(function(){ ta.focus(); }, 50); }
+  var st = document.getElementById('quick-add-memory-status');
+  if (st) st.textContent = '';
+}
+
+function closeQuickAddMemory() {
+  var m = document.getElementById('quick-add-memory-modal');
+  if (m) m.style.display = 'none';
+}
+
+async function submitQuickAddMemory() {
+  var ta = document.getElementById('quick-add-memory-text');
+  var sel = document.getElementById('quick-add-memory-salience');
+  var st = document.getElementById('quick-add-memory-status');
+  if (!ta) return;
+  var content = ta.value.trim();
+  if (!content) { if (st) { st.textContent = 'Type something first'; st.style.color = '#ef4444'; } return; }
+  if (st) { st.textContent = 'Saving…'; st.style.color = 'var(--text-muted)'; }
+  try {
+    var r = await apiFetch('/api/memory/quick-add', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ content: content, salience_hint: sel ? Number(sel.value) : 1.0 }),
+    });
+    var d = await r.json();
+    if (!r.ok || !d.ok) { if (st) { st.textContent = d.error || 'Failed'; st.style.color = '#ef4444'; } return; }
+    if (st) { st.textContent = 'Saved to ' + d.file; st.style.color = '#10b981'; }
+    setTimeout(function() {
+      closeQuickAddMemory();
+      if (typeof refreshRecentWrites === 'function') refreshRecentWrites();
+      if (typeof refreshMemory === 'function') refreshMemory();
+    }, 600);
+  } catch (err) {
+    if (st) { st.textContent = String(err); st.style.color = '#ef4444'; }
+  }
+}
+
+// Compact coverage strip at the top of the Memory tab. Shows BM25 (always
+// available), sparse TF-IDF coverage, and dense neural coverage with a
+// one-click backfill when coverage is incomplete. Pulls from /api/memory/health.
+async function refreshCoverageStrip() {
+  var el = document.getElementById('memory-coverage-strip');
+  if (!el) return;
+  try {
+    var r = await apiFetch('/api/memory/health');
+    var d = await r.json();
+    if (!d.ok || !d.health) { el.innerHTML = ''; return; }
+    var h = d.health;
+    var total = (h.chunks && h.chunks.total) || 0;
+    var de = h.denseEmbeddings || { withDense: 0, total: total, currentModel: '', ready: false };
+    var sparseCovered = (h.chunks && h.chunks.withSparseEmbedding != null) ? h.chunks.withSparseEmbedding : null;
+    var densePct = de.total > 0 ? Math.round((de.withDense / de.total) * 100) : 0;
+    var sparsePct = (sparseCovered != null && total > 0) ? Math.round((sparseCovered / total) * 100) : null;
+    var denseColor = densePct >= 95 ? '#10b981' : densePct >= 50 ? '#f59e0b' : '#ef4444';
+    var modelLabel = de.currentModel ? de.currentModel.split('/').pop() : '';
+    var html = '<div style="display:flex;align-items:center;gap:14px;padding:10px 14px;background:var(--bg-secondary);border:1px solid var(--border);border-radius:8px;flex-wrap:wrap;font-size:12px">';
+    html += '<span style="color:var(--text-muted)">Coverage:</span>';
+    html += '<span><span style="color:#10b981">●</span> BM25 ' + total.toLocaleString() + '</span>';
+    if (sparsePct != null) html += '<span><span style="color:#10b981">●</span> Sparse ' + sparsePct + '%</span>';
+    html += '<span><span style="color:' + denseColor + '">●</span> Dense ' + densePct + '%'
+      + (modelLabel ? ' <span style="color:var(--text-muted)">(' + esc(modelLabel) + ')</span>' : '') + '</span>';
+    if (de.total > 0 && de.withDense < de.total) {
+      var missing = de.total - de.withDense;
+      html += '<span style="margin-left:auto;display:flex;gap:6px">'
+        + '<span style="color:var(--text-muted)">' + missing.toLocaleString() + ' missing</span>'
+        + '<button class="btn-sm" onclick="memoryHealthAction(\\'reembed-dense\\', { limit: 200 })">Backfill 200</button>'
+        + '<button class="btn-sm" onclick="memoryHealthAction(\\'reembed-dense\\', { limit: 2000 })">Backfill 2000</button>'
+        + '</span>';
+    } else if (de.total > 0) {
+      html += '<span style="margin-left:auto;color:var(--text-muted)">All chunks indexed</span>';
+    }
+    html += '</div>';
+    el.innerHTML = html;
+  } catch (err) {
+    el.innerHTML = '';
+  }
+}
+
 async function refreshRecentWrites() {
   var el = document.getElementById('panel-recent-writes');
   if (!el) return;

package/dist/config/effective-config.js

@@ -24,7 +24,7 @@ const SPECS = [
     { key: 'DEFAULT_MODEL_TIER', group: 'models', jsonPath: 'models.default', default: 'sonnet' },
     { key: 'HAIKU_MODEL', group: 'models', jsonPath: 'models.haiku', default: 'claude-haiku-4-5-20251001' },
     { key: 'SONNET_MODEL', group: 'models', jsonPath: 'models.sonnet', default: 'claude-sonnet-4-6' },
-    { key: 'OPUS_MODEL', group: 'models', jsonPath: 'models.opus', default: 'claude-opus-4-6' },
+    { key: 'OPUS_MODEL', group: 'models', jsonPath: 'models.opus', default: 'claude-opus-4-7' },
     // Budgets
     { key: 'BUDGET_HEARTBEAT_USD', group: 'budgets', jsonPath: 'budgets.heartbeat', default: 0.50 },
     { key: 'BUDGET_CRON_T1_USD', group: 'budgets', jsonPath: 'budgets.cronT1', default: 2.00 },

package/dist/config.d.ts

@@ -84,10 +84,6 @@ export declare const TASK_BUDGET_TOKENS: {
 };
 export declare const DEFAULT_MODEL_TIER: keyof Models;
 export declare const MODEL: string;
-/** Enable 1M context window for Sonnet (beta). Toggle via ENABLE_1M_CONTEXT=true in .env or dashboard. */
-export declare let ENABLE_1M_CONTEXT: boolean;
-/** Update 1M context flag at runtime (called from dashboard settings API). */
-export declare function setEnable1MContext(value: boolean): void;
 export declare const DISCORD_TOKEN: string;
 export declare const DISCORD_OWNER_ID: string;
 export declare const DISCORD_WATCHED_CHANNELS: string[];

package/dist/config.js

@@ -179,7 +179,7 @@ function getSecret(envKey, keychainService) {
 export const MODELS = {
     haiku: getEnvOrJson('HAIKU_MODEL', json.models?.haiku, 'claude-haiku-4-5-20251001'),
     sonnet: getEnvOrJson('SONNET_MODEL', json.models?.sonnet, 'claude-sonnet-4-6'),
-    opus: getEnvOrJson('OPUS_MODEL', json.models?.opus, 'claude-opus-4-6'),
+    opus: getEnvOrJson('OPUS_MODEL', json.models?.opus, 'claude-opus-4-7'),
 };
 // ── Budget caps (USD per query) ──────────────────────────────────────
 // User-tunable via `clementine config set BUDGET_<NAME>_USD <value>`
@@ -236,12 +236,6 @@ export const TASK_BUDGET_TOKENS = {
 };
 export const DEFAULT_MODEL_TIER = (getEnvOrJson('DEFAULT_MODEL_TIER', json.models?.default, 'sonnet'));
 export const MODEL = MODELS[DEFAULT_MODEL_TIER] ?? MODELS.sonnet;
-/** Enable 1M context window for Sonnet (beta). Toggle via ENABLE_1M_CONTEXT=true in .env or dashboard. */
-export let ENABLE_1M_CONTEXT = getEnv('ENABLE_1M_CONTEXT', 'false').toLowerCase() === 'true';
-/** Update 1M context flag at runtime (called from dashboard settings API). */
-export function setEnable1MContext(value) {
-    ENABLE_1M_CONTEXT = value;
-}
 // ── Discord ──────────────────────────────────────────────────────────
 export const DISCORD_TOKEN = getSecret('DISCORD_TOKEN');
 export const DISCORD_OWNER_ID = getEnv('DISCORD_OWNER_ID', '0');

package/dist/gateway/outcome-grader.js

@@ -135,7 +135,13 @@ export async function gradeRun(entry, gateway, jobPrompt) {
     try {
         raw = await gateway.handleCronJob(`grade:${entry.jobName}`, prompt, 1, // tier 1
         3, // maxTurns — tight
-        'haiku');
+        'haiku', undefined, // workDir
+        'standard', // mode
+        undefined, // maxHours
+        undefined, // timeoutMs
+        undefined, // successCriteria
+        undefined, // agentSlug
+        { disableAllTools: true });
     }
     catch (err) {
         logger.warn({ err, jobName: entry.jobName }, 'Outcome grader LLM call failed');

package/dist/gateway/router.d.ts

@@ -152,7 +152,9 @@ export declare class Gateway {
     handleMessage(sessionKey: string, text: string, onText?: OnTextCallback, model?: string, maxTurns?: number, onToolActivity?: OnToolActivityCallback, onProgress?: OnProgressCallback): Promise<string>;
     private _handleMessageInner;
     handleHeartbeat(standingInstructions: string, changesSummary?: string, timeContext?: string, dedupContext?: string, profile?: import('../types.js').AgentProfile | null): Promise<string>;
-    handleCronJob(jobName: string, jobPrompt: string, tier?: number, maxTurns?: number, model?: string, workDir?: string, mode?: 'standard' | 'unleashed', maxHours?: number, timeoutMs?: number, successCriteria?: string[], agentSlug?: string): Promise<string>;
+    handleCronJob(jobName: string, jobPrompt: string, tier?: number, maxTurns?: number, model?: string, workDir?: string, mode?: 'standard' | 'unleashed', maxHours?: number, timeoutMs?: number, successCriteria?: string[], agentSlug?: string, opts?: {
+        disableAllTools?: boolean;
+    }): Promise<string>;
     /**
      * Process a team message as an autonomous task — same multi-phase execution
      * as cron unleashed jobs, so agents can work until done instead of being

package/dist/gateway/router.js

@@ -1345,7 +1345,7 @@ export class Gateway {
             releaseLane();
         }
     }
-    async handleCronJob(jobName, jobPrompt, tier = 1, maxTurns, model, workDir, mode = 'standard', maxHours, timeoutMs, successCriteria, agentSlug) {
+    async handleCronJob(jobName, jobPrompt, tier = 1, maxTurns, model, workDir, mode = 'standard', maxHours, timeoutMs, successCriteria, agentSlug, opts) {
         const releaseLane = await lanes.acquire('cron');
         try {
             logger.info(`Running cron job: ${jobName}${workDir ? ` in ${workDir}` : ''}${mode === 'unleashed' ? ' (unleashed)' : ''}${agentSlug && agentSlug !== 'clementine' ? ` as ${agentSlug}` : ''}`);
@@ -1357,7 +1357,7 @@ export class Gateway {
                     response = await this.assistant.runUnleashedTask(jobName, jobPrompt, tier, maxTurns, model, workDir, maxHours, agentSlug);
                 }
                 else {
-                    response = await this.assistant.runCronJob(jobName, jobPrompt, tier, maxTurns, model, workDir, timeoutMs, successCriteria, agentSlug);
+                    response = await this.assistant.runCronJob(jobName, jobPrompt, tier, maxTurns, model, workDir, timeoutMs, successCriteria, agentSlug, opts);
                 }
                 // Re-baseline integrity checksums after cron job (may write to vault)
                 scanner.refreshIntegrity();

package/dist/index.js

@@ -247,8 +247,6 @@ function printBanner(channels, profiles, cronJobs, graphEnabled = false) {
     const modelColor = modelColors[modelName] ?? CYAN;
     // Feature tags
     const tags = [];
-    if (config.ENABLE_1M_CONTEXT)
-        tags.push('1M context');
     if (config.GROQ_API_KEY)
         tags.push('voice');
     if (config.GOOGLE_API_KEY)
@@ -694,6 +692,24 @@ async function asyncMain() {
     const heartbeat = new HeartbeatScheduler(gateway, dispatcher);
     const cronScheduler = new CronScheduler(gateway, dispatcher);
     heartbeat.setCronScheduler(cronScheduler);
+    // Warm the dense embedding model in the background at boot so the first
+    // search/backfill doesn't pay the load cost. Failure here is non-fatal —
+    // search degrades to BM25 + sparse TF-IDF.
+    void (async () => {
+        try {
+            const embeddings = await import('./memory/embeddings.js');
+            const ready = await embeddings.probeDenseReady();
+            if (ready) {
+                logger.info('Dense embedding model warmed at boot');
+            }
+            else {
+                logger.warn('Dense embedding model failed to warm — search will use BM25 + sparse TF-IDF only');
+            }
+        }
+        catch (err) {
+            logger.warn({ err }, 'Dense embedding warmup threw');
+        }
+    })();
     // Builder runner — wire MCP invoke handler so canvas test runs can hit
     // real read-only MCP tools (gmail.list_unread, github.list_prs, etc.).
     // Stdio clients are pooled per server with idle teardown.

package/dist/memory/embeddings.d.ts

@@ -8,10 +8,11 @@
  *     chunks.embedding (BLOB).
  *
  *  2. Dense neural (preferred when available, async). Uses
- *     `@xenova/transformers` to run a local sentence-embedding model
- *     (default: Snowflake/snowflake-arctic-embed-m-v1.5, 768-dim) entirely
- *     on-device. Stored in chunks.embedding_dense (BLOB) with
- *     chunks.embedding_dense_model tracking which model produced it.
+ *     `@huggingface/transformers` (v4+) to run a local sentence-embedding
+ *     model (default: Snowflake/snowflake-arctic-embed-m-v1.5, 768-dim)
+ *     entirely on-device via onnxruntime-node. Stored in
+ *     chunks.embedding_dense (BLOB) with chunks.embedding_dense_model
+ *     tracking which model produced it.
  *
  * Runtime behavior:
  *  - At store insert time, sync TF-IDF is computed (cheap, no I/O).

package/dist/memory/embeddings.js

@@ -8,10 +8,11 @@
  *     chunks.embedding (BLOB).
  *
  *  2. Dense neural (preferred when available, async). Uses
- *     `@xenova/transformers` to run a local sentence-embedding model
- *     (default: Snowflake/snowflake-arctic-embed-m-v1.5, 768-dim) entirely
- *     on-device. Stored in chunks.embedding_dense (BLOB) with
- *     chunks.embedding_dense_model tracking which model produced it.
+ *     `@huggingface/transformers` (v4+) to run a local sentence-embedding
+ *     model (default: Snowflake/snowflake-arctic-embed-m-v1.5, 768-dim)
+ *     entirely on-device via onnxruntime-node. Stored in
+ *     chunks.embedding_dense (BLOB) with chunks.embedding_dense_model
+ *     tracking which model produced it.
  *
  * Runtime behavior:
  *  - At store insert time, sync TF-IDF is computed (cheap, no I/O).
@@ -219,12 +220,16 @@ async function getDensePipeline() {
             mkdirSync(MODEL_CACHE_DIR, { recursive: true });
         }
         catch { /* non-fatal */ }
-        const transformers = (await import('@xenova/transformers'));
+        const transformers = (await import('@huggingface/transformers'));
         transformers.env.cacheDir = MODEL_CACHE_DIR;
         transformers.env.allowLocalModels = true;
+        transformers.env.allowRemoteModels = true;
         const modelId = getDenseModelId();
-        logger.info({ modelId, cacheDir: MODEL_CACHE_DIR }, 'Loading dense embedding model (first use downloads ~440MB)');
-        const pipe = await transformers.pipeline('feature-extraction', modelId);
+        // dtype: 'q8' uses the int8-quantized ONNX file (model_quantized.onnx,
+        // ~110MB) instead of the fp32 file (~440MB). Quality loss is negligible
+        // for retrieval and load is dramatically faster on CPU.
+        logger.info({ modelId, cacheDir: MODEL_CACHE_DIR, dtype: 'q8' }, 'Loading dense embedding model');
+        const pipe = await transformers.pipeline('feature-extraction', modelId, { dtype: 'q8' });
         denseLoadState = true;
         logger.info({ modelId }, 'Dense embedding model loaded');
         return pipe;

package/dist/tools/memory-tools.js

@@ -180,51 +180,6 @@ export function registerMemoryTools(server) {
             }
         }
     });
-    // ── 0b. team_scratchpad ────────────────────────────────────────────────
-    //
-    // Cross-agent shared scratchpad. Unlike working_memory (per-agent), this
-    // is a single shared markdown file every agent can read and append to.
-    // Use cases: live coordination ("Sasha is drafting the brief, Ross hold
-    // outbound for 30m"), cross-agent context drops, async hand-offs that
-    // don't warrant a full goal_create or task_add. Append tags every entry
-    // with the author's agent slug + ISO timestamp so the trail stays clear.
-    const TEAM_SCRATCHPAD_FILE = path.join(BASE_DIR, 'team-scratchpad.md');
-    server.tool('team_scratchpad', getToolDescription('team_scratchpad') ?? 'Cross-agent shared scratchpad for live team coordination. All agents read/write the same file. Use for hand-offs, "I am working on X", short-term context drops. For durable facts, use memory_write/MEMORY.md instead.', {
-        action: z.enum(['read', 'append', 'replace', 'clear']).describe('What to do with the team scratchpad'),
-        content: z.string().optional().describe('Text to append or replace with (required for append/replace)'),
-    }, async ({ action, content }) => {
-        const author = ACTIVE_AGENT_SLUG ?? 'clementine';
-        switch (action) {
-            case 'read': {
-                if (!existsSync(TEAM_SCRATCHPAD_FILE)) {
-                    return textResult('Team scratchpad is empty.');
-                }
-                return textResult(readFileSync(TEAM_SCRATCHPAD_FILE, 'utf-8'));
-            }
-            case 'append': {
-                if (!content)
-                    return textResult('Error: content is required for append.');
-                const stamp = new Date().toISOString();
-                const entry = `\n- **[${author}@${stamp}]** ${content}\n`;
-                const existing = existsSync(TEAM_SCRATCHPAD_FILE) ? readFileSync(TEAM_SCRATCHPAD_FILE, 'utf-8') : '# Team Scratchpad\n\nShared across all agents. Append tags entries with author + timestamp.\n';
-                writeFileSync(TEAM_SCRATCHPAD_FILE, existing + entry);
-                return textResult(`Appended to team scratchpad as ${author}.`);
-            }
-            case 'replace': {
-                if (!content)
-                    return textResult('Error: content is required for replace.');
-                const stamp = new Date().toISOString();
-                const header = `# Team Scratchpad\n\n_Replaced by ${author} at ${stamp}._\n\n`;
-                writeFileSync(TEAM_SCRATCHPAD_FILE, header + content + '\n');
-                return textResult(`Team scratchpad replaced by ${author}.`);
-            }
-            case 'clear': {
-                if (existsSync(TEAM_SCRATCHPAD_FILE))
-                    unlinkSync(TEAM_SCRATCHPAD_FILE);
-                return textResult('Team scratchpad cleared.');
-            }
-        }
-    });
     // ── 1. memory_read ─────────────────────────────────────────────────────
     server.tool('memory_read', getToolDescription('memory_read') ?? "Read a note from the Obsidian vault. Shortcuts: 'today', 'yesterday', 'memory', 'tasks', 'heartbeat', 'cron', 'soul'. Or pass a relative path or note name.", {
         name: z.string().describe('Note name, path, or shortcut'),
@@ -507,43 +462,6 @@ export function registerMemoryTools(server) {
         }
         return textResult(`Unknown action: ${action}`);
     });
-    // ── 2b. memory_record_procedure ────────────────────────────────────────
-    server.tool('memory_record_procedure', getToolDescription('memory_record_procedure') ?? 'Record a learned workflow as a durable procedure. Use when you notice a repeating multi-step task ("how Nate ships a release", "how to handle inbound replies"). Stored under 00-System/procedures/ with category=procedure and trigger verbs that surface it later. Different from memory_write/MEMORY.md: those store facts, this stores reusable HOW-TO. From Mem0\'s 2026 procedural-memory pattern.', {
-        title: z.string().describe('Short procedure title (becomes filename slug)'),
-        steps: z.string().describe('Numbered steps or markdown body describing how to perform the task'),
-        triggers: z.array(z.string()).min(1).describe('Verb phrases (e.g. ["ship release", "publish to npm"]) that should surface this procedure when the user query contains them. Lowercase preferred.'),
-        notes: z.string().optional().describe('Optional context: when to use, when NOT to use, gotchas'),
-    }, async ({ title, steps, triggers, notes }) => {
-        const slug = title.toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-+|-+$/g, '').slice(0, 80);
-        if (!slug)
-            return textResult('Error: title must contain alphanumerics');
-        const proceduresDir = path.join(SYSTEM_DIR, 'procedures');
-        mkdirSync(proceduresDir, { recursive: true });
-        const filePath = path.join(proceduresDir, `${slug}.md`);
-        const triggersYaml = triggers.map((t) => `  - ${JSON.stringify(t.toLowerCase())}`).join('\n');
-        const body = [
-            '---',
-            `title: ${JSON.stringify(title)}`,
-            'category: procedure',
-            'triggers:',
-            triggersYaml,
-            `created_at: ${new Date().toISOString()}`,
-            ACTIVE_AGENT_SLUG ? `agent_slug: ${JSON.stringify(ACTIVE_AGENT_SLUG)}` : '',
-            '---',
-            '',
-            `# ${title}`,
-            '',
-            '## Steps',
-            '',
-            steps.trim(),
-            '',
-            ...(notes ? ['## Notes', '', notes.trim(), ''] : []),
-        ].filter((line) => line !== '').join('\n') + '\n';
-        writeFileSync(filePath, body, 'utf-8');
-        const rel = path.relative(VAULT_DIR, filePath);
-        await incrementalSync(rel, ACTIVE_AGENT_SLUG ?? undefined);
-        return textResult(`Recorded procedure: ${rel} (triggers: ${triggers.join(', ')})`);
-    });
     // ── 3. memory_search ───────────────────────────────────────────────────
     server.tool('memory_search', getToolDescription('memory_search') ?? 'FTS5 search across all vault notes. Returns matching chunks with relevance scores. Optional category/topic filters narrow results.', {
         query: z.string().describe('Search text'),

package/dist/tools/tool-meta.js

@@ -11,15 +11,10 @@
 const TOOL_META = {
     // ── Memory & Vault ────────────────────────────────────────────────
     working_memory: {
-        description: 'Per-agent persistent scratchpad — only YOU see it. Survives across conversations. Use for current project context, TODOs, reminders, or anything you need to remember for next time. Actions: read, append, replace, clear. ALWAYS read before replacing. For cross-agent coordination, use team_scratchpad instead.',
+        description: 'Per-agent persistent scratchpad — only YOU see it. Survives across conversations. Use for current project context, TODOs, reminders, or anything you need to remember for next time. Actions: read, append, replace, clear. ALWAYS read before replacing.',
         exampleUsage: 'Before starting complex work, read working_memory to check for context from prior sessions.',
         returnHint: 'Full working memory contents (markdown text).',
     },
-    team_scratchpad: {
-        description: 'Cross-agent shared scratchpad — every agent on the team reads and writes the same file. Use for live coordination, hand-offs, "I am working on X — back off until Y", short-lived context drops. For durable facts that should outlive coordination noise, use memory_write to MEMORY.md instead. Append tags entries with author slug + timestamp.',
-        exampleUsage: 'Before starting outbound work, read team_scratchpad to see if another agent has already claimed a prospect or paused outreach.',
-        returnHint: 'Full scratchpad contents with per-entry author + ISO timestamp.',
-    },
     memory_search: {
         description: 'Full-text search across all vault notes. Best for finding specific keywords or phrases. For broader semantic matching, use memory_recall instead. Results include file path, section heading, and relevance score.',
         exampleUsage: 'Use when the user asks "what did we discuss about X" or you need to find a specific note.',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "clementine-agent",
-  "version": "1.13.3",
+  "version": "1.15.0",
   "description": "Clementine — Personal AI Assistant (TypeScript)",
   "type": "module",
   "main": "dist/index.js",
@@ -21,15 +21,15 @@
     "postinstall": "node scripts/postinstall.js 2>/dev/null || true"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.2.119",
+    "@anthropic-ai/claude-agent-sdk": "^0.2.126",
     "@anthropic-ai/sdk": "^0.91.0",
     "@composio/claude-agent-sdk": "^0.8.1",
     "@composio/core": "^0.8.1",
+    "@huggingface/transformers": "^4.2.0",
     "@inquirer/prompts": "^7.0.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@slack/bolt": "^4.2.0",
     "@types/multer": "^2.1.0",
-    "@xenova/transformers": "^2.17.2",
     "better-sqlite3": "^11.7.0",
     "commander": "^13.1.0",
     "cron-parser": "^5.5.0",