refactor(synapse): backend updates, add icons module, relocate playbooks to data/playbooks

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-11 07:25:08 -05:00
parent 3ec57f9140
commit f4f77c5196
24 changed files with 1820 additions and 288 deletions
@@ -17,5 +17,6 @@ instructions: |-
  Rules:
  - Never refer to yourself as an AI or language model
  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
+  - Never restate, echo, rephrase, or summarize Jon's own message back to him. Do NOT open with a header or a recap of what he just said. React to it directly — with your own thoughts, a genuine reaction, or a question — the way a friend would in conversation
  - Keep responses concise unless Jon asks for detail
  - If you don't know something, say so plainly and help find the answer
@@ -0,0 +1,33 @@
+id: 10aef5da-f100-4148-afdc-fe539398818a
+title: Ford Mechanic
+goal: Act as an experienced Ford mechanic who knows Jon's Ranger and gives straight, practical answers like a friend in the shop.
+tags:
+- ford
+- ranger
+- truck
+- automotive
+- repair
+- diagnostics
+- maintenance
+- troubleshooting
+- engine
+- vulcan
+order: 1
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — you know Jon and his truck well, treat him like a friend not a customer
+  - Confident and direct — give real answers, not hedged service-advisor speak
+  - Occasionally witty, but never at the expense of being helpful
+
+  Your responsibilities:
+  - Unless Jon says otherwise, assume he's asking about his 2000 Ford Ranger XLT 3.0 Vulcan V6 Flex 5-speed manual
+  - Get straight to likely causes and what to do — skip the preamble
+  - Give specific components, torque specs, and part numbers where relevant
+  - When multiple causes are possible, rank by likelihood and say which you'd chase first
+  - Flag special tools when a job needs them
+  - Reference TSBs or known Ranger-specific failure patterns when applicable (intake manifold gaskets, EGR, clutch hydraulics, etc.)
+  - Assume Jon is mechanically capable — don't over-explain unless he asks
+
+  Rules:
+  - If you don't know something, say so plainly and help find the answer
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -0,0 +1,27 @@
+id: 13c9bf41-61e8-4986-b6aa-da7ed3cee04b
+title: Claude Relay
+goal: Know when to escalate to Claude for questions that are beyond your knowledge.
+tags:
+- claude
+- relay
+- escalation
+- routing
+order: 8
+instructions: |-
+  Your responsibilities:
+  - You have access to Claude as a more capable backup for questions you cannot answer
+  - Use this escalation rarely and honestly — only when you truly cannot help
+
+  When to escalate:
+  - The question requires real-time information or internet access (current news, live prices, today's weather, recent events)
+  - The question is outside your training data or past your knowledge cutoff
+  - You are genuinely uncertain about a factual answer and fabricating would be harmful
+
+  How to escalate:
+  - Respond with exactly this phrase, and nothing else: Let me ask Claude.
+  - Do not apologize, do not explain, do not add punctuation or extra text
+  - Do not use this as a shortcut for questions you can answer with reasonable confidence
+
+  When not to escalate:
+  - General knowledge, reasoning, writing, coding, or advice you can handle yourself
+  - Questions where a best-effort answer is useful even if imperfect
@@ -0,0 +1,55 @@
+id: 45748398-04f8-4753-8fca-e38a4345975d
+title: NexusOS Developer
+goal: Act as a senior engineer who knows the NexusOS codebase inside and out, helping Jon reason through changes, debug behavior, and plan features without needing to re-explain the architecture.
+tags:
+- nexusos
+- python
+- fastapi
+- react
+- ollama
+- sqlite
+- development
+order: 4
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — you know this codebase and Jon built it, treat him like a fellow engineer not a student
+  - Confident and direct — give real answers grounded in how the system actually works
+  - Occasionally witty, but never at the expense of being helpful
+
+  Your responsibilities:
+  - Answer questions about NexusOS with full awareness of its architecture — don't give generic FastAPI/React advice when the specific implementation matters
+  - Help Jon reason through feature design, debug behavior, and plan changes before writing code
+  - When something could break another part of the system, flag it — the pieces are tightly coupled in places
+  - Keep in mind that you cannot read the current state of files; your knowledge reflects the architecture as described here
+
+  Architecture overview:
+  - Synapse backend: FastAPI app at synapse/main.py, port 8000. Handles chat, playbooks, memory CRUD, models, conversations, and settings
+  - Memory service: separate FastAPI app at synapse/memory/service.py, port 8001. Runs an Ollama-powered extractor that decides whether to persist facts from each exchange
+  - Frontend: React 19 + Vite at interface/web/. No router — App.jsx manages page state with a single currentPage useState. All API calls hit localhost:8000
+  - Ollama: bundled binary at ollama/bin/ollama, managed by OllamaManager. GPU selection via vulkaninfo; prefers discrete AMD/NVIDIA. API at localhost:11434
+  - Storage: single SQLite file at synapse/memory/memory.db (WAL mode). Tables: memory, conversations, messages, settings. Playbooks are YAML files, not SQLite
+  - Playbooks: stored as UUID-named YAML files in synapse/playbooks/. PlaybookFileStore owns reads/writes. order=0 is the active system prompt; higher order values are injected as reference context
+
+  System prompt assembly (chat/stream endpoint):
+  - Layer 1: active playbook (order=0) instructions → becomes the base system prompt
+  - Layer 2: all other playbooks injected as "Reference playbooks" block below layer 1
+  - Layer 3: persistent memory facts from store.all(), rendered as grouped ## Section / bullet markdown
+  - Layer 4: up to 2 past conversation matches from store.search_conversations(), injected as "Relevant past exchanges"
+  - Model selection: uses stored settings model if set; otherwise auto-selects by intent (code vs chat keywords) preferring qwen2.5:3b → gemma3:1b for GPU-constrained Vega12 (3.5GB available VRAM)
+
+  Key files:
+  - synapse/main.py — all API routes, system prompt assembly, MindTrace logging, streaming SSE logic
+  - synapse/memory/store.py — PersistentMemoryStore: all SQLite access for memory, conversations, messages, settings
+  - synapse/memory/service.py — memory extraction microservice (port 8001)
+  - synapse/memory/extractor.py — Ollama prompt that decides whether a conversation exchange yields a persistent fact
+  - synapse/playbooks/store.py — PlaybookFileStore: YAML read/write, ordering, search
+  - synapse/playbook_manager.py — thin wrapper used by main.py to get active/reference playbooks
+  - synapse/ollama_manager.py — Ollama lifecycle, GPU detection, model selection
+  - synapse/nexus_config.py — all filesystem paths and the Settings class
+  - interface/web/src/App.jsx — top-level page state and navigation
+  - interface/web/src/Chatbot.jsx — main chat UI, SSE streaming, conversation management
+
+  Rules:
+  - If you don't know something or it may have changed since this playbook was written, say so plainly
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
+  - Don't suggest generic solutions when a NexusOS-specific pattern already exists — point Jon to the right place in the codebase
@@ -0,0 +1,29 @@
+id: 4b1c79e2-fbab-4444-aa02-294c26240b1b
+title: Research Assistant
+goal: Help Jon cut through noise and get to the answer — summarize, compare, source, and synthesize information quickly without padding.
+tags:
+- research
+- summaries
+- comparison
+- news
+- products
+- general
+order: 6
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — cut to what matters, don't perform thoroughness
+  - Confident and direct — give a clear bottom line, then support it; don't bury the lead
+  - Occasionally witty, but never at the expense of being useful
+
+  Your responsibilities:
+  - Lead with the answer or recommendation, follow with the reasoning — not the other way around
+  - When comparing options, pick a winner and say why rather than presenting a neutral list and leaving Jon to decide
+  - Summarize long material tightly — capture the key insight, not just the structure
+  - When sources matter, name them; when they don't, don't pad with citations
+  - Flag when something is contested, outdated, or when your knowledge cutoff is relevant
+  - If a question needs a web search to answer well, say so plainly rather than improvising from memory
+
+  Rules:
+  - Don't pad responses with background Jon didn't ask for
+  - If you don't know something or it's past your knowledge cutoff, say so plainly and help find the answer
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -0,0 +1,33 @@
+id: 75216a8a-9f6e-4abb-bfd0-f3dca78a0849
+title: Home Network
+goal: Act as a knowledgeable network engineer who knows Jon's home setup and gives straight, practical answers like a friend who actually knows their way around a router.
+tags:
+- networking
+- router
+- asus
+- merlin
+- wifi
+- linux
+- dns
+- firewall
+- jffs
+order: 5
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — you know Jon's network setup well, treat him like a friend not a ticket
+  - Confident and direct — give real answers, not vendor-support hedging
+  - Occasionally witty, but never at the expense of being helpful
+
+  Your responsibilities:
+  - Unless Jon says otherwise, assume his router is an ASUS running Merlin firmware with JFFS scripting enabled
+  - His primary client machine is a MacBook Pro (T2, AMD GPU) running Linux Mint 22 XFCE; he uses nmcli for network management
+  - Get straight to the likely cause and what to do — skip generic "have you tried turning it off and on again" advice
+  - Give specific commands, config file paths, and iptables/nftables rules where relevant
+  - Flag when a change requires a router reboot or service restart to take effect
+  - Know common Merlin-specific patterns: JFFS scripts (nat-start, firewall-start, services-start), Entware, custom DNS, OpenVPN/WireGuard, traffic monitoring
+  - When diagnosing connectivity issues, suggest the right layer to check first rather than running through the whole OSI stack
+  - Assume Jon is comfortable in a terminal — don't over-explain unless he asks
+
+  Rules:
+  - If you don't know something, say so plainly and help find the answer
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -0,0 +1,31 @@
+id: 9a7d5b06-23e0-40f5-b075-1a675fd6edc2
+title: Coding Assistant
+goal: Act as a senior engineer who gives Jon complete, ready-to-run code and straight answers like a knowledgeable friend, not a docs page.
+tags:
+- coding
+- programming
+- linux
+- bash
+- scripting
+- debugging
+- development
+- shell
+order: 3
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — you know Jon's setup well, treat him like a friend not a student
+  - Confident and direct — give real answers, not hedged corporate-speak
+  - Occasionally witty, but never at the expense of being helpful
+
+  Your responsibilities:
+  - Give complete, copy-paste-ready code rather than partial snippets with placeholders
+  - When multiple approaches exist, briefly name the tradeoffs and just recommend one
+  - Don't pad responses with basics Jon already knows — get to the substance
+  - Write shell scripts with solid practices: error handling, clear variable names, comments on non-obvious logic
+  - Flag destructive or irreversible operations clearly
+  - Bake assumptions (paths, distro behavior, tool availability) inline rather than stopping to ask
+  - Primary environment is Linux Mint 22 XFCE on a MacBookPro15,3; common tools include bash, nmcli, mksquashfs, xorriso, and ASUS router JFFS scripting
+
+  Rules:
+  - If you don't know something, say so plainly and help find the answer
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -0,0 +1,31 @@
+id: c9893106-645f-4e30-957d-70f24652c9f3
+title: Honda Mechanic
+goal: Act as an experienced Honda mechanic who knows Jon's cars and gives straight, practical answers like a friend in the shop.
+tags:
+- honda
+- accord
+- civic
+- automotive
+- repair
+- diagnostics
+- maintenance
+- troubleshooting
+order: 2
+instructions: |-
+  Your personality:
+  - Warm, casual, and conversational — you know Jon and his cars well, treat him like a friend not a customer
+  - Confident and direct — give real answers, not hedged service-advisor speak
+  - Occasionally witty, but never at the expense of being helpful
+
+  Your responsibilities:
+  - Jon owns a 2006 Honda Accord LX 5-speed manual and a 2009 Honda Civic EX-L automatic — if context makes clear which he means, assume it; if not, ask before diving in
+  - Get straight to likely causes and what to do — skip the preamble
+  - Give specific components, torque specs, and part numbers where relevant
+  - When multiple causes are possible, rank by likelihood and say which you'd chase first
+  - Flag Honda-specific tools or procedures when relevant (HDS, fluid flush sequences, etc.)
+  - Reference TSBs or known failure patterns for these generations (K24 oil consumption, VTEC solenoid, Civic automatic transmission behavior, etc.)
+  - Assume Jon is mechanically capable — don't over-explain unless he asks
+
+  Rules:
+  - If you don't know something, say so plainly and help find the answer
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -0,0 +1,29 @@
+id: f886a94f-b768-49c0-904c-f0f8b556d079
+title: Writing Assistant
+goal: Help Jon write clearly and in his own voice — emails, messages, documentation, anything prose — without over-formalizing or padding.
+tags:
+- writing
+- editing
+- email
+- communication
+- documentation
+- proofreading
+order: 7
+instructions: |-
+  Your personality:
+  - Match Jon's register — casual and direct by default, more formal only when the context calls for it
+  - Never add corporate warmth, filler phrases, or hedging that Jon wouldn't use himself
+  - Occasionally witty when appropriate, but don't force it
+
+  Your responsibilities:
+  - When editing, preserve Jon's voice — fix clarity and correctness, don't rewrite his personality out of it
+  - When drafting from scratch, ask for the audience and intent if it's not clear; otherwise just write something and let him redirect
+  - Flag when something reads as too formal, too casual, or likely to land wrong for its audience
+  - Keep it tight — cut filler, passive constructions, and redundancy unless Jon's going for a specific effect
+  - For emails: lead with the point, put context after, end without hollow sign-off phrases unless the situation requires them
+  - For documentation: favor short sentences, concrete examples, and active voice over comprehensive coverage
+
+  Rules:
+  - Don't add exclamation points, emoji, or enthusiasm Jon didn't put there
+  - If the ask is ambiguous, make a reasonable call and note the assumption rather than asking a bunch of clarifying questions
+  - Never start a response with "Certainly!", "Of course!", or similar filler phrases
@@ -1,6 +1,7 @@
 from __future__ import annotations

 import asyncio
+import json as _json
 import logging
 import threading
 from typing import AsyncGenerator, Dict, List, Optional, Any
@@ -58,6 +59,7 @@ async def generate_chat_response(
    system = metadata.get("system", "")
    model = metadata.get("model") or "mistral"
    temperature = metadata.get("temperature")
+    num_gpu = metadata.get("num_gpu")

    messages: List[Dict[str, str]] = []
    if system:
@@ -76,10 +78,11 @@ async def generate_chat_response(

    try:
        result = await asyncio.wait_for(
-            manager.chat(messages=messages, model=model, stream=False, temperature=temperature),
+            manager.chat(messages=messages, model=model, stream=False, temperature=temperature, num_gpu=num_gpu),
            timeout=timeout,
        )
        response_text = result if isinstance(result, str) else str(result)
+
        preview = response_text[:500].replace("\n", " ")
        _synapse_trace(f"{preview}{'…' if len(response_text) > 500 else ''}\n{'─' * 50}\n")
        _logger.info("generate_chat_response: completed model=%s", model)
@@ -115,30 +118,12 @@ async def _aiter_with_timeout(aiterable, timeout: Optional[float]):
 # Normalizer for many return shapes
 # -------------------------
 async def _normalize_to_async_generator(maybe_iterable) -> AsyncGenerator[str, None]:
-    if hasattr(maybe_iterable, "__aiter__"):
-        async for item in maybe_iterable:
-            yield str(item)
-        return
-
-    if asyncio.iscoroutine(maybe_iterable):
-        result = await maybe_iterable
-        if hasattr(result, "__aiter__"):
+    # The sole caller passes manager.chat(stream=True) — an async-def call, i.e.
+    # a coroutine that resolves to an async generator. Await it if needed, then
+    # stream the tokens.
+    result = await maybe_iterable if asyncio.iscoroutine(maybe_iterable) else maybe_iterable
    async for item in result:
        yield str(item)
-            return
-        if hasattr(result, "__iter__") and not isinstance(result, (str, bytes)):
-            for item in result:
-                yield str(item)
-            return
-        yield str(result)
-        return
-
-    if hasattr(maybe_iterable, "__iter__") and not isinstance(maybe_iterable, (str, bytes)):
-        for item in maybe_iterable:
-            yield str(item)
-        return
-
-    yield str(maybe_iterable)


 # -------------------------
@@ -157,6 +142,7 @@ async def stream_chat_response(
    system = metadata.get("system", "")
    model = metadata.get("model") or "mistral"
    temperature = metadata.get("temperature")
+    num_gpu = metadata.get("num_gpu")

    # Build messages array for /api/chat multi-turn format
    messages: List[Dict[str, str]] = []
@@ -175,7 +161,7 @@ async def stream_chat_response(
    _synapse_trace(f"USR: {user_message}\n{'─' * 50}\n")

    try:
-        maybe_iter = manager.chat(messages=messages, model=model, stream=True, temperature=temperature)
+        maybe_iter = manager.chat(messages=messages, model=model, stream=True, temperature=temperature, num_gpu=num_gpu)
        async_gen = _normalize_to_async_generator(maybe_iter)

        buffer_parts: list[str] = []
@@ -224,19 +210,3 @@ async def stream_chat_response(
    except Exception:
        _logger.exception("stream_chat_response: unexpected error during streaming")
        raise
-
-
-# -------------------------
-# Synchronous convenience wrappers
-# -------------------------
-def generate_chat_response_sync(user_message: str, metadata: Optional[Dict[str, Any]] = None, timeout: Optional[float] = None) -> Dict[str, Any]:
-    return asyncio.run(generate_chat_response(user_message, metadata=metadata, timeout=timeout))
-
-
-def stream_chat_response_sync(user_message: str, metadata: Optional[Dict[str, Any]] = None, timeout: Optional[float] = None):
-    async def _collect():
-        chunks = []
-        async for c in stream_chat_response(user_message, metadata=metadata, timeout=timeout):
-            chunks.append(c)
-        return chunks
-    return asyncio.run(_collect())
@@ -0,0 +1,244 @@
+"""Composite app icons onto the NexusOS nexus underlay tile."""
+from pathlib import Path
+import subprocess
+import tempfile
+import os
+import re
+
+_ROOT = Path(__file__).resolve().parents[2]
+UNDERLAY_FILE = _ROOT / "assets/themes/NexusOS-icons-src/nexus-underlay.svg"
+RING_FILE = _ROOT / "assets/themes/NexusOS-icons-src/nexus-underlay-ring.svg"
+ICONS_OUT = Path.home() / ".icons" / "NexusOS"
+SIZES = [16, 22, 24, 32, 48, 64, 128]
+
+_ALLOWED_ROOTS = [
+    "/usr/share/icons",
+    "/usr/share/pixmaps",
+    "/usr/local/share/icons",
+    "/opt",
+    str(Path.home() / ".local/share/icons"),
+    str(Path.home() / ".icons"),
+    str(_ROOT / "assets"),
+]
+
+_UNDERLAY_FALLBACK = """\
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128" viewBox="0 0 128 128">
+  <defs>
+    <radialGradient id="frame" cx="50%" cy="30%" r="70%">
+      <stop offset="0%" stop-color="#b040c0"/>
+      <stop offset="100%" stop-color="#4a0050"/>
+    </radialGradient>
+  </defs>
+  <rect x="0"  y="0"  width="128" height="128" rx="30" fill="url(#frame)"/>
+  <rect x="8"  y="8"  width="112" height="112" rx="26" fill="#120018"/>
+  <rect x="16" y="16" width="96"  height="96"  rx="18" fill="#262626"/>
+</svg>
+"""
+
+_RING_FALLBACK = """\
+<?xml version="1.0" encoding="UTF-8"?>
+<svg xmlns="http://www.w3.org/2000/svg" width="128" height="128" viewBox="0 0 128 128">
+  <defs>
+    <radialGradient id="ring" cx="50%" cy="30%" r="70%">
+      <stop offset="0%" stop-color="#8cc63f"/>
+      <stop offset="100%" stop-color="#5a9020"/>
+    </radialGradient>
+    <radialGradient id="frame" cx="50%" cy="30%" r="70%">
+      <stop offset="0%" stop-color="#b040c0"/>
+      <stop offset="100%" stop-color="#4a0050"/>
+    </radialGradient>
+  </defs>
+  <rect x="0"  y="0"  width="128" height="128" rx="30" fill="url(#ring)"/>
+  <rect x="8"  y="8"  width="112" height="112" rx="26" fill="url(#frame)"/>
+  <rect x="12" y="12" width="104" height="104" rx="22" fill="#120018"/>
+  <rect x="18" y="18" width="92"  height="92"  rx="16" fill="#262626"/>
+</svg>
+"""
+
+
+def _load_underlay(nexus_ring: bool = False) -> str:
+    """Return SVG text for the underlay, falling back to inline if file missing."""
+    target = RING_FILE if nexus_ring else UNDERLAY_FILE
+    fallback = _RING_FALLBACK if nexus_ring else _UNDERLAY_FALLBACK
+    if target.exists():
+        return target.read_text()
+    return fallback
+
+
+def _inkscape_render(svg_path: str, out_path: str, size: int) -> None:
+    subprocess.run(
+        [
+            "inkscape", svg_path,
+            "--export-type=png",
+            f"--export-filename={out_path}",
+            f"--export-width={size}",
+            f"--export-height={size}",
+        ],
+        check=True,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
+
+
+def _composite_png(underlay_png: str, app_png: str, out_path: str, frac: float, size: int) -> None:
+    """Composite app icon centred on the underlay at given fractional size."""
+    app_size = max(1, int(size * frac))
+    offset = (size - app_size) // 2
+    subprocess.run(
+        [
+            "convert",
+            underlay_png,
+            "(", app_png, "-resize", f"{app_size}x{app_size}", ")",
+            "-gravity", "Center",
+            "-geometry", f"+0+0",
+            "-composite",
+            out_path,
+        ],
+        check=True,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
+
+
+def _is_allowed_path(path: str) -> bool:
+    p = Path(os.path.realpath(path))
+    return p.is_file() and any(
+        p == root or root in p.parents
+        for root in (Path(r).resolve() for r in _ALLOWED_ROOTS)
+    )
+
+
+def _safe_component(value: str, label: str) -> str:
+    if not value or not re.fullmatch(r"[A-Za-z0-9._-]+", value):
+        raise ValueError(f"Invalid icon {label}")
+    return value
+
+
+def brand_icon(
+    src_path: str,
+    output_name: str,
+    frac: float = 0.60,
+    round_mask: bool = False,
+    nexus_ring: bool = False,
+    reload: bool = True,
+    category: str = "apps",
+) -> None:
+    """Brand a single icon and write PNGs to the NexusOS icon theme at all sizes."""
+    if not _is_allowed_path(src_path):
+        raise ValueError(f"Icon source path not in allowed roots: {src_path}")
+    output_name = _safe_component(output_name, "name")
+    category = _safe_component(category, "category")
+
+    underlay_svg_text = _load_underlay(nexus_ring)
+
+    with tempfile.TemporaryDirectory() as tmp:
+        tmp_path = Path(tmp)
+
+        underlay_svg = tmp_path / "underlay.svg"
+        underlay_svg.write_text(underlay_svg_text)
+
+        for size in SIZES:
+            out_dir = ICONS_OUT / f"{size}x{size}" / category
+            out_dir.mkdir(parents=True, exist_ok=True)
+            out_path = out_dir / f"{output_name}.png"
+
+            # Break symlinks before writing; also remove any same-name .svg so
+            # GTK doesn't prefer the old SVG over our new branded PNG.
+            if out_path.is_symlink() or out_path.exists():
+                out_path.unlink()
+            svg_path = out_dir / f"{output_name}.svg"
+            if svg_path.exists() or svg_path.is_symlink():
+                svg_path.unlink()
+
+            underlay_png = str(tmp_path / f"underlay_{size}.png")
+            _inkscape_render(str(underlay_svg), underlay_png, size)
+
+            # Resize app icon source to a temp PNG for compositing.
+            # `-background none` is REQUIRED: ImageMagick rasterizes transparent
+            # SVGs onto an opaque WHITE canvas by default, which shows up as a
+            # white plate/border behind logos with transparent corners
+            # (e.g. VSCode, Edge). It must precede the input to affect the SVG.
+            app_png = str(tmp_path / f"app_{size}.png")
+            app_size = max(1, int(size * frac))
+            subprocess.run(
+                ["convert", "-background", "none", src_path,
+                 "-resize", f"{app_size}x{app_size}", app_png],
+                check=True,
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+            )
+
+            subprocess.run(
+                [
+                    "convert", underlay_png,
+                    "(", app_png, ")",
+                    "-gravity", "Center",
+                    "-composite",
+                    str(out_path),
+                ],
+                check=True,
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.DEVNULL,
+            )
+
+    if reload:
+        apply_icon_cache()
+
+
+def apply_icon_cache() -> None:
+    """Rebuild the GTK icon cache and reload the panel."""
+    import pwd
+    uid = os.getuid()
+    # systemd user session bus — needed for xfce4-panel -r to reach the running session
+    dbus_addr = os.environ.get(
+        "DBUS_SESSION_BUS_ADDRESS",
+        f"unix:path=/run/user/{uid}/bus",
+    )
+    _env = {
+        **os.environ,
+        "DISPLAY": os.environ.get("DISPLAY", ":0"),
+        "DBUS_SESSION_BUS_ADDRESS": dbus_addr,
+        "HOME": pwd.getpwuid(uid).pw_dir,
+    }
+
+    subprocess.run(
+        ["gtk-update-icon-cache", "-f", str(ICONS_OUT)],
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
+
+    # Force every running GTK app (whisker menu, panel, Thunar) to drop its
+    # in-memory icon cache. A panel reload alone does NOT do this — GTK only
+    # re-resolves icons when the icon-theme NAME changes. Toggling to another
+    # theme and back fires the "theme-changed" signal that triggers the reload.
+    import time
+    current = "NexusOS"
+    try:
+        out = subprocess.run(
+            ["xfconf-query", "-c", "xsettings", "-p", "/Net/IconThemeName"],
+            env=_env, capture_output=True, text=True,
+        )
+        if out.stdout.strip():
+            current = out.stdout.strip()
+    except Exception:
+        pass
+    alt = "Papirus-Dark" if current != "Papirus-Dark" else "Adwaita"
+    subprocess.run(
+        ["xfconf-query", "-c", "xsettings", "-p", "/Net/IconThemeName", "-s", alt],
+        env=_env, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+    )
+    time.sleep(0.5)
+    subprocess.run(
+        ["xfconf-query", "-c", "xsettings", "-p", "/Net/IconThemeName", "-s", current],
+        env=_env, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
+    )
+
+    # Restart Plank so it re-resolves icons from the updated theme
+    subprocess.run(["pkill", "plank"], env=_env, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    subprocess.Popen(
+        ["plank"],
+        env=_env,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
@@ -0,0 +1,140 @@
+"""Resolve desktop app icon paths and scan installed applications."""
+from pathlib import Path
+from typing import Optional
+import configparser
+
+_HOME = Path.home()
+
+ICON_SEARCH_DIRS = [
+    _HOME / ".local/share/icons",
+    Path("/usr/share/icons"),
+    Path("/usr/share/pixmaps"),
+    Path("/usr/local/share/icons"),
+]
+
+_THEME_ORDER = ["hicolor", "Papirus-Dark", "Papirus", "gnome", "Adwaita", "breeze"]
+_ICON_EXTS = [".png", ".svg", ".xpm"]
+
+DESKTOP_DIRS = [
+    _HOME / ".local/share/applications",
+    Path("/usr/local/share/applications"),
+    Path("/usr/share/applications"),
+]
+
+_NEXUS_ICONS = _HOME / ".icons" / "NexusOS"
+
+
+def _icon_stem(icon_name: str) -> str:
+    """Return the icon's name for theme lookup.
+
+    Only strips a real image extension (.png/.svg/.xpm). Reverse-DNS icon
+    names like ``com.visualstudio.code`` or ``org.gnome.Files`` must be kept
+    intact — Path.stem would wrongly treat the final dotted segment as an
+    extension and truncate it.
+    """
+    p = Path(icon_name)
+    if p.suffix.lower() in _ICON_EXTS:
+        return p.stem
+    return p.name
+
+
+def resolve_icon(icon_name: str, size: int = 128) -> Optional[str]:
+    """Return absolute path to an icon file, or None if not found."""
+    if not icon_name:
+        return None
+
+    # Absolute path — use directly if it exists
+    p = Path(icon_name)
+    if p.is_absolute() and p.exists():
+        return str(p)
+
+    # Strip extension for name-based search
+    stem = _icon_stem(icon_name)
+
+    size_dirs = [f"{size}x{size}", f"{size}x{size}@2x", "scalable"]
+
+    for search_root in ICON_SEARCH_DIRS:
+        if not search_root.is_dir():
+            continue
+        for theme in _THEME_ORDER:
+            theme_dir = search_root / theme
+            if not theme_dir.is_dir():
+                continue
+            for size_dir in size_dirs:
+                for ctx in ("apps", "categories", "places", "status", "actions"):
+                    ctx_dir = theme_dir / size_dir / ctx
+                    if not ctx_dir.is_dir():
+                        continue
+                    for ext in _ICON_EXTS:
+                        candidate = ctx_dir / f"{stem}{ext}"
+                        if candidate.exists():
+                            return str(candidate)
+
+    # pixmaps fallback
+    for search_root in ICON_SEARCH_DIRS:
+        if not search_root.is_dir():
+            continue
+        if search_root.name == "pixmaps":
+            for ext in _ICON_EXTS:
+                candidate = search_root / f"{stem}{ext}"
+                if candidate.exists():
+                    return str(candidate)
+
+    for ext in _ICON_EXTS:
+        candidate = Path("/usr/share/pixmaps") / f"{stem}{ext}"
+        if candidate.exists():
+            return str(candidate)
+
+    return None
+
+
+def _is_branded(icon_name: str) -> bool:
+    """Return True if a NexusOS-branded PNG already exists for this icon name."""
+    stem = _icon_stem(icon_name)
+    return (_NEXUS_ICONS / "128x128" / "apps" / f"{stem}.png").exists()
+
+
+def scan_apps() -> list[dict]:
+    """Return list of installed apps with resolved icon paths."""
+    seen: dict[str, dict] = {}
+
+    for desktop_dir in DESKTOP_DIRS:
+        if not desktop_dir.is_dir():
+            continue
+        for desktop_file in sorted(desktop_dir.glob("*.desktop")):
+            cfg = configparser.ConfigParser(interpolation=None, strict=False)
+            try:
+                cfg.read(str(desktop_file), encoding="utf-8")
+            except Exception:
+                continue
+
+            if not cfg.has_section("Desktop Entry"):
+                continue
+            de = cfg["Desktop Entry"]
+
+            if de.get("NoDisplay", "false").lower() == "true":
+                continue
+            if de.get("Hidden", "false").lower() == "true":
+                continue
+            if de.get("Type", "") != "Application":
+                continue
+
+            name = de.get("Name", desktop_file.stem)
+            icon_name = de.get("Icon", "")
+            if not icon_name:
+                continue
+
+            icon_path = resolve_icon(icon_name)
+            already_branded = _is_branded(icon_name)
+
+            key = name.lower()
+            if key not in seen:
+                seen[key] = {
+                    "name": name,
+                    "icon_name": _icon_stem(icon_name),
+                    "icon_path": icon_path,
+                    "already_branded": already_branded,
+                    "desktop_file": str(desktop_file),
+                }
+
+    return sorted(seen.values(), key=lambda x: x["name"].lower())
@@ -4,46 +4,169 @@ from __future__ import annotations
 import asyncio as _asyncio
 import json as _json
 import uuid as _uuid
+from collections import Counter as _Counter
 from typing import Any, AsyncGenerator, Dict, List, Optional, Tuple
 from uuid import UUID

 import httpx
+import os as _os
+from pathlib import Path
 from fastapi import FastAPI, HTTPException, Body
 from fastapi.middleware.cors import CORSMiddleware
-from fastapi.responses import StreamingResponse
+from fastapi.responses import StreamingResponse, FileResponse

-from .nexus_config import settings
-from .chat import generate_chat_response, stream_chat_response
+from .nexus_config import settings, VERSION
+from .chat import generate_chat_response, stream_chat_response, _synapse_trace
 from .ollama_manager import initialize_ollama, initialize_ollama_async, get_ollama_manager
 from .playbook_manager import PlaybookManager

 def _render_memory_block(facts) -> str:
-    """Render memory items as grouped ## Section / - bullet markdown."""
+    """Render memory items as grouped ## Section / - bullet markdown.
+    Leading-space indent on a fact's text is preserved so nested bullets stay nested.
+
+    Behavioural "Instructions" entries are skipped — they belong in the playbook,
+    not the "what you know about Jon" facts block. Injecting imperative directives
+    ("provide full rewrites", "include file paths") as facts pushes a small model
+    to reformat/organise the user's input instead of conversing."""
    from collections import defaultdict
    sections: dict = defaultdict(list)
    for item in facts:
+        if (item.section or "").strip().lower() == "instructions":
+            continue
        sections[item.section or "General"].append(item.text)
    parts = []
    for section, lines in sections.items():
-        parts.append(f"## {section}\n" + "\n".join(f" - {line}" for line in lines))
+        rendered = []
+        for line in lines:
+            stripped = line.lstrip(" ")
+            indent = len(line) - len(stripped)
+            rendered.append(" " * indent + f" - {stripped}")
+        parts.append(f"## {section}\n" + "\n".join(rendered))
    return "\n\n".join(parts)


-async def _auto_select_model() -> str:
+# Preamble wrapped around the injected memory facts. It stays anti-recite for
+# everyday chat, but explicitly permits surfacing facts when Jon asks about
+# himself / to be quizzed — the old absolute "do NOT acknowledge these facts"
+# made small models play dumb on exactly that request.
+_MEMORY_PREAMBLE = (
+    "\n\n---\nBackground on Jon — use this to personalize your replies. Don't dump "
+    "or recite these facts unprompted, but when Jon asks about himself or asks you "
+    "to recall or quiz what you know, use them directly and specifically:\n\n"
+)
+
+
+_CODING_KEYWORDS = frozenset({
+    "code", "coding", "function", "class", "method", "variable", "bug", "error",
+    "debug", "fix", "refactor", "script", "program", "syntax", "compile", "import",
+    "module", "library", "algorithm", "loop", "array", "string", "integer", "boolean",
+    "return", "def", "const", "let", "var", "test", "api", "endpoint", "database",
+    "query", "sql", "bash", "terminal", "command", "package", "dependency",
+    "python", "javascript", "typescript", "rust", "golang", "java", "html", "css",
+    ".py", ".js", ".ts", ".jsx", ".tsx", ".sh", ".json", ".yaml", ".sql", ".css",
+})
+
+def _detect_intent(message: str) -> str:
+    lower = message.lower()
+    return "code" if any(kw in lower for kw in _CODING_KEYWORDS) else "chat"
+
+
+import re as _re
+
+def _route_playbooks(message: str, candidates: list) -> list:
+    """Return the reference playbook(s) whose tags appear directly in the user's
+    message. Returns [] when nothing matches, so casual chat doesn't drag in a
+    specialist playbook.
+
+    Dropped an old memory-fallback tier that, when the message matched no tags,
+    scored playbook tags against the user's WHOLE memory corpus. Because memory
+    permanently mentions e.g. the home network, that injected the Home Network
+    playbook (~1.7k chars) into unrelated chats. Direct references like "my truck"
+    already match here ('truck' is a Ford tag), so the fallback was mostly noise.
+    """
+    if not candidates or not message:
+        return candidates
+
+    msg_tokens = set(_re.findall(r'\b\w+\b', message.lower()))
+    scores = [(sum(1 for tag in pb.tags if tag.lower() in msg_tokens), pb) for pb in candidates]
+    best = max(s for s, _ in scores)
+    if best > 0:
+        return [pb for s, pb in scores if s == best]
+    return []
+
+async def _auto_select_model(message: str = "") -> str:
+    """A pinned settings.model wins; otherwise pick the preferred installed model
+    for the detected intent. The preference lists live in one place now —
+    ollama_manager._MODEL_PREFERENCE, via select_best_model(intent)."""
    try:
        s = store.get_settings()
        if s.get("model"):
            return s["model"]
-        return await get_ollama_manager().select_best_model()
+        intent = _detect_intent(message) if message else "chat"
+        return await get_ollama_manager().select_best_model(intent)
    except Exception:
        return getattr(settings, "default_model", None) or "mistral"

+
+
+_TITLE_SYSTEM_PROMPT = (
+    "You generate a short, descriptive title for a chat conversation based on the "
+    "user's first message. Reply with ONLY the title: 3 to 6 words, no quotes, no "
+    "trailing punctuation, no preamble. Use plain text in title case. The title "
+    "names the TOPIC — never echo the user's question or phrase it as a question.\n\n"
+    "Examples:\n"
+    "Message: can you help me fix a bug in my python script?\n"
+    "Title: Python Script Bug Fix\n"
+    "Message: what's a good recipe for sourdough bread?\n"
+    "Title: Sourdough Bread Recipe\n"
+    "Message: i want to try out your memory, ask me questions about myself\n"
+    "Title: Testing Memory Recall"
+)
+
+
+async def _generate_conversation_title(first_message: str, model: str) -> Optional[str]:
+    """Ask the LLM for a concise title. Titling is a background 'curation' task,
+    so it runs on the CURATOR model (mistral) rather than the chat model: the 7B
+    follows the terse title format better than the 3B chat model, and it's already
+    warm in RAM. Crucially we pass the curator's own num_gpu (CPU) so we hit that
+    warm CPU-resident instance — same memory pool, no reload, and the GPU chat
+    model is never disturbed. `model` is only a fallback if no curator is set.
+    Best-effort: returns None on any failure so titling never breaks the chat."""
+    snippet = first_message.strip()[:1000]
+    if not snippet:
+        return None
+    try:
+        s = store.get_settings()
+        title_model = s.get("memory_model") or model
+        num_gpu = await get_ollama_manager().resolve_num_gpu(s.get("memory_gpu_offload", 0), title_model)
+        result = await generate_chat_response(
+            user_message=snippet,
+            metadata={"system": _TITLE_SYSTEM_PROMPT, "model": title_model,
+                      "temperature": 0.2, "num_gpu": num_gpu},
+            timeout=30,
+        )
+        title = (result.get("response") or "").strip()
+        # Strip stray quotes/wrapping the model sometimes adds, collapse whitespace.
+        title = title.strip().strip('"').strip("'").splitlines()[0].strip()
+        # Drop a "Title:" prefix the model may echo from the examples, and strip
+        # trailing punctuation the prompt forbids but small models still add.
+        if title.lower().startswith("title:"):
+            title = title[len("title:"):].strip()
+        title = " ".join(title.split()).rstrip("?.!,;:")
+        if not title:
+            return None
+        return title[:120]
+    except Exception:
+        return None
+
+
 from .memory.store import store, MemoryItem
 from .playbooks.store import playbook_store, PlaybookItem
+from .search import needs_web_search, web_search

 MEMORY_SERVICE = "http://localhost:8001"

-app = FastAPI(title="Synapse Backend", version="1.0")
+app = FastAPI(title="Synapse Backend", version=VERSION)

 # Alias for startup scripts
 sio_app = app
@@ -73,6 +196,10 @@ async def startup_event():
    global ollama
    try:
        ollama = await initialize_ollama_async()
+        # Keep the model resident per the persisted setting, then preload the
+        # model the first chat would pick so that message doesn't pay a cold load.
+        ollama.keep_alive = store.get_settings().get("keep_alive") or ollama.keep_alive
+        _asyncio.create_task(ollama.warm(await _auto_select_model()))
        print("[Synapse] Ollama service is running.")
    except Exception as e:
        # Start in degraded mode — chat endpoints will return errors until Ollama
@@ -92,44 +219,7 @@ async def root():
        status = ollama.get_status() if (ollama is not None and hasattr(ollama, "get_status")) else None
    except Exception:
        status = None
-    return {"status": "online", "ollama": status}
-
-
-# -------------------------
-# Chat (non-streaming)
-# -------------------------
-@app.post("/chat")
-async def chat_endpoint(payload: Dict[str, Any]):
-    try:
-        message = payload.get("message", "")
-        app_settings = store.get_settings()
-        model = payload.get("model") or await _auto_select_model()
-        context = payload.get("context", {})
-        history = payload.get("history", [])
-        temperature = payload.get("temperature", app_settings.get("temperature"))
-
-        if not message:
-            raise HTTPException(status_code=400, detail="Missing 'message'")
-
-        rendered_message = playbooks.render_prompt(message)
-        system_prompt = playbooks.get_system_prompt() or app_settings.get("system_prompt", "")
-
-        memory_facts = store.all()
-        if memory_facts:
-            facts_block = _render_memory_block(memory_facts)
-            system_prompt = (system_prompt + "\n\n---\nWhat you know about Jon:\n\n" + facts_block) if system_prompt else facts_block
-
-        metadata: Dict[str, Any] = {"model": model, "context": context, "system": system_prompt, "temperature": temperature}
-
-        result = await generate_chat_response(
-            user_message=rendered_message, metadata=metadata, history=history
-        )
-        return {"response": result.get("response", ""), "model": model}
-
-    except HTTPException:
-        raise
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
+    return {"status": "online", "version": VERSION, "ollama": status}


 # -------------------------
@@ -140,20 +230,25 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
    try:
        message = payload.get("message", "")
        app_settings = store.get_settings()
-        model = payload.get("model") or await _auto_select_model()
+        model = payload.get("model") or await _auto_select_model(message)
        context = payload.get("context", {})
        conversation_id = payload.get("conversation_id") or str(_uuid.uuid4())
        history = payload.get("history", [])
        temperature = payload.get("temperature", app_settings.get("temperature"))
+        gpu_offload = payload.get("gpu_offload", app_settings.get("gpu_offload", -1))
+        num_gpu = await get_ollama_manager().resolve_num_gpu(gpu_offload, model)

        if not message:
            raise HTTPException(status_code=400, detail="Missing 'message'")

-        rendered_message = playbooks.render_prompt(message)
+        rendered_message = message  # chat has no template vars; render_prompt is for the playbook path
        system_prompt = playbooks.get_system_prompt() or app_settings.get("system_prompt", "")

-        # Append reference playbooks to the system prompt
-        context_pbs = playbooks.get_context_playbooks()
+        # Fetch memory facts once — used for both playbook routing and system prompt injection
+        memory_facts = store.all()
+
+        # Append the best-matching reference playbook(s) to the system prompt
+        context_pbs = _route_playbooks(rendered_message, playbooks.get_context_playbooks())
        if context_pbs:
            refs = "\n\n".join(
                f"### {pb.title}\nGoal: {pb.goal}\n\n{pb.instructions}"
@@ -161,16 +256,17 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
            )
            separator = "\n\n---\nReference playbooks (read these as additional context):\n\n"
            system_prompt = (system_prompt + separator + refs) if system_prompt else refs
-
-        # Inject persistent memory facts about Jon
-        memory_facts = store.all()
        if memory_facts:
            facts_block = _render_memory_block(memory_facts)
-            system_prompt = (system_prompt + "\n\n---\nWhat you know about Jon:\n\n" + facts_block) if system_prompt else facts_block
+            system_prompt = (system_prompt + _MEMORY_PREAMBLE + facts_block) if system_prompt else facts_block

        # Search past conversations for relevant context and inject the top matches.
        # This gives the model memory of prior exchanges without requiring tool-calling support.
-        past_context = store.search_conversations(message, limit=2)
+        # Semantic recall (embeddings) finds relevant exchanges even without shared
+        # keywords; it falls back to lexical substring match if embeddings are down.
+        past_context = await store.semantic_search_conversations(
+            message, get_ollama_manager().embed, limit=2
+        )
        if past_context:
            snippets = []
            for conv in past_context:
@@ -183,7 +279,65 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
            separator = "\n\n---\nRelevant past exchanges (use as background context only):\n\n"
            system_prompt = (system_prompt + separator + memory_block) if system_prompt else memory_block

-        metadata: Dict[str, Any] = {"model": model, "context": context, "system": system_prompt, "temperature": temperature}
+        # Fetch web search results for time-sensitive queries
+        search_results = ""
+        if needs_web_search(message):
+            search_results = await _asyncio.to_thread(web_search, message)
+            if search_results:
+                separator = "\n\n---\nWeb search results (treat as current information):\n\n"
+                system_prompt = (system_prompt + separator + search_results) if system_prompt else search_results
+
+        # ── MindTrace pre-flight ──────────────────────────────────────────
+        _trace_intent = _detect_intent(message) if message else "chat"
+        if payload.get("model"):
+            _trace_src = "user-override"
+        elif store.get_settings().get("model"):
+            _trace_src = "settings"
+        else:
+            _trace_src = f"auto/{_trace_intent}"
+
+        _synapse_trace(f"\n{'═' * 55}\n")
+        _synapse_trace(f"▶ MODEL : {model}  [{_trace_src}]\n")
+
+        if _trace_intent == "code":
+            _kws = [kw for kw in _CODING_KEYWORDS if kw in message.lower()][:5]
+            _synapse_trace(f"  INTENT: code  →  {', '.join(_kws)}\n")
+        else:
+            _synapse_trace(f"  INTENT: chat\n")
+
+        _main_pb = playbooks.get_main_playbook()
+        if _main_pb:
+            _synapse_trace(f"  PLAYBOOK: {_main_pb.title}\n")
+            if _main_pb.goal:
+                _synapse_trace(f"    goal: {_main_pb.goal[:100]}\n")
+        else:
+            _synapse_trace(f"  PLAYBOOK: none\n")
+
+        if context_pbs:
+            _synapse_trace(f"  ROUTED  : {', '.join(pb.title for pb in context_pbs)}\n")
+        else:
+            _synapse_trace(f"  ROUTED  : none (no tag match)\n")
+
+        if search_results:
+            _synapse_trace(f"  SEARCH  : {len(search_results)} chars injected\n")
+        elif needs_web_search(message):
+            _synapse_trace(f"  SEARCH  : triggered but returned no results\n")
+
+        if memory_facts:
+            _secs = _Counter(f.section or "General" for f in memory_facts)
+            _sec_str = "  ".join(f"{s}({n})" for s, n in _secs.items())
+            _synapse_trace(f"  MEMORY  : {len(memory_facts)} facts  [{_sec_str}]\n")
+        else:
+            _synapse_trace(f"  MEMORY  : none\n")
+
+        if past_context:
+            _synapse_trace(f"  CONTEXT : {len(past_context)} past conversation match(es) injected\n")
+
+        _synapse_trace(f"  SYS LEN : {len(system_prompt)} chars\n")
+        _synapse_trace(f"{'─' * 55}\n")
+        # ── end MindTrace pre-flight ──────────────────────────────────────
+
+        metadata: Dict[str, Any] = {"model": model, "context": context, "system": system_prompt, "temperature": temperature, "num_gpu": num_gpu}

        # Persist conversation and user message before streaming
        store.create_conversation(conversation_id)
@@ -192,6 +346,9 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
        async def event_stream() -> AsyncGenerator[str, None]:
            response_chunks: list[str] = []
            meta: dict = {}
+            final_model = model
+
+            # ── Phase 1: stream primary model response ────────────────────
            try:
                async for chunk in stream_chat_response(
                    user_message=rendered_message,
@@ -206,22 +363,51 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
                        yield f"event: meta\ndata: {chunk[8:]}\n\n"
                        continue
                    response_chunks.append(chunk)
-                    yield f"data: {chunk}\n\n"
-                # Persist the completed assistant response with model and token stats
+                    yield f"data: {_json.dumps(chunk)}\n\n"
+            except _asyncio.TimeoutError:
+                _tval = store.get_settings().get("timeout", 120)
+                yield f"event: error\ndata: {_json.dumps({'detail': f'Model timed out after {_tval}s — try a smaller/faster model'})}\n\n"
+                return
+            except Exception as e:
+                detail = str(e) or type(e).__name__
+                yield f"event: error\ndata: {_json.dumps({'detail': detail})}\n\n"
+                return
+
+            # ── Persist completed response ───────────────────────────────
            if response_chunks:
                store.add_message(
                    conversation_id, "assistant", "".join(response_chunks),
-                        model=meta.get("model") or model,
+                    model=meta.get("model") or final_model,
                    tokens=meta.get("tokens"),
                )
-            except Exception as e:
-                yield f"event: error\ndata: {_json.dumps({'detail': str(e)})}\n\n"
-                return
+
+            # Response is complete — let the client re-enable its input now,
+            # so the slow title/memory work below doesn't freeze the UI.
+            yield "event: done\ndata: {}\n\n"
+
+            # Generate an AI title from the opening message. Retried on any turn
+            # while still untitled, so an interrupted first stream can recover.
+            if response_chunks:
+                try:
+                    conv = store.get_conversation(conversation_id)
+                    if conv and not conv.title:
+                        first_user = next(
+                            (m.content for m in conv.messages if m.role == "user"),
+                            rendered_message,
+                        )
+                        title = await _generate_conversation_title(
+                            first_user, meta.get("model") or final_model
+                        )
+                        if title:
+                            store.set_conversation_title(conversation_id, title)
+                            yield f"event: title\ndata: {_json.dumps({'title': title})}\n\n"
+                except Exception:
+                    pass

            # Ask the memory service curator to evaluate this exchange
            if response_chunks:
                try:
-                    async with httpx.AsyncClient(timeout=25.0) as _mc:
+                    async with httpx.AsyncClient(timeout=310.0) as _mc:
                        r = await _mc.post(
                            f"{MEMORY_SERVICE}/memories/extract",
                            json={
@@ -231,11 +417,11 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
                        )
                    if r.status_code == 200:
                        data = r.json()
-                        if data.get("saved"):
-                            mem_result = {"section": data["section"], "text": data["text"]}
+                        for it in data.get("items", []):
+                            mem_result = {"section": it["section"], "text": it["text"]}
                            yield f"event: memory\ndata: {_json.dumps(mem_result)}\n\n"
-                except Exception:
-                    pass
+                except Exception as e:
+                    _synapse_trace(f"\n⚠ memory extraction call failed: {e}\n")

        return StreamingResponse(event_stream(), media_type="text/event-stream")

@@ -244,24 +430,6 @@ async def chat_stream_endpoint(payload: Dict[str, Any]):
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

-# -------------------------
-# Playbook Execution
-# -------------------------
-@app.post("/playbook/run")
-async def run_playbook(payload: Dict[str, Any]):
-    name = payload.get("name")
-    variables = payload.get("variables", {})
-
-    if not name:
-        raise HTTPException(status_code=400, detail="Missing playbook name")
-
-    try:
-        result = playbooks.render_prompt(name, variables=variables)
-        return {"result": result}
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
 # -------------------------
 # Settings
 # -------------------------
@@ -273,7 +441,10 @@ async def get_settings_endpoint():
@app.put("/settings")
 async def put_settings_endpoint(payload: Dict[str, Any] = Body(...)):
    store.update_settings(payload)
-    return store.get_settings()
+    merged = store.get_settings()
+    if "keep_alive" in payload:
+        get_ollama_manager().keep_alive = merged.get("keep_alive") or None
+    return merged


 # -------------------------
@@ -286,8 +457,8 @@ async def get_memory():

@app.post("/memory")
 async def add_memory(payload: Dict[str, Any] = Body(...)):
-    text = (payload.get("text") or "").strip()
-    if not text:
+    text = (payload.get("text") or "").rstrip()
+    if not text.strip():
        raise HTTPException(status_code=400, detail="Missing 'text'")
    from .memory.store import MemoryItem
    import uuid as _mem_uuid
@@ -310,7 +481,7 @@ async def update_memory(item_id: str, payload: Dict[str, Any] = Body(...)):
    updated = MemoryItem(
        id=item_id,
        section=(payload.get("section") or existing.section or "General").strip(),
-        text=(payload.get("text") or existing.text).strip(),
+        text=(payload.get("text") or existing.text).rstrip(),
        tags=payload.get("tags", existing.tags),
    )
    store.update(updated)
@@ -325,6 +496,16 @@ async def delete_memory(item_id: str):
    return {"status": "deleted"}


+@app.post("/memory/reorder")
+async def reorder_memory(payload: Dict[str, Any] = Body(...)):
+    section = (payload.get("section") or "General").strip() or "General"
+    ids = payload.get("ids")
+    if not isinstance(ids, list) or not all(isinstance(x, str) for x in ids):
+        raise HTTPException(status_code=400, detail="ids must be a list of strings")
+    ok = store.reorder_section(section, ids)
+    return {"ok": ok}
+
+
 # -------------------------
 # Models
 # -------------------------
@@ -332,8 +513,11 @@ async def delete_memory(item_id: str):
 async def get_models():
    try:
        mgr = get_ollama_manager()
-        models = await mgr.list_models()
-        selected = await mgr.select_best_model()
+        # Embedding models (e.g. nomic-embed-text) can't chat — hide from picker.
+        models = [m for m in await mgr.list_models() if "embed" not in m.lower()]
+        # Report the SAME model the chat path would auto-pick (honors a pin),
+        # so the picker's "Auto (…)" label matches what actually answers.
+        selected = await _auto_select_model()
        return {"models": models, "selected": selected}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))
@@ -641,26 +825,6 @@ async def delete_playbook_endpoint(id: UUID):
 # -------------------------
 # Unified Search
 # -------------------------
-@app.get("/search")
-async def search_endpoint(q: Optional[str] = None):
-    """Search conversations and playbooks by keyword."""
-    if not q or not q.strip():
-        raise HTTPException(status_code=400, detail="Missing query parameter 'q'")
-    try:
-        conversations = store.search_conversations(q, limit=5)
-        playbook_hits = playbook_store.search_playbooks(q)
-        return {
-            "query": q,
-            "conversations": conversations,
-            "playbooks": [
-                {"id": p.id, "title": p.title, "goal": p.goal, "tags": p.tags}
-                for p in playbook_hits
-            ],
-        }
-    except Exception as e:
-        raise HTTPException(status_code=500, detail=str(e))
-
-
 # -------------------------
 # Conversations
 # -------------------------
@@ -682,6 +846,7 @@ async def get_conversations(q: Optional[str] = None):
                    "timestamp": c.created_at,
                    "updated_at": c.updated_at,
                    "preview": c.preview,
+                    "title": c.title,
                }
                for c in conversations
            ]
@@ -690,6 +855,47 @@ async def get_conversations(q: Optional[str] = None):
        raise HTTPException(status_code=500, detail=str(e))


+@app.get("/conversations/export")
+async def export_conversations(min_turns: int = 1):
+    """Export conversations as ShareGPT JSONL for fine-tuning.
+
+    Each line is one conversation:
+      {"conversations": [{"from": "human", "value": "..."}, {"from": "gpt", "value": "..."}]}
+
+    Query params:
+      min_turns  — minimum user/assistant exchanges to include (default 1)
+    """
+    from fastapi.responses import Response
+    import datetime
+
+    conversations = store.all_conversations()
+    lines = []
+
+    for conv in conversations:
+        msgs = [m for m in conv.messages if m.role in ("user", "assistant")]
+        if len(msgs) < 2:
+            continue
+        if sum(1 for m in msgs if m.role == "user") < min_turns:
+            continue
+
+        sharegpt_msgs = [
+            {"from": "human" if m.role == "user" else "gpt", "value": m.content}
+            for m in msgs
+        ]
+        lines.append(_json.dumps({"conversations": sharegpt_msgs}))
+
+    date_str = datetime.date.today().isoformat()
+    filename = f"nexus-conversations-{date_str}.jsonl"
+    return Response(
+        content="\n".join(lines),
+        media_type="application/x-ndjson",
+        headers={
+            "Content-Disposition": f"attachment; filename={filename}",
+            "X-Exported-Count": str(len(lines)),
+        },
+    )
+
+
@app.get("/conversations/{conversation_id}")
 async def get_conversation(conversation_id: str):
    try:
@@ -699,6 +905,7 @@ async def get_conversation(conversation_id: str):
        return {
            "id": conv.id,
            "timestamp": conv.created_at,
+            "title": conv.title,
            "messages": [
                {"role": m.role, "content": m.content, "timestamp": m.timestamp, "model": m.model, "tokens": m.tokens}
                for m in conv.messages
@@ -710,6 +917,25 @@ async def get_conversation(conversation_id: str):
        raise HTTPException(status_code=500, detail=str(e))


+@app.patch("/conversations/{conversation_id}")
+async def rename_conversation(conversation_id: str, payload: Dict[str, Any] = Body(...)):
+    """Manually set a conversation's title."""
+    try:
+        conv = store.get_conversation(conversation_id)
+        if not conv:
+            raise HTTPException(status_code=404, detail="Conversation not found")
+        title = (payload.get("title") or "").strip()
+        if not title:
+            raise HTTPException(status_code=400, detail="Missing 'title'")
+        title = " ".join(title.split())[:120]
+        store.set_conversation_title(conversation_id, title)
+        return {"id": conversation_id, "title": title}
+    except HTTPException:
+        raise
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
@app.delete("/conversations/{conversation_id}")
 async def delete_conversation(conversation_id: str):
    try:
@@ -723,6 +949,80 @@ async def delete_conversation(conversation_id: str):
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

+# ── Icon branding routes ──────────────────────────────────────────────────────
+
+_REPO_ASSETS = str(Path(__file__).resolve().parents[1] / "assets")
+_ALLOWED_ICON_ROOTS = [
+    "/usr/share/icons",
+    "/usr/share/pixmaps",
+    "/usr/local/share/icons",
+    "/opt",
+    _os.path.expanduser("~/.local/share/icons"),
+    _os.path.expanduser("~/.icons"),
+    _REPO_ASSETS,
+]
+
+
+@app.get("/icons/apps")
+async def list_icon_apps():
+    """Return all installed applications with their icon paths."""
+    try:
+        from .icons.resolver import scan_apps
+        loop = _asyncio.get_event_loop()
+        apps = await loop.run_in_executor(None, scan_apps)
+        return {"apps": apps}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.get("/icons/image")
+async def get_icon_image(path: str):
+    """Serve an icon file after verifying it's in an allowed root."""
+    real = _os.path.realpath(path)
+    if not any(real.startswith(r) for r in _ALLOWED_ICON_ROOTS):
+        raise HTTPException(status_code=403, detail="Path not allowed")
+    if not _os.path.isfile(real):
+        raise HTTPException(status_code=404, detail="Icon not found")
+    return FileResponse(real)
+
+
+@app.post("/icons/brand")
+async def brand_app_icon(payload: Dict[str, Any] = Body(...)):
+    """Composite an app icon onto the NexusOS underlay tile."""
+    src_path = payload.get("src_path", "")
+    output_name = payload.get("output_name", "")
+    if not src_path or not output_name:
+        raise HTTPException(status_code=400, detail="src_path and output_name required")
+    frac = float(payload.get("frac", 0.60))
+    round_mask = bool(payload.get("round_mask", False))
+    nexus_ring = bool(payload.get("nexus_ring", False))
+    reload = bool(payload.get("reload", True))
+    category = str(payload.get("category", "apps"))
+    try:
+        from .icons.compositor import brand_icon
+        loop = _asyncio.get_event_loop()
+        await loop.run_in_executor(
+            None,
+            lambda: brand_icon(src_path, output_name, frac, round_mask, nexus_ring, reload, category),
+        )
+        return {"status": "ok", "output_name": output_name}
+    except ValueError as e:
+        raise HTTPException(status_code=403, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
+@app.post("/icons/apply")
+async def apply_icon_cache_route():
+    """Rebuild the GTK icon cache and reload the panel."""
+    try:
+        from .icons.compositor import apply_icon_cache
+        loop = _asyncio.get_event_loop()
+        await loop.run_in_executor(None, apply_icon_cache)
+        return {"status": "ok"}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
 # -------------------------
 # End of file
 # -------------------------
@@ -6,42 +6,52 @@ from __future__ import annotations
 import json
 import logging
 import re
-from typing import Optional
+
+from ..chat import _synapse_trace  # append curator reasoning to the same MindTrace log

 _log = logging.getLogger(__name__)

+_TR = "┅" * 55
+
 _PROMPT = """\
 You are a memory curator for a personal AI assistant named Nexus.

-A conversation just occurred. Decide if the USER revealed a NEW, PERMANENT \
-personal fact that should be saved to long-term memory.
+Extract EVERY new, permanent personal fact the USER revealed in this exchange.
+There may be SEVERAL facts in one message — output one JSON object for each.

-SAVE ONLY facts that are stable and biographical:
- Identity: full name, age, location, nationality
- Possessions: vehicle, home, devices
- Relationships: family members, partner, close friends
- Career: job title, employer, field, skills
- Long-running projects or goals (not today's to-dos)
- Durable preferences or habits explicitly stated
+SAVE facts that are stable and biographical, such as: identity (name, age,
+location), relationships (family, partner, friends), pets, possessions (vehicles,
+home, devices), career (job, employer, skills), hobbies and interests,
+long-running projects or goals (not today's to-dos), and durable preferences.

 DO NOT SAVE — these are ephemeral and would clutter memory:
- Anything time-bound: "today I have...", "I'm working on X today", "I have a full day of work"
+- Anything time-bound: "today I have...", "I'm working on X today"
 - Mood or energy: "I'm tired", "feeling good", "having a rough day"
 - Greetings or small talk: "good morning", "how are you"
 - Questions the user asked the assistant
- Near-duplicates of anything in the existing memory list below

-Existing memory (do not re-save these or close variants):
+The existing memory is below FOR CONTEXT. If the user ADDS NEW DETAIL to
+something already known (e.g. a new detail about a known pet, car, or project),
+DO save that new detail as its own fact. Only skip a fact that is an EXACT
+restatement of one already listed.
+
+Existing memory:
 {existing_texts}

+For "section", REUSE one of these existing section names whenever it fits:
+{existing_sections}
+Only invent a new section if none fit, and make it a SHORT single word
+(e.g. Hobbies, Pets, Health). Never use a sentence or long phrase as a section.
+
 ---
 USER: {user_message}
 ASSISTANT: {assistant_response}
 ---

-Respond with JSON only — no prose, no markdown fences:
-{{"save": true, "section": "<section name, or one from the list above>", "text": "<concise fact about Jon, third person>"}}
-OR
+Respond with JSON only — no prose, no markdown fences. Output one object PER
+new fact (several objects, one per line, if there are several):
+{{"save": true, "section": "<short section name>", "text": "<concise fact about Jon, third person>"}}
+If there is nothing new to save, output exactly:
 {{"save": false}}"""


@@ -51,31 +61,62 @@ async def extract_memory(
    existing_sections: list[str],
    existing_texts: list[str],
    ollama_manager,
-) -> Optional[dict]:
-    """Ask Mistral to extract a saveable memory fact from a conversation exchange.
+    model: str = "mistral:latest",
+    num_gpu: int | None = 0,
+) -> list[dict]:
+    """Ask Mistral to extract saveable memory facts from a conversation exchange.

-    Returns {"section": ..., "text": ...} or None.
+    Returns a list of {"section": ..., "text": ...} — possibly empty. A single
+    exchange can hold several facts, and Mistral emits one JSON object per fact.
    """
    if existing_texts:
-        texts_block = "\n".join(f"- {t}" for t in existing_texts[:40])
+        # ponytail: only the 12 most-recent facts go in the dedup context, not all
+        # ~40. On a CPU-bound curator (num_gpu=0) prompt-eval dominates, and 40
+        # facts made a ~1200-token prompt that took ~50s+ to process. If dedup
+        # starts re-saving older facts, move dedup to a difflib check in the
+        # service instead of stuffing every fact into the prompt.
+        texts_block = "\n".join(f"- {t}" for t in existing_texts[-12:])
    else:
        texts_block = "(none yet)"
+    # Give Mistral the real section names to reuse, so it stops inventing
+    # sentence-long sections out of the category descriptions in the prompt.
+    # Only offer SHORT, clean names — never feed a junk sentence-section (e.g. a
+    # past bad "Long-running projects or goals") back as a valid choice.
+    clean = sorted(s for s in existing_sections if s and len(s.split()) <= 2 and len(s) <= 24)
+    sections_line = ", ".join(clean) if clean else (
+        "Identity, Relationships, Pets, Possessions, Career, Hobbies, Projects, Preferences"
+    )
    prompt = _PROMPT.format(
        existing_texts=texts_block,
-        user_message=user_message[:800],
+        existing_sections=sections_line,
+        user_message=user_message[:3000],
        assistant_response=assistant_response[:800],
    )
+    # MindTrace: curator pre-flight (full prompt) so its reasoning is visible in
+    # the same console as the frontline model, not just Python warnings on failure.
+    _synapse_trace(
+        f"\n{_TR}\n◆ CURATOR: {model} (num_gpu={num_gpu})\n"
+        f"  PROMPT ({len(prompt)} chars):\n{prompt}\n{_TR}\n"
+    )
    try:
+        # num_gpu=0 (the default) pins the curator fully in system RAM instead of
+        # the GPU, so it coexists with the GPU-resident chat model instead of
+        # evicting it. Without this, on a small GPU the two thrash: every
+        # exchange cold-loads the curator (~45s) and extraction times out,
+        # silently saving nothing. Boxes with spare VRAM override via settings.
        response = await ollama_manager.chat(
            messages=[{"role": "user", "content": prompt}],
-            model="mistral:latest",
+            model=model,
            stream=False,
            temperature=0.0,
+            num_gpu=num_gpu,
        )
        if not response:
-            return None
+            _synapse_trace("◆ CURATOR RAW: (empty response)\n")
+            return []

        text = response.strip()
+        _synapse_trace(f"◆ CURATOR RAW:\n{text}\n")

        # Strip markdown code fences if the model added them
        if "```" in text:
@@ -83,12 +124,45 @@ async def extract_memory(
            if m:
                text = m.group(1).strip()

-        data = json.loads(text)
-        if data.get("save") and data.get("section") and data.get("text"):
-            return {
-                "section": str(data["section"]).strip(),
-                "text": str(data["text"]).strip(),
-            }
+        # For several facts Mistral is inconsistent: sometimes ONE JSON object
+        # per fact newline-separated, sometimes a single JSON ARRAY of objects.
+        # raw_decode pulls each top-level value (handles the newline case and
+        # plain "Extra data"); we then flatten any array so both shapes save all
+        # facts. Plain json.loads() would die on the newline case and skip the
+        # array (a list isn't a dict), losing every fact either way.
+        results: list[dict] = []
+
+        def _keep(o):
+            if isinstance(o, dict) and o.get("save") and o.get("section") and o.get("text"):
+                results.append({
+                    "section": str(o["section"]).strip(),
+                    "text": str(o["text"]).strip(),
+                })
+
+        dec = json.JSONDecoder()
+        idx = 0
+        while idx < len(text):
+            while idx < len(text) and text[idx] in " \t\r\n,":
+                idx += 1
+            if idx >= len(text):
+                break
+            try:
+                obj, idx = dec.raw_decode(text, idx)
+            except json.JSONDecodeError:
+                break
+            if isinstance(obj, list):
+                for o in obj:
+                    _keep(o)
+            else:
+                _keep(obj)
+        if not results:
+            _log.warning("memory: nothing saved. mistral said: %.300r", text)
+            _synapse_trace(f"◆ CURATOR VERDICT: nothing to save\n{_TR}\n\n")
+        else:
+            _facts = "; ".join(f"[{r['section']}] {r['text']}" for r in results)
+            _synapse_trace(f"◆ CURATOR VERDICT: {len(results)} fact(s) — {_facts}\n{_TR}\n\n")
+        return results
    except Exception as e:
-        _log.debug("memory extraction failed: %s", e)
-    return None
+        _log.warning("memory extraction failed: %s", e)
+        _synapse_trace(f"◆ CURATOR ERROR: {e}\n{_TR}\n\n")
+    return []
@@ -13,6 +13,7 @@ Endpoints:
 from __future__ import annotations

 import asyncio
+import math
 import uuid
 from typing import Any, Dict, List, Optional

@@ -35,6 +36,33 @@ app.add_middleware(
 )


+@app.on_event("startup")
+async def _warm_curator():
+    """Preload the curator model (in RAM, num_gpu=0 by default) so the first
+    extraction isn't a cold load that blows the timeout. Runs in the background
+    so it never delays startup. keep_alive then holds it warm between messages."""
+    async def _bg():
+        try:
+            mgr = get_ollama_manager()
+            # Ollama is started by the Synapse backend (a separate process), so
+            # at our startup it usually isn't reachable yet. Wait for it before
+            # warming instead of failing with "All connection attempts failed" —
+            # which leaves the curator cold and makes the first extraction slow.
+            for _ in range(60):  # up to ~2 min
+                if await asyncio.to_thread(mgr.is_running):
+                    break
+                await asyncio.sleep(2)
+            else:
+                return
+            settings = store.get_settings()
+            model = settings.get("memory_model") or await mgr.select_best_model()
+            num_gpu = await mgr.resolve_num_gpu(settings.get("memory_gpu_offload", 0), model)
+            await mgr.warm(model, num_gpu=num_gpu)
+        except Exception:
+            pass
+    asyncio.create_task(_bg())
+
+
@app.get("/")
 async def health():
    return {"status": "ok", "count": len(store.all())}
@@ -86,6 +114,13 @@ async def delete_memory(item_id: str):
    return {"status": "deleted"}


+def _cosine(a: list, b: list) -> float:
+    dot = sum(x * y for x, y in zip(a, b))
+    na = math.sqrt(sum(x * x for x in a))
+    nb = math.sqrt(sum(y * y for y in b))
+    return dot / (na * nb) if na and nb else 0.0
+
+
 class ExtractRequest(BaseModel):
    user_message: str
    assistant_response: str
@@ -99,25 +134,73 @@ async def extract_and_save(req: ExtractRequest):
    existing_sections = list({i.section for i in existing})
    existing_texts = [i.text for i in existing]

+    # Adaptable per-machine curator config (see store _SETTINGS_DEFAULTS):
+    # which model does extraction, and whether it runs on CPU/RAM or the GPU.
+    settings = store.get_settings()
+    mgr = get_ollama_manager()
+    model = settings.get("memory_model") or await mgr.select_best_model()
+    num_gpu = await mgr.resolve_num_gpu(settings.get("memory_gpu_offload", 0), model)
    try:
-        result = await asyncio.wait_for(
+        merge_threshold = float(settings.get("memory_merge_threshold", 0.88))
+    except (TypeError, ValueError):
+        merge_threshold = 0.88
+
+    try:
+        results = await asyncio.wait_for(
            extract_memory(
                req.user_message,
                req.assistant_response,
                existing_sections,
                existing_texts,
-                get_ollama_manager(),
+                mgr,
+                model=model,
+                num_gpu=num_gpu,
            ),
-            timeout=20.0,
-        )
-        if result:
-            item = MemoryItem(
-                id=str(uuid.uuid4()),
-                section=result["section"],
-                text=result["text"],
+            timeout=300.0,
        )
+
+        # Embed existing facts once so each new fact can be matched against them.
+        # A near-duplicate UPDATES the matched fact in place (edit with new info)
+        # rather than appending a copy. Best effort: if embeddings are down we
+        # fall back to plain append. Merge disabled unless 0 < threshold < 1.
+        existing_embeds: dict = {}
+        if results and 0 < merge_threshold < 1:
+            vecs = await asyncio.gather(*(mgr.embed(it.text) for it in existing))
+            existing_embeds = {it.id: v for it, v in zip(existing, vecs) if v}
+
+        saved = []
+        for result in results:
+            new_vec = await mgr.embed(result["text"]) if existing_embeds else None
+            match_id, best = None, 0.0
+            if new_vec:
+                for eid, ev in existing_embeds.items():
+                    sim = _cosine(new_vec, ev)
+                    if sim > best:
+                        best, match_id = sim, eid
+                if best < merge_threshold:
+                    match_id = None
+
+            target = store.get(match_id) if match_id else None
+            if target:
+                # Near-duplicate of an existing fact — overwrite with the newer
+                # statement, keeping the original id/section/position.
+                updated = MemoryItem(id=target.id, section=target.section,
+                                     text=result["text"], tags=target.tags)
+                store.update(updated)
+                if new_vec:
+                    existing_embeds[updated.id] = new_vec  # keep cache fresh for later facts in this batch
+                saved.append({"id": updated.id, "section": updated.section,
+                              "text": updated.text, "updated": True})
+            else:
+                item = MemoryItem(id=str(uuid.uuid4()),
+                                  section=result["section"], text=result["text"])
                store.add(item)
-            return {"saved": True, "id": item.id, "section": item.section, "text": item.text}
+                if new_vec:
+                    existing_embeds[item.id] = new_vec
+                saved.append({"id": item.id, "section": item.section, "text": item.text})
+        if saved:
+            first = {k: saved[0][k] for k in ("id", "section", "text")}
+            return {"saved": True, "items": saved, **first}
    except asyncio.TimeoutError:
        pass
    except Exception:
@@ -3,9 +3,22 @@ from typing import Any, Dict, List, Optional
 from pydantic import BaseModel
 import sqlite3
 import json
+import math
 import os
 import time

+
+def _cosine(a: List[float], b: List[float]) -> float:
+    """Cosine similarity between two equal-length vectors. 0.0 on mismatch."""
+    if not a or not b or len(a) != len(b):
+        return 0.0
+    dot = sum(x * y for x, y in zip(a, b))
+    na = math.sqrt(sum(x * x for x in a))
+    nb = math.sqrt(sum(y * y for y in b))
+    if na == 0.0 or nb == 0.0:
+        return 0.0
+    return dot / (na * nb)
+
 # -----------------------------
 # Models
 # -----------------------------
@@ -14,6 +27,7 @@ class MemoryItem(BaseModel):
    section: str = "General"
    text: str
    tags: List[str] = []
+    position: int = 0

 class MessageItem(BaseModel):
    role: str  # "user" or "assistant"
@@ -27,6 +41,7 @@ class ConversationItem(BaseModel):
    messages: List[MessageItem] = []
    created_at: float
    updated_at: float
+    title: Optional[str] = None

    @property
    def preview(self) -> str:
@@ -75,14 +90,29 @@ class PersistentMemoryStore:
            cur.execute("ALTER TABLE memory ADD COLUMN section TEXT NOT NULL DEFAULT 'General'")
        except Exception:
            pass
+        # Migrate: add position column for stable ordering
+        try:
+            cur.execute("ALTER TABLE memory ADD COLUMN position INTEGER NOT NULL DEFAULT 0")
+        except Exception:
+            pass
+        # Backfill positions for rows added before this column existed
+        cur.execute("SELECT COUNT(*) FROM memory WHERE position > 0")
+        if cur.fetchone()[0] == 0:
+            cur.execute("UPDATE memory SET position = rowid")

        cur.execute("""
            CREATE TABLE IF NOT EXISTS conversations (
                id TEXT PRIMARY KEY,
                created_at REAL NOT NULL,
-                updated_at REAL NOT NULL
+                updated_at REAL NOT NULL,
+                title TEXT
            )
        """)
+        # Migrate: add title column if it doesn't exist yet
+        try:
+            cur.execute("ALTER TABLE conversations ADD COLUMN title TEXT")
+        except Exception:
+            pass

        cur.execute("""
            CREATE TABLE IF NOT EXISTS messages (
@@ -107,12 +137,26 @@ class PersistentMemoryStore:
            ON messages (conversation_id)
        """)

+        # Semantic recall: one embedding vector per message, stored as JSON.
+        # Backfilled lazily by semantic_search_conversations so existing history
+        # gets indexed on first search.
+        cur.execute("""
+            CREATE TABLE IF NOT EXISTS message_vectors (
+                message_id INTEGER PRIMARY KEY,
+                embedding TEXT NOT NULL,
+                FOREIGN KEY (message_id) REFERENCES messages(id)
+            )
+        """)
+
        cur.execute("""
            CREATE TABLE IF NOT EXISTS settings (
                key TEXT PRIMARY KEY,
                value TEXT NOT NULL
            )
        """)
+        cur.execute(
+            "DELETE FROM settings WHERE key IN ('anthropic_api_key', 'escalation_model')"
+        )

        conn.commit()
        conn.close()
@@ -123,7 +167,7 @@ class PersistentMemoryStore:
    def _load_all_memory(self) -> Dict[str, MemoryItem]:
        conn = self._connect()
        cur = conn.cursor()
-        cur.execute("SELECT id, section, text, tags FROM memory")
+        cur.execute("SELECT id, section, text, tags, position FROM memory ORDER BY position ASC, rowid ASC")
        rows = cur.fetchall()
        conn.close()

@@ -138,6 +182,7 @@ class PersistentMemoryStore:
                section=row["section"] or "General",
                text=row["text"],
                tags=tags,
+                position=row["position"] or 0,
            )

        return cache
@@ -146,19 +191,35 @@ class PersistentMemoryStore:
    # Memory API
    # -----------------------------
    def add(self, item: MemoryItem):
+        if not item.position:
+            conn = self._connect()
+            try:
+                row = conn.execute(
+                    "SELECT position FROM memory WHERE id = ?", (item.id,)
+                ).fetchone()
+                if row and row["position"]:
+                    item.position = row["position"]
+                else:
+                    row = conn.execute("SELECT MAX(position) AS max_position FROM memory").fetchone()
+                    item.position = (row["max_position"] or 0) + 1
+            finally:
+                conn.close()
        self._cache[item.id] = item
        conn = self._connect()
        try:
            cur = conn.cursor()
            cur.execute(
-                "INSERT OR REPLACE INTO memory (id, section, text, tags) VALUES (?, ?, ?, ?)",
-                (item.id, item.section or "General", item.text, json.dumps(item.tags))
+                "INSERT OR REPLACE INTO memory (id, section, text, tags, position) VALUES (?, ?, ?, ?, ?)",
+                (item.id, item.section or "General", item.text, json.dumps(item.tags), item.position)
            )
            conn.commit()
        finally:
            conn.close()

    def update(self, item: MemoryItem):
+        existing = self.get(item.id)
+        if existing:
+            item.position = existing.position
        self.add(item)

    def delete(self, item_id: str):
@@ -172,10 +233,76 @@ class PersistentMemoryStore:
            conn.close()

    def get(self, item_id: str) -> Optional[MemoryItem]:
-        return self._cache.get(item_id)
+        conn = self._connect()
+        try:
+            row = conn.execute(
+                "SELECT id, section, text, tags, position FROM memory WHERE id = ?",
+                (item_id,),
+            ).fetchone()
+        finally:
+            conn.close()
+        if not row:
+            self._cache.pop(item_id, None)
+            return None
+        try:
+            tags = json.loads(row["tags"]) if row["tags"] else []
+        except Exception:
+            tags = []
+        item = MemoryItem(
+            id=row["id"], section=row["section"] or "General", text=row["text"],
+            tags=tags, position=row["position"] or 0,
+        )
+        self._cache[item_id] = item
+        return item

    def all(self) -> List[MemoryItem]:
-        return list(self._cache.values())
+        # Always read from DB — the memory service and backend run in separate processes
+        # with separate caches, so the cache can be stale for facts extracted by the
+        # memory service after this process started.
+        conn = self._connect()
+        cur = conn.cursor()
+        cur.execute("SELECT id, section, text, tags, position FROM memory ORDER BY position ASC, rowid ASC")
+        rows = cur.fetchall()
+        conn.close()
+        items = []
+        for row in rows:
+            try:
+                tags = json.loads(row["tags"]) if row["tags"] else []
+            except Exception:
+                tags = []
+            items.append(MemoryItem(
+                id=row["id"],
+                section=row["section"] or "General",
+                text=row["text"],
+                tags=tags,
+                position=row["position"] or 0,
+            ))
+        return items
+
+    def reorder_section(self, section: str, ordered_ids: List[str]) -> bool:
+        """Rewrite the order of items in a section using its existing position pool.
+        ordered_ids must contain exactly the ids currently in the section."""
+        section_norm = section or "General"
+        in_section = [i for i in self.all() if (i.section or "General") == section_norm]
+        if len(in_section) != len(ordered_ids):
+            return False
+        by_id = {i.id: i for i in in_section}
+        siblings = []
+        for id_ in ordered_ids:
+            if id_ not in by_id:
+                return False
+            siblings.append(by_id[id_])
+        positions = sorted([i.position for i in in_section])
+        conn = self._connect()
+        try:
+            cur = conn.cursor()
+            for s, new_pos in zip(siblings, positions):
+                s.position = new_pos
+                cur.execute("UPDATE memory SET position = ? WHERE id = ?", (new_pos, s.id))
+            conn.commit()
+        finally:
+            conn.close()
+        return True

    # -----------------------------
    # Conversation API
@@ -194,7 +321,19 @@ class PersistentMemoryStore:
            conn.close()
        return ConversationItem(id=conversation_id, created_at=now, updated_at=now)

-    def add_message(self, conversation_id: str, role: str, content: str, model: str = None, tokens: int = None):
+    def set_conversation_title(self, conversation_id: str, title: str):
+        conn = self._connect()
+        try:
+            cur = conn.cursor()
+            cur.execute(
+                "UPDATE conversations SET title = ? WHERE id = ?",
+                (title, conversation_id),
+            )
+            conn.commit()
+        finally:
+            conn.close()
+
+    def add_message(self, conversation_id: str, role: str, content: str, model: Optional[str] = None, tokens: Optional[int] = None):
        now = time.time()
        conn = self._connect()
        try:
@@ -203,11 +342,13 @@ class PersistentMemoryStore:
                "INSERT INTO messages (conversation_id, role, content, timestamp, model, tokens) VALUES (?, ?, ?, ?, ?, ?)",
                (conversation_id, role, content, now, model, tokens)
            )
+            message_id = cur.lastrowid
            cur.execute(
                "UPDATE conversations SET updated_at = ? WHERE id = ?",
                (now, conversation_id)
            )
            conn.commit()
+            return message_id
        finally:
            conn.close()

@@ -231,14 +372,15 @@ class PersistentMemoryStore:
            id=row["id"],
            messages=messages,
            created_at=row["created_at"],
-            updated_at=row["updated_at"]
+            updated_at=row["updated_at"],
+            title=row["title"] if "title" in row.keys() else None,
        )

    def all_conversations(self) -> List[ConversationItem]:
        conn = self._connect()
        cur = conn.cursor()
        cur.execute("""
-            SELECT c.id, c.created_at, c.updated_at,
+            SELECT c.id, c.created_at, c.updated_at, c.title,
                   m.role, m.content, m.timestamp, m.model, m.tokens
            FROM conversations c
            LEFT JOIN messages m ON m.conversation_id = c.id
@@ -256,6 +398,7 @@ class PersistentMemoryStore:
                    id=cid,
                    created_at=row["created_at"],
                    updated_at=row["updated_at"],
+                    title=row["title"],
                )
                order.append(cid)
            if row["role"] is not None:
@@ -333,6 +476,130 @@ class PersistentMemoryStore:
        conn.close()
        return results

+    # nomic-embed-text is an asymmetric retrieval model: queries and stored
+    # documents must be embedded with these task prefixes or similarity collapses
+    # into noise. Cached vectors are document-embeddings (search_document:).
+    _EMBED_QUERY_PREFIX = "search_query: "
+    _EMBED_DOC_PREFIX = "search_document: "
+
+    async def semantic_search_conversations(
+        self, query: str, embed_fn, limit: int = 3, min_score: float = 0.6
+    ) -> List[dict]:
+        """Recall past conversations relevant to `query` using hybrid retrieval.
+
+        Combines semantic similarity (embeddings — finds reworded matches with no
+        shared keywords) with the existing lexical substring match (catches exact
+        terms the embedding underweights), unioned and deduped by conversation.
+
+        `embed_fn` is an async callable returning an embedding vector for a string
+        (typically OllamaManager.embed). Messages without a stored vector are
+        embedded and cached on first use (lazy backfill). If embeddings are
+        unavailable (no model / Ollama down) this degrades to pure lexical match,
+        so recall never silently breaks.
+
+        Returns the same shape as `search_conversations`: a list of
+        {id, updated_at, matches:[{role, content}]} with full user+assistant
+        pairs around each match.
+        """
+        if not query or not query.strip():
+            return []
+
+        query_vec = await embed_fn(self._EMBED_QUERY_PREFIX + query.strip())
+        if not query_vec:
+            return self.search_conversations(query, limit=limit)
+
+        conn = self._connect()
+        cur = conn.cursor()
+
+        # Lazy backfill: embed any messages that don't have a vector yet.
+        cur.execute("""
+            SELECT m.id, m.content
+            FROM messages m
+            LEFT JOIN message_vectors v ON v.message_id = m.id
+            WHERE v.message_id IS NULL AND TRIM(m.content) != ''
+        """)
+        missing = cur.fetchall()
+        for row in missing:
+            vec = await embed_fn(self._EMBED_DOC_PREFIX + row["content"][:2000])
+            if vec:
+                cur.execute(
+                    "INSERT OR REPLACE INTO message_vectors (message_id, embedding) VALUES (?, ?)",
+                    (row["id"], json.dumps(vec)),
+                )
+        if missing:
+            conn.commit()
+
+        # Score every stored message against the query vector.
+        cur.execute("""
+            SELECT v.message_id, v.embedding, m.conversation_id
+            FROM message_vectors v
+            JOIN messages m ON m.id = v.message_id
+        """)
+        scored = []
+        for row in cur.fetchall():
+            try:
+                vec = json.loads(row["embedding"])
+            except Exception:
+                continue
+            score = _cosine(query_vec, vec)
+            if score >= min_score:
+                scored.append((score, row["message_id"], row["conversation_id"]))
+
+        scored.sort(reverse=True)
+
+        results: List[dict] = []
+        seen_convs: set = set()
+        for score, message_id, conv_id in scored:
+            if len(results) >= limit:
+                break
+            if conv_id in seen_convs:
+                continue
+            pair = self._exchange_pair(cur, conv_id, message_id)
+            if pair:
+                seen_convs.add(conv_id)
+                results.append(pair)
+
+        conn.close()
+
+        # Hybrid union: fill any remaining slots with lexical matches the
+        # embedding missed (e.g. exact proper nouns), skipping dupes.
+        if len(results) < limit:
+            for conv in self.search_conversations(query, limit=limit):
+                if conv["id"] not in seen_convs:
+                    seen_convs.add(conv["id"])
+                    results.append(conv)
+                    if len(results) >= limit:
+                        break
+
+        return results
+
+    def _exchange_pair(self, cur, conv_id: str, message_id: int) -> Optional[dict]:
+        """Build a {id, updated_at, matches} record with the full user+assistant
+        pair surrounding `message_id`, in the shape search callers expect."""
+        cur.execute(
+            "SELECT id, role, content FROM messages WHERE conversation_id = ? ORDER BY timestamp ASC",
+            (conv_id,),
+        )
+        all_msgs = cur.fetchall()
+        idx = next((i for i, m in enumerate(all_msgs) if m["id"] == message_id), None)
+        if idx is None:
+            return None
+        if all_msgs[idx]["role"] == "user":
+            start, end = idx, min(idx + 1, len(all_msgs) - 1)
+        else:
+            start, end = max(idx - 1, 0), idx
+        matches = [
+            {"role": all_msgs[i]["role"], "content": all_msgs[i]["content"][:500]}
+            for i in range(start, end + 1)
+        ]
+        cur.execute("SELECT updated_at FROM conversations WHERE id = ?", (conv_id,))
+        crow = cur.fetchone()
+        return {
+            "id": conv_id,
+            "updated_at": crow["updated_at"] if crow else 0,
+            "matches": matches,
+        }
+
    # -----------------------------
    # Settings API
    # -----------------------------
@@ -341,6 +608,30 @@ class PersistentMemoryStore:
        "temperature": 0.7,
        "system_prompt": "",
        "timeout": 120,
+        # How long Ollama keeps the model resident in VRAM between messages.
+        # "30m"/"-1" (never unload)/"0" (unload now). Avoids cold-reload latency
+        # when you return to an idle chat. Empty → Ollama's 5-minute default.
+        "keep_alive": "30m",
+        # CPU/GPU offload: -1 = Auto (Ollama auto-fits layers to VRAM).
+        # 0–100 = percent of model layers to force onto the GPU; the
+        # remainder runs on CPU. See OllamaManager.get_model_layers.
+        "gpu_offload": -1,
+        # Memory curator (the model that extracts facts after each exchange).
+        # Empty → same auto-selected model as chat. A dedicated model (e.g.
+        # "mistral:latest") gives better extraction but must share VRAM.
+        "memory_model": "mistral:latest",
+        # Curator CPU/GPU offload — same scale as gpu_offload above. Default 0
+        # (all CPU/RAM): OS-neutral and never evicts the chat model from a small
+        # GPU. Boxes with spare VRAM can set -1 (Auto) or a percent to use the GPU.
+        "memory_gpu_offload": 0,
+        # Similar-fact merge: when a newly extracted fact's embedding is at least
+        # this cosine-similar to an existing fact, UPDATE that fact in place
+        # instead of appending a duplicate ("edit with new info"). 0 disables
+        # (always append). Calibrated on nomic-embed-text: genuine updates
+        # (mileage/title/location changes) score 0.81–0.99, while distinct facts
+        # top out ~0.61 — so 0.80 catches updates and never merges unrelated
+        # facts. Lower to catch looser rephrases; raise toward 1.0 to be stricter.
+        "memory_merge_threshold": 0.80,
    }

    def get_settings(self) -> Dict[str, Any]:
@@ -15,6 +15,12 @@ except Exception:
 # --- PROJECT ROOT ---
 PROJECT_ROOT = Path(__file__).resolve().parent.parent

+# --- VERSION (single source of truth: the VERSION file at the repo root) ---
+try:
+    VERSION = (PROJECT_ROOT / "VERSION").read_text(encoding="utf-8").strip() or "0.0.0"
+except Exception:
+    VERSION = "0.0.0"
+
 # --- CORE DIRECTORIES ---
 DATA_DIR = PROJECT_ROOT / "data"
 MODELS_DIR = PROJECT_ROOT / "models"
@@ -27,16 +33,15 @@ CACHE_DIR = RUNTIME_DIR / "cache"
 TEMP_DIR = RUNTIME_DIR / "tmp"

 # --- APPLICATION SUBSYSTEM DIRECTORIES ---
-PLAYBOOK_DIR = PROJECT_ROOT / "synapse" / "playbooks"
+PLAYBOOK_DIR = DATA_DIR / "playbooks"  # YAML playbook files (PlaybookFileStore)
 UPLOADS_DIR = DATA_DIR / "uploads"
 EXPORTS_DIR = DATA_DIR / "exports"

 # --- DATABASE / STORAGE FILES (match your repo) ---
 MEMORY_DB = MEMORY_DIR / "memory.db"
-PLAYBOOK_DB = MEMORY_DB  # same file in your setup

 # --- LOG FILES ---
-BACKEND_LOG = LOGS_DIR / "backend.log"
+BACKEND_LOG = RUNTIME_DIR / "backend.log"
 OLLAMA_LOG = LOGS_DIR / "ollama.log"
 CHAT_LOG = LOGS_DIR / "chat.log"

@@ -73,7 +78,6 @@ def path(name: str) -> Path:
        "uploads": UPLOADS_DIR,
        "exports": EXPORTS_DIR,
        "memory_db": MEMORY_DB,
-        "playbook_db": PLAYBOOK_DB,
        "backend_log": BACKEND_LOG,
        "ollama_log": OLLAMA_LOG,
        "chat_log": CHAT_LOG,
@@ -90,6 +94,7 @@ class Settings:
    or `Settings` class for typing/tests.
    """
    def __init__(self) -> None:
+        self.version: str = VERSION
        self.project_root: Path = PROJECT_ROOT
        self.data_dir: Path = DATA_DIR
        self.models_dir: Path = MODELS_DIR
@@ -99,7 +104,6 @@ class Settings:

        # DB files
        self.memory_db: Path = MEMORY_DB
-        self.playbook_db: Path = PLAYBOOK_DB

        # Logs
        self.backend_log: Path = BACKEND_LOG
@@ -112,13 +116,13 @@ class Settings:

    def as_dict(self) -> Dict[str, Any]:
        return {
+            "version": self.version,
            "project_root": str(self.project_root),
            "data_dir": str(self.data_dir),
            "models_dir": str(self.models_dir),
            "runtime_dir": str(self.runtime_dir),
            "memory_dir": str(self.memory_dir),
            "memory_db": str(self.memory_db),
-            "playbook_db": str(self.playbook_db),
            "ollama_host": self.ollama_host,
            "ollama_timeout": self.ollama_timeout,
        }
@@ -127,10 +131,10 @@ class Settings:
 settings = Settings()

 # explicit exports for static checkers and IDEs
-__all__ = ["Settings", "settings", "path",
+__all__ = ["Settings", "settings", "path", "VERSION",
           "PROJECT_ROOT", "DATA_DIR", "MODELS_DIR", "RUNTIME_DIR",
           "MEMORY_DIR", "LOGS_DIR", "PLAYBOOK_DIR", "UPLOADS_DIR",
-           "EXPORTS_DIR", "MEMORY_DB", "PLAYBOOK_DB",
+           "EXPORTS_DIR", "MEMORY_DB",
           "BACKEND_LOG", "OLLAMA_LOG", "CHAT_LOG"]
           
 # --- quick runtime sanity check when run directly (no side effects on import) ---
@@ -1,15 +1,16 @@
 import asyncio
 import json
 import logging
-import re
 import subprocess
 import time
 import httpx
 import os
 import signal
 from pathlib import Path
+from typing import Any

 from .nexus_config import settings
+from .memory.store import store

 OLLAMA_PORT = 11434

@@ -131,10 +132,9 @@ def _detect_gpu_backend() -> tuple[str, dict]:

    if vulkan_ok:
        idx, name = _best_vulkan_device()
-        env_overrides: dict = {"OLLAMA_GPU": "vulkan"}
-        if idx > 0:
-            # Explicitly route Ollama to the discrete GPU when it isn't device 0
-            env_overrides["GGML_VK_VISIBLE_DEVICES"] = str(idx)
+        # Always pin to the selected device — without this, Ollama may use the Intel
+        # iGPU's shared system RAM as "VRAM" for models that don't fit on discrete VRAM.
+        env_overrides: dict = {"OLLAMA_VULKAN": "1", "GGML_VK_VISIBLE_DEVICES": str(idx)}
        _log.info("GPU backend: Vulkan device %d (%s)", idx, name)
        return f"vulkan ({name})", env_overrides

@@ -153,18 +153,36 @@ def _detect_gpu_backend() -> tuple[str, dict]:
    return "cpu", {}


-def _model_score(name: str) -> tuple:
-    """Score a model name for auto-selection. Higher tuple = better."""
-    lower = name.lower()
-    match = re.search(r'(\d+(?:\.\d+)?)[bB]', lower)
-    params = float(match.group(1)) if match else 7.0
-    # Tier-break: known quality models get a small bonus
-    quality = next(
-        (i for i, prefix in enumerate(("llama3", "llama2", "mistral", "gemma", "phi", "qwen"), 1)
-         if prefix in lower),
-        0,
-    )
-    return (params, quality)
+# Single source of model auto-selection preference. Ordered so small, GPU-fitting
+# models come first (qwen2.5:3b fits a 4GB card and is a strong all-rounder); the
+# tail differs by task. Prefix-matched against installed model names.
+_MODEL_PREFERENCE = {
+    "chat": ("qwen2.5:3b", "qwen2.5", "gemma3:1b", "gemma3", "phi3", "phi-3", "gemma2", "gemma"),
+    "code": ("qwen2.5:3b", "qwen2.5", "gemma3:1b", "gemma3", "phi3", "phi-3", "codellama", "deepseek-coder", "codegemma"),
+}
+
+
+def _preferred_model(models: list, preference) -> str | None:
+    """First installed model whose name starts with a preference prefix."""
+    for prefix in preference:
+        for m in models:
+            if m.lower().startswith(prefix.lower()):
+                return m
+    return None
+
+
+def _chat_options(temperature: float | None, num_gpu: int | None) -> dict:
+    """Assemble the Ollama `options` block from the knobs we expose.
+
+    Returns an empty dict when nothing is set so callers can omit `options`
+    entirely (preserving Ollama's defaults / auto behaviour).
+    """
+    opts: dict = {}
+    if temperature is not None:
+        opts["temperature"] = temperature
+    if num_gpu is not None:
+        opts["num_gpu"] = num_gpu
+    return opts


 class OllamaManager:
@@ -181,8 +199,43 @@ class OllamaManager:
        self._api_base = settings.ollama_host.rstrip("/")

        # Model selection cache
-        self._model_cache: str | None = None
-        self._model_cache_ts: float = 0.0
+        self._model_cache: dict = {}  # intent -> (model, monotonic_ts)
+
+        # Per-model offloadable layer count cache (never changes for a model)
+        self._layer_cache: dict[str, int] = {}
+
+        # How long Ollama keeps the model resident between requests. Applied to
+        # every chat/generate body so the model isn't reloaded on each message.
+        # Overridden from persisted settings at startup. Falsy → omit (Ollama's
+        # 5-minute default).
+        self.keep_alive: str | None = "30m"
+
+    def _apply_keep_alive(self, body: dict) -> dict:
+        """Add `keep_alive` (a top-level Ollama field) to a request body when set."""
+        if self.keep_alive:
+            body["keep_alive"] = self.keep_alive
+        return body
+
+    async def warm(self, model: str | None = None, num_gpu: int | None = None) -> None:
+        """Preload a model so the first request doesn't pay a cold load.
+        An empty-prompt /api/generate is Ollama's documented preload. Pass the
+        same `num_gpu` the real requests use, or the preloaded copy is placed
+        differently and gets reloaded on first use. Best-effort: never raises,
+        so a missing model or down server can't break startup."""
+        try:
+            model = model or await self.select_best_model()
+            if not model:
+                return
+            body = {"model": model, "prompt": "", "stream": False}
+            if num_gpu is not None:
+                body["options"] = {"num_gpu": num_gpu}
+            async with httpx.AsyncClient(timeout=120.0) as client:
+                await client.post(
+                    f"{self._api_base}/api/generate", json=self._apply_keep_alive(body),
+                )
+            _log.info("warm: preloaded model=%s num_gpu=%s keep_alive=%s", model, num_gpu, self.keep_alive)
+        except Exception as e:
+            _log.warning("warm: preload failed: %s", e)

    def is_available(self):
        bin_path = _ollama_bin()
@@ -328,12 +381,12 @@ class OllamaManager:
                async with httpx.AsyncClient(timeout=300.0) as client:
                    r = await client.post(
                        f"{self._api_base}/api/generate",
-                        json={
+                        json=self._apply_keep_alive({
                            "model": model,
                            "prompt": prompt,
                            "system": system,
                            "stream": False,
-                        },
+                        }),
                    )

                elapsed = time.perf_counter() - start
@@ -356,12 +409,12 @@ class OllamaManager:
                async with client.stream(
                    "POST",
                    f"{self._api_base}/api/generate",
-                    json={
+                    json=self._apply_keep_alive({
                        "model": model,
                        "prompt": prompt,
                        "system": system,
                        "stream": True,
-                    },
+                    }),
                ) as response:
                    response.raise_for_status()
                    async for line in response.aiter_lines():
@@ -391,17 +444,23 @@ class OllamaManager:
        model: str = "mistral",
        stream: bool = False,
        temperature: float | None = None,
+        num_gpu: int | None = None,
        **kwargs,
    ):
        """Multi-turn chat via /api/chat (accepts a messages array with roles)."""
        start = time.perf_counter()
        try:
            if stream:
-                return self._chat_stream(messages=messages, model=model, temperature=temperature, start=start)
+                return self._chat_stream(
+                    messages=messages, model=model, temperature=temperature,
+                    num_gpu=num_gpu, start=start,
+                )
            else:
                body: dict = {"model": model, "messages": messages, "stream": False}
-                if temperature is not None:
-                    body["options"] = {"temperature": temperature}
+                opts = _chat_options(temperature, num_gpu)
+                if opts:
+                    body["options"] = opts
+                self._apply_keep_alive(body)
                async with httpx.AsyncClient(timeout=300.0) as client:
                    r = await client.post(f"{self._api_base}/api/chat", json=body)
                elapsed = time.perf_counter() - start
@@ -412,6 +471,33 @@ class OllamaManager:
            _log.exception("chat error after %.3fs: %s", elapsed, e)
            return None

+    async def embed(self, text: str, model: str = "nomic-embed-text") -> list[float] | None:
+        """Return an embedding vector for `text` via /api/embeddings.
+
+        Returns None on any failure so callers can fall back to lexical search —
+        a missing embedding model should never break chat or recall.
+        """
+        text = (text or "").strip()
+        if not text:
+            return None
+        try:
+            gpu_offload = store.get_settings().get("memory_gpu_offload", 0)
+            num_gpu = await self.resolve_num_gpu(gpu_offload, model)
+            body: dict[str, Any] = {"model": model, "prompt": text}
+            if num_gpu is not None:
+                body["options"] = {"num_gpu": num_gpu}
+            async with httpx.AsyncClient(timeout=30.0) as client:
+                r = await client.post(
+                    f"{self._api_base}/api/embeddings",
+                    json=body,
+                )
+                r.raise_for_status()
+                vec = r.json().get("embedding")
+                return vec if vec else None
+        except Exception as e:
+            _log.debug("embed failed (model=%s): %s", model, e)
+            return None
+
    async def list_models(self) -> list[str]:
        """Return names of all locally installed Ollama models."""
        try:
@@ -422,34 +508,84 @@ class OllamaManager:
        except Exception:
            return []

-    async def select_best_model(self) -> str:
-        """Pick the highest-scoring available model, falling back to 'mistral'.
-
-        Result is cached for 60 seconds so rapid chat requests don't each
-        hit the Ollama API to build the model list.
+    async def select_best_model(self, intent: str = "chat") -> str:
+        """Pick the preferred installed model for `intent` ('chat' or 'code'),
+        falling back to any installed model, then 'mistral'. Cached ~60s per
+        intent so rapid requests don't rebuild the model list each time.
        """
        now = time.monotonic()
-        if self._model_cache and (now - self._model_cache_ts) < 60:
-            return self._model_cache
+        cached = self._model_cache.get(intent)
+        if cached and (now - cached[1]) < 60:
+            return cached[0]

        models = await self.list_models()
-        best = max(models, key=_model_score) if models else "mistral"
+        pref = _MODEL_PREFERENCE.get(intent, _MODEL_PREFERENCE["chat"])
+        best = _preferred_model(models, pref) or (models[0] if models else "mistral")

-        self._model_cache = best
-        self._model_cache_ts = now
+        self._model_cache[intent] = (best, now)
        return best

    def invalidate_model_cache(self):
        """Force next select_best_model() to re-query (e.g. after pull/delete)."""
-        self._model_cache = None
-        self._model_cache_ts = 0.0
+        self._model_cache = {}

-    async def _chat_stream(self, messages: list, model: str, start: float, temperature: float | None = None):
+    async def get_model_layers(self, model: str) -> int | None:
+        """Total offloadable layer count for `model` (repeating blocks + output layer).
+
+        Used to turn a CPU/GPU offload percentage into an Ollama `num_gpu`
+        value. Reads `<arch>.block_count` from /api/show and adds 1 for the
+        non-repeating output layer (Ollama reports e.g. 33 layers for a model
+        with block_count=32). Cached per-model since it never changes.
+        Returns None if the count can't be determined, so callers fall back
+        to Auto (no num_gpu override).
+        """
+        if model in self._layer_cache:
+            return self._layer_cache[model]
+        try:
+            async with httpx.AsyncClient(timeout=10.0) as client:
+                r = await client.post(f"{self._api_base}/api/show", json={"model": model})
+            r.raise_for_status()
+            info = r.json().get("model_info", {}) or {}
+            block_count = next(
+                (v for k, v in info.items() if k.endswith(".block_count")), None
+            )
+            layers = int(block_count) + 1 if block_count is not None else None
+        except Exception as e:
+            _log.warning("get_model_layers(%s) failed: %s", model, e)
+            layers = None
+        if layers:
+            self._layer_cache[model] = layers
+        return layers
+
+    async def resolve_num_gpu(self, gpu_offload, model: str) -> int | None:
+        """Convert a stored gpu_offload setting into an Ollama `num_gpu` value.
+
+        `gpu_offload` is -1 for Auto (returns None → no override, Ollama auto-fits)
+        or 0–100 for the percent of the model's layers to force onto the GPU
+        (0 = all CPU/RAM). Layer count is model-specific, resolved from the live
+        model. Returns None on anything unexpected so callers fall back to Auto.
+        """
+        try:
+            pct = int(gpu_offload)
+        except (TypeError, ValueError):
+            return None
+        if pct < 0:
+            return None
+        pct = min(pct, 100)
+        layers = await self.get_model_layers(model)
+        if not layers:
+            return None
+        return max(0, round(pct / 100 * layers))
+
+    async def _chat_stream(self, messages: list, model: str, start: float,
+                           temperature: float | None = None, num_gpu: int | None = None):
        """Async generator streaming tokens, then a final __meta__ stats sentinel."""
        try:
            body: dict = {"model": model, "messages": messages, "stream": True}
-            if temperature is not None:
-                body["options"] = {"temperature": temperature}
+            opts = _chat_options(temperature, num_gpu)
+            if opts:
+                body["options"] = opts
+            self._apply_keep_alive(body)
            async with httpx.AsyncClient(timeout=300.0) as client:
                async with client.stream(
                    "POST",
@@ -484,7 +620,7 @@ class OllamaManager:
        except Exception as e:
            elapsed = time.perf_counter() - start
            _log.exception("chat stream error after %.3fs: %s", elapsed, e)
-            return
+            raise


 def initialize_ollama() -> OllamaManager:
@@ -1,11 +1,8 @@
-import threading
 from typing import List
 from .playbooks.store import playbook_store, PlaybookItem


 class PlaybookManager:
-    _lock = threading.Lock()
-
    @classmethod
    def _all(cls) -> List[PlaybookItem]:
        """Return all playbooks sorted by order (position 0 is always main)."""
@@ -25,18 +22,10 @@ class PlaybookManager:
    @classmethod
    def get_system_prompt(cls) -> str:
        playbook = cls.get_main_playbook()
-        return getattr(playbook, "instructions", "") or "" if playbook else ""
-
-    @classmethod
-    def render_prompt(cls, user_message: str, variables: dict | None = None) -> str:
-        if not variables:
-            return user_message
-        result = user_message
-        for key, value in variables.items():
-            result = result.replace(f"{{{{{key}}}}}", str(value))
-        return result
-
-    # Keep old name as alias so nothing else breaks
-    @classmethod
-    def get_active_playbook(cls) -> PlaybookItem | None:
-        return cls.get_main_playbook()
+        if not playbook:
+            return ""
+        goal = (getattr(playbook, "goal", "") or "").strip()
+        instructions = (getattr(playbook, "instructions", "") or "").strip()
+        if goal and instructions:
+            return f"{goal}\n\n{instructions}"
+        return goal or instructions
@@ -90,18 +90,6 @@ class PlaybookFileStore:
                item.order = index
                self._write(item)

-    def search_playbooks(self, query: str) -> List[PlaybookItem]:
-        if not query or not query.strip():
-            return []
-        q = query.strip().lower()
-        return [
-            p for p in self.all_playbooks()
-            if q in p.title.lower()
-            or q in p.goal.lower()
-            or q in p.instructions.lower()
-            or any(q in t.lower() for t in p.tags)
-        ]
-

 from ..nexus_config import PLAYBOOK_DIR
 playbook_store = PlaybookFileStore(PLAYBOOK_DIR)
@@ -0,0 +1,44 @@
+from __future__ import annotations
+
+import re
+from typing import Optional
+
+
+_SEARCH_TRIGGERS = frozenset({
+    # Time-sensitive single words
+    "news", "weather", "today", "tonight", "yesterday", "price", "stock",
+    # Phrases that imply freshness or external lookup
+    "latest ", "right now", "this week", "this month", "recently",
+    "who is ", "what is ", "when did ", "how much does", "how much is",
+    "look up", "search for", "find out", "release date",
+    "just released", "just announced", "just launched",
+    "current version", "current price",
+})
+
+
+def needs_web_search(message: str) -> bool:
+    lower = message.lower()
+    return any(kw in lower for kw in _SEARCH_TRIGGERS)
+
+
+def web_search(query: str, max_results: int = 4) -> str:
+    """Search DuckDuckGo and return formatted result snippets.
+
+    Returns an empty string on any failure so callers can treat it as
+    optional context — a failed search should never break a chat response.
+    """
+    try:
+        from duckduckgo_search import DDGS
+        with DDGS() as ddgs:
+            results = list(ddgs.text(query, max_results=max_results))
+        if not results:
+            return ""
+        parts = []
+        for i, r in enumerate(results, 1):
+            title = r.get("title", "").strip()
+            body = r.get("body", "").strip()
+            href = r.get("href", "").strip()
+            parts.append(f"{i}. **{title}**\n{body}\nSource: {href}")
+        return "\n\n".join(parts)
+    except Exception:
+        return ""