feat: playbook tools, RAG, vision, voice, chat controls, faster boot

- Tool-using playbooks: read-only tool registry (search_memory,
  search_history, search_documents, list_models, get_time), per-playbook
  allowlist, agentic loop, and a "running tool" status indicator.
- Document ingest / RAG: documents table + chunker + embed/cosine retrieval
  reusing the existing stack; Documents page (upload/paste, viewer, delete);
  top-k chunks injected into the system prompt.
- Vision chat: attach images, base64 into /api/chat.
- Voice I/O: Web Speech dictation + read-aloud (browser-native, no backend).
- Chat controls: stop, regenerate, edit-and-resend; num_ctx knob in Settings.
- Per-playbook model override.
- History polish: per-conversation + bulk ShareGPT export.
- Faster ncp start: UI up first, Ollama warms in the background, with a
  per-phase timing readout.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

synapse/chat.py

@@ -8,6 +8,11 @@ from typing import AsyncGenerator, Dict, List, Optional, Any
 
 from .nexus_config import settings, DEFAULT_CHAT_MODEL
 from .ollama_manager import get_ollama_manager
+from . import tools as _tools
+
+# Cap on tool-call round-trips before the final answer — stops a confused small
+# model from looping forever.
+MAX_TOOL_STEPS = 5
 
 
 # -------------------------
@@ -126,6 +131,38 @@ async def _normalize_to_async_generator(maybe_iterable) -> AsyncGenerator[str, N
         yield str(item)
 
 
+async def _run_tool_loop(manager, messages, model, tool_schemas, temperature, num_gpu):
+    """Let the model call read-only tools before the final streamed answer.
+
+    Mutates `messages` IN PLACE, appending the assistant tool-call turns and
+    their `role:"tool"` results, and yields a `__status__<tool>` sentinel before
+    each tool runs (surfaced to the UI as a "running tool" indicator).
+    Non-streamed — tool calls arrive as whole messages. Degrades to an untouched
+    `messages` if the model can't do tool calling.
+
+    ponytail: the turn that finally returns content is thrown away and the answer
+    is re-generated by the streaming turn (one wasted call). Simpler than
+    streaming a maybe-already-complete message; revisit if latency matters.
+    """
+    for _ in range(MAX_TOOL_STEPS):
+        msg = await manager.chat(
+            messages=messages, model=model, stream=False,
+            temperature=temperature, num_gpu=num_gpu, tools=tool_schemas,
+        )
+        if not isinstance(msg, dict):
+            break  # None/error or no tool support -> fall back to plain stream
+        calls = msg.get("tool_calls")
+        if not calls:
+            break
+        messages.append(msg)
+        for c in calls:
+            fn = c.get("function", {})
+            name = fn.get("name", "")
+            yield f"__status__{name}"
+            result = await _tools.dispatch(name, fn.get("arguments"))
+            messages.append({"role": "tool", "content": result})
+
+
 # -------------------------
 # Streaming implementation
 # -------------------------
@@ -143,6 +180,7 @@ async def stream_chat_response(
     model = metadata.get("model") or DEFAULT_CHAT_MODEL
     temperature = metadata.get("temperature")
     num_gpu = metadata.get("num_gpu")
+    num_ctx = metadata.get("num_ctx")
     think = metadata.get("think", False)
 
     # Build messages array for /api/chat multi-turn format
@@ -151,7 +189,21 @@ async def stream_chat_response(
         messages.append({"role": "system", "content": system})
     for msg in (history or []):
         messages.append({"role": msg["role"], "content": msg["content"]})
-    messages.append({"role": "user", "content": user_message})
+    user_msg: Dict[str, Any] = {"role": "user", "content": user_message}
+    images = metadata.get("images")  # base64 strings (no data: prefix) for vision models
+    if images:
+        user_msg["images"] = images
+    messages.append(user_msg)
+
+    # Tool-using playbooks: run read-only tool calls, then stream the final answer
+    # with their results already in the messages array.
+    tool_schemas = metadata.get("tools")
+    if tool_schemas:
+        try:
+            async for status in _run_tool_loop(manager, messages, model, tool_schemas, temperature, num_gpu):
+                yield status
+        except Exception:
+            _logger.exception("tool loop failed; streaming without tools")
 
     _logger.info("stream_chat_response: starting stream (model=%s, turns=%d, timeout=%s)", model, len(messages), timeout)
 
@@ -162,7 +214,7 @@ 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, num_gpu=num_gpu, think=think)
+        maybe_iter = manager.chat(messages=messages, model=model, stream=True, temperature=temperature, num_gpu=num_gpu, think=think, num_ctx=num_ctx)
         async_gen = _normalize_to_async_generator(maybe_iter)
 
         buffer_parts: list[str] = []