Add task brief research agent: Stage 1 deep-research + canvas draft emission

- run_task_brief_research() runner with brief-specific tool set and max_steps=12
- New agents: client_agent (list_clients, get_client) and relations_agent (query_relations)
- search_associative tool wrapping MemoryMiddleware semantic search
- BRIEF_RESEARCH_TOOLS constant: read-only task/project/note/timeline + memory + client/relations
- canvas block extraction in output_formatter (splits visible text from <canvas> draft)
- device_ws.py: task_brief_research request type; emits canvas_draft mutation on stream_end
- Stage 2 briefMode: briefing_context injected into floating system prompt when present
- briefingContext kwarg wired through compile_prompt call chain

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/agents/client_agent.py

@@ -0,0 +1,52 @@
+"""Client agent — read-only tools for the clients table."""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.ws_context import execute_on_client
+
+
+@tool
+async def list_clients(search: str = "", limit: int = 20) -> str:
+    """List clients, optionally filtered by a name/email substring search.
+
+    search: optional substring to match against client name or email.
+    limit: max rows to return (default 20).
+    """
+    filters: dict[str, Any] = {"limit": limit}
+    if search:
+        filters["search"] = search
+
+    result = await execute_on_client(action="select", table="clients", filters=filters)
+    rows = result.get("rows", [])
+    if not rows:
+        return "No clients found."
+    lines = [
+        f"- {r.get('name', '?')} (id: {r.get('id')}, email: {r.get('email', '')}, "
+        f"company: {r.get('company', '')})"
+        for r in rows
+    ]
+    return f"Found {len(rows)} client(s):\n" + "\n".join(lines)
+
+
+@tool
+async def get_client(id: str) -> str:
+    """Get full details for one client by UUID.
+
+    id: the client's UUID.
+    """
+    if not id:
+        return "Client id is required."
+
+    result = await execute_on_client(action="get", table="clients", data={"id": id})
+    row = result.get("row") or result.get("rows", [None])[0] if result else None
+    if not row:
+        return f"Client '{id}' not found."
+    return f"Client details:\n{json.dumps(row, ensure_ascii=False, indent=2)}"
+
+
+CLIENT_TOOLS: list[Any] = [list_clients, get_client]

app/agents/relations_agent.py

@@ -0,0 +1,63 @@
+"""Relations agent — read-only tool wrapping MemoryMiddleware.query_relations."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langchain_core.tools import tool
+
+from app.core.memory_middleware import MemoryMiddleware
+from app.db import async_session
+
+# Injected at tool-factory time by _brief_research_tools(); not a module-level global.
+# Each tool closure captures the user_id bound at factory time.
+
+
+def make_query_relations_tool(user_id: str, trace_id: str | None = None) -> Any:
+    """Return a query_relations tool bound to *user_id*."""
+
+    @tool
+    async def query_relations(
+        subject_label: str = "",
+        predicate: str = "",
+        object_label: str = "",
+        limit: int = 10,
+    ) -> str:
+        """Query the relational memory graph for entity relationships.
+
+        Returns rows where subject ↔ predicate ↔ object match the given filters.
+        All parameters are optional — omit to retrieve all relations up to limit.
+
+        subject_label: entity label on the left side (e.g. a client name, "Acme Corp").
+        predicate: relationship type (e.g. "mentioned_in", "works_at", "related_to").
+        object_label: entity label on the right side (e.g. a project name, "Website Redesign").
+        limit: max rows to return (default 10).
+        """
+        import logging
+        logger = logging.getLogger(__name__)
+        logger.info(
+            "relations_agent: query_relations trace=%s user=%s subject=%r predicate=%r object=%r",
+            trace_id or "-", user_id, subject_label, predicate, object_label,
+        )
+
+        async with async_session() as db:
+            memory = MemoryMiddleware(db)
+            rows = await memory.query_relations(
+                user_id=user_id,
+                subject=subject_label or None,
+                predicate=predicate or None,
+                object_=object_label or None,
+                limit=limit,
+            )
+
+        if not rows:
+            return "No relational memory entries found for the given filters."
+
+        lines = [
+            f"- {r.subject_label} —[{r.predicate}]→ {r.object_label}"
+            + (f" (confidence: {r.confidence:.2f})" if r.confidence is not None else "")
+            for r in rows
+        ]
+        return f"Found {len(rows)} relation(s):\n" + "\n".join(lines)
+
+    return query_relations