feat: live agent answers (SSE streaming) + «Чат» tab on phones

- POST /dialog/{token}/events/stream: Server-Sent Events with the agent's steps («Смотрю вашу карту ЕМИАС…»,
  «Ищу в клинических рекомендациях…»), the answer as the model writes it (OpenRouter stream=true), then the
  final replies (same as /events, incl. the forbidden-output filter and sources). X-Accel-Buffering: no.
- Web: free-text messages use the stream — a live bubble with steps, typing dots and a cursor; falls back to
  the plain request if the stream breaks.
- Phones: the bottom tab bar gets «Чат» (it had no way back to the chat).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app/api/routes.py

@@ -8,9 +8,14 @@ disclosed — the dashboard polls every 2–3 s), `ack_alert` (in `repo.ack_aler
 `app/api/fixtures.py` is kept for tests and as the demo cabinet reference (doctor id).
 """
 
+import asyncio
+import json
+import logging
 import secrets
+from typing import Any
 
 from fastapi import APIRouter, HTTPException
+from fastapi.responses import StreamingResponse
 from sqlalchemy import select
 
 from app.api import demo as console
@@ -42,10 +47,12 @@ from app.db.identity import Patient
 from app.db.session import get_clinical_session, get_identity_session
 from app.dialog.contracts import Event, get_engine
 from app.integrations.base import integrations
+from app.llm.agent import stream as agent_stream
 from app.scheduler import jobs as scheduler
 from app.services.anonymous_care import display_name as anonymous_name
 from app.services.monitoring import refresh_status
 
+log = logging.getLogger(__name__)
 router = APIRouter(prefix="/api/v1")
 router.include_router(max_notifications.router)
 router.include_router(questionnaire_router)
@@ -219,6 +226,49 @@ async def dialog_event(token: str, body: DialogIn) -> DialogOut:
     return DialogOut(replies=await get_engine().handle(token, body))
 
 
+@router.post("/dialog/{token}/events/stream", tags=["dialog"])
+async def dialog_event_stream(token: str, body: DialogIn) -> StreamingResponse:
+    """Same as /events, as Server-Sent Events: `status` (agent steps), `delta` (answer as it is written),
+    `reset`, then `done` with the final replies (identical to /events). `error` if the engine failed."""
+    await _require_patient(token)
+    queue: asyncio.Queue[tuple[str, Any]] = asyncio.Queue()
+    marker = agent_stream.STREAM.set(queue)  # copied into the task's context below
+    try:
+
+        async def run() -> None:
+            try:
+                replies = await get_engine().handle(token, body)
+                queue.put_nowait(("done", [r.model_dump() for r in replies]))
+            except Exception:  # noqa: BLE001 — reported to the client, which falls back to /events
+                log.exception("dialog stream failed")
+                queue.put_nowait(("error", None))
+
+        task = asyncio.create_task(run())
+    finally:
+        agent_stream.STREAM.reset(marker)
+
+    async def events():
+        try:
+            while True:
+                try:
+                    kind, payload = await asyncio.wait_for(queue.get(), timeout=15)
+                except TimeoutError:
+                    yield ": keepalive\n\n"
+                    continue
+                yield f"event: {kind}\ndata: {json.dumps(payload, ensure_ascii=False)}\n\n"
+                if kind in ("done", "error"):
+                    return
+        finally:
+            if not task.done():
+                await task
+
+    return StreamingResponse(
+        events(),
+        media_type="text/event-stream",
+        headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"},  # nginx: do not buffer SSE
+    )
+
+
 @router.get("/dialog/{token}/outbox", response_model=DialogOut, tags=["dialog"])
 async def dialog_outbox(token: str) -> DialogOut:
     """System-initiated replies (reminders, check-ins) waiting for the web console."""

app/llm/agent/__init__.py

@@ -21,6 +21,7 @@ from pydantic import BaseModel
 
 from app.config import Settings, get_settings
 from app.core.types import ScrubbedText
+from app.llm.agent import stream
 from app.llm.agent.prompts import SYSTEM_PROMPT
 from app.llm.agent.providers import (
     GigaChatToolProvider,
@@ -127,6 +128,7 @@ async def answer_chat(
         return None
     if is_small_talk(question):
         return AgentAnswer(text=SMALL_TALK_REPLY, sources=[], tools=[], provider="small_talk")
+    stream.emit("status", "Думаю над вопросом…")
     ctx = ToolContext(token=token)
     messages: list[dict[str, Any]] = [
         {"role": "system", "content": SYSTEM_PROMPT + await _patient_brief(ctx)},
@@ -159,6 +161,7 @@ async def answer_chat(
                     }
                 )
                 for c in turn.tool_calls[:MAX_TOOL_CALLS_PER_TURN]:
+                    stream.emit("status", stream.TOOL_STATUS.get(c.name, "Проверяю данные…"))
                     result = await run_tool(ctx, c.name, c.arguments)
                     messages.append({"role": "tool", "tool_call_id": c.id, "name": c.name, "content": result})
             log.warning("chat agent: step limit reached tools=%s", ",".join(ctx.used))

app/llm/agent/providers.py

@@ -16,6 +16,7 @@ import httpx
 from pydantic import BaseModel, Field
 
 from app.config import Settings
+from app.llm.agent import stream
 
 log = logging.getLogger(__name__)
 
@@ -53,6 +54,8 @@ class OpenRouterProvider:
         self._client = client
 
     async def chat(self, messages: list[dict[str, Any]], tools: list[dict[str, Any]]) -> AssistantTurn:
+        if stream.active():
+            return await self._chat_stream(messages, tools)
         body: dict[str, Any] = {"model": self._model, "messages": messages, "tools": tools, "temperature": 0.1}
         body["max_tokens"] = 900
         if not any(m["role"] == "tool" for m in messages):
@@ -174,6 +177,78 @@ class GigaChatToolProvider:
         return AssistantTurn(content=msg.content or None, tool_calls=calls)
 
 
+# ------------------------------------------------------------------------------------------------ streaming
+
+
+async def _iter_sse(resp: httpx.Response):
+    async for line in resp.aiter_lines():
+        if not line.startswith("data:"):
+            continue  # comments («: OPENROUTER PROCESSING») and blank lines
+        data = line[5:].strip()
+        if data == "[DONE]":
+            return
+        try:
+            yield json.loads(data)
+        except json.JSONDecodeError:
+            continue
+
+
+async def _openrouter_stream(self: "OpenRouterProvider", messages: list[dict[str, Any]], tools: list[dict[str, Any]]):
+    """Same request with `stream: true`: content deltas go to the live channel, tool calls are assembled."""
+    body: dict[str, Any] = {
+        "model": self._model,
+        "messages": messages,
+        "tools": tools,
+        "temperature": 0.1,
+        "max_tokens": 900,
+        "stream": True,
+    }
+    if not any(m["role"] == "tool" for m in messages):
+        body["tool_choice"] = "required"
+    headers = {"Authorization": f"Bearer {self._key}", "X-Title": "PTSRaiAgent prototype (synthetic data)"}
+    content: list[str] = []
+    calls: dict[int, dict[str, Any]] = {}
+    try:
+        client = self._client or httpx.AsyncClient(timeout=self._timeout, proxy=self._proxy)
+        try:
+            async with client.stream("POST", self._url, json=body, headers=headers, timeout=self._timeout) as resp:
+                if resp.status_code != 200:
+                    await resp.aread()
+                    raise ProviderError(f"http {resp.status_code}")
+                async for chunk in _iter_sse(resp):
+                    choice = (chunk.get("choices") or [{}])[0]
+                    delta = choice.get("delta") or {}
+                    if delta.get("content"):
+                        content.append(delta["content"])
+                        stream.emit("delta", delta["content"])
+                    for tc in delta.get("tool_calls") or []:
+                        slot = calls.setdefault(tc.get("index", 0), {"id": "", "name": "", "args": ""})
+                        slot["id"] = tc.get("id") or slot["id"]
+                        fn = tc.get("function") or {}
+                        slot["name"] += fn.get("name") or ""
+                        slot["args"] += fn.get("arguments") or ""
+        finally:
+            if self._client is None:
+                await client.aclose()
+    except httpx.HTTPError as exc:
+        raise ProviderError(type(exc).__name__) from exc
+    out = []
+    for i in sorted(calls):
+        c = calls[i]
+        try:
+            args = json.loads(c["args"] or "{}")
+        except json.JSONDecodeError:
+            args = {}
+        out.append(ToolCall(id=c["id"] or c["name"] or f"call{i}", name=c["name"], arguments=args))
+    if out and content:
+        stream.emit("reset", None)  # a preface before tool calls is not the answer
+    log.info("openrouter stream ok calls=%d chars=%d", len(out), sum(map(len, content)))
+    return AssistantTurn(content="".join(content) or None, tool_calls=out)
+
+
+OpenRouterProvider._chat_stream = _openrouter_stream  # type: ignore[attr-defined]
+
+
 # ------------------------------------------------------------------------------------------------ offline mock
 
 _DRUG_WORDS = re.compile(

app/llm/agent/stream.py

@@ -0,0 +1,32 @@
+"""Live progress of the chat agent for streaming channels (web SSE).
+
+The SSE endpoint sets a queue in the context; the agent pushes events while it works:
+    status — a tool step in plain Russian («Смотрю вашу карту ЕМИАС…»)
+    delta  — a chunk of the answer as the model writes it
+    reset  — the streamed text so far was a preface before tool calls: clear it
+Channels without a queue (Telegram, tests) are unaffected. Deltas are the raw model text: the final reply
+(forbidden-output filter, sources footer) always comes in the closing `done` event and replaces it.
+"""
+
+import asyncio
+from contextvars import ContextVar
+from typing import Any
+
+STREAM: ContextVar[asyncio.Queue[tuple[str, Any]] | None] = ContextVar("agent_stream", default=None)
+
+TOOL_STATUS = {
+    "get_patient_record": "Смотрю вашу карту ЕМИАС…",
+    "check_drug_for_patient": "Проверяю препарат по вашей карте: болезни, непереносимости, другие лекарства…",
+    "get_drug_info": "Открываю инструкцию к препарату (ГРЛС)…",
+    "search_guidelines": "Ищу в клинических рекомендациях Минздрава…",
+}
+
+
+def active() -> bool:
+    return STREAM.get() is not None
+
+
+def emit(kind: str, payload: Any) -> None:
+    q = STREAM.get()
+    if q is not None:
+        q.put_nowait((kind, payload))