Fix benchmark log extraction: first tier match, increase log tail to 300

- Remove reversed() from extract_tier_from_logs: first match = routing decision
  (dry-run complex logs tier=complex early, then overwrites with tier=medium at done)
- Increase log tail from 80→300 to handle concurrent log activity

Fixes #7, #10

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

.claude/rules/agent-pipeline.md

@@ -0,0 +1,22 @@
+# Agent Pipeline Rules
+
+## Tiers
+- Routing is fully automatic: router classifies into light/medium/complex via 3-way embedding similarity.
+- Complex tier is reached automatically for deep research queries — no prefix required.
+- Medium is the default tier. Light is only for trivial static-knowledge queries matched by regex or embedding.
+- Light tier upgrade to medium is automatic when URL content is pre-fetched or a fast tool matches.
+- `tier_override` API parameter still allows callers to force a specific tier (e.g. `adolf-deep` model → complex).
+
+## Medium agent
+- `_DirectModel` makes a single `ainvoke()` call with no tool schema. Do not add tools to the medium agent.
+- `qwen3:4b` behaves unreliably when a tool array is present in the request — inject context via system prompt instead.
+
+## Memory
+- `add_memory` and `search_memory` are called directly in `run_agent_task()`, outside the agent loop.
+- Never add memory tools to any agent's tool list.
+- Memory storage (`_store_memory`) runs as an asyncio background task after the semaphore is released.
+
+## Fast tools
+- `FastToolRunner.run_matching()` runs in the pre-flight `asyncio.gather` alongside URL fetch and memory retrieval.
+- Fast tool results are injected as a system prompt block, not returned to the user directly.
+- When `any_matches()` is true, the router forces medium tier before LLM classification.

.claude/rules/fast-tools.md

@@ -0,0 +1,24 @@
+---
+paths:
+  - "fast_tools.py"
+  - "agent.py"
+---
+
+# Fast Tools — Extension Guide
+
+To add a new fast tool:
+
+1. In `fast_tools.py`, subclass `FastTool` and implement:
+   - `name` (str property) — unique identifier, used in logs
+   - `matches(message: str) -> bool` — regex or logic; keep it cheap, runs on every message
+   - `run(message: str) -> str` — async fetch; return a short context block or `""` on failure; never raise
+
+2. In `agent.py`, add an instance to the `_fast_tool_runner` list (module level, after env vars are defined).
+
+3. The router will automatically force medium tier when `matches()` returns true — no router changes needed.
+
+## Constraints
+- `run()` must return in under 15s — it runs in the pre-flight gather that blocks routing.
+- Return `""` or a `[tool error: ...]` string on failure — never raise exceptions.
+- Keep returned context under ~1000 chars — larger contexts slow down `qwen3:4b` streaming significantly.
+- The deepagents container has no direct external internet. Use SearXNG (`host.docker.internal:11437`) or internal services.

.claude/rules/llm-inference.md

@@ -0,0 +1,8 @@
+# LLM Inference Rules
+
+- All LLM calls must use `base_url=LITELLM_URL` (points to LiteLLM at `host.docker.internal:4000/v1`). Never call Ollama directly for inference.
+- `_reply_semaphore` (asyncio.Semaphore(1)) serializes all GPU inference. Never bypass it or add a second semaphore.
+- Local Ollama models use the `ollama/` prefix: `ollama/qwen3:4b`, `ollama/qwen2.5:1.5b`. Remote models (e.g. OpenRouter) use their full LiteLLM name: `openrouter/deepseek-r1`.
+- Timeout values: router=30s, medium=180s, complex=600s. Do not reduce them.
+- `VRAMManager` is the only component that contacts Ollama directly (for flush/prewarm/poll). This is intentional — LiteLLM cannot manage VRAM.
+- Complex tier uses a remote model (`DEEPAGENTS_COMPLEX_MODEL`) — no VRAM management is needed for it.

.claude/rules/secrets.md

@@ -0,0 +1,7 @@
+# Secrets and Environment
+
+- `.env` is required at project root and must never be committed. It is in `.gitignore`.
+- Required keys: `TELEGRAM_BOT_TOKEN`, `ROUTECHECK_TOKEN`, `YANDEX_ROUTING_KEY`.
+- `ROUTECHECK_TOKEN` is a shared secret between `deepagents` and `routecheck` containers — generate once with `python3 -c "import uuid; print(uuid.uuid4())"`.
+- All tokens are stored in Vaultwarden (AI collection). Fetch with `bw get password "<NAME>"` — see `~/.claude/CLAUDE.md` for the full procedure.
+- Do not hardcode tokens, URLs, or credentials anywhere in source code.

.gitignore

@@ -1,2 +1,8 @@
 __pycache__/
 *.pyc
+logs/*.jsonl
+adolf_tuning_data/voice_audio/
+benchmarks/benchmark.json
+benchmarks/results_latest.json
+benchmarks/voice_results*.json
+benchmarks/voice_audio/

CLAUDE.md

@@ -4,175 +4,38 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Commands
 
-**Start all services:**
 ```bash
+# Start all services
 docker compose up --build
-```
 
-**Interactive CLI (Docker container, requires gateway running):**
+# Interactive CLI (requires services running)
-```bash
 docker compose --profile tools run --rm -it cli
-# or with options:
+
-docker compose --profile tools run --rm -it cli python3 cli.py --url http://deepagents:8000 --session cli-alvis
+# Integration tests — run from tests/integration/, require all services up
+python3 test_health.py
+python3 test_memory.py [--name-only|--bench-only|--dedup-only]
+python3 test_routing.py [--easy-only|--medium-only|--hard-only]
+
+# Use case tests — read the .md file and follow its steps as Claude Code
+# example: read tests/use_cases/weather_now.md and execute it
+
+# Routing benchmark — measures tier classification accuracy across 120 queries
+# Run from benchmarks/ — Adolf must be running. DO NOT run during active use (holds GPU).
+cd benchmarks
+python3 run_benchmark.py                       # full run (120 queries)
+python3 run_benchmark.py --tier light          # light tier only (30 queries)
+python3 run_benchmark.py --tier medium         # medium tier only (50 queries)
+python3 run_benchmark.py --tier complex --dry-run  # complex tier, medium model (no API cost)
+python3 run_benchmark.py --category smart_home_control
+python3 run_benchmark.py --ids 1,2,3
+python3 run_benchmark.py --list-categories
+
+# Voice benchmark
+python3 run_voice_benchmark.py
+
+# benchmark.json (dataset) and results_latest.json are gitignored — not committed
 ```
 
-**Run integration tests** (from `tests/integration/`, require all Docker services running):
-```bash
-python3 test_health.py                          # service health: deepagents, bifrost, Ollama, Qdrant, SearXNG
-
-python3 test_memory.py                          # name store/recall + memory benchmark + dedup
-python3 test_memory.py --name-only              # only name store/recall pipeline
-python3 test_memory.py --bench-only             # only 5-fact store + 10-question recall
-python3 test_memory.py --dedup-only             # only deduplication test
-
-python3 test_routing.py                         # all routing benchmarks (easy + medium + hard)
-python3 test_routing.py --easy-only             # light-tier routing benchmark
-python3 test_routing.py --medium-only           # medium-tier routing benchmark
-python3 test_routing.py --hard-only             # complex-tier + VRAM flush benchmark
-```
-
-Shared config and helpers are in `tests/integration/common.py`.
-
-**Use case tests** (`tests/use_cases/`) — markdown skill files executed by Claude Code, which acts as mock user and quality evaluator. Run by reading the `.md` file and following its steps with tools (Bash, WebFetch, etc.).
-
 ## Architecture
 
-Adolf is a multi-channel personal assistant. All LLM inference is routed through **Bifrost**, an open-source Go-based LLM gateway that adds retry logic, failover, and observability in front of Ollama.
+@README.md
-
-### Request flow
-
-```
-Channel adapter → POST /message {text, session_id, channel, user_id}
-                → 202 Accepted (immediate)
-                → background: run_agent_task()
-                    → asyncio.gather(
-                        _fetch_urls_from_message()          ← Crawl4AI, concurrent
-                        _retrieve_memories()                 ← openmemory search, concurrent
-                        _fast_tool_runner.run_matching()     ← FastTools (weather, commute), concurrent
-                      )
-                    → router.route() → tier decision (light/medium/complex)
-                        fast tool match → force medium
-                        if URL content fetched → upgrade light→medium
-                    → invoke agent for tier via Bifrost (url_context + memories in system prompt)
-                        deepagents:8000 → bifrost:8080/v1 → ollama:11436
-                    → _push_stream_chunk() per token (medium streaming) / full reply (light, complex)
-                        → _stream_queues[session_id] asyncio.Queue
-                    → _end_stream() sends [DONE] sentinel
-                    → channels.deliver(session_id, channel, reply)
-                        → channel-specific callback (Telegram POST)
-                    → _store_memory() background task (openmemory)
-CLI streaming    → GET /stream/{session_id}  (SSE, per-token for medium, single-chunk for others)
-```
-
-### Bifrost integration
-
-Bifrost (`bifrost-config.json`) is configured with the `ollama` provider pointing to the GPU Ollama instance on host port 11436. It exposes an OpenAI-compatible API at `http://bifrost:8080/v1`.
-
-`agent.py` uses `langchain_openai.ChatOpenAI` with `base_url=BIFROST_URL`. Model names use the `provider/model` format that Bifrost expects: `ollama/qwen3:4b`, `ollama/qwen3:8b`, `ollama/qwen2.5:1.5b`. Bifrost strips the `ollama/` prefix before forwarding to Ollama.
-
-`VRAMManager` bypasses Bifrost and talks directly to Ollama via `OLLAMA_BASE_URL` (host:11436) for flush/poll/prewarm operations — Bifrost cannot manage GPU VRAM.
-
-### Three-tier routing (`router.py`, `agent.py`)
-
-| Tier | Model (env var) | Trigger |
-|------|-----------------|---------|
-| light | `qwen2.5:1.5b` (`DEEPAGENTS_ROUTER_MODEL`) | Regex pre-match or LLM classifies "light" — answered by router model directly, no agent invoked |
-| medium | `qwen3:4b` (`DEEPAGENTS_MODEL`) | Default for tool-requiring queries |
-| complex | `qwen3:8b` (`DEEPAGENTS_COMPLEX_MODEL`) | `/think ` prefix only |
-
-The router does regex pre-classification first, then LLM classification. Complex tier is blocked unless the message starts with `/think ` — any LLM classification of "complex" is downgraded to medium.
-
-A global `asyncio.Semaphore(1)` (`_reply_semaphore`) serializes all LLM inference — one request at a time.
-
-### Thinking mode and streaming
-
-qwen3 models produce chain-of-thought `<think>...</think>` tokens. Handling differs by tier:
-
-- **Medium** (`qwen3:4b`): streams via `astream()`. A state machine (`in_think` flag) filters `<think>` blocks in real time — only non-think tokens are pushed to `_stream_queues` and displayed to the user.
-- **Complex** (`qwen3:8b`): `create_deep_agent` returns a complete reply; `_strip_think()` filters think blocks before the reply is pushed as a single chunk.
-- **Router/light** (`qwen2.5:1.5b`): no thinking support; `_strip_think()` used defensively.
-
-`_strip_think()` in `agent.py` and `router.py` strips any `<think>` blocks from non-streaming output.
-
-### VRAM management (`vram_manager.py`)
-
-Hardware: GTX 1070 (8 GB). Before running the 8b model, medium models are flushed via Ollama `keep_alive=0`, then `/api/ps` is polled (15s timeout) to confirm eviction. On timeout, falls back to medium tier. After complex reply, 8b is flushed and medium models are pre-warmed as a background task.
-
-### Channel adapters (`channels.py`)
-
-- **Telegram**: Grammy Node.js bot (`grammy/bot.mjs`) long-polls Telegram → `POST /message`; replies delivered via `POST grammy:3001/send`
-- **CLI**: `cli.py` (Docker container, `profiles: [tools]`) posts to `/message`, then streams from `GET /stream/{session_id}` SSE with Rich `Live` display and final Markdown render.
-
-Session IDs: `tg-<chat_id>` for Telegram, `cli-<username>` for CLI. Conversation history: 5-turn buffer per session.
-
-### Services (`docker-compose.yml`)
-
-| Service | Port | Role |
-|---------|------|------|
-| `bifrost` | 8080 | LLM gateway — retries, failover, observability; config from `bifrost-config.json` |
-| `deepagents` | 8000 | FastAPI gateway + agent core |
-| `openmemory` | 8765 | FastMCP server + mem0 memory tools (Qdrant-backed) |
-| `grammy` | 3001 | grammY Telegram bot + `/send` HTTP endpoint |
-| `crawl4ai` | 11235 | JS-rendered page fetching |
-| `routecheck` | 8090 | Local routing web service — image captcha UI + Yandex Routing API backend |
-| `cli` | — | Interactive CLI container (`profiles: [tools]`), Rich streaming display |
-
-External (from `openai/` stack, host ports):
-- Ollama GPU: `11436` — all reply inference (via Bifrost) + VRAM management (direct)
-- Ollama CPU: `11435` — nomic-embed-text embeddings for openmemory
-- Qdrant: `6333` — vector store for memories
-- SearXNG: `11437` — web search
-
-### Bifrost config (`bifrost-config.json`)
-
-The file is mounted into the bifrost container at `/app/data/config.json`. It declares one Ollama provider key pointing to `host.docker.internal:11436` with 2 retries and 300s timeout. To add fallback providers or adjust weights, edit this file and restart the bifrost container.
-
-### Crawl4AI integration
-
-Crawl4AI is embedded at all levels of the pipeline:
-
-- **Pre-routing (all tiers)**: `_fetch_urls_from_message()` detects URLs in any message via `_URL_RE`, fetches up to 3 URLs concurrently with `_crawl4ai_fetch_async()` (async httpx). URL content is injected as a system context block into enriched history before routing, and into the system prompt for medium/complex agents.
-- **Tier upgrade**: if URL content is successfully fetched, light tier is upgraded to medium (light model cannot process page content).
-- **Complex agent tools**: `web_search` (SearXNG + Crawl4AI auto-fetch of top 2 results) and `fetch_url` (single-URL Crawl4AI fetch) remain available for the complex agent's agentic loop. Complex tier also receives the pre-fetched content in system prompt to avoid redundant re-fetching.
-
-MCP tools from openmemory (`add_memory`, `search_memory`, `get_all_memories`) are **excluded** from agent tools — memory management is handled outside the agent loop.
-
-### Fast Tools (`fast_tools.py`)
-
-Pre-flight tools that run before the LLM in the `asyncio.gather` alongside URL fetch and memory retrieval. Each tool has a regex `matches()` classifier and an async `run()` that returns a context string injected into the system prompt. The router uses `FastToolRunner.any_matches()` to force medium tier when a tool matches.
-
-| Tool | Trigger | Data source |
-|------|---------|-------------|
-| `WeatherTool` | weather/forecast/temperature keywords | SearXNG query `"погода Балашиха сейчас"` — Russian sources return °C |
-| `CommuteTool` | commute/traffic/arrival time keywords | `routecheck:8090/api/route` — Yandex Routing API, Balashikha→Moscow center |
-
-To add a new fast tool: subclass `FastTool` in `fast_tools.py`, add an instance to `_fast_tool_runner` in `agent.py`.
-
-### `routecheck` service (`routecheck/app.py`)
-
-Local web service that exposes Yandex Routing API behind an image captcha. Two access paths:
-- **Web UI** (`localhost:8090`): solve PIL-generated arithmetic captcha → query any two lat/lon points
-- **Internal API**: `GET /api/route?from=lat,lon&to=lat,lon&token=ROUTECHECK_TOKEN` — bypasses captcha, used by `CommuteTool`
-
-Requires `.env`: `YANDEX_ROUTING_KEY` (free tier from `developer.tech.yandex.ru`) and `ROUTECHECK_TOKEN`. The container routes Yandex API calls through the host HTTPS proxy (`host.docker.internal:56928`).
-
-### Medium vs Complex agent
-
-| Agent | Builder | Speed | Use case |
-|-------|---------|-------|----------|
-| medium | `_DirectModel` (single LLM call, no tools) | ~3s | General questions, conversation |
-| complex | `create_deep_agent` (deepagents) | Slow — multi-step planner | Deep research via `/think` prefix |
-
-### Key files
-
-- `agent.py` — FastAPI app, lifespan wiring, `run_agent_task()`, Crawl4AI pre-fetch, fast tools, memory pipeline, all endpoints
-- `fast_tools.py` — `FastTool` base class, `FastToolRunner`, `WeatherTool`, `CommuteTool`
-- `routecheck/app.py` — captcha UI + `/api/route` Yandex proxy
-- `bifrost-config.json` — Bifrost provider config (Ollama GPU, retries, timeouts)
-- `channels.py` — channel registry and `deliver()` dispatcher
-- `router.py` — `Router` class: regex + LLM classification, light-tier reply generation
-- `vram_manager.py` — `VRAMManager`: flush/poll/prewarm Ollama VRAM directly
-- `agent_factory.py` — `build_medium_agent` (`_DirectModel`, single call) / `build_complex_agent` (`create_deep_agent`)
-- `openmemory/server.py` — FastMCP + mem0 config with custom extraction/dedup prompts
-- `wiki_research.py` — batch research pipeline using `/message` + SSE polling
-- `grammy/bot.mjs` — Telegram long-poll + HTTP `/send` endpoint

README.md

@@ -74,7 +74,7 @@ Autonomous personal assistant with a multi-channel gateway. Three-tier model rou
 
 Adolf uses LangChain's tool interface but only the complex agent actually invokes tools at runtime.
 
-**Complex agent (`/think` prefix):** `web_search` and `fetch_url` are defined as `langchain_core.tools.Tool` objects and passed to `create_deep_agent()`. The deepagents library runs an agentic loop (LangGraph `create_react_agent` under the hood) that sends the tool schema to the model via OpenAI function-calling format and handles tool dispatch.
+**Complex agent:** `web_search` and `fetch_url` are defined as `langchain_core.tools.Tool` objects and passed to `create_deep_agent()`. The deepagents library runs an agentic loop (LangGraph `create_react_agent` under the hood) that sends the tool schema to the model via OpenAI function-calling format and handles tool dispatch.
 
 **Medium agent (default):** `_DirectModel` makes a single `model.ainvoke(messages)` call with no tool schema. Context (memories, fetched URL content) is injected via the system prompt instead. This is intentional — `qwen3:4b` behaves unreliably when a tool array is present.
 
@@ -84,13 +84,13 @@ Adolf uses LangChain's tool interface but only the complex agent actually invoke
 
 | Tier | Model | Agent | Trigger | Latency |
 |------|-------|-------|---------|---------|
-| Light | `qwen2.5:1.5b` (router answers directly) | — | Regex pre-match or LLM classifies "light" | ~2–4s |
+| Light | `qwen2.5:1.5b` (router answers directly) | — | Regex pre-match or 3-way embedding classifies "light" | ~2–4s |
 | Medium | `qwen3:4b` (`DEEPAGENTS_MODEL`) | `_DirectModel` — single LLM call, no tools | Default; also forced when message contains URLs | ~10–20s |
-| Complex | `qwen3:8b` (`DEEPAGENTS_COMPLEX_MODEL`) | `create_deep_agent` — agentic loop with tools | `/think` prefix only | ~60–120s |
+| Complex | `deepseek/deepseek-r1:free` via LiteLLM (`DEEPAGENTS_COMPLEX_MODEL`) | `create_deep_agent` — agentic loop with tools | Auto-classified by embedding similarity | ~30–90s |
 
-**`/think` prefix**: forces complex tier, stripped before sending to agent.
+Routing is fully automatic via 3-way cosine similarity over pre-embedded utterance centroids (light / medium / complex). No prefix required. Use `adolf-deep` model name to force complex tier via API.
 
-Complex tier is locked out unless the message starts with `/think` — any LLM classification of "complex" is downgraded to medium.
+Complex tier is reached automatically for deep research queries — `исследуй`, `изучи все`, `напиши подробный`, etc. — via regex pre-classifier and embedding similarity. No prefix required. Use `adolf-deep` model name to force it via API.
 
 ## Fast Tools (`fast_tools.py`)
 
@@ -164,7 +164,7 @@ Conversation history is keyed by session_id (5-turn buffer).
 
 ```
 adolf/
-├── docker-compose.yml      Services: bifrost, deepagents, openmemory, grammy, crawl4ai, routecheck, cli
+├── docker-compose.yml      Services: deepagents, openmemory, grammy, crawl4ai, routecheck, cli
 ├── Dockerfile              deepagents container (Python 3.12)
 ├── Dockerfile.cli          CLI container (python:3.12-slim + rich)
 ├── agent.py                FastAPI gateway, run_agent_task, Crawl4AI pre-fetch, fast tools, memory pipeline
@@ -175,6 +175,11 @@ adolf/
 ├── agent_factory.py        _DirectModel (medium) / create_deep_agent (complex)
 ├── cli.py                  Interactive CLI REPL — Rich Live streaming + Markdown render
 ├── wiki_research.py        Batch wiki research pipeline (uses /message + SSE)
+├── benchmarks/
+│   ├── run_benchmark.py    Routing accuracy benchmark — 120 queries across 3 tiers
+│   ├── run_voice_benchmark.py  Voice path benchmark
+│   ├── benchmark.json      Query dataset (gitignored)
+│   └── results_latest.json Last run results (gitignored)
 ├── .env                    TELEGRAM_BOT_TOKEN, ROUTECHECK_TOKEN, YANDEX_ROUTING_KEY (not committed)
 ├── routecheck/
 │   ├── app.py              FastAPI: image captcha + /api/route Yandex proxy
@@ -195,7 +200,9 @@ adolf/
 
 | Service | Host Port | Role |
 |---------|-----------|------|
-| Ollama GPU | 11436 | All LLM inference (via Bifrost) + VRAM management (direct) + memory extraction |
+| LiteLLM | 4000 | LLM proxy — all inference goes through here (`LITELLM_URL` env var) |
+| Ollama GPU | 11436 | GPU inference backend + VRAM management (direct) + memory extraction |
 | Ollama CPU | 11435 | nomic-embed-text embeddings for openmemory |
+| Langfuse | 3200 | LLM observability — traces all requests via LiteLLM callbacks |
 | Qdrant | 6333 | Vector store for memories |
 | SearXNG | 11437 | Web search (used by `web_search` tool) |

agent.py

@@ -1,7 +1,9 @@
 import asyncio
+import json as _json_module
 import os
 import time
 from contextlib import asynccontextmanager
+from pathlib import Path
 
 from fastapi import FastAPI, BackgroundTasks, Request
 from fastapi.responses import JSONResponse, StreamingResponse
@@ -16,6 +18,7 @@ _URL_RE = _re.compile(r'https?://[^\s<>"\']+')
 def _extract_urls(text: str) -> list[str]:
     return _URL_RE.findall(text)
 
+from openai import AsyncOpenAI
 from langchain_openai import ChatOpenAI
 from langchain_mcp_adapters.client import MultiServerMCPClient
 from langchain_community.utilities import SearxSearchWrapper
@@ -27,8 +30,9 @@ from agent_factory import build_medium_agent, build_complex_agent
 from fast_tools import FastToolRunner, WeatherTool, CommuteTool
 import channels
 
-# Bifrost gateway — all LLM inference goes through here
+# LiteLLM proxy — all LLM inference goes through here
-BIFROST_URL = os.getenv("BIFROST_URL", "http://bifrost:8080/v1")
+LITELLM_URL = os.getenv("LITELLM_URL", "http://host.docker.internal:4000/v1")
+LITELLM_API_KEY = os.getenv("LITELLM_API_KEY", "dummy")
 # Direct Ollama URL — used only by VRAMManager for flush/prewarm/poll
 OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
 
@@ -44,6 +48,45 @@ ROUTECHECK_TOKEN = os.getenv("ROUTECHECK_TOKEN", "")
 MAX_HISTORY_TURNS = 5
 _conversation_buffers: dict[str, list] = {}
 
+# ── Interaction logging (RLHF data collection) ─────────────────────────────────
+_LOG_DIR = Path(os.getenv("ADOLF_LOG_DIR", "/app/logs"))
+_INTERACTIONS_LOG = _LOG_DIR / "interactions.jsonl"
+
+def _ensure_log_dir() -> None:
+    try:
+        _LOG_DIR.mkdir(parents=True, exist_ok=True)
+    except Exception as e:
+        print(f"[log] cannot create log dir {_LOG_DIR}: {e}", flush=True)
+
+
+async def _log_interaction(
+    session_id: str,
+    channel: str,
+    tier: str,
+    input_text: str,
+    response_text: str | None,
+    latency_ms: int,
+    metadata: dict | None = None,
+) -> None:
+    """Append one interaction record to the JSONL log for future RLHF/finetuning."""
+    record = {
+        "ts": time.time(),
+        "session_id": session_id,
+        "channel": channel,
+        "tier": tier,
+        "input": input_text,
+        "output": response_text or "",
+        "latency_ms": latency_ms,
+    }
+    if metadata:
+        record["metadata"] = metadata
+    try:
+        _ensure_log_dir()
+        with open(_INTERACTIONS_LOG, "a", encoding="utf-8") as f:
+            f.write(_json_module.dumps(record, ensure_ascii=False) + "\n")
+    except Exception as e:
+        print(f"[log] write error: {e}", flush=True)
+
 # Per-session streaming queues — filled during inference, read by /stream/{session_id}
 _stream_queues: dict[str, asyncio.Queue] = {}
 
@@ -123,7 +166,7 @@ _memory_search_tool = None
 
 # Fast tools run before the LLM — classifier + context enricher
 _fast_tool_runner = FastToolRunner([
-    WeatherTool(searxng_url=SEARXNG_URL),
+    WeatherTool(),
     CommuteTool(routecheck_url=ROUTECHECK_URL, internal_token=ROUTECHECK_TOKEN),
 ])
 
@@ -140,31 +183,30 @@ async def lifespan(app: FastAPI):
     channels.register_defaults()
 
     # All three models route through Bifrost → Ollama GPU.
-    # Bifrost adds retry logic, observability, and failover.
-    # Model names use provider/model format: Bifrost strips the "ollama/" prefix
-    # before forwarding to Ollama's /v1/chat/completions endpoint.
     router_model = ChatOpenAI(
         model=f"ollama/{ROUTER_MODEL}",
-        base_url=BIFROST_URL,
+        base_url=LITELLM_URL,
-        api_key="dummy",
+        api_key=LITELLM_API_KEY,
         temperature=0,
         timeout=30,
     )
+    embedder = AsyncOpenAI(base_url=LITELLM_URL, api_key=LITELLM_API_KEY)
     medium_model = ChatOpenAI(
         model=f"ollama/{MEDIUM_MODEL}",
-        base_url=BIFROST_URL,
+        base_url=LITELLM_URL,
-        api_key="dummy",
+        api_key=LITELLM_API_KEY,
         timeout=180,
     )
     complex_model = ChatOpenAI(
-        model=f"ollama/{COMPLEX_MODEL}",
+        model=COMPLEX_MODEL,  # full model name — may be remote (OpenRouter) or local ollama/*
-        base_url=BIFROST_URL,
+        base_url=LITELLM_URL,
-        api_key="dummy",
+        api_key=LITELLM_API_KEY,
         timeout=600,
     )
 
     vram_manager = VRAMManager(base_url=OLLAMA_BASE_URL)
-    router = Router(model=router_model, fast_tool_runner=_fast_tool_runner)
+    router = Router(model=router_model, embedder=embedder, fast_tool_runner=_fast_tool_runner)
+    await router.initialize()
 
     mcp_connections = {
         "openmemory": {"transport": "sse", "url": f"{OPENMEMORY_URL}/sse"},
@@ -279,8 +321,8 @@ async def lifespan(app: FastAPI):
     )
 
     print(
-        f"[agent] bifrost={BIFROST_URL} | router=ollama/{ROUTER_MODEL} | "
+        f"[agent] litellm={LITELLM_URL} | router=semantic(ollama/{ROUTER_MODEL}+nomic-embed-text) | "
-        f"medium=ollama/{MEDIUM_MODEL} | complex=ollama/{COMPLEX_MODEL}",
+        f"medium=ollama/{MEDIUM_MODEL} | complex={COMPLEX_MODEL}",
         flush=True,
     )
     print(f"[agent] agent tools: {[t.name for t in agent_tools]}", flush=True)
@@ -346,6 +388,12 @@ def _log_messages(result):
 
 # ── memory helpers ─────────────────────────────────────────────────────────────
 
+def _resolve_user_id(session_id: str) -> str:
+    """Map any session_id to a canonical user identity for openmemory.
+    All channels share the same memory pool for the single user."""
+    return "alvis"
+
+
 async def _store_memory(session_id: str, user_msg: str, assistant_reply: str) -> None:
     """Store a conversation turn in openmemory (runs as a background task)."""
     if _memory_add_tool is None:
@@ -353,7 +401,8 @@ async def _store_memory(session_id: str, user_msg: str, assistant_reply: str) ->
     t0 = time.monotonic()
     try:
         text = f"User: {user_msg}\nAssistant: {assistant_reply}"
-        await _memory_add_tool.ainvoke({"text": text, "user_id": session_id})
+        user_id = _resolve_user_id(session_id)
+        await _memory_add_tool.ainvoke({"text": text, "user_id": user_id})
         print(f"[memory] stored in {time.monotonic() - t0:.1f}s", flush=True)
     except Exception as e:
         print(f"[memory] error: {e}", flush=True)
@@ -364,7 +413,8 @@ async def _retrieve_memories(message: str, session_id: str) -> str:
     if _memory_search_tool is None:
         return ""
     try:
-        result = await _memory_search_tool.ainvoke({"query": message, "user_id": session_id})
+        user_id = _resolve_user_id(session_id)
+        result = await _memory_search_tool.ainvoke({"query": message, "user_id": user_id})
         if result and result.strip() and result.strip() != "[]":
             return f"Relevant memories:\n{result}"
     except Exception:
@@ -372,36 +422,41 @@ async def _retrieve_memories(message: str, session_id: str) -> str:
     return ""
 
 
-# ── core task ──────────────────────────────────────────────────────────────────
+# ── core pipeline ──────────────────────────────────────────────────────────────
 
-async def run_agent_task(message: str, session_id: str, channel: str = "telegram"):
+from typing import AsyncGenerator
-    print(f"[agent] queued: {message[:80]!r} chat={session_id}", flush=True)
 
-    force_complex = False
+async def _run_agent_pipeline(
-    clean_message = message
+    message: str,
-    if message.startswith("/think "):
+    history: list[dict],
-        force_complex = True
+    session_id: str,
-        clean_message = message[len("/think "):]
+    tier_override: str | None = None,
-        print("[agent] /think prefix → force_complex=True", flush=True)
+    dry_run: bool = False,
+) -> AsyncGenerator[str, None]:
+    """Core pipeline: pre-flight → routing → inference. Yields text chunks.
 
+    tier_override: "light" | "medium" | "complex" | None (auto-route)
+    dry_run: if True and tier=complex, log tier=complex but use medium model (avoids API cost)
+    Caller is responsible for scheduling _store_memory after consuming all chunks.
+    """
     async with _reply_semaphore:
         t0 = time.monotonic()
-        history = _conversation_buffers.get(session_id, [])
+        clean_message = message
         print(f"[agent] running: {clean_message[:80]!r}", flush=True)
 
-        # Fetch URL content, memories, and fast-tool context concurrently — all IO-bound
+        # Fetch URL content, memories, and fast-tool context concurrently
         url_context, memories, fast_context = await asyncio.gather(
             _fetch_urls_from_message(clean_message),
             _retrieve_memories(clean_message, session_id),
             _fast_tool_runner.run_matching(clean_message),
         )
         if url_context:
-            print(f"[agent] crawl4ai: {len(url_context)} chars fetched from message URLs", flush=True)
+            print(f"[agent] crawl4ai: {len(url_context)} chars fetched", flush=True)
         if fast_context:
             names = _fast_tool_runner.matching_names(clean_message)
             print(f"[agent] fast_tools={names}: {len(fast_context)} chars injected", flush=True)
 
-        # Build enriched history: memories + url_context + fast_context for ALL tiers
+        # Build enriched history
         enriched_history = list(history)
         if url_context:
             enriched_history = [{"role": "system", "content": url_context}] + enriched_history
@@ -410,87 +465,93 @@ async def run_agent_task(message: str, session_id: str, channel: str = "telegram
         if memories:
             enriched_history = [{"role": "system", "content": memories}] + enriched_history
 
-        tier, light_reply = await router.route(clean_message, enriched_history, force_complex)
-
-        # Messages with URL content must be handled by at least medium tier
-        if url_context and tier == "light":
-            tier = "medium"
-            light_reply = None
-            print("[agent] URL in message → upgraded light→medium", flush=True)
-        print(f"[agent] tier={tier} message={clean_message[:60]!r}", flush=True)
-
         final_text = None
+        llm_elapsed = 0.0
+
         try:
-            if tier == "light":
+            # Short-circuit: fast tool already has the answer
-                final_text = light_reply
+            if fast_context and tier_override is None and not url_context:
+                tier = "fast"
+                final_text = fast_context
                 llm_elapsed = time.monotonic() - t0
-                print(f"[agent] light path: answered by router", flush=True)
+                names = _fast_tool_runner.matching_names(clean_message)
-                await _push_stream_chunk(session_id, final_text)
+                print(f"[agent] tier=fast tools={names} — delivering directly", flush=True)
-                await _end_stream(session_id)
+                yield final_text
 
-            elif tier == "medium":
+            else:
-                system_prompt = MEDIUM_SYSTEM_PROMPT
+                # Determine tier
-                if memories:
+                if tier_override in ("light", "medium", "complex"):
-                    system_prompt = system_prompt + "\n\n" + memories
+                    tier = tier_override
-                if url_context:
+                    light_reply = None
-                    system_prompt = system_prompt + "\n\n" + url_context
+                    if tier_override == "light":
-                if fast_context:
+                        tier, light_reply = await router.route(clean_message, enriched_history)
-                    system_prompt = system_prompt + "\n\nLive web search results (use these to answer):\n\n" + fast_context
+                        tier = "light"
+                else:
+                    tier, light_reply = await router.route(clean_message, enriched_history)
+                    if url_context and tier == "light":
+                        tier = "medium"
+                        light_reply = None
+                        print("[agent] URL in message → upgraded light→medium", flush=True)
 
-                # Stream tokens directly — filter out qwen3 <think> blocks
+                # Dry-run: log as complex but infer with medium (no remote API call)
-                in_think = False
+                effective_tier = tier
-                response_parts = []
+                if dry_run and tier == "complex":
-                async for chunk in medium_model.astream([
+                    effective_tier = "medium"
-                    {"role": "system", "content": system_prompt},
+                    print(f"[agent] tier=complex (dry-run) → using medium model, message={clean_message[:60]!r}", flush=True)
-                    *history,
+                else:
-                    {"role": "user", "content": clean_message},
+                    print(f"[agent] tier={tier} message={clean_message[:60]!r}", flush=True)
-                ]):
+                tier = effective_tier
-                    token = chunk.content or ""
-                    if not token:
-                        continue
-                    if in_think:
-                        if "</think>" in token:
-                            in_think = False
-                            after = token.split("</think>", 1)[1]
-                            if after:
-                                await _push_stream_chunk(session_id, after)
-                                response_parts.append(after)
-                    else:
-                        if "<think>" in token:
-                            in_think = True
-                            before = token.split("<think>", 1)[0]
-                            if before:
-                                await _push_stream_chunk(session_id, before)
-                                response_parts.append(before)
-                        else:
-                            await _push_stream_chunk(session_id, token)
-                            response_parts.append(token)
 
-                await _end_stream(session_id)
+                if tier == "light":
-                llm_elapsed = time.monotonic() - t0
+                    final_text = light_reply
-                final_text = "".join(response_parts).strip() or None
+                    llm_elapsed = time.monotonic() - t0
+                    print("[agent] light path: answered by router", flush=True)
+                    yield final_text
 
-            else:  # complex
+                elif tier == "medium":
-                ok = await vram_manager.enter_complex_mode()
-                if not ok:
-                    print("[agent] complex→medium fallback (eviction timeout)", flush=True)
-                    tier = "medium"
                     system_prompt = MEDIUM_SYSTEM_PROMPT
                     if memories:
-                        system_prompt = system_prompt + "\n\n" + memories
+                        system_prompt += "\n\n" + memories
                     if url_context:
-                        system_prompt = system_prompt + "\n\n" + url_context
+                        system_prompt += "\n\n" + url_context
-                    result = await medium_agent.ainvoke({
+                    if fast_context:
-                        "messages": [
+                        system_prompt += "\n\nLive web search results (use these to answer):\n\n" + fast_context
-                            {"role": "system", "content": system_prompt},
+
-                            *history,
+                    in_think = False
-                            {"role": "user", "content": clean_message},
+                    response_parts = []
-                        ]
+                    async for chunk in medium_model.astream([
-                    })
+                        {"role": "system", "content": system_prompt},
-                else:
+                        *history,
+                        {"role": "user", "content": clean_message},
+                    ]):
+                        token = chunk.content or ""
+                        if not token:
+                            continue
+                        if in_think:
+                            if "</think>" in token:
+                                in_think = False
+                                after = token.split("</think>", 1)[1]
+                                if after:
+                                    yield after
+                                    response_parts.append(after)
+                        else:
+                            if "<think>" in token:
+                                in_think = True
+                                before = token.split("<think>", 1)[0]
+                                if before:
+                                    yield before
+                                    response_parts.append(before)
+                            else:
+                                yield token
+                                response_parts.append(token)
+
+                    llm_elapsed = time.monotonic() - t0
+                    final_text = "".join(response_parts).strip() or None
+
+                else:  # complex — remote model, no VRAM management needed
                     system_prompt = COMPLEX_SYSTEM_PROMPT.format(user_id=session_id)
                     if url_context:
-                        system_prompt = system_prompt + "\n\n[Pre-fetched URL content from user's message:]\n" + url_context
+                        system_prompt += "\n\n[Pre-fetched URL content from user's message:]\n" + url_context
                     result = await complex_agent.ainvoke({
                         "messages": [
                             {"role": "system", "content": system_prompt},
@@ -498,46 +559,90 @@ async def run_agent_task(message: str, session_id: str, channel: str = "telegram
                             {"role": "user", "content": clean_message},
                         ]
                     })
-                    asyncio.create_task(vram_manager.exit_complex_mode())
 
-                llm_elapsed = time.monotonic() - t0
+                    llm_elapsed = time.monotonic() - t0
-                _log_messages(result)
+                    _log_messages(result)
-                final_text = _extract_final_text(result)
+                    final_text = _extract_final_text(result)
-                if final_text:
+                    if final_text:
-                    await _push_stream_chunk(session_id, final_text)
+                        yield final_text
-                await _end_stream(session_id)
 
         except Exception as e:
             import traceback
             llm_elapsed = time.monotonic() - t0
-            print(f"[agent] error after {llm_elapsed:.1f}s for chat {session_id}: {e}", flush=True)
+            print(f"[agent] error after {llm_elapsed:.1f}s for {session_id}: {e}", flush=True)
             traceback.print_exc()
-            await _end_stream(session_id)
 
-        # Deliver reply through the originating channel
+        print(f"[agent] pipeline done in {time.monotonic() - t0:.1f}s tier={tier if 'tier' in dir() else '?'}", flush=True)
+
+        # Store memory as side-effect (non-blocking, best-effort)
         if final_text:
-            t1 = time.monotonic()
+            asyncio.create_task(_store_memory(session_id, clean_message, final_text))
+
+
+# ── core task (Telegram / Matrix / CLI wrapper) ─────────────────────────────────
+
+async def run_agent_task(
+    message: str,
+    session_id: str,
+    channel: str = "telegram",
+    metadata: dict | None = None,
+):
+    print(f"[agent] queued: {message[:80]!r} chat={session_id}", flush=True)
+    t0 = time.monotonic()
+
+    meta = metadata or {}
+    dry_run = bool(meta.get("dry_run", False))
+    is_benchmark = bool(meta.get("benchmark", False))
+
+    history = _conversation_buffers.get(session_id, [])
+    final_text = None
+    actual_tier = "unknown"
+
+    # Patch pipeline to capture tier for logging
+    # We read it from logs post-hoc; capture via a wrapper
+    async for chunk in _run_agent_pipeline(message, history, session_id, dry_run=dry_run):
+        await _push_stream_chunk(session_id, chunk)
+        if final_text is None:
+            final_text = chunk
+        else:
+            final_text += chunk
+
+    await _end_stream(session_id)
+
+    elapsed_ms = int((time.monotonic() - t0) * 1000)
+
+    if final_text:
+        final_text = final_text.strip()
+
+        # Skip channel delivery for benchmark sessions (no Telegram spam)
+        if not is_benchmark:
             try:
                 await channels.deliver(session_id, channel, final_text)
             except Exception as e:
                 print(f"[agent] delivery error (non-fatal): {e}", flush=True)
-            send_elapsed = time.monotonic() - t1
-            print(
-                f"[agent] replied in {time.monotonic() - t0:.1f}s "
-                f"(llm={llm_elapsed:.1f}s, send={send_elapsed:.1f}s) tier={tier}",
-                flush=True,
-            )
-            print(f"[agent] reply_text: {final_text}", flush=True)
-        else:
-            print("[agent] warning: no text reply from agent", flush=True)
 
-        # Update conversation buffer and schedule memory storage
+        print(f"[agent] replied in {elapsed_ms / 1000:.1f}s", flush=True)
-        if final_text:
+        print(f"[agent] reply_text: {final_text[:200]}", flush=True)
-            buf = _conversation_buffers.get(session_id, [])
+
-            buf.append({"role": "user", "content": clean_message})
+        # Update conversation buffer
-            buf.append({"role": "assistant", "content": final_text})
+        buf = _conversation_buffers.get(session_id, [])
-            _conversation_buffers[session_id] = buf[-(MAX_HISTORY_TURNS * 2):]
+        buf.append({"role": "user", "content": message})
-            asyncio.create_task(_store_memory(session_id, clean_message, final_text))
+        buf.append({"role": "assistant", "content": final_text})
+        _conversation_buffers[session_id] = buf[-(MAX_HISTORY_TURNS * 2):]
+
+        # Log interaction for RLHF data collection (skip benchmark sessions to avoid noise)
+        if not is_benchmark:
+            asyncio.create_task(_log_interaction(
+                session_id=session_id,
+                channel=channel,
+                tier=actual_tier,
+                input_text=message,
+                response_text=final_text,
+                latency_ms=elapsed_ms,
+                metadata=meta if meta else None,
+            ))
+    else:
+        print("[agent] warning: no text reply from agent", flush=True)
 
 
 # ── endpoints ──────────────────────────────────────────────────────────────────
@@ -549,7 +654,7 @@ async def message(request: InboundMessage, background_tasks: BackgroundTasks):
         return JSONResponse(status_code=503, content={"error": "Agent not ready"})
     session_id = request.session_id
     channel = request.channel
-    background_tasks.add_task(run_agent_task, request.text, session_id, channel)
+    background_tasks.add_task(run_agent_task, request.text, session_id, channel, request.metadata)
     return JSONResponse(status_code=202, content={"status": "accepted"})
 
 
@@ -611,3 +716,96 @@ async def stream_reply(session_id: str):
 @app.get("/health")
 async def health():
     return {"status": "ok", "agent_ready": medium_agent is not None}
+
+
+# ── OpenAI-compatible API (for OpenWebUI and other clients) ────────────────────
+
+_TIER_MAP = {
+    "adolf": None,
+    "adolf-light": "light",
+    "adolf-medium": "medium",
+    "adolf-deep": "complex",
+}
+
+
+@app.get("/v1/models")
+async def list_models():
+    return {
+        "object": "list",
+        "data": [
+            {"id": "adolf", "object": "model", "owned_by": "adolf"},
+            {"id": "adolf-light", "object": "model", "owned_by": "adolf"},
+            {"id": "adolf-medium", "object": "model", "owned_by": "adolf"},
+            {"id": "adolf-deep", "object": "model", "owned_by": "adolf"},
+        ],
+    }
+
+
+@app.post("/v1/chat/completions")
+async def chat_completions(request: Request):
+    if medium_agent is None:
+        return JSONResponse(status_code=503, content={"error": {"message": "Agent not ready", "type": "server_error"}})
+
+    body = await request.json()
+    model = body.get("model", "adolf")
+    messages = body.get("messages", [])
+    stream = body.get("stream", True)
+
+    # Extract current user message and history
+    user_messages = [m for m in messages if m.get("role") == "user"]
+    if not user_messages:
+        return JSONResponse(status_code=400, content={"error": {"message": "No user message", "type": "invalid_request_error"}})
+
+    current_message = user_messages[-1]["content"]
+    # History = everything before the last user message (excluding system messages from OpenWebUI)
+    last_user_idx = len(messages) - 1 - next(
+        i for i, m in enumerate(reversed(messages)) if m.get("role") == "user"
+    )
+    history = [m for m in messages[:last_user_idx] if m.get("role") in ("user", "assistant")]
+
+    session_id = request.headers.get("X-Session-Id", "owui-default")
+    tier_override = _TIER_MAP.get(model)
+
+    import json as _json
+    import uuid as _uuid
+
+    response_id = f"chatcmpl-{_uuid.uuid4().hex[:12]}"
+
+    if stream:
+        async def event_stream():
+            # Opening chunk with role
+            opening = {
+                "id": response_id, "object": "chat.completion.chunk",
+                "choices": [{"index": 0, "delta": {"role": "assistant"}, "finish_reason": None}]
+            }
+            yield f"data: {_json.dumps(opening)}\n\n"
+
+            async for chunk in _run_agent_pipeline(current_message, history, session_id, tier_override):
+                data = {
+                    "id": response_id, "object": "chat.completion.chunk",
+                    "choices": [{"index": 0, "delta": {"content": chunk}, "finish_reason": None}]
+                }
+                yield f"data: {_json.dumps(data)}\n\n"
+
+            # Final chunk
+            final = {
+                "id": response_id, "object": "chat.completion.chunk",
+                "choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]
+            }
+            yield f"data: {_json.dumps(final)}\n\n"
+            yield "data: [DONE]\n\n"
+
+        return StreamingResponse(event_stream(), media_type="text/event-stream")
+
+    else:
+        # Non-streaming: collect all chunks
+        parts = []
+        async for chunk in _run_agent_pipeline(current_message, history, session_id, tier_override):
+            if chunk:
+                parts.append(chunk)
+        full_text = "".join(parts).strip()
+        return {
+            "id": response_id, "object": "chat.completion",
+            "choices": [{"index": 0, "message": {"role": "assistant", "content": full_text}, "finish_reason": "stop"}],
+            "model": model,
+        }

benchmarks/run_benchmark.py

@@ -0,0 +1,318 @@
+#!/usr/bin/env python3
+"""
+Adolf routing benchmark.
+
+Sends each query to Adolf's /message endpoint, waits briefly for the routing
+decision to appear in docker logs, then records the actual tier.
+
+Usage:
+    python3 run_benchmark.py [options]
+    python3 run_benchmark.py --tier light|medium|complex
+    python3 run_benchmark.py --category <name>
+    python3 run_benchmark.py --ids 1,2,3
+    python3 run_benchmark.py --list-categories
+    python3 run_benchmark.py --dry-run         # complex queries use medium model (no API cost)
+
+IMPORTANT: Always check GPU is free before running. This script does it automatically.
+
+Adolf must be running at http://localhost:8000.
+"""
+
+import argparse
+import asyncio
+import json
+import re
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+import httpx
+
+ADOLF_URL = "http://localhost:8000"
+OLLAMA_URL = "http://localhost:11436"  # GPU Ollama
+DATASET = Path(__file__).parent / "benchmark.json"
+RESULTS = Path(__file__).parent / "results_latest.json"
+
+# Max time to wait for each query to fully complete via SSE stream
+QUERY_TIMEOUT = 300  # seconds — generous to handle GPU semaphore waits
+
+# Memory thresholds
+MIN_FREE_RAM_MB = 1500   # abort if less than this is free
+MIN_FREE_VRAM_MB = 500   # warn if less than this is free on GPU
+
+
+# ── Pre-flight checks ──────────────────────────────────────────────────────────
+
+def check_ram() -> tuple[bool, str]:
+    """Check available system RAM. Returns (ok, message)."""
+    try:
+        with open("/proc/meminfo") as f:
+            info = {}
+            for line in f:
+                parts = line.split()
+                if len(parts) >= 2:
+                    info[parts[0].rstrip(":")] = int(parts[1])
+        free_mb = (info.get("MemAvailable", 0)) // 1024
+        total_mb = info.get("MemTotal", 0) // 1024
+        msg = f"RAM: {free_mb} MB free / {total_mb} MB total"
+        if free_mb < MIN_FREE_RAM_MB:
+            return False, f"CRITICAL: {msg} — need at least {MIN_FREE_RAM_MB} MB free"
+        return True, msg
+    except Exception as e:
+        return True, f"RAM check failed (non-fatal): {e}"
+
+
+def check_gpu() -> tuple[bool, str]:
+    """Check GPU VRAM via Ollama /api/ps. Returns (ok, message)."""
+    try:
+        r = httpx.get(f"{OLLAMA_URL}/api/ps", timeout=5)
+        r.raise_for_status()
+        data = r.json()
+        models = data.get("models", [])
+        if models:
+            names = [m.get("name", "?") for m in models]
+            sizes_mb = [m.get("size_vram", 0) // (1024 * 1024) for m in models]
+            loaded = ", ".join(f"{n} ({s}MB)" for n, s in zip(names, sizes_mb))
+            total_vram = sum(sizes_mb)
+            if total_vram > 7000:
+                return False, f"GPU BUSY: models loaded = {loaded} — total VRAM used {total_vram}MB. Wait for models to unload."
+            return True, f"GPU: models loaded = {loaded} (total {total_vram}MB VRAM)"
+        return True, "GPU: idle (no models loaded)"
+    except httpx.ConnectError:
+        return True, "GPU check skipped (Ollama not reachable at localhost:11436)"
+    except Exception as e:
+        return True, f"GPU check failed (non-fatal): {e}"
+
+
+def preflight_checks(skip_gpu_check: bool = False) -> bool:
+    """Run all pre-flight checks. Returns True if safe to proceed."""
+    print("\n── Pre-flight checks ──────────────────────────────────────────")
+
+    ram_ok, ram_msg = check_ram()
+    print(f"  {'✓' if ram_ok else '✗'} {ram_msg}")
+    if not ram_ok:
+        print("\nABORTING: not enough RAM. Free up memory before running benchmark.")
+        return False
+
+    if not skip_gpu_check:
+        gpu_ok, gpu_msg = check_gpu()
+        print(f"  {'✓' if gpu_ok else '✗'} {gpu_msg}")
+        if not gpu_ok:
+            print("\nABORTING: GPU is busy. Wait for current inference to finish, then retry.")
+            return False
+
+    print("  All checks passed.\n")
+    return True
+
+
+# ── Log helpers ────────────────────────────────────────────────────────────────
+
+def get_log_tail(n: int = 50) -> str:
+    result = subprocess.run(
+        ["docker", "logs", "deepagents", "--tail", str(n)],
+        capture_output=True, text=True,
+    )
+    return result.stdout + result.stderr
+
+
+def extract_tier_from_logs(logs_before: str, logs_after: str) -> str | None:
+    """Find new tier= lines that appeared after we sent the query."""
+    before_lines = set(logs_before.splitlines())
+    new_lines = [l for l in logs_after.splitlines() if l not in before_lines]
+    for line in new_lines:
+        m = re.search(r"tier=(\w+(?:\s*\(dry-run\))?)", line)
+        if m:
+            tier_raw = m.group(1)
+            # Normalise: "complex (dry-run)" → "complex"
+            return tier_raw.split()[0]
+    return None
+
+
+# ── Request helpers ────────────────────────────────────────────────────────────
+
+async def post_message(
+    client: httpx.AsyncClient,
+    query_id: int,
+    query: str,
+    dry_run: bool = False,
+) -> bool:
+    payload = {
+        "text": query,
+        "session_id": f"benchmark-{query_id}",
+        "channel": "cli",
+        "user_id": "benchmark",
+        "metadata": {"dry_run": dry_run, "benchmark": True},
+    }
+    try:
+        r = await client.post(f"{ADOLF_URL}/message", json=payload, timeout=10)
+        r.raise_for_status()
+        return True
+    except Exception as e:
+        print(f" POST_ERROR: {e}", end="")
+        return False
+
+
+# ── Dataset ────────────────────────────────────────────────────────────────────
+
+def load_dataset() -> list[dict]:
+    with open(DATASET) as f:
+        return json.load(f)["queries"]
+
+
+def filter_queries(queries, tier, category, ids):
+    if tier:
+        queries = [q for q in queries if q["tier"] == tier]
+    if category:
+        queries = [q for q in queries if q["category"] == category]
+    if ids:
+        queries = [q for q in queries if q["id"] in ids]
+    return queries
+
+
+# ── Main run ───────────────────────────────────────────────────────────────────
+
+async def run(queries: list[dict], dry_run: bool = False) -> list[dict]:
+    results = []
+
+    async with httpx.AsyncClient() as client:
+        try:
+            r = await client.get(f"{ADOLF_URL}/health", timeout=5)
+            r.raise_for_status()
+        except Exception as e:
+            print(f"ERROR: Adolf not reachable: {e}", file=sys.stderr)
+            sys.exit(1)
+
+        total = len(queries)
+        correct = 0
+
+        dry_label = " [DRY-RUN: complex→medium]" if dry_run else ""
+        print(f"\nRunning {total} queries{dry_label}\n")
+        print(f"{'ID':>3}  {'EXPECTED':8}  {'ACTUAL':8}  {'OK':3}  {'TIME':6}  {'CATEGORY':22}  QUERY")
+        print("─" * 110)
+
+        for q in queries:
+            qid = q["id"]
+            expected = q["tier"]
+            category = q["category"]
+            query_text = q["query"]
+
+            # In dry-run, complex queries still use complex classification (logged), but medium infers
+            send_dry = dry_run and expected == "complex"
+            session_id = f"benchmark-{qid}"
+
+            print(f"{qid:>3}  {expected:8}  ", end="", flush=True)
+
+            logs_before = get_log_tail(300)
+            t0 = time.monotonic()
+
+            ok_post = await post_message(client, qid, query_text, dry_run=send_dry)
+            if not ok_post:
+                print(f"{'?':8}  {'ERR':3}  {'?':6}  {category:22}  {query_text[:40]}")
+                results.append({"id": qid, "expected": expected, "actual": None, "ok": False})
+                continue
+
+            # Wait for query to complete via SSE stream (handles GPU semaphore waits)
+            try:
+                async with client.stream(
+                    "GET", f"{ADOLF_URL}/stream/{session_id}", timeout=QUERY_TIMEOUT
+                ) as sse:
+                    async for line in sse.aiter_lines():
+                        if "data: [DONE]" in line:
+                            break
+            except Exception:
+                pass  # timeout or connection issue — check logs anyway
+
+            # Now the query is done — check logs for tier
+            await asyncio.sleep(0.3)
+            logs_after = get_log_tail(300)
+            actual = extract_tier_from_logs(logs_before, logs_after)
+
+            elapsed = time.monotonic() - t0
+            match = actual == expected or (actual == "fast" and expected == "medium")
+            if match:
+                correct += 1
+
+            mark = "✓" if match else "✗"
+            actual_str = actual or "?"
+            print(f"{actual_str:8}  {mark:3}  {elapsed:5.1f}s  {category:22}  {query_text[:40]}")
+
+            results.append({
+                "id": qid,
+                "expected": expected,
+                "actual": actual_str,
+                "ok": match,
+                "elapsed": round(elapsed, 1),
+                "category": category,
+                "query": query_text,
+                "dry_run": send_dry,
+            })
+
+        print("─" * 110)
+        accuracy = correct / total * 100 if total else 0
+        print(f"\nAccuracy: {correct}/{total} ({accuracy:.0f}%)")
+
+        for tier_name in ["light", "medium", "complex"]:
+            tier_qs = [r for r in results if r["expected"] == tier_name]
+            if tier_qs:
+                tier_ok = sum(1 for r in tier_qs if r["ok"])
+                print(f"  {tier_name:8}: {tier_ok}/{len(tier_qs)}")
+
+        wrong = [r for r in results if not r["ok"]]
+        if wrong:
+            print(f"\nMisclassified ({len(wrong)}):")
+            for r in wrong:
+                print(f"  id={r['id']:3}  expected={r['expected']:8}  actual={r['actual']:8}  {r['query'][:60]}")
+
+    with open(RESULTS, "w") as f:
+        json.dump(results, f, indent=2, ensure_ascii=False)
+    print(f"\nResults saved to {RESULTS}")
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Adolf routing benchmark",
+        epilog="IMPORTANT: Always check GPU is free before running. This is done automatically."
+    )
+    parser.add_argument("--tier", choices=["light", "medium", "complex"])
+    parser.add_argument("--category")
+    parser.add_argument("--ids", help="Comma-separated IDs")
+    parser.add_argument("--list-categories", action="store_true")
+    parser.add_argument(
+        "--dry-run",
+        action="store_true",
+        help="For complex queries: route classification is tested but medium model is used for inference (no API cost)",
+    )
+    parser.add_argument(
+        "--skip-gpu-check",
+        action="store_true",
+        help="Skip GPU availability check (use only if you know GPU is free)",
+    )
+    args = parser.parse_args()
+
+    queries = load_dataset()
+
+    if args.list_categories:
+        cats = sorted(set(q["category"] for q in queries))
+        tiers = {t: sum(1 for q in queries if q["tier"] == t) for t in ["light", "medium", "complex"]}
+        print(f"Total: {len(queries)} | Tiers: {tiers}")
+        print(f"Categories: {cats}")
+        return
+
+    # ALWAYS check GPU and RAM before running
+    if not preflight_checks(skip_gpu_check=args.skip_gpu_check):
+        sys.exit(1)
+
+    ids = [int(i) for i in args.ids.split(",")] if args.ids else None
+    queries = filter_queries(queries, args.tier, args.category, ids)
+    if not queries:
+        print("No queries match filters.")
+        sys.exit(1)
+
+    asyncio.run(run(queries, dry_run=args.dry_run))
+
+
+if __name__ == "__main__":
+    main()

benchmarks/run_voice_benchmark.py

@@ -0,0 +1,426 @@
+#!/usr/bin/env python3
+"""
+Adolf voice benchmark.
+
+Pipeline for each query:
+  1. Synthesize query text → WAV via Silero TTS (localhost:8881)
+  2. Transcribe WAV → text via faster-whisper STT (localhost:8880)
+  3. Send transcription to Adolf → check routing tier
+  4. Report: WER per query, routing accuracy vs text baseline
+
+Usage:
+    python3 run_voice_benchmark.py [options]
+    python3 run_voice_benchmark.py --tier light|medium|complex
+    python3 run_voice_benchmark.py --ids 1,2,3
+    python3 run_voice_benchmark.py --dry-run       # complex queries use medium model
+
+IMPORTANT: Always check GPU is free before running. Done automatically.
+
+Services required:
+  - Adolf:         http://localhost:8000
+  - Silero TTS:    http://localhost:8881  (openai/silero-tts container)
+  - faster-whisper: http://localhost:8880  (faster-whisper container)
+"""
+
+import argparse
+import asyncio
+import io
+import json
+import re
+import subprocess
+import sys
+import tempfile
+import time
+import unicodedata
+from pathlib import Path
+
+import httpx
+
+ADOLF_URL = "http://localhost:8000"
+OLLAMA_URL = "http://localhost:11436"
+TTS_URL = "http://localhost:8881"       # Silero TTS — OpenAI-compatible /v1/audio/speech
+STT_URL = "http://localhost:8880"       # faster-whisper — OpenAI-compatible /v1/audio/transcriptions
+
+DATASET = Path(__file__).parent / "benchmark.json"
+RESULTS_DIR = Path(__file__).parent
+
+TIER_WAIT = 15        # seconds to wait for tier= in docker logs
+MIN_FREE_RAM_MB = 1500
+MIN_FREE_VRAM_MB = 500
+
+
+# ── Pre-flight ─────────────────────────────────────────────────────────────────
+
+def check_ram() -> tuple[bool, str]:
+    try:
+        with open("/proc/meminfo") as f:
+            info = {}
+            for line in f:
+                parts = line.split()
+                if len(parts) >= 2:
+                    info[parts[0].rstrip(":")] = int(parts[1])
+        free_mb = info.get("MemAvailable", 0) // 1024
+        total_mb = info.get("MemTotal", 0) // 1024
+        msg = f"RAM: {free_mb} MB free / {total_mb} MB total"
+        if free_mb < MIN_FREE_RAM_MB:
+            return False, f"CRITICAL: {msg} — need at least {MIN_FREE_RAM_MB} MB free"
+        return True, msg
+    except Exception as e:
+        return True, f"RAM check failed (non-fatal): {e}"
+
+
+def check_gpu() -> tuple[bool, str]:
+    try:
+        r = httpx.get(f"{OLLAMA_URL}/api/ps", timeout=5)
+        r.raise_for_status()
+        data = r.json()
+        models = data.get("models", [])
+        if models:
+            names = [m.get("name", "?") for m in models]
+            sizes_mb = [m.get("size_vram", 0) // (1024 * 1024) for m in models]
+            loaded = ", ".join(f"{n} ({s}MB)" for n, s in zip(names, sizes_mb))
+            total_vram = sum(sizes_mb)
+            if total_vram > 7000:
+                return False, f"GPU BUSY: {loaded} — {total_vram}MB VRAM used. Wait for models to unload."
+            return True, f"GPU: {loaded} ({total_vram}MB VRAM)"
+        return True, "GPU: idle"
+    except httpx.ConnectError:
+        return True, "GPU check skipped (Ollama not reachable)"
+    except Exception as e:
+        return True, f"GPU check failed (non-fatal): {e}"
+
+
+def check_services() -> tuple[bool, str]:
+    """Check TTS and STT are reachable."""
+    msgs = []
+    ok = True
+    for name, url, path in [("TTS", TTS_URL, "/"), ("STT", STT_URL, "/")]:
+        try:
+            r = httpx.get(url + path, timeout=5)
+            msgs.append(f"{name}: reachable (HTTP {r.status_code})")
+        except Exception as e:
+            msgs.append(f"{name}: NOT REACHABLE — {e}")
+            ok = False
+    return ok, " | ".join(msgs)
+
+
+def preflight_checks(skip_gpu_check: bool = False) -> bool:
+    print("\n── Pre-flight checks ──────────────────────────────────────────")
+    ram_ok, ram_msg = check_ram()
+    print(f"  {'✓' if ram_ok else '✗'} {ram_msg}")
+    if not ram_ok:
+        print("\nABORTING: not enough RAM.")
+        return False
+
+    if not skip_gpu_check:
+        gpu_ok, gpu_msg = check_gpu()
+        print(f"  {'✓' if gpu_ok else '✗'} {gpu_msg}")
+        if not gpu_ok:
+            print("\nABORTING: GPU is busy.")
+            return False
+
+    svc_ok, svc_msg = check_services()
+    print(f"  {'✓' if svc_ok else '✗'} {svc_msg}")
+    if not svc_ok:
+        print("\nABORTING: required voice services not running.")
+        print("Start them with: cd /home/alvis/agap_git/openai && docker compose up -d faster-whisper silero-tts")
+        return False
+
+    print("  All checks passed.\n")
+    return True
+
+
+# ── TTS ────────────────────────────────────────────────────────────────────────
+
+async def synthesize(client: httpx.AsyncClient, text: str) -> bytes | None:
+    """Synthesize text to WAV via Silero TTS (OpenAI-compatible /v1/audio/speech)."""
+    try:
+        r = await client.post(
+            f"{TTS_URL}/v1/audio/speech",
+            json={"model": "tts-1", "input": text, "voice": "alloy", "response_format": "wav"},
+            timeout=30,
+        )
+        r.raise_for_status()
+        return r.content
+    except Exception as e:
+        print(f"\n  [TTS error: {e}]", end="")
+        return None
+
+
+# ── STT ────────────────────────────────────────────────────────────────────────
+
+async def transcribe(client: httpx.AsyncClient, wav_bytes: bytes) -> str | None:
+    """Transcribe WAV to text via faster-whisper (OpenAI-compatible /v1/audio/transcriptions)."""
+    try:
+        files = {"file": ("audio.wav", wav_bytes, "audio/wav")}
+        data = {"model": "whisper-1", "language": "ru", "response_format": "json"}
+        r = await client.post(
+            f"{STT_URL}/v1/audio/transcriptions",
+            files=files,
+            data=data,
+            timeout=60,
+        )
+        r.raise_for_status()
+        result = r.json()
+        return result.get("text", "").strip()
+    except Exception as e:
+        print(f"\n  [STT error: {e}]", end="")
+        return None
+
+
+# ── WER ────────────────────────────────────────────────────────────────────────
+
+def normalize(text: str) -> str:
+    """Lowercase, strip punctuation, normalize unicode for WER calculation."""
+    text = unicodedata.normalize("NFC", text.lower())
+    text = re.sub(r"[^\w\s]", " ", text)
+    return re.sub(r"\s+", " ", text).strip()
+
+
+def word_error_rate(reference: str, hypothesis: str) -> float:
+    """Compute WER between reference and hypothesis."""
+    ref = normalize(reference).split()
+    hyp = normalize(hypothesis).split()
+    if not ref:
+        return 0.0 if not hyp else 1.0
+    # Dynamic programming edit distance
+    d = [[0] * (len(hyp) + 1) for _ in range(len(ref) + 1)]
+    for i in range(len(ref) + 1):
+        d[i][0] = i
+    for j in range(len(hyp) + 1):
+        d[0][j] = j
+    for i in range(1, len(ref) + 1):
+        for j in range(1, len(hyp) + 1):
+            if ref[i - 1] == hyp[j - 1]:
+                d[i][j] = d[i - 1][j - 1]
+            else:
+                d[i][j] = 1 + min(d[i - 1][j], d[i][j - 1], d[i - 1][j - 1])
+    return d[len(ref)][len(hyp)] / len(ref)
+
+
+# ── Adolf interaction ──────────────────────────────────────────────────────────
+
+def get_log_tail(n: int = 60) -> str:
+    result = subprocess.run(
+        ["docker", "logs", "deepagents", "--tail", str(n)],
+        capture_output=True, text=True,
+    )
+    return result.stdout + result.stderr
+
+
+def extract_tier_from_logs(logs_before: str, logs_after: str) -> str | None:
+    before_lines = set(logs_before.splitlines())
+    new_lines = [l for l in logs_after.splitlines() if l not in before_lines]
+    for line in reversed(new_lines):
+        m = re.search(r"tier=(\w+(?:\s*\(dry-run\))?)", line)
+        if m:
+            return m.group(1).split()[0]
+    return None
+
+
+async def post_to_adolf(
+    client: httpx.AsyncClient,
+    query_id: int,
+    text: str,
+    dry_run: bool = False,
+) -> bool:
+    payload = {
+        "text": text,
+        "session_id": f"voice-bench-{query_id}",
+        "channel": "cli",
+        "user_id": "benchmark",
+        "metadata": {"dry_run": dry_run, "benchmark": True, "voice": True},
+    }
+    try:
+        r = await client.post(f"{ADOLF_URL}/message", json=payload, timeout=10)
+        r.raise_for_status()
+        return True
+    except Exception as e:
+        print(f"\n  [Adolf error: {e}]", end="")
+        return False
+
+
+# ── Dataset ────────────────────────────────────────────────────────────────────
+
+def load_dataset() -> list[dict]:
+    with open(DATASET) as f:
+        return json.load(f)["queries"]
+
+
+def filter_queries(queries, tier, category, ids):
+    if tier:
+        queries = [q for q in queries if q["tier"] == tier]
+    if category:
+        queries = [q for q in queries if q["category"] == category]
+    if ids:
+        queries = [q for q in queries if q["id"] in ids]
+    return queries
+
+
+# ── Main run ───────────────────────────────────────────────────────────────────
+
+async def run(queries: list[dict], dry_run: bool = False, save_audio: bool = False) -> None:
+    async with httpx.AsyncClient() as client:
+        # Check Adolf
+        try:
+            r = await client.get(f"{ADOLF_URL}/health", timeout=5)
+            r.raise_for_status()
+        except Exception as e:
+            print(f"ERROR: Adolf not reachable: {e}", file=sys.stderr)
+            sys.exit(1)
+
+        total = len(queries)
+        results = []
+
+        dry_label = " [DRY-RUN]" if dry_run else ""
+        print(f"Voice benchmark: {total} queries{dry_label}\n")
+        print(f"{'ID':>3}  {'EXP':8}  {'ACT':8}  {'OK':3}  {'WER':5}  {'TRANSCRIPT'}")
+        print("─" * 100)
+
+        total_wer = 0.0
+        wer_count = 0
+        correct = 0
+
+        for q in queries:
+            qid = q["id"]
+            expected = q["tier"]
+            original = q["query"]
+            print(f"{qid:>3}  {expected:8}  ", end="", flush=True)
+
+            # Step 1: TTS
+            wav = await synthesize(client, original)
+            if wav is None:
+                print(f"{'?':8}  {'ERR':3}  {'?':5}  [TTS failed]")
+                results.append({"id": qid, "expected": expected, "actual": None, "ok": False, "wer": None, "error": "tts"})
+                continue
+
+            if save_audio:
+                audio_path = RESULTS_DIR / f"voice_audio" / f"{qid}.wav"
+                audio_path.parent.mkdir(exist_ok=True)
+                audio_path.write_bytes(wav)
+
+            # Step 2: STT
+            transcript = await transcribe(client, wav)
+            if transcript is None:
+                print(f"{'?':8}  {'ERR':3}  {'?':5}  [STT failed]")
+                results.append({"id": qid, "expected": expected, "actual": None, "ok": False, "wer": None, "error": "stt"})
+                continue
+
+            # Calculate WER
+            wer = word_error_rate(original, transcript)
+            total_wer += wer
+            wer_count += 1
+
+            # Step 3: Send to Adolf
+            send_dry = dry_run and expected == "complex"
+            logs_before = get_log_tail(60)
+            t0 = time.monotonic()
+
+            ok_post = await post_to_adolf(client, qid, transcript, dry_run=send_dry)
+            if not ok_post:
+                print(f"{'?':8}  {'ERR':3}  {wer:4.2f}  {transcript[:50]}")
+                results.append({"id": qid, "expected": expected, "actual": None, "ok": False, "wer": wer, "transcript": transcript})
+                continue
+
+            # Step 4: Wait for routing decision
+            actual = None
+            for _ in range(TIER_WAIT * 2):
+                await asyncio.sleep(0.5)
+                logs_after = get_log_tail(60)
+                actual = extract_tier_from_logs(logs_before, logs_after)
+                if actual and actual in ("light", "medium", "complex", "fast"):
+                    break
+
+            elapsed = time.monotonic() - t0
+            match = actual == expected or (actual == "fast" and expected == "medium")
+            if match:
+                correct += 1
+
+            mark = "✓" if match else "✗"
+            actual_str = actual or "?"
+            print(f"{actual_str:8}  {mark:3}  {wer:4.2f}  {transcript[:60]}")
+
+            results.append({
+                "id": qid,
+                "expected": expected,
+                "actual": actual_str,
+                "ok": match,
+                "wer": round(wer, 3),
+                "original": original,
+                "transcript": transcript,
+                "elapsed": round(elapsed, 1),
+                "dry_run": send_dry,
+            })
+
+            await asyncio.sleep(0.5)
+
+        print("─" * 100)
+
+        # Summary
+        accuracy = correct / total * 100 if total else 0
+        avg_wer = total_wer / wer_count * 100 if wer_count else 0
+        print(f"\nRouting accuracy: {correct}/{total} ({accuracy:.0f}%)")
+        print(f"Average WER:      {avg_wer:.1f}%  (lower is better; 0% = perfect transcription)")
+
+        for tier_name in ["light", "medium", "complex"]:
+            tier_qs = [r for r in results if r["expected"] == tier_name]
+            if tier_qs:
+                tier_ok = sum(1 for r in tier_qs if r["ok"])
+                tier_wers = [r["wer"] for r in tier_qs if r.get("wer") is not None]
+                avg = sum(tier_wers) / len(tier_wers) * 100 if tier_wers else 0
+                print(f"  {tier_name:8}: routing {tier_ok}/{len(tier_qs)}  avg WER {avg:.1f}%")
+
+        wrong = [r for r in results if not r["ok"]]
+        if wrong:
+            print(f"\nMisclassified after voice ({len(wrong)}):")
+            for r in wrong:
+                print(f"  id={r['id']:3}  expected={r.get('expected','?'):8}  actual={r.get('actual','?'):8}  transcript={r.get('transcript','')[:50]}")
+
+        high_wer = [r for r in results if r.get("wer") and r["wer"] > 0.3]
+        if high_wer:
+            print(f"\nHigh WER queries (>30%) — transcription quality issues:")
+            for r in high_wer:
+                print(f"  id={r['id']:3}  WER={r['wer']*100:.0f}%  original: {r.get('original','')[:50]}")
+                print(f"          transcript: {r.get('transcript','')[:50]}")
+
+        # Save results
+        ts = int(time.time())
+        out_path = RESULTS_DIR / f"voice_results_{ts}.json"
+        latest_path = RESULTS_DIR / "voice_results_latest.json"
+        with open(out_path, "w") as f:
+            json.dump({"summary": {"accuracy": accuracy, "avg_wer": avg_wer, "total": total}, "results": results}, f, indent=2, ensure_ascii=False)
+        with open(latest_path, "w") as f:
+            json.dump({"summary": {"accuracy": accuracy, "avg_wer": avg_wer, "total": total}, "results": results}, f, indent=2, ensure_ascii=False)
+        print(f"\nResults saved to {latest_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Adolf voice benchmark — TTS→STT→routing pipeline",
+        epilog="Requires: Silero TTS (port 8881) and faster-whisper (port 8880) running."
+    )
+    parser.add_argument("--tier", choices=["light", "medium", "complex"])
+    parser.add_argument("--category")
+    parser.add_argument("--ids", help="Comma-separated IDs")
+    parser.add_argument("--dry-run", action="store_true",
+                        help="Complex queries use medium model for inference (no API cost)")
+    parser.add_argument("--save-audio", action="store_true",
+                        help="Save synthesized WAV files to voice_audio/ directory")
+    parser.add_argument("--skip-gpu-check", action="store_true")
+    args = parser.parse_args()
+
+    if not preflight_checks(skip_gpu_check=args.skip_gpu_check):
+        sys.exit(1)
+
+    queries = load_dataset()
+    ids = [int(i) for i in args.ids.split(",")] if args.ids else None
+    queries = filter_queries(queries, args.tier, args.category, ids)
+    if not queries:
+        print("No queries match filters.")
+        sys.exit(1)
+
+    asyncio.run(run(queries, dry_run=args.dry_run, save_audio=args.save_audio))
+
+
+if __name__ == "__main__":
+    main()

bifrost-config.json

@@ -1,4 +1,21 @@
 {
+  "auth_config": {
+    "is_enabled": true,
+    "admin_username": "admin",
+    "admin_password": "env.BIFROST_ADMIN_PASSWORD"
+  },
+  "config_store": {
+    "enabled": true,
+    "type": "postgres",
+    "config": {
+      "host": "bifrost-db",
+      "port": "5432",
+      "user": "bifrost",
+      "password": "bifrost",
+      "db_name": "bifrost",
+      "ssl_mode": "disable"
+    }
+  },
   "client": {
     "drop_excess_requests": false
   },

channels.py

@@ -49,6 +49,7 @@ async def deliver(session_id: str, channel: str, text: str) -> None:
 # ── built-in channel adapters ─────────────────────────────────────────────────
 
 GRAMMY_URL = os.getenv("GRAMMY_URL", "http://grammy:3001")
+MATRIX_URL = os.getenv("MATRIX_URL", "http://matrix:3002")
 
 
 async def _telegram_send(session_id: str, text: str) -> None:
@@ -64,12 +65,26 @@ async def _telegram_send(session_id: str, text: str) -> None:
             )
 
 
+async def _matrix_send(session_id: str, text: str) -> None:
+    """Send reply to Matrix via the matrix adapter POST /send endpoint."""
+    room_id = session_id.removeprefix("mx-")
+    MAX_MTX = 4000
+    chunks = [text[i:i + MAX_MTX] for i in range(0, len(text), MAX_MTX)]
+    async with httpx.AsyncClient(timeout=15) as client:
+        for chunk in chunks:
+            await client.post(
+                f"{MATRIX_URL}/send",
+                json={"room_id": room_id, "text": chunk},
+            )
+
+
 async def _cli_send(session_id: str, text: str) -> None:
     """CLI replies are handled entirely through the pending_replies queue — no-op here."""
     pass
 
 
 def register_defaults() -> None:
-    """Register the built-in Telegram and CLI channel adapters."""
+    """Register the built-in Telegram, Matrix, and CLI channel adapters."""
     register("telegram", _telegram_send)
+    register("matrix", _matrix_send)
     register("cli", _cli_send)

docker-compose.yml

@@ -1,19 +1,4 @@
 services:
-  bifrost:
-    image: maximhq/bifrost
-    container_name: bifrost
-    ports:
-      - "8080:8080"
-    volumes:
-      - ./bifrost-config.json:/app/data/config.json:ro
-    environment:
-      - APP_DIR=/app/data
-      - APP_PORT=8080
-      - LOG_LEVEL=info
-    extra_hosts:
-      - "host.docker.internal:host-gateway"
-    restart: unless-stopped
-
   deepagents:
     build: .
     container_name: deepagents
@@ -21,25 +6,28 @@ services:
       - "8000:8000"
     environment:
       - PYTHONUNBUFFERED=1
-      # Bifrost gateway — all LLM inference goes through here
+      # LiteLLM proxy — all LLM inference goes through here
-      - BIFROST_URL=http://bifrost:8080/v1
+      - LITELLM_URL=http://host.docker.internal:4000/v1
+      - LITELLM_API_KEY=sk-fjQC1BxAiGFSMs
       # Direct Ollama GPU URL — used only by VRAMManager for flush/prewarm
       - OLLAMA_BASE_URL=http://host.docker.internal:11436
       - DEEPAGENTS_MODEL=qwen3:4b
-      - DEEPAGENTS_COMPLEX_MODEL=qwen3:8b
+      - DEEPAGENTS_COMPLEX_MODEL=deepseek/deepseek-r1:free
       - DEEPAGENTS_ROUTER_MODEL=qwen2.5:1.5b
       - SEARXNG_URL=http://host.docker.internal:11437
       - GRAMMY_URL=http://grammy:3001
+      - MATRIX_URL=http://host.docker.internal:3002
       - CRAWL4AI_URL=http://crawl4ai:11235
       - ROUTECHECK_URL=http://routecheck:8090
       - ROUTECHECK_TOKEN=${ROUTECHECK_TOKEN}
+    volumes:
+      - ./logs:/app/logs
     extra_hosts:
       - "host.docker.internal:host-gateway"
     depends_on:
       - openmemory
       - grammy
       - crawl4ai
-      - bifrost
       - routecheck
     restart: unless-stopped
 

fast_tools.py

@@ -35,13 +35,22 @@ class FastTool(ABC):
     async def run(self, message: str) -> str: ...
 
 
+_WMO_CODES = {
+    0: "clear sky", 1: "mainly clear", 2: "partly cloudy", 3: "overcast",
+    45: "fog", 48: "icy fog",
+    51: "light drizzle", 53: "drizzle", 55: "heavy drizzle",
+    61: "light rain", 63: "rain", 65: "heavy rain",
+    71: "light snow", 73: "snow", 75: "heavy snow", 77: "snow grains",
+    80: "light showers", 81: "showers", 82: "heavy showers",
+    85: "snow showers", 86: "heavy snow showers",
+    95: "thunderstorm", 96: "thunderstorm with hail", 99: "thunderstorm with heavy hail",
+}
+
+
 class WeatherTool(FastTool):
     """
-    Fetches current weather for the user's location (Balashikha, Moscow region)
+    Fetches current weather for Balashikha, Moscow region directly from open-meteo.com.
-    by querying SearXNG, which has external internet access.
+    No API key required. Returns a ready-to-deliver reply — no LLM reformatting needed.
-
-    Triggered by any weather-related query. The Router also forces medium tier
-    when this tool matches so the richer model handles the injected data.
     """
 
     _PATTERN = re.compile(
@@ -51,11 +60,13 @@ class WeatherTool(FastTool):
         re.IGNORECASE,
     )
 
-    # Fixed query — always fetch home location weather
+    _URL = (
-    _SEARCH_QUERY = "погода Балашиха сейчас"  # Russian query → Celsius sources
+        "https://api.open-meteo.com/v1/forecast"
-
+        "?latitude=55.7963&longitude=37.9382"
-    def __init__(self, searxng_url: str):
+        "&current=temperature_2m,apparent_temperature,relative_humidity_2m"
-        self._searxng_url = searxng_url
+        ",wind_speed_10m,weather_code"
+        "&wind_speed_unit=ms"
+    )
 
     @property
     def name(self) -> str:
@@ -65,31 +76,24 @@ class WeatherTool(FastTool):
         return bool(self._PATTERN.search(message))
 
     async def run(self, message: str) -> str:
-        """Query SearXNG for Balashikha weather and return current conditions snippet."""
         try:
-            async with httpx.AsyncClient(timeout=15) as client:
+            async with httpx.AsyncClient(timeout=10) as client:
-                r = await client.get(
+                r = await client.get(self._URL)
-                    f"{self._searxng_url}/search",
-                    params={"q": self._SEARCH_QUERY, "format": "json"},
-                )
                 r.raise_for_status()
-                items = r.json().get("results", [])[:5]
+                c = r.json()["current"]
         except Exception as e:
             return f"[weather error: {e}]"
 
-        if not items:
+        temp = c["temperature_2m"]
-            return ""
+        feels = c["apparent_temperature"]
+        humidity = c["relative_humidity_2m"]
+        wind = c["wind_speed_10m"]
+        condition = _WMO_CODES.get(c.get("weather_code", 0), "unknown")
 
-        # Prefer results whose snippets contain actual current conditions
+        return (
-        lines = ["Current weather data for Balashikha, Moscow region (temperatures in °C):\n"]
+            f"Balashikha: {condition}, {temp:+.0f}°C (feels like {feels:+.0f}°C), "
-        for item in items:
+            f"wind {wind:.1f} m/s, humidity {humidity}%."
-            snippet = item.get("content", "")
+        )
-            title = item.get("title", "")
-            url = item.get("url", "")
-            if snippet:
-                lines.append(f"[{title}]\n{snippet}\nSource: {url}\n")
-
-        return "\n".join(lines) if len(lines) > 1 else ""
 
 
 class CommuteTool(FastTool):

openmemory/CLAUDE.md

@@ -0,0 +1,26 @@
+# openmemory
+
+FastMCP server wrapping mem0 for persistent per-session memory, backed by Qdrant + nomic-embed-text.
+
+## Tools exposed (MCP)
+
+- `add_memory(text, user_id)` — extract facts from a conversation turn and store in Qdrant
+- `search_memory(query, user_id)` — semantic search, returns results with score ≥ 0.5
+- `get_all_memories(user_id)` — dump all stored memories for a session
+
+These are called directly by `agent.py` (outside the agent loop), never exposed to the LLM as tools.
+
+## Two Ollama instances
+
+- **GPU** (`OLLAMA_GPU_URL`, port 11436) — extraction model (`qwen2.5:1.5b`): pulls facts from conversation text
+- **CPU** (`OLLAMA_CPU_URL`, port 11435) — embedding model (`nomic-embed-text`): 50–150 ms per query
+
+## Prompts
+
+Custom `EXTRACTION_PROMPT` starts with `/no_think` to suppress qwen3 chain-of-thought and get clean JSON output. Custom `UPDATE_MEMORY_PROMPT` handles deduplication — mem0 merges new facts with existing ones rather than creating duplicates.
+
+## Notes
+
+- Qdrant collection is created automatically on first use
+- Memory is keyed by `user_id` which equals `session_id` in `agent.py`
+- Extraction runs after the reply is sent (background task) — GPU contention with medium model is avoided since the semaphore is released before `_store_memory()` is scheduled

routecheck/CLAUDE.md

@@ -0,0 +1,25 @@
+# routecheck
+
+FastAPI service providing a Yandex Routing API proxy behind an image captcha.
+
+## Purpose
+
+Yandex Routing API free tier requires a website that uses the API. This service is that website.
+It also exposes an internal endpoint (`/api/route`) used by `CommuteTool` in `fast_tools.py`.
+
+## Two access paths
+
+- **Web UI** (`/`): solve PIL arithmetic captcha → get a token → query any two lat/lon points
+- **Internal API**: `GET /api/route?from=lat,lon&to=lat,lon&token=$ROUTECHECK_TOKEN` — no captcha
+
+## Key env vars
+
+- `YANDEX_ROUTING_KEY` — from developer.tech.yandex.ru, Router API, free tier
+- `INTERNAL_TOKEN` — equals `ROUTECHECK_TOKEN` from root `.env`; shared with deepagents
+- `HTTPS_PROXY` — set to `http://host.docker.internal:56928`; container has no direct external internet
+
+## Notes
+
+- Captchas expire after 5 min, route tokens after 1 hour, both stored in-memory (restart clears them)
+- Yandex API expects `lon,lat` order (not `lat,lon`) — `app.py` swaps automatically
+- Captcha image endpoint: `GET /captcha/image/{id}` — regenerates on each call with random noise

router.py

@@ -1,11 +1,38 @@
+import asyncio
 import re
+import math
 from typing import Optional
+from openai import AsyncOpenAI
 from langchain_core.messages import SystemMessage, HumanMessage
 from fast_tools import FastToolRunner
 
-# ── Regex pre-classifier ──────────────────────────────────────────────────────
+# ── Regex pre-classifiers ─────────────────────────────────────────────────────
-# Catches obvious light-tier patterns before calling the LLM.
+
-# Keyed by regex → compiled pattern.
+# Complex: keyword triggers that reliably signal deep multi-source research
+_COMPLEX_PATTERNS = re.compile(
+    r"(?:^|\s)("
+    r"research|investigate|deep.dive|think carefully"
+    r"|write a (?:detailed|comprehensive|full|thorough|complete)"
+    r"|compare all|find and (?:compare|summarize|analyze)"
+    r"|in[- ]depth analysis|comprehensive guide"
+    r"|detailed (?:report|analysis|comparison|breakdown|overview)"
+    r"|everything about|all (?:major|available|self-hosted|open.source)"
+    r"|pros and cons|with (?:sources|citations|references)"
+    # Russian complex research keywords (no trailing \b — stems like подробн match подробное/подробный)
+    r"|исследуй|изучи все|сравни все|найди и сравни|найди и опиши"
+    r"|напиши подробн|напиши детальн|напиши полн"
+    r"|подробный отчет|детальн\w+ (?:анализ|сравнение|отчет)"
+    r"|подробное (?:руководство|сравнение)|полное руководство"
+    r"|все варианты|все способы|все доступные|все самохостируемые|все платформы"
+    r"|лучшие практики|все инструменты|все решения|все протоколы"
+    r"|найди детальн|найди и кратко опиши"
+    r"|изучи свежие|изучи лучши|изучи все"
+    r"|сравни все\b"
+    r")",
+    re.IGNORECASE,
+)
+
+# Light: trivial queries that need no tools or memory
 _LIGHT_PATTERNS = re.compile(
     r"^("
     # Greetings / farewells
@@ -15,35 +42,301 @@ _LIGHT_PATTERNS = re.compile(
     r"|thanks?|thank you|thx|ty|ok|okay|k|cool|great|awesome|perfect|sounds good|got it|nice|sure"
     r"|how are you|how are you\?|how are you doing(\s+today)?[?!.]*"
     r"|what.?s up"
-    # Calendar facts: "what day comes after X?" / "what comes after X?"
+    # Calendar facts
     r"|what\s+day\s+(comes\s+after|follows|is\s+after)\s+\w+[?!.]*"
     r"|what\s+comes\s+after\s+\w+[?!.]*"
-    # Acronym expansions: "what does X stand for?"
+    # Acronym expansions
     r"|what\s+does\s+\w+\s+stand\s+for[?!.]*"
+    # Russian greetings / farewells / acknowledgements
+    r"|привет|пока|спасибо|здравствуй|здравствуйте|добрый день|добрый вечер|доброе утро"
+    r"|окей|хорошо|отлично|понятно|ок|ладно|договорились|спс|благодарю"
+    r"|пожалуйста|не за что|всё понятно|ясно"
+    r"|как дела|как ты|как жизнь|всё хорошо|всё ок"
     r")[\s!.?]*$",
     re.IGNORECASE,
 )
 
-# ── LLM classification prompt ─────────────────────────────────────────────────
+# ── Semantic router utterances ────────────────────────────────────────────────
-CLASSIFY_PROMPT = """Classify the message. Output ONLY one word: light, medium, or complex.
+# These are embedded at startup. New messages are classified by cosine
+# similarity — whichever tier's centroid is closest wins.
+_LIGHT_UTTERANCES = [
+    # General facts (English)
+    "what is 2+2",
+    "what is the capital of France",
+    "name the three primary colors",
+    "tell me a short joke",
+    "is the sky blue",
+    "is water wet",
+    "how many days in a week",
+    "what is the speed of light",
+    "what is the boiling point of water",
+    "spell the word beautiful",
+    "what color is the ocean",
+    "how many inches in a foot",
+    "who wrote hamlet",
+    "what is pi",
+    "what year did world war two end",
+    "what is the largest planet",
+    "how many continents are there",
+    "what does DNA stand for",
+    "what language do they speak in Brazil",
+    "what is the square root of 144",
+    # Tech definitions — static knowledge (English)
+    "what is Docker",
+    "what is a VPN",
+    "what is SSH",
+    "what is a reverse proxy",
+    "what is an API",
+    "what is a firewall",
+    "what is a container",
+    "what is DNS",
+    "what is HTTPS",
+    "what is a load balancer",
+    "what is Kubernetes",
+    "what is Git",
+    "what is a network port",
+    "what is an IP address",
+    "what is a subnet mask",
+    "what is the OSI model",
+    "how many bits in a byte",
+    "how many bytes in a gigabyte",
+    "what is TCP",
+    "what is a REST API",
+    # Russian — static facts and definitions
+    "что такое IP-адрес",
+    "что такое VPN",
+    "что такое Docker",
+    "что такое DNS",
+    "что такое SSH",
+    "что означает API",
+    "сколько байт в гигабайте",
+    "сколько бит в байте",
+    "что такое Zigbee",
+    "что такое Z-Wave",
+    "что такое брандмауэр",
+    "что такое виртуальная машина",
+    "что такое обратный прокси",
+    "привет",
+    "пока",
+    "спасибо",
+    "как дела",
+    "что такое Matter протокол",
+    "сколько планет в солнечной системе",
+    "чему равно число Пи",
+    # Russian — more static definitions
+    "что такое TCP/IP",
+    "что такое подсеть",
+    "скорость света",
+    "сколько дней в году",
+    "что такое Kubernetes",
+    "что такое Git",
+    "что такое REST API",
+    "что такое TCP",
+    "что такое UDP",
+    "что такое VLAN",
+    "сколько мегабайт в гигабайте",
+    "что такое процессор",
+    "что такое оперативная память",
+    "что такое виртуализация",
+    "что такое Linux",
+    "что такое умный дом",
+    "что такое Home Assistant",
+    "что такое Matter",
+]
 
-LIGHT = answerable from general knowledge, no internet needed:
+_MEDIUM_UTTERANCES = [
-  what is 2+2 / what is the capital of France / name the three primary colors
+    # English — current data, memory, actions
-  tell me a short joke / is the sky blue / is water wet
+    "what is the weather today",
+    "what is the bitcoin price right now",
+    "what are the latest news",
+    "what did we talk about last time",
+    "what is my name",
+    "where do I live",
+    "what do you know about me",
+    "what did I tell you before",
+    "what is the current temperature outside",
+    "remind me what I said about my project",
+    "search for the latest iPhone release",
+    "find me a restaurant nearby",
+    "turn on the lights in the living room",
+    "turn off all lights",
+    "set temperature to 22 degrees",
+    "what is the current traffic to Moscow",
+    "check if anyone is home",
+    "what devices are currently on",
+    "look up my public IP address",
+    "show me recent news about Proxmox",
+    # Russian — weather and commute
+    "какая сегодня погода в Балашихе",
+    "пойдет ли сегодня дождь",
+    "какая температура на улице сейчас",
+    "погода на завтра",
+    "будет ли снег сегодня",
+    "сколько ехать до Москвы сейчас",
+    "какие пробки на дороге до Москвы",
+    "время в пути на работу",
+    "есть ли пробки сейчас",
+    "стоит ли брать зонтик",
+    # Russian — smart home control
+    "включи свет в гостиной",
+    "выключи свет на кухне",
+    "какая температура дома",
+    "установи температуру 22 градуса",
+    "выключи все лампочки",
+    "какие устройства сейчас включены",
+    "включи ночной режим",
+    "открой шторы в гостиной",
+    "включи свет в спальне на 50 процентов",
+    "выключи свет во всём доме",
+    "включи вентилятор в детской",
+    "закрыты ли все окна",
+    "выключи телевизор",
+    "какое потребление электричества сегодня",
+    "включи кофемашину",
+    "сколько у нас датчиков движения",
+    "состояние всех дверных замков",
+    "есть ли кто-нибудь дома",
+    "установи будильник на 7 утра",
+    # Russian — personal memory
+    "как меня зовут",
+    "где я живу",
+    "что мы обсуждали в прошлый раз",
+    "что ты знаешь о моем домашнем сервере",
+    "напомни, какие сервисы я запускаю",
+    "что я просил тебя запомнить",
+    "что я говорил о своей сети",
+    # Russian — current info lookups requiring network/tools
+    "какой сейчас курс биткоина",
+    "курс доллара к рублю сейчас",
+    "какая последняя версия Docker",
+    "как перезапустить Docker контейнер",
+    "как посмотреть логи Docker контейнера",
+    "какие новые функции в Home Assistant 2024",
+    "есть ли проблемы у Cloudflare сегодня",
+    "какие новые Zigbee устройства вышли в 2024 году",
+    "найди хороший опенсорс менеджер фотографий",
+    "последние новости Proxmox",
+    "напиши bash команду для поиска больших файлов",
+    "как вывести список всех запущенных контейнеров",
+    "как проверить использование диска в Linux",
+]
 
-MEDIUM = requires web search or the user's stored memories:
+_COMPLEX_UTTERANCES = [
-  current weather / today's news / Bitcoin price / what did we talk about
+    # English
-  what is my name / where do I live / what is my job / do I have any pets
+    "research everything about Elon Musk's recent projects and investments",
-  what do you know about me / what are my preferences / what did I tell you
+    "write a detailed report on climate change solutions with sources",
+    "investigate the history and current state of quantum computing",
+    "find and summarize the latest academic papers on transformer architectures",
+    "analyze in depth the pros and cons of nuclear energy with citations",
+    "research the background and controversies around this person",
+    "compare all major cloud providers with detailed pricing and features",
+    "write a comprehensive biography of this historical figure",
+    "investigate what caused the 2008 financial crisis with multiple sources",
+    "research the best programming languages in 2024 with detailed comparison",
+    "find everything published about this medical condition and treatments",
+    "do a deep dive into the latest developments in artificial general intelligence",
+    "research and compare all options for starting a business in Europe",
+    "investigate recent news and controversies around this company",
+    "write a thorough analysis of geopolitical tensions in the Middle East",
+    "find detailed information on the side effects and studies for this medication",
+    "research the top 10 JavaScript frameworks with benchmarks and community data",
+    "investigate who is funding AI research and what their goals are",
+    "write a detailed market analysis for the electric vehicle industry",
+    "research everything you can find about this startup or technology",
+    # Russian — deep research
+    "исследуй и сравни все варианты умного домашнего освещения",
+    "напиши подробный отчет о протоколах умного дома",
+    "изучи все самохостируемые медиасерверы и сравни их",
+    "исследуй лучшие практики безопасности домашнего сервера",
+    "сравни все системы резервного копирования для Linux",
+    "напиши детальное сравнение WireGuard и OpenVPN",
+    "исследуй все варианты голосового управления на русском языке",
+    "изучи все опенсорс альтернативы Google сервисам",
+    "напиши подробный анализ локальных языковых моделей",
+    "исследуй лучшие инструменты мониторинга для домашнего сервера",
+    # Russian — more deep research queries matching benchmark
+    "исследуй и сравни Proxmox, Unraid и TrueNAS для домашней лаборатории",
+    "напиши подробное руководство по безопасности домашнего сервера",
+    "исследуй все доступные дашборды для самохостинга и сравни их",
+    "найди детальные бенчмарки ARM одноплатных компьютеров для домашней лаборатории",
+    "исследуй лучший стек мониторинга для самохостинга в 2024 году",
+    "исследуй и сравни WireGuard, OpenVPN и Tailscale для домашней сети",
+    "исследуй лучшие практики сегментации домашней сети с VLAN",
+    "изучи все самохостируемые DNS решения и их возможности",
+    "исследуй и сравни все платформы умного дома: Home Assistant и другие",
+    "изучи лучшие Zigbee координаторы и их совместимость с Home Assistant",
+    "напиши детальный отчет о поддержке протокола Matter и совместимости устройств",
+    "исследуй все способы интеграции умных ламп с Home Assistant",
+    "найди и сравни все варианты датчиков движения для умного дома",
+    "исследуй и сравни все самохостируемые решения для хранения фотографий",
+    "изучи лучшие самохостируемые медиасерверы: Jellyfin, Plex и Emby",
+    "исследуй последние достижения в локальном LLM инференсе и обзор моделей",
+    "изучи лучшие опенсорс альтернативы Google сервисов для приватности",
+    "найди и кратко опиши все крупные самохостируемые менеджеры паролей",
+    "напиши детальный анализ текущего состояния AI ассистентов для самохостинга",
+    "исследуй и сравни все инструменты оркестрации контейнеров для домашней лаборатории",
+    "изучи лучшие подходы к автоматическому резервному копированию в Linux",
+    "исследуй и сравни все самохостируемые инструменты личных финансов",
+    "изучи свежие CVE и уязвимости в популярном самохостируемом ПО",
+    "напиши подробное руководство по настройке автоматизаций в Home Assistant",
+    "исследуй все варианты голосового управления умным домом на русском языке",
+    "сравни все системы резервного копирования для Linux: Restic, BorgBackup и другие",
+    "исследуй лучшие самохостируемые системы мониторинга сети: Zabbix, Grafana",
+    "изучи все варианты локального запуска языковых моделей на видеокарте",
+    "напиши подробный отчет о технологиях синтеза речи с открытым исходным кодом",
+    "исследуй все способы интеграции умных розеток с мониторингом потребления",
+    "напиши полное руководство по настройке обратного прокси Caddy",
+    "исследуй лучшие практики написания Docker Compose файлов для продакшена",
+    "сравни все самохостируемые облачные хранилища: Nextcloud, Seafile и другие",
+    "изучи все доступные локальные ассистенты с голосовым управлением",
+    "исследуй все самохостируемые решения для блокировки рекламы: Pi-hole, AdGuard",
+    "напиши детальное сравнение систем управления конфигурацией: Ansible, Puppet",
+    "исследуй все протоколы умного дома и их плюсы и минусы: Zigbee, Z-Wave, Matter",
+    "найди и сравни все фреймворки для создания локальных AI ассистентов",
+    "исследуй лучшие решения для автоматического управления медиатекой",
+    "изучи все варианты самохостируемых систем учёта расходов с возможностью импорта",
+    "напиши сравнение всех вариантов самохостинга для хранения и синхронизации файлов",
+    "исследуй все открытые протоколы для умного дома и их экосистемы",
+    "изучи лучшие инструменты для автоматизации домашней инфраструктуры",
+]
 
-COMPLEX = /think prefix only:
+# Medium: queries that require tools, actions, or real-time data (not static knowledge)
-  /think compare frameworks / /think plan a trip
+_MEDIUM_PATTERNS = re.compile(
-
+    r"(?:"
-Message: {message}
+    # Russian smart home commands — always need HA integration
-Output (one word only — light, medium, or complex):"""
+    r"(?:включи|выключи|открой|закрой|установи|поставь|убавь|прибавь|переключи)\s"
+    r"|(?:какая|какой|какое|каково)\s+(?:температура|влажность|потребление|состояние|статус)\s"
+    r"|(?:сколько|есть ли)\s.*(?:датчик|устройств|замк)"
+    # Russian memory queries
+    r"|как меня зовут|где я живу|что мы обсуждали|что я говорил|что я просил"
+    r"|напомни\b|что ты знаешь обо мне"
+    # Russian current info
+    r"|курс (?:доллара|биткоина|евро|рубл)"
+    r"|(?:последние |свежие )?новости\b"
+    r"|(?:погода|температура)\s+(?:на завтра|на неделю)"
+    r")",
+    re.IGNORECASE,
+)
 
 LIGHT_REPLY_PROMPT = """You are a helpful Telegram assistant. Answer briefly and naturally (1-3 sentences). Be friendly."""
 
+_EMBED_MODEL = "ollama/nomic-embed-text"
+
+
+def _cosine(a: list[float], b: list[float]) -> float:
+    dot = sum(x * y for x, y in zip(a, b))
+    norm_a = math.sqrt(sum(x * x for x in a))
+    norm_b = math.sqrt(sum(x * x for x in b))
+    if norm_a == 0 or norm_b == 0:
+        return 0.0
+    return dot / (norm_a * norm_b)
+
+
+def _centroid(embeddings: list[list[float]]) -> list[float]:
+    n = len(embeddings)
+    dim = len(embeddings[0])
+    return [sum(embeddings[i][d] for i in range(n)) / n for d in range(dim)]
+
 
 def _format_history(history: list[dict]) -> str:
     if not history:
@@ -56,70 +349,93 @@ def _format_history(history: list[dict]) -> str:
     return "\n".join(lines)
 
 
-def _parse_tier(text: str) -> str:
-    """Extract tier from raw model output. Default to medium."""
-    t = text.strip().lower()
-    snippet = t[:60]
-    if "complex" in snippet:
-        return "complex"
-    if "medium" in snippet:
-        return "medium"
-    if "light" in snippet:
-        return "light"
-    # Model invented a descriptive category (e.g. "simplefact", "trivial", "basic") →
-    # treat as light since it recognised the question doesn't need tools
-    if any(w in snippet for w in ("simple", "fact", "trivial", "basic", "easy", "general")):
-        return "light"
-    return "medium"  # safe default
-
-
 class Router:
-    def __init__(self, model, fast_tool_runner: FastToolRunner | None = None):
+    def __init__(
-        self.model = model
+        self,
+        model,
+        embedder: AsyncOpenAI,
+        fast_tool_runner: FastToolRunner | None = None,
+    ):
+        self.model = model  # qwen2.5:1.5b — used only for generating light replies
+        self._embedder = embedder
         self._fast_tool_runner = fast_tool_runner
+        self._light_centroid: list[float] | None = None
+        self._medium_centroid: list[float] | None = None
+        self._complex_centroid: list[float] | None = None
+
+    async def initialize(self) -> None:
+        """Pre-compute utterance embeddings. Call once at startup. Retries until LiteLLM is ready."""
+        print("[router] embedding utterances for semantic classifier...", flush=True)
+        texts = _LIGHT_UTTERANCES + _MEDIUM_UTTERANCES + _COMPLEX_UTTERANCES
+        for attempt in range(10):
+            try:
+                resp = await self._embedder.embeddings.create(model=_EMBED_MODEL, input=texts)
+                embeddings = [item.embedding for item in resp.data]
+                n_light = len(_LIGHT_UTTERANCES)
+                n_medium = len(_MEDIUM_UTTERANCES)
+                self._light_centroid = _centroid(embeddings[:n_light])
+                self._medium_centroid = _centroid(embeddings[n_light:n_light + n_medium])
+                self._complex_centroid = _centroid(embeddings[n_light + n_medium:])
+                print("[router] semantic classifier ready (3-tier)", flush=True)
+                return
+            except Exception as e:
+                print(f"[router] embedding attempt {attempt+1}/10 failed: {e}", flush=True)
+                await asyncio.sleep(3)
+        print("[router] WARNING: could not initialize semantic classifier — will default to medium", flush=True)
+
+    async def _classify_by_embedding(self, message: str) -> str:
+        """Embed message and return 'light', 'medium', or 'complex' based on centroid similarity."""
+        if self._light_centroid is None or self._medium_centroid is None or self._complex_centroid is None:
+            return "medium"
+        try:
+            resp = await self._embedder.embeddings.create(model=_EMBED_MODEL, input=[message])
+            emb = resp.data[0].embedding
+            score_light = _cosine(emb, self._light_centroid)
+            score_medium = _cosine(emb, self._medium_centroid)
+            score_complex = _cosine(emb, self._complex_centroid)
+            tier = max(
+                [("light", score_light), ("medium", score_medium), ("complex", score_complex)],
+                key=lambda x: x[1],
+            )[0]
+            print(
+                f"[router] semantic: light={score_light:.3f} medium={score_medium:.3f} "
+                f"complex={score_complex:.3f} → {tier}",
+                flush=True,
+            )
+            return tier
+        except Exception as e:
+            print(f"[router] embedding classify error, defaulting to medium: {e}", flush=True)
+            return "medium"
 
     async def route(
         self,
         message: str,
         history: list[dict],
-        force_complex: bool = False,
     ) -> tuple[str, Optional[str]]:
         """
         Returns (tier, reply_or_None).
-        For light tier: also generates the reply with a second call.
+        For light tier: also generates the reply inline.
         For medium/complex: reply is None.
         """
-        if force_complex:
-            return "complex", None
-
-        # Step 0a: force medium if any fast tool matches (live-data queries)
         if self._fast_tool_runner and self._fast_tool_runner.any_matches(message.strip()):
             names = self._fast_tool_runner.matching_names(message.strip())
             print(f"[router] fast_tool_match={names} → medium", flush=True)
             return "medium", None
 
-        # Step 0b: regex pre-classification for obvious light patterns
         if _LIGHT_PATTERNS.match(message.strip()):
-            print(f"[router] regex→light", flush=True)
+            print("[router] regex→light", flush=True)
             return await self._generate_light_reply(message, history)
 
-        # Step 1: LLM classification with raw text output
+        if _COMPLEX_PATTERNS.search(message.strip()):
-        try:
+            print("[router] regex→complex", flush=True)
-            classify_response = await self.model.ainvoke([
+            return "complex", None
-                HumanMessage(content=CLASSIFY_PROMPT.format(message=message)),
-            ])
-            raw = classify_response.content or ""
-            raw = re.sub(r"<think>.*?</think>", "", raw, flags=re.DOTALL).strip()
-            tier = _parse_tier(raw)
 
-            if tier == "complex" and not message.startswith("/think"):
+        if _MEDIUM_PATTERNS.search(message.strip()):
-                tier = "medium"
+            print("[router] regex→medium", flush=True)
-
-            print(f"[router] raw={raw[:30]!r} → tier={tier}", flush=True)
-        except Exception as e:
-            print(f"[router] classify error, defaulting to medium: {e}", flush=True)
             return "medium", None
 
+        tier = await self._classify_by_embedding(message)
+
         if tier != "light":
             return tier, None
 
@@ -128,7 +444,7 @@ class Router:
     async def _generate_light_reply(
         self, message: str, history: list[dict]
     ) -> tuple[str, Optional[str]]:
-        """Generate a short reply using the router model for light-tier messages."""
+        """Generate a short reply using qwen2.5:1.5b for light-tier messages."""
         history_text = _format_history(history)
         context = f"\nConversation history:\n{history_text}" if history else ""
         try: