# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Commands

**Start all services:**
```bash
docker compose up --build
```

**Interactive CLI (requires gateway running):**
```bash
python3 cli.py [--url http://localhost:8000] [--session cli-alvis] [--timeout 400]
```

**Run integration tests:**
```bash
python3 test_pipeline.py [--chat-id CHAT_ID]

# Selective sections:
python3 test_pipeline.py --bench-only      # routing + memory benchmarks only (sections 10–13)
python3 test_pipeline.py --easy-only       # light-tier routing benchmark
python3 test_pipeline.py --medium-only     # medium-tier routing benchmark
python3 test_pipeline.py --hard-only       # complex-tier + VRAM flush benchmark
python3 test_pipeline.py --memory-only     # memory store/recall/dedup benchmark
python3 test_pipeline.py --no-bench        # service health + single name store/recall only
```

## 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.

### 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
                      )
                    → router.route() → tier decision (light/medium/complex)
                        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
                    → channels.deliver(session_id, channel, reply)
                        → pending_replies[session_id] queue (SSE)
                        → channel-specific callback (Telegram POST, CLI no-op)
                    → _store_memory() background task (openmemory)
CLI/wiki polling → GET /reply/{session_id}  (SSE, blocks until reply)
```

### 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

qwen3 models produce chain-of-thought `<think>...</think>` tokens via Ollama's OpenAI-compatible endpoint. Adolf controls this via system prompt prefixes:

- **Medium** (`qwen2.5:1.5b`): no thinking mode in this model; fast ~3s calls
- **Complex** (`qwen3:8b`): no prefix — thinking enabled by default, used for deep research
- **Router** (`qwen2.5:1.5b`): no thinking support in this model

`_strip_think()` in `agent.py` and `router.py` strips any `<think>` blocks from model output before returning to users.

### 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` posts to `/message`, then blocks on `GET /reply/{session_id}` SSE

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 |

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.

### 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, memory pipeline, all endpoints
- `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