Restructure CLAUDE.md per official Claude Code recommendations

CLAUDE.md: 178→25 lines — commands + @ARCHITECTURE.md import only

Rules split into .claude/rules/ (load at startup, topic-scoped):
  llm-inference.md  — Bifrost-only, semaphore, model name format, timeouts
  agent-pipeline.md — tier rules, no tools in medium, memory outside loop
  fast-tools.md     — extension guide (path-scoped: fast_tools.py + agent.py)
  secrets.md        — .env keys, Vaultwarden, no hardcoding

Path-scoped rule: fast-tools.md only loads when editing fast_tools.py or agent.py

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

.claude/rules/agent-pipeline.md

@@ -0,0 +1,20 @@
+# Agent Pipeline Rules
+
+## Tiers
+- Complex tier requires `/think ` prefix. Any LLM classification of "complex" is downgraded to medium. Do not change this.
+- Medium is the default tier. Light is only for trivial static-knowledge queries matched by regex or LLM.
+- Light tier upgrade to medium is automatic when URL content is pre-fetched or a fast tool matches.
+
+## 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 +1,7 @@
+# LLM Inference Rules
+
+- All LLM calls must use `base_url=BIFROST_URL` with model name `ollama/<model>`. Never call Ollama directly for inference.
+- `_reply_semaphore` (asyncio.Semaphore(1)) serializes all GPU inference. Never bypass it or add a second semaphore.
+- Model names in code always use the `ollama/` prefix: `ollama/qwen3:4b`, `ollama/qwen3:8b`, `ollama/qwen2.5:1.5b`.
+- Timeout values: router=30s, medium=180s, complex=600s. Do not reduce them — GPU inference under load is slow.
+- `VRAMManager` is the only component that contacts Ollama directly (for flush/prewarm/poll). This is intentional — Bifrost cannot manage VRAM.

.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.

CLAUDE.md

@@ -11,31 +11,15 @@ docker compose up --build
 # Interactive CLI (requires services running)
 docker compose --profile tools run --rm -it cli
 
-# Integration tests (run from tests/integration/, requires all services)
+# 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
-# e.g.: read tests/use_cases/weather_now.md and execute it
+# example: read tests/use_cases/weather_now.md and execute it
 ```
 
-## Key Conventions
-
-- **Models via Bifrost only** — all LLM calls use `base_url=BIFROST_URL` with `ollama/<model>` prefix. Never call Ollama directly for inference.
-- **One inference at a time** — `_reply_semaphore` serializes GPU use. Do not bypass it.
-- **No tools in medium agent** — `_DirectModel` is a plain `ainvoke()` call. Context is injected via system prompt. `qwen3:4b` is unreliable with tool schemas.
-- **Fast tools are pre-flight** — `FastToolRunner` runs before routing and before any LLM call. Results are injected as context, not returned to the user directly.
-- **Memory outside agent loop** — `add_memory`/`search_memory` are called directly, never passed to agent tool lists.
-- **Complex tier is opt-in** — `/think ` prefix only. LLM classification of "complex" is always downgraded to medium.
-- **`.env` is required** — `TELEGRAM_BOT_TOKEN`, `ROUTECHECK_TOKEN`, `YANDEX_ROUTING_KEY`. Never commit it.
-
-## Adding a Fast Tool
-
-1. Subclass `FastTool` in `fast_tools.py` — implement `name`, `matches(message) → bool`, `run(message) → str`
-2. Add instance to `_fast_tool_runner` list in `agent.py`
-3. The router will automatically force medium tier when `matches()` returns true
-
 ## Architecture
 
 @ARCHITECTURE.md