Add session hygiene rules and lumen_model_speed to AGENTS.md

AGENTS.md

@@ -9,7 +9,7 @@
 
 ## What this repo is
 - Lumen: a Telegram bot (persona styled after Claude) running as a single FastAPI + aiogram webhook service on a Hugging Face Space (Docker). LLM backbone: Google Gemini + free OpenRouter models with per-message routing.
-- Entry point is bot.py (orchestrator, ~5000 lines). Focused modules sit next to it: lumen_router_config.py (model routing), lumen_formatting.py (markdown to Telegram HTML), lumen_security.py (injection/leak defenses), lumen_images.py, lumen_tts.py, lumen_tiktok.py, lumen_telegram_transport.py, lumen_state_storage.py, lumen_typing_pace.py, plus system_prompt.py.
+- Entry point is bot.py (orchestrator, ~5000 lines). Focused modules sit next to it: lumen_router_config.py (model routing), lumen_formatting.py (markdown to Telegram HTML), lumen_security.py (injection/leak defenses), lumen_images.py, lumen_tts.py, lumen_tiktok.py, lumen_telegram_transport.py, lumen_state_storage.py, lumen_typing_pace.py, lumen_model_speed.py (measured model latency), plus system_prompt.py.
 
 ## Commands (verified against CI, .github/workflows/ci.yml)
 - Full gate: pip install -r requirements.txt -r requirements-dev.txt; pyflakes bot.py lumen_*.py system_prompt.py conftest.py test_*.py; pytest -q; pip-audit -r requirements.txt
@@ -19,3 +19,8 @@
 ## Gotchas
 - Every push to main that passes CI auto-deploys to the live HF Space (.github/workflows/sync-to-hf.yml waits for the CI workflow, then force-pushes the checked commit). A green CI run is a deploy.
 - Versions in requirements.txt are pinned with == (HF rebuilds the image from scratch on every deploy; see the file header comment). Bumping a version is a deliberate decision followed by a manual smoke test (/diag + main commands), not a side effect.
+
+## Session hygiene (guide the owner proactively, they don't track this)
+- Effort: at the start of each task, state which effort fits and why, then proceed on their OK. Low: read/explain/rename; Medium (default): single-module fixes, tests, docs; High: streaming/security/refactor; XHigh: audits, architecture, prod incidents.
+- New chat: suggest starting one when the topic changes completely or the session mixes several unrelated tasks.
+- Compactness: suggest `/compact` (or a fresh chat) when the session gets long — many files read, repeated re-reads, or signs of lost context.