User prefers MANUAL /think on|off — NO auto-routing. Thinking ON for planning/audit/review, OFF for execute/code/terminal.
§
User expectation: ALWAYS web research before guessing/changing flags or configs.
§
Local network IP: 192.168.1.50. Open WebUI runs on :3081.
§
User reverse-engineers Android APKs for security analysis (jadx, SMALI patching, repack/sign). Has patched RevenueCat subscription checks, probes Firebase Firestore backends, audits debrid streaming app infrastructure.
§
User works with debrid/streaming apps (Real-Debrid, TorBox, AllDebrid) and scraper/catalog architectures as part of Android TV development.
§
Unspooled: ~/Unspooled/ (Kotlin/Compose TV). ultimate-scraper (port 3091), nexstream-scraper (port 3089, legacy), Flask backend (port 8800), free-scraper (port 3092), FlareSolverr.
§
Scraper Suite at ~/Scraper Suite/: free-scraper (Node, :3092, 6 embed extractors), FlareSolverr, foxhound, Scrapling, proxy-scraper.
§
Hermes config: stub_mode:true, toolsets:7 core, web.search:searxng, web.extract:firecrawl. compression:deepseek-v4-flash threshold 0.3 protect_last_n 10.
§
Fleet: qwen36-35b-mtp, gemma-26b-200k, ornith-35b-q6-mtp, qwythos-9b-mtp-q6, ornith-9b-mtp-q5. Flags in llama-server-flags skill. OWUI v0.10.2.
§
Tailscale funnel domain for Debian server: debian-ai.tailf83ed5.ts.net. Funnel via `sudo tailscale funnel --bg <port>`.
§
Ultimate Scraper v5 (:3091): TorrServer-backed, TMDB title/year/SE matching, seeders>quality>codec>size. Addon: debian-ai.tailf83ed5.ts.net/manifest.json
§
US v5 matching: TMDB-backed, reboot/sequel detect ±3yr, season pack accept. Res (4K=100,1080p=75), codec (HEVC=90,AV1=80,AVC=70), HDR/DV/Atmos parse.
§
TorrServer: 4GB cache, 200 conn, ResponsiveMode=true, auto-cleanup 15min idle. Config via BoltDB BEFORE restart. :8090.
§
TMDB key (v3): ab74812bfa9665ebd944ffbbb5b57d76. Used for metadata matching during scraping (title/year verification), NOT for catalog UI.
§
Real-Debrid: instantAvailability deprecated (error 37). Use addMagnet + selectFiles + poll info status instead.
§
RTX 2080 Ti MoE tuning: only --n-cpu-moe matters for prefill. GGML_CUDA_GRAPH_OPT, CUBLAS_16F, PDL, PREFETCH_EXPERTS, ubatch/poll sweeps — zero effect with --no-mmap. Bottleneck is PCIe Gen3 bandwidth. Qwen3.6-35B: n-cpu-moe=26 = 434 t/s (+20%, 1.1GB free). n-cpu-moe=24 = 463 t/s but 198MB headroom only.
§
User expects evidence, not conclusions. When a tweak doesn't help, verify mechanism activated (logs, env vars) and give researched explanation — don't stop at 'didn't work.' Pushes through negative results.