July 11, 2026
I Deployed the Homepage and Immediately 404’d Myself
Pushed real homepage copy live to replace the old "coming soon" placeholder, dragged the site folder into Cloudflare Pages' upload box, hit deploy, and got a clean success message. Went to check the live site: 404 Not Found.
The tell was that even the deployment's own unique preview URL 404’d, not just the production alias — which ruled out a DNS or domain-routing issue and pointed straight at the upload itself. Turned out dragging the whole project folder nested the file a level deep instead of landing index.html at the site root, so Cloudflare had nothing to serve at /.
Fix: redeploy dragging just the index.html file itself, not the folder around it. Confirmed live a few minutes later.
July 4, 2026
The 2048-Token Ceiling Nobody Warns You About
Web search tool-calling appeared to silently fail — the model just answered from general knowledge, with zero trace in the logs and no round-trip delay to suggest a search had even been attempted. No errors, nothing to grep for.
The actual cause had nothing to do with the search tool itself: Ollama's default 2048-token context window was silently overflowing the moment Open WebUI injected the full tool definitions into the prompt. Once that ceiling gets hit, the model just stops seeing the tools it's supposed to call — no warning, no error, just quiet failure.
Fix: set num_ctx explicitly to 8192 per-model, or set OLLAMA_CONTEXT_LENGTH=8192 globally. Also worth knowing: Web Search needs two separate checkboxes enabled in the model editor (Capabilities and Default Features) — checking only one leaves it silently disabled too.
July 4, 2026
Wayland Ate My VNC Server
Tried to set up headless remote access to the SKYe machine with TigerVNC — control it from another PC with no monitor attached. vncserver :1 kept failing with a socket-listener error claiming a server was already running, even though vncserver -list and ps aux both showed nothing.
Not a stale lock file. The machine auto-logs into a live GNOME Wayland session on boot, and TigerVNC is X11-based — it can't spin up its own virtual display because Wayland already owns the display stack. Fundamental incompatibility, not a config bug.
Fix: removed TigerVNC entirely and switched to GNOME Remote Desktop, which is Wayland-native and mirrors the actual live session over RDP instead of faking a separate X11 display. (Bonus gotcha on the Windows client side: a CredSSP encryption-oracle mismatch needed a registry tweak to connect at all — acceptable only because it's a trusted home LAN.)
July 2, 2026
Why 14B Beat a Smaller and an Equally-Sized Alternative
Picking the local model to run SKYe's brain on came down to three options on an 8GB-VRAM RTX 2070 Super: Llama 3.1 8B, Qwen 2.5 7B, and Qwen 2.5 14B.
Llama 3.1 8B was accurate and fast but unreliable at function-calling — it often just wouldn't trigger a knowledge-base search when it should have. Qwen 2.5 7B was faster than the 14B, but had a split-brain problem: it needed Legacy mode to reliably auto-retrieve from the knowledge base, but Native mode to reliably call custom tools — and couldn't do both in the same session.
Qwen 2.5 14B (4-bit) fits in just under 8GB of VRAM with no CPU spillover, and is the only one of the three that handles retrieval and custom tool-calling reliably at the same time under a single mode. Slower than both alternatives — but the only one that doesn't force a tradeoff mid-conversation, which made it the easy call.