Sidelook 0.17.0: local models, the Bench button, three words under the conversation

What changed

• Three words under the conversation: New chat, Clear context, Compact. New chat is the old Clear conversation from Settings, moved where you need it. Clear context leaves the messages on screen and sends none of them with the next message. Compact is one send that trades the earlier messages for a summary under 150 words, and a line saying what it saved.
• A meter line under them: how full the 24,000 characters the next send can carry is, what the last send cost in tokens, and what this chat has cost. Cached input is counted beside each number, never inside it. The percent turns warm past 80.
• The model's token counts reach the page: Codex reports its cached input, Claude reports both its cache read and its cache write. They also go in the What goes panel.
• Local models. Whatever LM Studio or Ollama holds on this computer appears in the model list under the runtime's name, while the runtime is running. Sidelook talks to it through Codex's open-source provider, in the same read-only sandbox as a subscription model; nothing leaves the machine and nothing is metered. Setup checks that the runtime answers and the model is still there, and offers Start LM Studio server when it is not. Proven with Qwen3 8B in LM Studio on 2026-09-06; Ollama is wired the same way but has not been run yet.
• An LM Studio model is loaded with a 32,768-token context before the first request, because Codex's own prompt is about 12,700 tokens and LM Studio's default load of 4,096 refuses it. The studio shows "Loading into memory" while that happens, about ten seconds for an 8B model.
• Local models get the schema inside their instructions and the JSON object read out of the reply, because LM Studio does not enforce Codex's output schema; low effort turns a local model's reasoning off (8 s a turn on Qwen3 8B instead of 30 to 50); the GPU share at load is computed from the card and the model so the 32k cache fits (every layer on an 8 GB card swapped at 5 tokens/s); a local chat has 300 s, the same as a build. A local model that answers in words instead of JSON is run once more; the conversation then takes words as the reply, a build fails closed. A suggestion the model invents reads as none.
• Bench ↗ sits in the panel header and opens the studio. It was a line inside Settings, two presses away; that line is gone.
• The studio opens to fit the monitor Sidelook lives on: 85% of the working area, capped at 1480x900 and floored at 760x520. Drag it to another size and it opens at that size next time, saved in studio-size.json beside the dock position. The panel and the dock still remember nothing.
• The studio reflows instead of clipping. Above 1180 it is unchanged. From 900 to 1180 the chat column steps aside and a Chat button in the toolbar slides it over the right edge of the stage; Escape, ← Panel, or a window wide enough for the column closes it. Under 900 the rail stacks above the stage with its own scroll, the toolbar wraps and Share window and Live build shorten to Share and Live. Nothing scrolls sideways down to 760.
• The browser check loads the studio at 1480, 1180, 1100, 900, 800 and 760, opens and closes the chat overlay at two of them, and fails on any horizontal overflow.

SHA-256 of Sidelook-0.17.0-Windows-x64.exe is in SHA256SUMS.txt. The exe is unsigned; the bundled Node and Codex are publisher-verified at build. See docs/MODELS.md for local models.