LocalPilot v1.2.9

Fixed

• Auto now stays Auto. An Auto chat was silently pinned to whichever model answered its first turn, so later turns no longer followed the default model, image routing or busy-model spreading. Every turn of an Auto chat is now routed again; the tab shows which model answered. A command's one-off model: no longer changes the chat's model either.
• Parallel Auto chats use several models again. Because of the pinning above, a second turn in a busy Auto chat queued behind the first model instead of moving to a comparable idle one.
• Faster first token. Warm-ups now load the model with the same context size (num_ctx) the chat sends. Before, Ollama loaded the weights twice. Compaction, inline edits, completions and the VS Code language-model provider now send the configured keepAlive and the same window too, so they no longer cut the model's stay in memory short or force a reload.

Changed

• Models are picked up automatically. While the window is focused, LocalPilot re-checks your servers in the background. Every 45 s it looks for models pulled or removed in a terminal. Every 10 s while none are found, a server started after VS Code shows up within seconds. Models uninstalled outside LocalPilot are cleared from settings and chats, which go back to Auto.
• Auto prefers models already in memory. Ollama's loaded models (/api/ps) break ties between comparable models for Auto picks, and busy-model spreading prefers an idle model that is already loaded (no load time, no extra memory). A clearly better model still wins.
• Opening the chat view starts loading its model. A model evicted from memory is warmed again right away instead of after the throttle. Changing unrelated settings no longer rebuilds providers or re-reads every model's metadata.

Install

code --install-extension localpilot-1.2.9.vsix

Or Extensions → ··· → Install from VSIX… in VS Code. Documentation: https://isathish.github.io/localpilot/

Full changelog: v1.2.8...v1.2.9