LocalPilot v1.1.9

Added

• Model gallery — search and install anything from the Ollama library. A new Settings → Model gallery page (LocalPilot: Browse Ollama Model Gallery…, also in the LocalPilot menu) browses the whole public library on ollama.com, not just the curated catalog: search as you type, filter by Tools / Vision / Thinking / Embedding, sort by popularity / newest / name, and page through results. Choose size lists every tag of a model with its download size, context length, inputs and age, marked fits / tight / too big for this machine, latest and installed; plain sizes come first and quantisation / MLX / bf16 variants sit behind show all. Install pulls through your local Ollama server with live progress and cancel, and the model shows up under Models, in the picker and in routing when done. Pull by name accepts anything ollama pull does, including user/model:tag and hf.co/org/repo:Q4_K_M GGUF repositories; references are validated on both sides. Cloud-only models and tags are shown but marked not downloadable; the page offers Start Ollama / Install Ollama when no server is running. The Models catalog links to it.

Fixed

• No more duplicated replies. When a small model announced an action it never took ("Delegated to 2 background workers. I'll report back when they finish.") and the agent loop pushed it to act, the second answer streamed straight onto the first one ("…finish.Understood as:…"). The loop now emits a nudge event: the chat dims the retracted sentence under a retracted label, explains in one line why the model is being asked again, and the real answer starts in a fresh block — live and when the chat is reopened. A reply that re-announces a hand-off after the workers already finished is nudged to report their results instead of being treated as a new claim, and a fan-out that ran counts as work done so the real report is never flagged as a claim. "Understood as: …" is shown as a quiet kicker rather than body text.
• One kind of "running" instead of seven spinners. Tabs, the background-agents strip and each worker lane show a small breathing dot; the only spinners left are the N Working card and the status line. In-progress checklist items use a filled marker instead of spinning. While workers run, the status line summarises (2 workers running · 1 done) and each lane carries its own current step — the same text is no longer shown twice.
• Worker lanes fit a sidebar. The agent name, task and current step are laid out as a grid: in a narrow panel the task gets its own line under the agent instead of being squeezed to nothing, and checklist items run with run_todos_parallel show their title (not the first 60 characters of the prompt) on the lane.
• CLI / headless runs can create files again. write_file on a new path failed with edit.createFile is not a function outside VS Code; edit_file with an empty old_text on a file that does not exist now creates it instead of throwing ENOENT, and a non-empty old_text on a missing file gets a clear hint to use write_file.

Install

code --install-extension localpilot-1.1.9.vsix

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

Full changelog: v1.1.8...v1.1.9