1.0.0-beta.40

published 21 October 2026

Added model-first /llmhost and /model hosting actions: /llmhost opens on Which model?, showing where each model can run right now, its quantization variants, how its GPU-memory budget splits across parallel conversations (conversations × context) and its default thinking level, and /model includes searchable run it → /llmhost actions (run it on this Mac → when a model fits a local Mac)
Added GPU rentals across Runpod, Vast.ai and Lambda in /llmhost: Runpod rentals, background setup workers and 5-second stock watches are joined by Vast.ai and Lambda rental adapters, which are new and verified against the vendors' published APIs on fixtures (live rental runs were not performed)
Added own-computer model servers (/llmhost → My computers and mercury host start|status|stop|restart|logs|uninstall|add|recipes): checks this computer or a saved SSH computer, picks an Ollama or GGUF recipe that fits its memory, verifies a local SHA256SUMS weight bank or confirms a download, and runs the engine as a user login service (launchd, systemd --user or Windows Scheduled Task) registered as compat/box-<name>/<model>
Added fast-load kept weights disks in /llmhost (Fast-load (keep weights near the GPU)): keeps revision-verified model weights on a Runpod network volume or global volume (Prepare now or At the next rental), verifies warm mounts by seeded sampling in a few seconds (with full SHA-256 verification on demand), and shows monthly storage billing and keep-or-delete options when terminating a pod
Added Tune before renting and on a running pod in /llmhost: sets quantization, conversations × context splits against the GPU-memory budget and default thinking level before confirming a rental, and reloads a running llama.cpp Runpod pod in place with automatic rollback if the reload fails
Added /effort for self-hosted and OpenAI-compatible models: lists and sends each model's supported reasoning levels (chat_template_kwargs on llama.cpp, top-level reasoning_effort on vLLM, SGLang, Ollama and Kimi K3), retrying once without the field if a server rejects it with HTTP 400
Added trillion-parameter recipes and multi-node Runpod Instant Clusters in /llmhost: 19 recipes across Kimi K3, Qwen3.8-2.4T-A95B (FP8 and NVFP4), DeepSeek V4 Pro, MiMo-V2.6-Pro-RL, Mistral Large 3 675B and GLM-5.3 (FP8 and BF16), showing hourly compute, per-node disk, weight-download and model-licence details and replaying Kimi K3 assistant reasoning and tool calls across turns
Added prebuilt CUDA 13 llama.cpp engine image (ghcr.io/whq02/mercury-llamacpp): GLM-5.3-Flash and Nemotron 3.5 Super VL 120B GGUF rentals use the preinstalled llama-server binary and skip source compilation during pod setup
Added live output tail for running ! shell commands in chat: shows a ticking running · … timer and the latest output lines (six at 120 columns, three at 80) while keeping carriage-return progress updates on a single line
Added /restart and one-time installed-build notices: a live chat shows a one-time notice between turns when a newer Mercury build is installed, and /restart lists the Monitor watches and background shells that will stop, switches the background daemon to the installed build and resumes the chat
Added oculus and headless session kinds: interactive terminal sessions run as oculus (permission asks and questions wait for your answer without timing out), while unattended mercury run, mercury runner, scheduled and coordinator sessions run as headless (/health reports Session kind and the Session Concourse marks headless rows)
Added connected server tools list and tool-schema detail view in /mcp: expanding a connected MCP server opens a filterable tool list and per-tool parameter schema view, with Esc stepping back one level at a time
Changed headless runs that ask an unanswered AskUserQuestion to end blocked with the question quoted on the outcome row, read from the run's recorded asks
Changed self-hosted llama.cpp, vLLM, SGLang, Ollama and LM Studio streaming: llama.cpp conversations reuse a consistent server slot (id_slot with cache_prompt: true), self-hosted servers poll liveness during long prompt prefills instead of timing out after 30 seconds, and streaming tool-call arguments count as live crewmate activity
Fixed daemon-hosted chats dropping post-compaction messages after /compact: the transcript writer scans all recorded IDs on disk without duplicating earlier history, and the chat keeps post-compaction messages in order across reattach, /restart and daemon restart
Fixed unreachable custom endpoints in /model opening an unrelated sign-in screen: keyless endpoints retry in place, unreachable custom endpoints show Retry this server → and Remove this endpoint → in /model, terminating a pod removes its provider entry, and opening /llmhost cleans up endpoints whose pods have ended
Fixed /restart waiting on its own session during daemon handover, repeating on an already-current session, or suggesting esc during a running turn: stops the session's runner before handover, checks the runner's build info first, and tells you to wait for the active turn to finish
Fixed the installed-build notice repeating after /clear or rendering collapsed: /clear carries over the shown-build marker from the previous chat, and the notice renders unfolded
Fixed Monitor watches waking the model on every 1-second tick: after the initial 200 ms burst, ongoing output is delivered at most once every 30 seconds with the line count and last line
Fixed Windows Monitor, run_in_background and inherit_session_env calls under the default shell engine: uses installed Git Bash when present and spawns per-command brush engine processes on Windows machines without bash.exe
Fixed /mcp sign-in and mid-session server additions: pressing ↵ on a needs sign-in row starts authentication from the active runner, and user-scope MCP servers added mid-session become available in the running chat right away
Fixed sessions resumed after a redeploy sending stale tool schemas: if a restored tool's live input_schema or build path has changed, Anthropic and OpenAI Responses requests send the current tool schema
Fixed compaction headroom and session startup checks: compaction reserves the largest token underestimate seen in the session so file-heavy turns do not immediately re-trigger compaction, keyless named endpoints (compat/<name>/<model>) work as the default model for new and /clear sessions, the full-slots notice on the home screen leads with how to free a slot (/concourse or /seats N), and a project inside a home-directory git repo scopes its startup git status check to the launch folder