1.0.6
published 6 November 2026
superseded by 1.0.7 · current release 1.0.12
Mercury CLI 1.0.6 adds first-class support for the TensorFold local inference engine (/tensorfold and the TENSORFOLD and TENSORFOLD HOST groups in /model) across Apple silicon, Linux CUDA, LAN pairing, rented Runpod and Vast.ai pods, and a Windows llama.cpp adapter (served by Mercury's engine until TensorFold ships for Windows), along with a local prompt profile and a configurable local helper model in /config. The Session Concourse coordinator is now off by default (mercury --coordinator-on turns it on), leaving a two-pane concourse with the selected session's live view on the left and the session list on the right, and /goal joins /mission with an end-state verdict at the end of each turn. The release also repairs OpenAI HTTP 4xx diagnostics and resume recovery, crew-board compaction replay, dropped large @-mentioned image notices, ApolloReview's model-facing schema, and four tool-description lines.
Added
- Added /tensorfold and the TENSORFOLD and TENSORFOLD HOST groups in /model for the TensorFold 1.2.2 inference engine on Apple silicon (Metal) and Linux NVIDIA GPUs (CUDA): install TensorFold through Homebrew on macOS or the verified release archive on Linux, offer TensorFold first on first-run sign-in and in /model on a fresh install of a capable machine, pull and serve the qualified 256k starting points (Nemotron Lightning, ordered first on Macs under 64 GiB of memory, including 48 GB Macs, and on pre-M5 Apple silicon; Qwen3.8 27B with its DFlash2 speculative drafter, ordered first on 64 GiB+ M5 Macs and Linux CUDA; and Qwen3.5 2B for 8-16 GB Macs) or any supported MLX, NVFP4 or EXL3 checkpoint on disk or from Hugging Face, step the context window across 32k, 64k, 128k and 256k against usable memory, copy the macOS GPU wired-memory sysctl command with w when needed, and read per-turn decode speed, time to first token, prefill, cache hit rate, reasoning tokens and draft acceptance in /usage.
- Added LAN sharing and pairing and cloud GPU pod rental under TENSORFOLD HOST: press n on a Mercury-started local server to share it on the local network with a scoped mercury[inference] key and a 6-digit one-time pairing code (or connect by host:port), or rent a Runpod or Vast.ai CUDA pod through /llmhost that verifies the TensorFold archive by SHA-256 and Ed25519 signature, pulls and serves over an SSH tunnel, shows live spend and token speed on its card, requests a stop at once with s (the row reads stopping until the vendor confirms, then stopped, when compute billing ends), and terminates the pod and its disk from the panel with t (or clears the record when a Vast.ai instance was already destroyed outside Mercury).
- Added the Windows x64 TensorFold adapter (dist/tensorfold-adapter.mjs, "served by Mercury's engine until TensorFold ships for Windows"): until a native Windows TensorFold binary is installed, /tensorfold and /model serve Qwen3.5 4B, Qwen3.5 2B and Nemotron Lightning on CPU through llama.cpp (build b11829 or newer, downloadable in place with SHA-256 verification), translating reasoning_effort and thinking_budget and naming Mercury's engine across the picker, card, chat feed, /health and the support string.
- Added the local prompt profile (MERCURY_LOCAL_PROMPT_PROFILE, on by default), the Helper model and Helper model for compaction settings in /config (off by default, supporting 2B on TensorFold or Mercury's engine, 4B on llama.cpp, or the session's local model), the TensorFold, support and Helper model rows in /health, and the c key in /tensorfold and /model to copy the one-line TensorFold support string.
- Added /goal as a second name for /mission: a standing session goal is split into its end state and constraints and checked at the end of each turn against the turn's closing text, the <<<GOAL-MET>>> completion line (elided on screen in the chat and concourse live tiles), tool-call and verification records, and working-tree progress, using JEV when enabled (at a 0.6 confidence floor) or the completion-line check, counting three consecutive checks with no tree change and no new verification record as unattainable on either check, and asking the model to end with one question to you when the goal is ambiguous or unattainable.
Changed
- Changed the Session Concourse coordinator to be off by default: mercury --coordinator-on turns it on (and --coordinator-off turns it off again). With the coordinator off, the concourse is two panes without the top panel (the selected session's live view and message box on the left, the session list on the right with a LAST USED column and x to stop or a to archive), n starts sessions directly, and /config and /health report the switch; when the coordinator is on in manager mode without a model, its note reads "manager mode needs a coordinator model: sign in to a provider or serve a local model".
- Changed /llmhost's private recipes file (<config home>/host/private-recipes.json, MERCURY_HOST_RECIPES_FILE or the saved choice) to join the shipped catalogue, with private rows ordered first and a private row replacing a shipped row with the same id, instead of replacing the shipped list, while malformed entries are skipped with a warning.
- Changed the README's recording to a true-black take of structural search and edit, in place of the session-board recording, under a new file name so cached copies are not served.
- Changed the bundled FilmCraft, PhotoCraft and LightCraft extensions to the upstream releases of 5 November: FilmCraft 0.17.1 (a window takes every click sent through ui_click again; timeline.move takes seconds, frame and timecode; an image-sequence export names its first and last stills), PhotoCraft 0.20.2 (TIFF options and pcraftFullComposite on doc_save and doc_export; a PSD or PSB above the memory budget opens only with approve_high_memory) and LightCraft 0.18.1 (the first launch with a library offers the lensfun lens database once, and lens.databaseStatus, lens.downloadDatabase and library.refreshLensProfiles join the commands); each lock carries the release's asset names, sizes and SHA-256 digests, and the three skills carry the new words.
Fixed
- Fixed four model-facing tool descriptions and deferred-roster announcements: Bash and PowerShell timeout descriptions name the background task id that TaskStop takes; WebFetch states that a page's body comes back as Markdown without the HTML head and title; Grep's output_mode names content as the mode that prints matching lines; and unattended and local deferred-tool lists include the one-line purpose for ChangeSet and Eval alongside Crewmate, Skill, Test, LspRead and AstSearch.
- Fixed ApolloReview's model-facing input schema so it exposes only summary, blockers, specFiles and runNote, keeping the consent card's decision and refineNote fields off the wire.
- Fixed @-mentioned image files that exceed the 256 KB mention limit or fail image resizing being dropped without a trace: Mercury now emits a dropped-image notice row in the chat ("Dropped <path>: <reason>") and tells the model to attach a smaller copy or Read the file with a crop.
- Fixed a crewmate's saved transcript chain breaking when an attachment row is skipped or an empty-content message is omitted on write, so resuming a crewmate preserves its earlier history.
- Fixed crew-board chat rows from earlier turns being retained and repainted after /compact or auto-compaction instead of compacting with their turns.
- Fixed OpenAI Responses requests failing on unpaired UTF-16 surrogates, opaque HTTP 4xx errors, or oversized replayed tool outputs on a resumed turn: request strings are sanitised before send, non-retryable 4xx errors surface error.detail and error.param and write a credential-scrubbed diagnostic capture in the session folder, and a resumed turn that hits a 4xx compacts oversized tool results once, persists those replacements across later turns, and retries.
- Fixed the Document tool's PDF render through the headless browser, which failed every call with "Failed to parse parameter value: 54pt": the page margin reaches the browser in inches now, and the tool's margin is still given in points.