Skip to content

Releases: Vkuparin/gamepause-for-local-ai

GamePause for Local AI 2.6.1

Choose a tag to compare

@github-actions github-actions released this 18 Oct 17:32
  • Fixed Hermes pausing when its gateway runs the model server. The gateway now closes while gaming and reopens afterward, pausing messaging and background tasks too. Keep Hermes Desktop open.
  • Hermes is enabled on new installs; existing choices are preserved.
  • Shortened Ollama/Hermes settings and fixed clipped text.

Quit GamePause before upgrading. Restore paused AI before downgrading. Downloads are unsigned; check SHA256SUMS.txt.

GamePause for Local AI 2.6.0

Choose a tag to compare

@github-actions github-actions released this 15 Oct 01:34
  • Hermes is a main provider beside LM Studio and Ollama. Enable it under Advanced > Hermes to pause its verified managed local runtime while gaming, leaving Hermes open. GamePause records recovery before using Hermes's native off switch, then restores the runtime and captured local model content, live context and supported load fields. A Hermes-only setup needs no LM Studio installation.
  • The dashboard fits three main providers and Other AI apps, with stable provider order, a Hermes settings page, read-only diagnostics and provider-scoped round-trip verification.
  • Settings migrate to version 5 with a byte-identical version-4 backup and Hermes disabled. Recovery stays at envelope version 3 and accepts four provider kinds. Original Hermes evidence remains unchanged through retries and partial failures.

A live native pause/start/model replay passed with Qwen3.8-27B-UD-Q4_K_M on CUDA llama.cpp b11370. The runtime returned enabled, the model returned with its original 147456-token live context and supported fields, saved preferences matched, and recovery cleared. The check also exposed and fixed retained exited-process handles and asynchronous model readiness. Multi-model replay, mixed live providers, sleep/reboot, installer upgrade and physical/accessibility UI checks remain untested. Hermes support currently targets a local default-profile backend; cloud, remote and independently managed servers are excluded. Conversations, KV cache and idle timing are not restored. Restore pending AI before downgrading; older builds cannot read version-5 settings or Hermes recovery.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.5.3

Choose a tag to compare

@github-actions github-actions released this 12 Oct 05:34
  • Advanced > Ollama has an opt-in Prevent Ollama from reloading models while gaming setting. For verified Windows Ollama Desktop 0.40.3, GamePause unloads models, closes Desktop and its managed server, then restarts it in hidden mode and restores supported models after gaming. Requests arriving during shutdown may be canceled. Unknown versions and unsupported installations keep ordinary model unloading; external restarts are left alone.
  • Desktop recovery keeps encrypted startup evidence, original executable hashes and process identities in the existing journal. Separate checkpoints cover shutdown, restart and verification. Replaced executables, changed server configuration, unrelated servers and partial failures retain recovery. Desktop is rechecked after a failed model restoration before retrying loads.

Validated with 337 library tests and nine CLI integration tests, formatting and Clippy. Live Ollama 0.40.3 checks covered native shutdown/restart, persisted model recovery, active inference, external restart reconciliation and a harmless custom-game scanner/engine cycle. Original residency was restored. The model check used one small local GGUF completion model with context 4096 and indefinite keep-alive. LM Studio and generic process behavior were checked with fixtures; live mixed-provider recovery, Windows reboot/sleep and installed-app upgrade were not exercised. Settings remain version 4 and recovery envelope version 3; opt-in Ollama captures use guarded snapshot version 2, and ordinary captures retain version 1. Restore pending AI and turn the new setting off before downgrading.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.5.2

Choose a tag to compare

@github-actions github-actions released this 07 Oct 20:42
  • LM Studio accepts an already-loaded copy of the same model and quantization, even under another instance name or with different settings. Duplicate saved instances do not force another copy. If the loadout differs, a Cancel-default prompt offers to unload current models and restore saved models, with a fresh check before every confirmed unload.
  • Logs now use a logs subfolder, dated filenames, daily and 1024 KiB size rotation, and retention of the 10 newest managed files. Activity opens that folder and reads the latest log. Repeated countdown messages are omitted.

Validated with 320 unit tests and 9 integration tests using isolated fixtures. Live provider operations and the installed app upgrade have not been exercised. Settings and recovery formats are unchanged.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.5.1

Choose a tag to compare

@github-actions github-actions released this 07 Oct 19:57
  • LM Studio recovery refuses to load a second copy when the saved instance is missing but the same model is already resident under another identifier. Conflicting copies and idle TTL mismatches retain recovery and stop automatic retries until Resume AI. Independent healthy models and providers can still finish recovery.
  • Other apps are stopped one process at a time behind saved checkpoints. GamePause checks creation time, executable path and start details before stopping; new or changed instances without a matching saved relaunch command are left running. Inaccessible or incomplete process inventories hold recovery. Drive-root working folders are preserved, including older saved captures.
  • An Ask answer applies only to the game instances shown in the prompt. Another Ask game or a relaunched instance needs a new answer.
  • Executable-icon work uses bounded queues and uploads, and evicts one old cache entry at a time. Gameplay suggestions, Ask prompts and provider visibility changes now request dashboard updates.
  • Worker logs identify the build version and process ID. Activity shows the running build, process ID, data folder and log-write failures.

Validated with automated mock-provider, process, recovery and UI-state tests. The reported restart/Hermes scenario was investigated through existing logs, without controlling live models. The new behavior has not been live-tested with LM Studio, Ollama, real llama.cpp or KoboldCpp; sleep/wake and Windows 10 remain untested. Settings and recovery formats are unchanged.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.5.0

Choose a tag to compare

@github-actions github-actions released this 07 Oct 05:22
  • No change in behavior. The source was reorganized into smaller files, and a test that failed by chance about once in thirty runs was fixed.
  • README screenshots and the recording were retaken with this version.

Checked on one Windows 11 PC with a real game: LM Studio (one 27B model) and Ollama 0.40.0 (one small model) paused and restored, the Ollama model with the keep-alive time it had left. Not yet tested: sleep and wake, uninstalling, a real Ollama reload slower than ten seconds, real llama.cpp or KoboldCpp, the Scoop manifest, and Windows 10.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.3.0

Choose a tag to compare

@github-actions github-actions released this 06 Oct 23:54
  • GOG games are found automatically, whether they were installed through GOG Galaxy or with a GOG offline installer. DLC is not listed as a game of its own.

Checked on one Windows 11 PC with two games installed through GOG Galaxy, one of them with DLC. Not yet tested: a game installed with a GOG offline installer.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.2.1

Choose a tag to compare

@github-actions github-actions released this 06 Oct 22:58
  • Task Manager and the program file's properties now show GamePause for Local AI. They still showed the old name, GamePause for LM Studio.

Nothing else changed since 2.2.0, and what was not yet tested there is still untested.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.2.0

Choose a tag to compare

@github-actions github-actions released this 06 Oct 22:25
  • A closed LM Studio is no longer started when a game launches. GamePause used to ask LM Studio's command-line tool what was loaded, and that tool starts LM Studio in the background when it is not running. LM Studio is now left alone while it is closed and owes no restore.
  • Opening GamePause from the tray, the shortcut or the resume prompt now brings back a minimized window. A fullscreen game minimizes the dashboard, and a minimized dashboard used to ignore the request, so nothing appeared.
  • Ask games: the prompt has Always pause and Never pause, which remember the answer by setting the game's rule to On or Off.
  • Xbox game folders are looked for every five minutes on fixed local disks, instead of every 30 seconds on every drive letter, so network and sleeping drives are no longer touched.
  • A command that starts another program no longer makes GamePause wait for that program: the first LM Studio command after LM Studio was closed used to time out after 25 seconds.
  • The memory-freed line in the status card lines up with the text below it.
  • The worker log records a game-list refresh only when the list changed.
  • Scoop bucket in this repository: scoop bucket add gamepause https://github.com/Vkuparin/gamepause-for-local-ai, then scoop install gamepause.
  • README: how GamePause fits always-on assistants such as Hermes Agent and OpenClaw.
  • Diagnostics no longer probe AI apps with default settings when the data folder path is not a folder.
  • The memory-freed line is shown only while AI is held paused, not during the restore.

Checked on one Windows 11 PC with a real game: LM Studio (one 27B model) and Ollama 0.40.0 (one small model) paused and restored, and the dashboard opening over the fullscreen game. Not yet tested: sleep and wake, uninstalling, a real Ollama reload slower than ten seconds, real llama.cpp or KoboldCpp, the Scoop manifest, and Windows 10.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.

GamePause for Local AI 2.1.0

Choose a tag to compare

@github-actions github-actions released this 06 Oct 21:36
  • The app is now called GamePause for Local AI, in the window title, installer and Start menu, and the repository moved to gamepause-for-local-ai. The program file, install folder and data folder keep the name GamePause, so settings and pending recovery carry over.
  • The status card lists only the AI apps this PC has: LM Studio when it is installed, Ollama when it is installed or running. An app you do not use no longer gets a line.
  • A listed app that is closed reads Not running. Ready used to appear for LM Studio and Ollama even when neither was running; it now means the program is running and will be paused when a game starts. This is read from the running-process list; the apps are still not contacted while idle.
  • GamePauseCLI.exe is no longer part of the download, and its --status and --games scripting commands are removed. Diagnostics and the round-trip test are in Advanced. GamePause.exe --restore retries a pending restore when GamePause is not running and reports the result in a message box.
  • Ollama 0.40.0 is tested alongside 0.35.1: a live pause and restore of one small model passed on both with this build. The longer reload limit from 2.0.1 is now also covered by an automated slow-server test; a real reload slower than ten seconds has still not been observed live.
  • README: a table of the LM Studio and Ollama versions GamePause was tested with.
  • Old planning and review documents moved to docs/history and are no longer packaged.
  • Releases are built and published by the tag workflow, with this changelog section as the release notes.
  • Upgrading from 2.0 keeps settings and the sign-in startup entry, replaces the Start menu shortcuts with ones under the new name, and removes files that are no longer shipped.

Checked on one Windows 11 PC: a silent upgrade from 2.0.1, and a full pause and restore of LM Studio (one 27B model) and Ollama 0.40.0 (one small model) around a real game. Not yet tested: uninstalling, the installer's interactive pages, sleep and wake, a real Ollama reload slower than ten seconds, real llama.cpp or KoboldCpp, and Windows 10.

Download the Setup.exe for a per-user installation or the windows-x64.zip for portable use. Quit a running GamePause from the tray before upgrading. The binaries are unsigned; compare downloads against SHA256SUMS.txt.