Skip to content

Companion Hub Desktop v0.2.74

Choose a tag to compare

@github-actions github-actions released this 21 Sep 23:02
· 6519 commits to 1383a9fda83c7616817881f3cd0449462f6b07e9 since this release

Companion Hub Desktop v0.2.74

Downloads

Platform Installer Notes
Windows (x64) .msi or -setup.exe Recommended for most Windows users
Windows (ARM64) .msi or -setup.exe Windows on ARM (Snapdragon and other ARM64 PCs)
macOS (Apple Silicon) .dmg M1/M2/M3/M4 Macs
macOS (Intel) .dmg Older Intel-based Macs
Linux (x64) .deb or .rpm Ubuntu/Debian or Fedora/RHEL
Linux (ARM64) .deb or .rpm Raspberry Pi, ARM servers

Requirements

  • Docker must be installed and running
  • Windows: Windows 10+ with WebView2 (ships with Windows 11)
  • macOS: macOS 11.0+
  • Linux: WebKitGTK 4.1

What's Changed

  • fleet: backends --ollama-max-loaded N manages OLLAMA_MAX_LOADED_MODELS in the runtime drop-in by @mrhegemon in #1553
  • pool: slot-aware placement — a full Ollama node goes behind one with a free slot (poolSlotAwareness, off by default) by @mrhegemon in #1554
  • Restricted Hub to only pruning its own Docker items by @barankyle in #1552
  • pool: a request's num_ctx picks which nodes may serve it, so the handout uses the fleet's largest context cap, not its smallest by @mrhegemon in #1555
  • apps: probe an app-declared readiness endpoint and surface its failing checks in runtime health by @mrhegemon in #1557
  • ci: refresh the app-lifecycle compose snapshots and isolate the resident-model cap test by @mrhegemon in #1558
  • pool: a request the caller abandoned keeps the node it was waiting on, instead of reading as "placement found nothing" by @mrhegemon in #1561
  • pool: a node's context cap is visible — per peer in pool status/pool peers, per node in a read-only fleet backends by @mrhegemon in #1562
  • build identity: a deployed Hub can say which build it is, and the release keeps the version honest by @mrhegemon in #1563
  • update: report CLI-vs-stack skew, honour the operative env file's image pin, and record what was deployed by @mrhegemon in #1564
  • fleet: backends decides Ollama ownership by the :11434 listener, not by a user unit's name; hub-pool docs on proxy-only queue depth by @mrhegemon in #1560
  • sandbox: grant NET_ADMIN to transmission-vpn and wg-easy by @mrhegemon in #1565
  • fix: root createTempDirectory prefix under the OS temp dir by @mrhegemon in #1569
  • inference: tier an NVIDIA card from the host probe when the container has no nvidia runtime by @mrhegemon in #1559
  • fleet: a name-matched user unit gives way only to a SERVING system unit; a free :11434 refuses as before (#1560 follow-up) by @mrhegemon in #1568
  • Let an editor or agent use this Hub, and reach engines the operator already runs by @mrhegemon in #1567
  • fix: post-update health probe ignores one-shot init exits, and auto-rollback always restores + brings the app back up by @mrhegemon in #1570

Full Changelog: v0.2.73...v0.2.74