Skip to content

GuideAnts v0.11

Choose a tag to compare

View all tags
@douglasware douglasware released this 17 Sep 16:17
· 20 commits to main since this release
v0.11
8ba1147
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Sends that don't wait, and a local-AI stack that boots. The API no longer crashes at startup when the llama runtime is the discard-loopback sentinel (the v0.10 startup regression); chat sends go optimistic with rollback and the redundant preflight round trips are gone; the load dialog is the sole model-load owner — a not-ready send gets an immediate 409 instead of a 15-minute load poll; and the installer wizard no longer dies on a bash parse error. llama.cpp server images are pinned to b10615 across all backends.

Get started

  1. Install Docker Desktop (Windows/macOS) or Docker Engine 24+ with Compose (Linux). Windows needs WSL2.
  2. Download guideants-installer-v0.11.zip from this release.
  3. Unzip and run:
  • Windows: double-click guideants.cmd
  • Linux / macOS: chmod +x guideants.sh && ./guideants.sh
  1. Choose database layout, AI backend, and optional services when prompted.
  2. Open http://localhost:5107/ — first account becomes Admin.

Existing installs: relaunch the installer; if :main moved past your pins, accept the update prompt to pick up this channel. Volumes are kept.

Highlights

Local AI startup & lifecycle

  • Startup crash fixed (v0.10 regression): the API no longer fails to start when LlamaCpp:BaseUrl is the discard-loopback sentinel (http://127.0.0.1:9/... = "local llama not in use"). The typed llama-admin client leaves BaseAddress unset for the sentinel instead of throwing (#138)
  • The load dialog is the sole model-load owner. POST /messages no longer starts or waits on a model load (removed auto-load + 15-min poll). Not-ready or invalid model gets an immediate 409 with live runtimeStatus; the client restores the draft and surfaces the LlamaRuntimeModal (#146)
  • Notebook-scoped chat alias is tracked, so lifecycle applies keep an in-use local chat model loaded instead of emitting llama.enabled=false (fixes "model not loaded" mid-inference) (#146)
  • llama.cpp server images pinned to b10615 across all backends; engine logs now pump to a log instead of DEVNULL (#146)

Chat send path

  • Sent messages render optimistically before runtime preflight, with rollback on failure; post-yield preflight failures surface as SSE errors broadcast to conversation observers (#141)
  • Redundant conversation-snapshot GET after send removed; private-path turns created in streaming status (drops two DB round trips); turn_created yielded before history setup (#141)
  • Attachment history loading batched into one query instead of one per message; assistant-switch attachment loop batched too (#141)
  • Stop button is no longer silently discarded during the preflight window — cancel state is reset before the optimistic dispatches, and a Stop-then-not-ready sequence can no longer leave the stream stuck cancelling (#141)
  • Model reasoning choices cached for 60s instead of a DbContext per chat run; chat-ready attachment renderings cached keyed by file id + LastModifiedUtc (#141)

Notebook & host mounts

  • Host mount trees are lazy-loaded: shallow scan with on-demand listing APIs, expanded branches grafted across polls (#139)
  • Notebook folder-tree sync fixes; ChangeQuant no longer deletes prior quants; stream/watchdog races that left aliases or SSE stuck are fixed (#139)
  • host-ssh skill supports macOS (#148)

Installer

  • Wizard bash crash fixed: ${#arr[@]} combined with a :- default was a parse error ("bad substitution") that made guideants.sh crash on every run right after backend selection (#144)
  • installer_set_local_image_env no longer trips set -e with a silent, message-free exit when COMPOSE_MODE=local and no GA_*_IMAGE overrides are set — installs now proceed through image checks and actually start the compose stack (#144)

Sample skills & media

  • Sample skills expanded: AudioCpp, host SSH, HTML, MSSQL, Qwen image, SearXNG, talking-head video, video production, and reusable templates (#142)
  • ASR/TTS lifecycle hardening: engine failure classification, recovery behavior, inference isolation (#142)
  • CPU, CUDA, ROCm, Vulkan, and hybrid media compose/build support (#142)

Settings

  • reasoning choices field in model settings is now stored in JSON format to match other settings (#149)

Notes

  • This release pins llama.cpp server images to b10615 — local-model workloads need the republished AI images (fresh install, or accept the installer update prompt).
  • Apple Silicon: prefer the slim AI backend (linux/amd64 under emulation).
  • Operator details: docs/release-runbook.md, docs/local-ai-lifecycle/, installer/README.md, deploy/azure/README.md.
Assets 3
Loading