Changelog

All changes to this project are documented in this file.

The format is based on Keep a Changelog and this project adheres to Semantic Versioning.

[Unreleased]

[0.2.5] - 2026-11-07 @ 07:09

Changed

• Website tweaks.

[0.2.4] - 2026-11-07 @ 07:00

Changed

• Reverted theme to fixed state.

[0.2.3] - 2026-11-07 @ 06:19

Summary

Upgrades no longer wait for an HTTP health response when restarting services.

Changed

• Expanded the environment setup guide into a consistent Debian 13 walkthrough for supported NVIDIA or CPU hardware, including llama.cpp builds, Qwen3.5 vision conversion, and first-run checks.

Fixed

• Skip the readiness health check during upgrades, including when services were stopped beforehand. Install readiness checks and upgrade failure handling for runtime validation and service start commands remain in place.

[0.2.1] - 2026-11-07 @ 05:41

Changed

• Rewrote the project documentation for readers: separate install, upgrade and uninstall guides, runtime examples, a clearer installation inventory, and a shorter project-structure page.

[0.2.0] - 2026-11-07 @ 04:52

Summary

Manage local inference through a browser control panel, with validated settings, live readiness and narrowly authorized service restarts. Honor explicit Jinja disable settings and keep installation tests reliable across releases.

Added

• Dark, responsive HTTP control panel based on the configuration/status sketch, using the project logo, installed GGUF and projector selectors, English message keys, atomic settings updates, restart confirmation and live readiness.
• Passwordless start/stop grants for exactly the HTTP and LLM systemd units in /etc/sudoers.d/llamasrv, validated before installation and tracked for removal during uninstall. The LLM restart uses sequential noninteractive stop/start.
• Isolated production Jinja2 runtime and writable service-owned configuration, while keeping application code and the sudoers policy owned by root.
• HTTP and offline installation acceptance coverage for configuration validation, exact sudo commands, policy ownership/removal and isolated deployed rendering; operator documentation for the browser and service permissions.

Fixed

• Explicitly disable llama.cpp chat templating with --no-jinja when runtime jinja is false; omitting --jinja left it enabled on builds such as wintermute’s.
• Keep disposable lifecycle-test versions independent of the repository’s current release metadata.

Changed

• Updated site logo, again.

[0.1.0] - 2026-11-07 @ 03:28

Summary

Run a configurable standalone llama-server on port 51001 alongside LlamaSrv’s HTTP service on port 51000. Defaults match ax3l’s Qwen3.5-4B vision setup; operator-owned runtime settings survive installation and upgrades.

Added

• Shared, validated runtime configuration for both listeners and the selected GGUF model, optional vision projector, context size, metrics and Jinja flags.
• llama-server launcher with fatal checks for missing or unusable external binaries and model files, and on-demand inference readiness in /health.
• Separate hardened systemd units, sequential service controls, and completed install/reinstall/upgrade/uninstall operations with configuration preservation, deployed-code checks and readiness-gated successful-version bookkeeping.
• GPU-free process/lifecycle tests and operator documentation for development on sally and manual installation on wintermute; Git and release operations remain operator-controlled.
• Standard-library threaded HTTP server started with python3 -m llamasrv, defaulting to 0.0.0.0:51000 from CMDB metadata, with JSON service information, /health, HEAD support, local-time request logs, bounded request timeouts, and graceful shutdown.
• Validated production JSON listener settings and CLI overrides, English CLI message catalogue, startup documentation, and HTTP/configuration/lifecycle acceptance checks.
• Pinned bmforge-github-pages skill with the local BMForge manager and lock file.
• Dark-themed Jekyll configuration for llamasrv.osoyalce.com, with English metadata, multilingual support, and the project’s MyCount identifier.
• Documentation homepage, maintenance guide, and original SVG project logo.
• GitHub Actions workflow to build and deploy the documentation from main.
• Ignore rules for generated Jekyll output and caches.
• Pinned bmforge-http skill for the planned application web interface.
• Pinned bmforge-project, bmforge-constants, bmforge-localization, and bmforge-installation baseline skills at BMForge revision 29a100c39997.
• Project engineering instructions and canonical LlamaSrv metadata, with initial version 0.0.0, subtype Application, and base directory /opt/prod/llamasrv.
• Standard BMForge release script, CMDB checker, and install, upgrade, and uninstall entry points.
• Development guide covering the scaffold, verification, and remaining application implementation work.

Changed

• Updated site logo.
• Keep the README focused on the project overview and documentation link.
• Exclude the application package from the documentation site output.

---

Documentation index