Harbor

🦭 TL;DR A whole menagerie in one release. 🦉 Chandra 2 joins the backend flock as a vLLM-served OCR model that reads documents and images into Markdown, HTML or JSON. 🐙 Boost grows a new arm with the codemode module, where a model writes one Python program that calls every Boost tool as a plain function, and harbor launch --codemode wires it into any launched CLI. 🦔 harbor down and harbor restart now touch only the named services, 🐜 the ollama-init sidecar shrinks from the multi-GB Ollama image to Alpine, 🐿️ services can ship their own default.env, and 🐝 harbor dev docs --check keeps generated docs honest. 🦉 Chandra Chandra 2 OCR by Datalab, served with vLLM behind an OpenAI-compatible API: turns documents and images into Markdown, HTML or JSON with tables, forms, math and reading order intact. Needs an NVIDIA GPU; 16GB is enough. harbor up chandra nvidia 🐙 Boost codemode Replaces the whole Boost tool catalog with a single execute_code tool. The model writes one Python program that calls Boost's tools as functions; Boost runs it in a sandboxed subprocess runner and round-trips every call to the real tool, with caps on timeout, output size and number of tool calls. harbor launch --codemode harbor boost modules add codemode 🐾 Misc - 🦔 harbor down and harbor restart are scoped to the named services; other running services are left untouched. - 🐜 Slim ollama-init sidecar: runs on alpine:3.20 instead of the multi-GB Ollama image and streams pulls through the server's /api/pull API. - 🐿️ Per-service defaults: services//default.env files are loaded with the profile. - 🦫 Multi-source .env repair is atomic and duplicate-free. - 🐝 harbor dev docs --check flags stale generated docs, and the release pipeline builds Boost before regenerating its docs. - 🐘 Chandra defaults tuned for 16GB GPUs. Full Changelog: https://github.com/av/harbor/compare/v0.5.6...v0.5.7