Koru introduces a novel approach to closed-loop automation, aiming to orchestrate workflows across the expanding semcod ecosystem.
It seeks to eliminate manual toil by creating feedback loops where code changes, test results, and operational metrics continuously inform one another.
The system monitors the state of semcod/* repositories, interprets telemetry, and initiates corrective actions before problems escalate.
One key innovation is driving integrated development environments even when conventional plugin pathways fail, using low-level display capture, visual query language, and ydotool.
This combination enables Koru to see the IDE interface, interpret UI elements, and send precise keyboard and mouse commands without IDE‑specific plugins.
For hands‑on experimentation, Koru offers a one‑shot chat drive mode ideal for manual testing and exploratory sessions.
By invoking a single command, users can prompt Koru to perform a predefined interaction sequence within the IDE chat, such as sending a message and acting on the response.
Maintaining a reliable distribution channel, Koru’s Makefile includes a publish target that bumps version, cleans dist, builds wheel, runs twine check, and uploads to PyPI.
The publish target ensures clean, verifiable packages while requiring explicit authentication via PYPI_API_TOKEN or TWINE_USERNAME/TWINE_PASSWORD.
Koru operates in two primary modes: koru auto (active execution) and koru observe up (passive telemetry).
Telemetry and observability are woven deeply into Koru’s design, emitting JSONL lines on stderr for integration with platforms like Elasticsearch, Loki, or Splunk.
Advanced features include LLM reflection via OpenRouter to react to the IDE chat, allowing the agent to read conversation, seek guidance, and execute actions.