Question 1

Can I run this on Linux, or is it macOS and Windows only?

Accepted Answer

The offline runtime only publishes macOS arm64 and Windows x64 payloads, so a Linux container cannot host the engine. What runs in Docker is the MCP gateway, covering the Python package, knowledge base, and memory, and it needs the server root variables pointed at a host that has the runtime installed. The bundled Dockerfile and compose file are labelled experimental, and exposing the gateway beyond localhost requires a token.

Question 2

How do I know the model is not inventing chart interpretations?

Accepted Answer

The README describes a fidelity check rather than trusting the model. A deterministic validator compares each interpretation sentence against machine-readable chart truth and labels it supported, invented, or contradicted, and feeding a wrong chart or an induced restatement is meant to force a red result, with benchmark cases locked in step with the tool registry. Provenance is layered on top: the knowledge base carries 409 entries across 31 domains, each citing its upstream file and version.

Question 3

What happens when a technique needs data the user has not given me?

Accepted Answer

The agent is stopped rather than allowed to guess. When a technique depends on time, location, timezone, gender, subject, house system, calendar, or casting method, a hard clarification gate intercepts the call and returns text the agent can forward straight to the user, and nothing proceeds until the user confirms. Where several techniques run in parallel, each conclusion must bind to a real passage and disagreements are disclosed rather than averaged.

Question 4

What are the realistic install headaches?

Accepted Answer

Several, and they are documented in a troubleshooting table rather than hidden. Setup needs uv and prepares Python 3.12 or newer itself. The PyPI channel is not open yet, so zero-install means running the published wheel through uvx. On Windows, a non-ASCII runtime path breaks the Chinese systems because the bundled Java and ephemeris use narrow-character file APIs, path length can force a custom runtime root, and proxy or VPN filters can kill the Java process outright. On macOS, downloaded binaries may carry quarantine attributes needing an attribute fix.

Question 5

How do I wire it into an AI client, and does anything leave the machine?

Accepted Answer

One command per client: `uv run horosa-skill setup --client <name>` writes the MCP configuration, and `serve` starts a local MCP endpoint on 127.0.0.1:8765 with stdio and streamable-http transports available for clients that prefer them. The computation is local and offline, so the techniques themselves do not require a hosted model call.

Question 6

What does a call actually return, and can I export a report?

Accepted Answer

Every call returns a uniform envelope, and techniques wired to the export protocol add export_snapshot with the body broken into structured paragraphs. `horosa_technique_report` renders DOCX, PDF or JSON from that material, with Markdown tables becoming real Word tables plus a navigation outline, table of contents and page numbers. response_view=titles or sections trims the payload while the full snapshot stays archived.

Question 7

What does the technique evidence card tell me?

Accepted Answer

The response carries a card naming the technique, the school convention it followed, who computed it, whether every paragraph is present, and the version chain behind the result, so a conclusion can be traced back to how it was produced. The same machinery flags cross-technique convention conflicts at the session level rather than letting two traditions silently disagree.

Horosa Skill

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community