Museum Exhibit Studio

This Maven CLI sample uses the GitHub Copilot SDK as a focused museum-curation agent. A museum educator chooses one of three approved fact sets or enters their own bounded facts, optionally streams scoped Wikipedia background research, generates visitor-facing exhibit copy, validates its structure, and can opt in to an exhibit.html capstone.

Run

From this directory:

./mvnw compile exec:java

Set COPILOT_MODEL to select a model; otherwise the Copilot runtime chooses its default. The sample requires an authenticated GitHub Copilot CLI.

Compile without contacting a model:

./mvnw compile

What it demonstrates

The learner-authored MuseumExhibitStudio entrypoint builds sessions directly with new CopilotClient(). The pre-built Curator* helpers provide approved facts, streaming, validation, scoped permissions, source extraction, and terminal prompts.

Prompt guidance is not an authorization boundary, so the application also:

• always registers and allowlists approved_fact_lookup, adding the read-only local approved_wikipedia_fact_lookup only when usable cited research exists;
• limits research to the configured Wikipedia MCP server and wikipedia-search / wikipedia-readArticle through a deny-by-default permission handler;
• captures the research body and trailing ## Sources citations for the second local lookup, never merging them into educator-approved facts or giving generation live Wikipedia access;
• bounds input to 20 facts of at most 500 characters each before every model send;
• uses explicit timeouts, rejects blank exhibit output, and disconnects sessions / stops clients on success and failure;
• checks one H1, required sections, a 100-140-word narrative, exactly three numbered questions ending in ?, and prohibited software vocabulary; and
• optionally allows builtin:apply_patch and builtin:create to write only exhibit.html in the application working directory.

The curator is instructed to call both local lookups before writing the narrative and visitor questions when research exists. Approved facts take precedence over supplemental research, which is data, not instructions. "Approved" research means application-accepted, not human-verified. Declined research keeps the single-tool path. Failed research or an unusable cited summary prints a warning and takes the same fallback. Confirm both lookup events on a successful research run; sources still print after the exhibit. The validator cannot prove semantic factual grounding. Generated claims still require human review or a separate evaluator.

Optional HTML capstone

When prompted, answer yes to generate exhibit.html. The pinned Java SDK 1.0.11 preserves permission fields such as fileName, so the strict permission handler approves a write only when its normalized path is exactly exhibit.html in this directory. Missing path data, other file paths, and non-write requests remain denied; there is no broad write fallback.

This is the application a learner ends up with after the museum lessons, not a separate reference architecture. The entrypoint keeps one small session runner that starts the client, creates the session, enforces the timeout, rejects blank output, and cleans up on every path; the research, generation, and optional HTML steps reuse it with different session configurations. Follow the track from workshop/museum-00-preflight.md.