Skip to content

FAQ

No. Discovery reads ADRs, RFCs, design notes, READMEs, and other Markdown without a required format. .decision/instructions.md can hold optional extraction hints.

No. Source Markdown is read-only input. The app writes its records to .decision/.

In your repository: Markdown plus validated YAML under .decision/. Commit it like any other text. SQLite and Qdrant hold rebuildable indexes in Docker volumes.

No. Deterministic/mock extraction runs without a model. Local OpenAI-compatible servers work without a key. Hosted providers need a credential through an environment variable, and that path is untested.

Will re-running extraction overwrite my reviews?

Section titled “Will re-running extraction overwrite my reviews?”

No. Human-reviewed records and their review chains stay byte-for-byte unchanged, and human judgments outrank inference.

Not yet measured beyond small fixtures and the bundled sample. A local Qwen3.8 model with reasoning off found all 7 decisions in the sample; the one-command demo itself uses mock rules tuned to the sample’s text. Measuring recall on more repositories is up next.

Can I expose it to my team over the network?

Section titled “Can I expose it to my team over the network?”

Not safely. The API and MCP endpoint have no authentication and bind to loopback. Remote access is an open design question.

No. Code analysis is a documented extension point with no implementation.

Yes. The Facts page lists the review log; each entry shows the prior state and an Undo action. Undo appends a record that restores the prior judgment, and is also available through POST /api/v1/reviews/{record_id}/undo.

Nothing in your repository until you save it. Option switches in the What-if panel are temporary previews. A saved scenario stores only its selections under .decision/scenarios/.

Apache-2.0. The source is at https://github.com/bcmyguest/switchplinth.