Use transient desktop chat

The v0.6 Chat workspace is a narrow, transient presentation over a reviewed provider profile. It is not a genealogy command surface, autonomous agent, or evidence system. Model output is advisory, not evidence, and must be verified against authoritative sources.

Do not paste real genealogy records, government identifiers, credentials, private notes, database contents, local paths, logs, or other sensitive material while learning the interface. Use fictional text and the minimum permitted data classes.

Start a conversation

  1. Open Chat.
  2. If the workspace says Configure a provider to start, choose Open provider settings and follow Configure a desktop provider and consent.
  3. Select the exact provider profile and model.
  4. Choose a reviewed purpose: Genealogy analysis, Source analysis, or Writing assistance.
  5. Review Provider and privacy scope. It identifies Provider, Model, Purpose, Data shared, and Consent.
  6. Choose New conversation only when the scope matches your intent.

For a loopback Ollama profile, the status is Local and consent reports Not required for local provider. Cloud profiles are Remote and require a current compatible grant; No compatible consent blocks conversation creation. provider=none stays network-free and cannot generate chat output.

Send and control a message

The workspace labels the session Transient conversation and Not saved. Type no more than 16,384 characters, then choose Send message or press Ctrl+Enter. Use Stop response to request cooperative cancellation. After a completed response, Regenerate response repeats the last eligible turn within the same reviewed scope.

The composer and transcript remain in memory only. The service bounds each session to 32 messages, 65,536 characters of context, 4,096 output tokens, a 120-second timeout, and one safe retry before output starts. Tools, files, databases, shells, plugins, external services, and autonomous actions are disabled.

The status can be Not started, Streaming, Stopping, Completed, Interrupted, or Failed. A screen reader receives one polite status announcement path; the transcript itself is not a live region.

Treat output as untrusted display data

Rendered Markdown uses a closed CommonMark/GFM allowlist. Raw HTML, images, embeds, implicit links, and executable actions are disabled. To retain text, choose Copy response; only plain text reaches the clipboard.

An explicit HTTPS link shows its normalized destination and must pass through a native Electron confirmation before opening. The renderer cannot navigate or call the network directly. Never treat a generated claim as evidence: compare it with the source transcription, GEDCOM provenance, and other authoritative records.

End or recover a conversation

Choose Close conversation when finished. Closing the session, tearing down the workspace, or shutting down the sidecar clears its in-memory content. Audit records retain reviewed identifiers, counters, usage, and one-way hashes—not prompt or response payloads.

If the stream is interrupted, use the displayed stable code and remediation. Do not repeatedly submit after output begins: replay and acknowledgement are bounded, and generation is never retried after visible output. A revoked consent, changed endpoint, missing credential, owner mismatch, cursor gap, or stalled renderer fails closed.

See Desktop reference for limits and error codes and Privacy and consent for the disclosure boundary.