Changelog

All notable changes to Donna are documented here. The format follows Keep a Changelog; versions follow SemVer.

[0.3.0] — 2026-06-24 — Sticky skills

A per-chat "Keep skills on" toggle so applied skills carry forward into follow-up turns without re-attaching.

Added

• Sticky skills — a Keep skills on switch in the in-chat composer toolbar. Turn it on and the skills you have attached to the current turn are kept in place for every follow-up message in that chat, with a quiet "Keeping N on" indicator (skill names on hover) when active.
• Design posture: off by default, per-chat scope (state lives on the chat row — a new chat always starts off), and audit-honest — every turn still records its own applied_skills so the receipt is never misleading.
• Backed by backend pin bump 658fdbc → 5ad9f9e (lq-ai PR #211: chats.sticky_skills text[] column + message-create set_sticky field).

0.2.0 — 2026-06-21 — Legal research + MCP

The legal research + MCP milestone — surfacing LQ-AI's CourtListener case-law research and Model Context Protocol tool support in Donna, governed and transparent.

Added

• Research workspace (/research) — search U.S. case law via CourtListener, read full opinions in the document panel, find in opinion, and verify citations in a block of text against the source.
• Governed chat tool-loop — the assistant can consult case law and connected tools mid-answer, pausing for your Approve / Deny before it runs one; a connect-on-demand prompt links an account inline when a tool needs it.
• External-source citations — a "sources consulted" panel + provenance pill beneath answers that consulted case law, each row linking to the case on CourtListener (distinct from the character-verified quote citations).
• MCP tool administration (/settings/mcp) — admins view operator-connected Model Context Protocol servers and enable individual tools (with read-only / destructive / needs-confirmation badges).
• Per-user Connections (/settings/connections) — connect your own account to OAuth-protected MCP tool servers.
• Built-in skill inspector (/skills/view/{name}) — read a built-in skill's content and its declared connector usage ("Uses: …", with an amber note when a connector isn't configured).
• In-app guidance — new About pages (Research; Tools & connections), plain-language explainers on each new surface, and clickable starter searches / example prompts.
• An empty-response fallback — a completed turn that returns no content now shows an honest message with a Retry, instead of a blank bubble.

Changed

• Backend pin advanced from c4d4482 to 658fdbc (lq-ai WS2–WS5: CourtListener research, the MCP client, the governed tool-loop, and transparency / external-source citations).

0.1.0 — 2026-06-07 — First public release

Donna's first public release: a friendly, document-forward frontend for the LQ.AI legal-AI engine. Inference can be free and local (Ollama) or fully user-controlled (bring-your-own provider keys) — the product is free and the backend is self-hostable. Donna talks to lq-ai only through its published API and vendors that backend as a pinned git submodule (pin c4d4482), so the whole product runs from one compose file.

Assistant

• Streaming chat with character-verified citation pills — hover for the source quote, click to open a docked document panel jumped to the exact cited passage (PDF rendering + verbatim text-layer highlighting; inline markdown/plain-text rendering for text documents).
• Per-turn receipts drawer exposing every retrieval, inference, and skill event behind an answer, including whether anonymization was applied.
• Prompt enhance on every composer (landing + in-chat), per-message file attachment, a saved-prompts inserter, a model/tier picker, and skill attach with a typed skill-inputs form.

Matters (projects)

• Scope chats to a matter with files, linked knowledge bases, attached skills, and free-form context. Privileged matters enforce a minimum inference tier in the composer.

Knowledge bases

• Create, link, and upload; documents auto-ingest for retrieval (RAG) with live status, hybrid-search (alpha) tuning, and per-file download.

Workflows

• Skills — author your own or fork built-ins; typed inputs with a per-message input form; slash aliases.
• Playbooks — negotiation positions applied to a contract → verdict scorecard + a consolidated redline view; browse, apply, generate a draft from your own documents (easy-gen), or author manually with a full position editor.
• Saved prompts — reusable snippets inserted at the cursor.
• Automations — runs Donna executes on its own: run-now, cron schedules, and KB-arrival watches, each opt-in per user. Every run leaves a transparency receipt (phases, tool calls, cost, terminal reason) and surfaces its results: findings, document-grade artifacts (memos written into the run's target knowledge base, openable inline or downloadable from the receipt), proposed memories, and recurring precedents — with a review queue (keep/dismiss/promote) and a notifications inbox.

Tabular review

• Ask the same questions across many documents → a cited, confidence-scored grid; per-column model-tier floors and ensemble verification; Excel/CSV export; run history with resume and cell→source citation navigation.

Settings

• Account & security (profile view + display-name edit, MFA disable), data export and scheduled deletion (with a pending-deletion banner), preferences (ambient trust pills, provenance visibility, automations opt-in), a read-only trust matrix, and model management: per-category routing, installed local (Ollama) models, and bring-your-own provider keys (admin, hot-applied, write-only).

Guidance

• An in-app guide at /about — instructional pages for every area plus interactive playgrounds explaining how the LQ-AI engine works.

Architecture

• A backend-for-frontend: the browser talks only to Donna's SvelteKit server, which holds the lq-ai JWT in httpOnly cookies, attaches Authorization: Bearer when proxying, and refreshes on 401 — no CORS, and the JWT never reaches client JavaScript.

Notable upstream-driven capabilities & fixes (lq-ai)

• Skill inputs reach the model for non-templated (built-in) skills (lq-ai #115).
• Per-message chat file attachment (lq-ai #116/#117); profile edit via PATCH /users/me (#118).
• Navigable tabular cell citations (#125) and per-column ensemble verification (#127).
• Autonomous run findings persisted/readable (#135); document-grade artifacts (#138).
• arq-worker skill-registry initialization so skill_ref automations resolve on the worker (#139) — previously every worker-side skill_ref run failed at startup.

Engineering

• Quality bar held throughout: npm run check 0/0, npm run lint fully green, 1318 unit/component tests, plus live Playwright e2e per feature. Every feature shipped through a brainstorm → spec → plan → subagent-driven execution → review → PR loop; specs and plans for every phase are archived under docs/superpowers/.