Changelog

Generated from the LlamaNexus-WebUI repo's own CHANGELOG.md - see source for the definitive version. Run node scripts/sync-changelogs.mjs to refresh this page.

[0.1.7] - 2026-07-19

Added

• A real "who's logged in" affordance: the sidebar footer's plain Settings button is now a user menu - a round avatar (initials from your username, or your email if you haven't set one) plus your name, opening a small menu with Settings and Log out. Logged out, it shows a generic avatar and a Sign in entry that opens a centered login dialog (LoginModal) instead of requiring a trip through Settings first. The login form itself (LoginForm) is shared between this new dialog and the existing Settings → Account tab, so there's one place owning the actual login logic

Fixed

• Hugging Face model downloads got stuck on "Starting…" forever and never progressed, even though search still worked - routes/huggingface.ts's /pull relay listened on req.on('close', ...) to abort if the client disconnected, but in modern Node that event fires as soon as the (small) POST body finishes being read, independent of the response - aborting the relay to the LlamaNexus proxy almost immediately on every request, with the resulting error silently swallowed. Listens on res.on('close', ...) instead, which only fires on a genuine client disconnect. Verified against a real proxy: progress now streams through correctly, and a real mid-download client disconnect still aborts cleanly
• The login form's submit button rendered small and left-aligned under the full-width email/password fields above it, since it sat in a plain <form> with no width set, unlike the flex/stretch layout other dialogs' buttons get from .confirmation-buttons. Now spans the same width as the fields above it

[0.1.6] - 2026-07-18

Added

• A new Account settings tab: log in with a LlamaNexus-Docs account (the same one used on the docs site) directly from Settings. The backend proxies the login to LlamaNexus-Docs server-side and persists the resulting token in backend/data/account.json - never sent to the browser, the same discipline already applied to HF_TOKEN. Verified end-to-end that the token never appears in any browser-visible response (GET /api/settings, GET /api/account). New accountApiBaseUrl setting (Connection group) points at your LlamaNexus-Docs deployment. Nothing is gated behind being logged in yet - this only adds the capability, ahead of a coming subscription/license system

[0.1.5] - 2026-07-18

Added

• Add build-docker.sh <tag> to build and tag the Docker image as makershop/llamanexus-webui:<tag> (--push also pushes it; IMAGE_NAME env var overrides the repository name)

Changed

• docker-compose.yml now runs a pre-built image by name/tag (makershop/llamanexus-webui:${TAG:-latest}) instead of building it itself (build: .) - build one first with build-docker.sh, then TAG=<tag> docker compose up -d. Verified end-to-end: build-docker.sh produces a correctly-tagged image, and compose starts a container from it (not a fresh build) that boots and serves correctly

[0.1.4] - 2026-07-18

Added

• The active chat now persists across a page reload instead of dropping back to "no session selected" - the last-selected session is remembered (localStorage) and re-joined on load, as long as it still exists
• The active chat gets the same left-edge accent "sidemark" the selected server row already had, instead of just a background-color highlight

Fixed

• The sidebar's Settings button used the Unicode ⚙ character, which rendered as a thin, barely-visible dot in the sidebar's serif font - replaced with a real solid gear/cog SVG icon (ring, 8 teeth, a punched-out center hole), and its label is centered again like before

[0.1.3] - 2026-07-18

Added

• Add Dockerfile + docker-compose.yml (docker compose up -d --build) - a multi-stage build so only backend's production dependencies (not TypeScript/Vite/React/esbuild) end up in the runtime image. docker-entrypoint.sh seeds bindHost: 0.0.0.0 into the persisted settings on first boot only, since the app's own default (127.0.0.1) would otherwise be unreachable from outside the container with no way to fix it via the Settings UI (you could never reach that UI to begin with). Compose defaults to network_mode: host so LAN server auto-discovery (UDP broadcast) works, with a commented-out ports: mapping as the alternative if you don't need it. Verified end-to-end (build, first-boot seeding, port reachability, settings persistence across a restart, SELinux-safe bind mounts via :z)

[0.1.2] - 2026-07-18

Fixed

• Settings' model list context size no longer says "tokens" - just the numbers (4,096 (32,768))
• The rotating "Thinking/Chewing..." word's character-by-character reveal was silently broken: every character mounted at once and the whole word faded in together over 0.5s (a leftover visible class had no matching CSS rule), instead of a real staggered typewriter effect. Characters now only mount once their turn in the reveal comes up, so the stagger is genuinely visible

Changed

• Per-character reveal is a bit slower (90ms/char, was 60ms) and once a word is fully typed it now pulses gently until the next rotation, instead of just sitting static
• Word changes now crossfade instead of cutting instantly or leaving a gap: the outgoing word fades out on its own overlay while the next word starts revealing immediately underneath, rather than waiting for the old one to fully disappear first

[0.1.1] - 2026-07-18

Added

• The sidebar is now user-resizable (drag its right edge) with a 200-480px range, and remembers both its width and collapsed/expanded state across reloads (localStorage)
• Ported the VS Code extension's rotating "Thinking / Chewing / Pondering..." word animation (same word list, character-by-character reveal, rotates every 5s) - shown in place of the role title before the first token arrives, then continues below the reply for as long as the turn is still in flight, covering gaps like a pause before a tool call
• Assistant replies (streaming and completed) show the actual model name in the role title instead of a literal "Assistant" placeholder
• User messages: "You" now aligns with the bubble's own left edge and the timestamp with its right edge (previously both were clustered together at the right, disconnected from the bubble's actual width); the copy button moved to its own row below the bubble, aligned to its right edge
• Set a default system prompt (visible/editable in Settings → Model & Generation)
• Set default temperature (0.8), max tokens (4096), and context size (4096)
• Add the discoveryPort setting (Connection group) - the LAN discovery UDP port was previously hardcoded, unlike the VS Code extension's own llamanexus.discoveryPort
• Picking a model from the composer's model picker now also updates Settings' "Model" field (renamed from "Default model", mirroring the VS Code extension's own /model picker writing straight to its global setting); removed the "Use default" entry from the picker (picking always sets a real value now); added a "Search Hugging Face for models..." entry with a search icon at the bottom of the picker, opening the same Hugging Face search/download modal used from Settings
• Settings' model list context text reformatted to match the chat view's context-usage bar (4,096 tokens (32,768)) instead of a differently-worded 4096 / 32768 trained suffix
• Add a model-removal button (red ✕) to Settings' model list - permanently deletes the model from the LlamaNexus proxy (file(s) + its router.preset.ini entry); requires LlamaNexus 0.1.21+

Fixed

• Pressing Stop while a reply was streaming left the Send/Stop button stuck on "Stop" and the thinking animation running forever - the backend silently dropped the abort instead of settling the turn. It now saves whatever text had streamed so far (deliberately skipping tool-call synthesis - a call still being constructed when interrupted must not run) so the frontend actually hears back and resets

[0.1.0] - 2026-07-18

Added

• Sessions: create, rename, delete, and switch between chat sessions; each persists its full history, response stats, and reasoning/thinking text as a JSON file server-side
• Streaming chat over WebSocket, with markdown-ish rendering, fenced code blocks, and response stats (prompt/response/total tokens, elapsed time, tokens/sec)
• Tool calling: read_file, write_file, patch_file, list_files, run_bash, all confined to a configurable workspaceRoot, each gated behind a Settings toggle and an in-app confirmation dialog before touching disk
• update_plan tool with a collapsible Plan panel above the composer, and a per-session Plan/Edit mode toggle that blocks disk/bash tools outright while in Plan Mode
• run_agent tool to delegate a self-contained sub-task to a separately-configured agent server, with its own bounded tool-calling loop going through the same confirmation flow as the main model
• LAN server auto-discovery (UDP heartbeat) surfaced as a clickable server list in the sidebar
• Forced per-chat server selection: the composer stays disabled with a prompt until a server is explicitly picked for that session - applies uniformly to every session, old and new, matching the VS Code extension's own never-auto-select behavior. Session.baseUrlOverride is the field backing this
• The model that most recently generated a reply is tracked server-side (Settings.lastUsedModel) and persisted across backend restarts, then used to pre-select the model for every newly created session
• RPC worker management (single/RPC mode toggle, add/remove workers, live roster) for the configured proxy, shown inline in the sidebar's server list; only the row matching the configured proxy gets these admin controls
• Hugging Face model search and download, with multi-shard quant grouping, a "✓ Downloaded" badge, and cancelable in-progress downloads
• An "Eject" button on the currently-loaded model in the model picker to unload it from the server and free its memory
• A collapsible per-session model-tuning "Parameters" sidebar (Temperature, Top K, Top P, Min P, Frequency/Presence Penalty, Repeat Penalty, Repeat Last N, Mirostat + Tau/Eta, Seed, Max Tokens, Stop Sequences)
• Slash commands (/model, /ctx-size, /worker, /compact) with a live-filtered popup menu; the first match is highlighted while typing, arrow keys move the highlight, and Enter completes the highlighted suggestion into the composer instead of sending
• The model list and slash-command popup close on Escape or a click outside the menu
• Automatic context compaction once usage crosses a configurable threshold, plus a manual /compact command
• A full granular Settings panel (Connection, Model & Generation, Tool Calling, Chat Appearance), backed by GET/PUT /api/settings
• In-app confirmation dialogs (deleting a chat, ejecting a model, disruptive proxy-restarting actions, tool-call writes/patches/bash) instead of native browser confirm() popups
• Toast-style system notes for background events (mode switches, plan revisions, auto-compaction)

Fixed

• Chat deletion used a native window.confirm() popup instead of the app's own in-app dialog - now consistent with every other confirmation in the app