Release notes

1.2.0-beta.22 — 2026-07-31

Smarter turn stamps and better conversation memory. Semantic routing stamps graduate under dual-gate eval, continuity follow-ups hold corrections and “second option” threads, and past findings/collabs retrieve into relevant context without a new vector store. Install from Downloads or GitHub Releases.

Added

• Memory retrieval gates — corpus + live embed eval (make memory-retrieval-corpus / make memory-eval).
• Findings/collab backfill — review and collab markdown indexed for past-context retrieval.
• Continuity scenarios — correction and entity soft-followup coverage.

Changed

• Stamp policy — reason-code / ontology graduation; fewer brittle text demotions when live dual-gate holds.
• Chunk/inject — softer boundaries and fuller scored-chunk injection under budget.

1.2.0-beta.21 — 2026-07-30

Setup wizard completion and beta auto-update checks fixed on installed builds. First-run setup now persists under hub auth, and beta installers check only the git-backed updater channel. Install from Downloads or GitHub Releases.

Fixed

• Setup wizard — mints a hub session and sends X-NJ-Session when saving so setup_completed sticks (no more wizard every launch).
• Settings PUT auth — shared settings merge uses the same mutation session headers.
• Beta auto-update — drop the dead legacy updater-beta endpoint; check only git-backed updater/beta/ manifests.

1.2.0-beta.20 — 2026-07-30

Ollama one-click install that works on Windows, macOS, and Linux. Linux uses a Polkit password dialog (pkexec), macOS uses an administrator prompt, and Windows uses winget / silent setup with UAC when needed — so Install Ollama from the app no longer fails silently in the GUI. Install from Downloads, GitHub Releases, or read the beta.20 article.

Fixed

• Linux — pkexec elevation for the official install script.
• macOS — osascript admin dialog when Ollama is not bundled.
• Windows — UAC RunAs fallback after silent installer / winget.
• Detection — wait for binary including /usr/bin/ollama.

1.2.0-beta.19 — 2026-07-29

Install-and-go Ollama, working first-run wizard, and cleaner chat. Linux/Windows can one-click install Ollama from the app (OLL chip / model library); the setup wizard no longer skips itself because defaults seed a provider; plus routing/continuity hardening, model-library polish, web search settings, and SUT/release-prep tooling. Install from Downloads or GitHub Releases.

Fixed

• Setup wizard gate — setup_needed / setup_completed so first launch shows the wizard.
• One-click Ollama — Install from OLL chip or model library without re-running setup; auto-starts runtime.
• Chat continuity — turn routing, handoffs, trust/mode, and response-echo hardening.

Added

• Installed models library — transfers/installed view in the model library.
• Web search settings — dedicated settings tab.
• SUT loop + test portfolio — Human→SUT→Judge harness and climb/soak map.

1.2.0-beta.18 — 2026-07-27

Sharp Mermaid Canvas, Maps artifacts, and composition packaging. Canvas Mermaid creates stay workspace-grounded (no meta bleed, no npx mermaid), zoom stays crisp, Maps land as nj.map, and Share Agent / tool-grant / runbook export plus the Composition Model article land for portable composition. Install from Downloads or GitHub Releases.

Added

• Mermaid Canvas shortcut — tree-grounded creates, open-artifact revisions, canvas-only tool surface.
• Maps canvas — interactive OSM maps with markers and routes (nj.map).
• Composition Model — design note, campaign creative, and site article.
• Share Agent / user-tools / runbook export — package agents, grants, and runbook definitions.

Fixed

• Canvas routing — pure canvas asks skip implement sessions; image-gen suppressed on canvas turns.
• Mermaid syntax + zoom — safe path labels; SVG resize zoom stays sharp (no CSS scale blur).

1.2.0-beta.17 — 2026-07-26

Neural Canvas, quieter session files, and a user-flow test harness. Agents can ship revisioned Canvas artifacts with trusted renderers; last-session.json is debug-only by default (SQLite remains durable history); and real-world user-flow scenarios land under scenarios/user-flows/ with multi-turn journeys available but opted out of the default suite until they stabilize. Install from Downloads or GitHub Releases.

Added

• Neural Canvas — revisioned agent artifacts (Markdown, Mermaid, code, tables, charts, timelines, images, graphs) with chat cards, workspace tabs, and approved export.
• Pack artifact renderers — packs declare media/glob → host renderer maps without executing pack UI code.
• User-flow scenarios — product journeys + optional multi-turn journey-* flows (make user-flow-scenarios). WIP entries skip in --all.
• Context Stack article — how mode → intent → memory → grounding → persona → budget builds each turn.

Changed

• Session snapshot — periodic last-session.json write/restore off by default; dump on demand from Settings or API.
• Implement routing — turn-goal and semantic routing hardening for live implement/parity scenarios.

1.2.0-beta.16 — 2026-07-23

Assistant, CLI, auth opener, and IDE graph/explorer fixes. Meeting-notes questions stay conversational with synced notes, Gemini CLI no longer scans from /, Google Meet Connect opens the browser, and knowledge-graph payloads are capped so the file explorer stays responsive. Install from Downloads or GitHub Releases.

Fixed

• Meeting notes answers — personal Assistant Q&A injects synced notes instead of web-search / MCP detours.
• Gemini CLI workdir — packaged hub uses a safe cwd; benign CLI stderr is filtered.
• Google Meet Connect — opener capability + copyable auth URL fallback.
• Knowledge graph freezes — server and client enforce dense-view node caps for large repos.
• File explorer — errors keep the tree visible; folder expand path matching is normalized.

1.2.0-beta.12 — 2026-07-22

Packaged hub login hotfix. Agent startup no longer full-exports large SQLite message archives, which could wedge /api/health and block desktop sign-in after install. Install from Downloads or GitHub Releases.

Fixed

• Hub not reachable after install — history bootstrap uses a bounded message page instead of exporting entire channels from ~/.neural-junkie/messages.db.

1.2.0-beta.11 — 2026-07-22

Durable orchestration, semantic turn routing, and desktop polish. Restart-safe collab claims, semantic routing cutover, run_command approvals, and UI fixes for multi-select, toasts, terminal scroll, and send feedback. Install from GitHub Releases (prefer beta.12 if the hub hung on login).

Added

• Durable SQLite orchestration — restart-safe claims and HITL gates.
• Semantic turn routing — classifier-backed intent cutover with article and eval coverage.

Known issue (fixed in beta.12)

• Large message archives — packaged hub could hang during agent history bootstrap; upgrade to beta.12.

1.2.0-beta.10 — 2026-07-21

Automatic updates and safer agent workflows. Tauri v2 brings signed background downloads and clean-restart installation to macOS and Windows, while agent questions now pause peers, deduplicate prompts, and continue the original task after an answer. Install from Downloads or GitHub Releases.

Added

• Automatic signed updates — background checks and downloads on macOS and Windows with safe restart installation.
• Release policy controls — staged rollout, critical deadlines, minimum supported versions, and explicit platform availability.
• Restart safety — protects unsaved drafts and active streams, collaborations, jobs, and terminal foreground work.

Changed

• Agent questions — equivalent questions coalesce, recent answers are reused, and agents resume the original turn.
• Offline update behavior — cached metadata can warn offline, but only verified installable bytes can enforce an update.
• Linux updates — .deb installation remains manual while package-upgrade behavior is validated.

1.2.0-beta.7 — 2026-07-20

Beta.6 hotfix. Packaged-app soak fixes for hub disconnect races, DM rate limits, Knowledge Graph workspace grounding, pack chip icons, and IDE hide-chat unmounting the editor. Install from Downloads or GitHub Releases.

Fixed

• WebSocket reconnect race — no more fake hub disconnect while typing after channel reconnect.
• DM / channel create 429 — rate limiter respects config; dedicated open-DM endpoint; client DM open serialized.
• Pack toolbar chip icons — CSP-safe hub icons via blob URLs.
• Hub connection chip — toolbar connectivity status.
• Knowledge Graph Ask agents — workspace path attached so semantic_search works.
• IDE hide chat — editor stays open when the main chat panel is hidden.
• Context retrieve placeholders — schema examples rejected as real refs.

1.2.0-beta.6 — 2026-07-17

A memory of its own code. Native repository knowledge graph, Model Arena, MQTT/Kafka stream subscriptions, PrismML Bonsai 27B, Room Chat, Homebrew distribution, domain pack v2 plumbing, plus workspace image preview and pack toolbar chip fixes. Install from Downloads, GitHub Releases, or read the beta.6 article.

Added

• Repository knowledge graph — native code graph index, /api/repo/graph* APIs, code_graph agent route, and Knowledge Graph workbench. See feature guide.
• Model Arena pack — head-to-head model matches with reliable move handling and observable status.
• Inference usage telemetry — per-turn inference accounting.
• Stream subscriptions — MQTT/Kafka triggers for runbooks, channels, or webhooks. See stream subscriptions article.
• PrismML Bonsai 27B — model library entry with mmproj, prompted Ollama updater, and min-version gates.
• Room Chat pack — ephemeral LAN rooms on a host hub.
• Homebrew distribution — macOS cask and Linux formula with automated tap bump on release.
• Domain pack v2 plumbing — CAD, AWS, and software-development sidecar/platform work.
• Platform foundations — turn pipeline, span tracing, and durable collab state.

Changed

• Homepage messaging — any-model, local-or-cloud lead; marketing refresh and SEO.
• Collab deliverable scoping — tighter scoping for minimal-repo findings.

Fixed

• Workspace image previews — hub data URLs for editor images; wider Tauri asset scope.
• Pack toolbar chips — chips appear immediately on install/enable.
• Collaboration regression gates — chat and collab-core/full gate hardening.

Known issues

• Live parity gate — test-parity-stable-restart still failing; tracked for a follow-up beta (not part of release CI).

1.2.0-beta.5 — 2026-07-06

The release where the loops close. ReAct tools on Gemma, per-turn routing trace, Runbooks v2, multi-repo workspace scope, collab hardening, LoRA v2 specialists, personal learning, Slack diagnostics, and layer-gate / fix-loop / test-growth automation. Install from Downloads, GitHub Releases, or read the beta.5 article.

Added

• ReAct tool wrapper — MCP tools on non-native models (e.g. gemma3:12b) via tagged <tool_call> parsing with Qwen swap fallback. See ReAct tools article.
• Per-turn routing trace (MVP) — telemetry drawer, routing badges, and trace panel for model tier, retrieval mode, and governance on every turn.
• Runbooks v2 — persisted definitions, run history, connector profiles, pack templates, and library UI. See RUNBOOKS_V2.md.
• Multi-repo workspace scope — project sets, workspace scope chip, and cross-repo hints in agent context.
• Connection settings — hub URL, server/network, automation, and connectors tabs.
• LoRA v2 compound specialists — full import/review/bootstrap/compose lifecycle across chat, collab, IDE, and packs.
• Personal learning — explicit user-confirmed memory with optional LoRA export.
• Slack setup diagnostics — checklist, smoke/diagnose endpoints, clearer OAuth errors.
• Release engineering loops — make layer-gate, make layer-fix-loop, make test-growth-loop.

Changed

• Collaboration hardening — plan parser fix, faster planning recap fallback, discussion watchdog, turn handoff retries, longer approve-plan harness timeout.

Fixed

• Approve-plan test flake — async review-asset writes drained before test cleanup.
• Pack sidecar IDE typing — gopls-clean sidecar env types.

1.2.0-beta.4 — 2026-06-28

Agent Runtime reliability, image gen, music pack, turn telemetry. Install from Downloads or GitHub Releases.

Added

• Agent Runtime Reliability — Fix Loop command policy, boot-fix grounding, reliable-tier routing, implementation session outcome cards.
• Ollama image generation — hub store, agent tools, settings panel.
• Music creation v1.0.2 — ACE-Step variants and inference tuning.
• Turn telemetry drawer — per-turn routing trace, tool activity, slash-command history.
• Editor agent trust — expanded trust modes and boot-fix routing.
• Thinking activity labels — structured in-progress labels during tool loops.

Changed

• Capability profiles — refreshed benchmarks and reliable-tier routing metadata.
• Implement scenarios — stronger preflight checks and required outcome assertions.

1.2.0-beta.3 — 2026-06-24

Site navigation polish and CI reliability. Install from Downloads or GitHub Releases.

Added

• Unified site nav — canonical header and footer explore strip across all marketing pages via scripts/site_nav.py.

Fixed

• CI release gate — Node 24 in workflows; PendingChangesPanel tests use mock git change store.

1.2.0-beta.2 — 2026-06-24

Marketing refresh, pack expansion, IDE v4.1. Install from Downloads or GitHub Releases. See Start here, Security, and Life sciences guide.

Added

• Marketing site refresh — hive-mind narrative, start-here onboarding, security page, life-sciences feature guide.
• AWS, incident-management, and web-browser domain packs — seven official packs in the store.
• Pack capability sidecars — per-pack hub routes when packs are enabled.
• IDE v4.1 — remote LSP relay and remote collab worktrees. See IDE v4 guide.

Changed

• Moderator merged into Assistant — chat guidance and safety-net on Assistant.
• Hub auth hardening — mutations require real sessions or API keys.

Fixed

• Cursor parity infrastructure — release-prep reliability, collab turn routing, implement scenario regressions.
• Domain pack catalog — stale hub catalog no longer hides new packs.

1.2.0-beta.1 — 2026-06-17

IDE v4 — full editor depth, remote workspaces. Install from Downloads or GitHub Releases. Open-source agent IDE with full Monaco LSP on local workspaces, remote SSH via the nj-remote sidecar, dev container attach plans, and tree-sitter symbol search — on top of IDE v3 Ask/Agent routing and implementation sessions. See IDE v4 guide, IDE_V4.md, and REMOTE_WORKSPACES.md.

Added

• Full Monaco LSP (local) — persistent gopls, rust-analyzer, pyright-langserver; WebSocket didOpen/didChange/didClose; diagnostics, hover, completion, go-to-definition, references, rename.
• WorkspaceBackend — pluggable local + remote FS, git, search, symbols, file changes, and @codebase via internal/workspacebackend/.
• nj-remote sidecar — HTTP FS, one-shot exec, PTY WebSocket for SSH hosts and containers (cmd/nj-remote).
• Remote SSH workspaces — desktop Remote SSH wizard, token persist, sidecar health checks on connect and hub startup.
• Remote terminal — hub proxies PTY to sidecar; desktop routes terminal tabs and remote exec.
• Dev containers — GET /api/workspaces/devcontainer-plan parses .devcontainer/devcontainer.json.
• tree-sitter symbols — optional symbol index when tree-sitter CLI is on PATH.
• IDE v3.5 polish — LSP-lite squiggles fallback, workspace-wide Problems panel, yolo editor trust auto-approve.

Known limitations (v4.1)

• Remote LSP — gopls runs on sidecar host paths; full remote LSP relay deferred.
• Dev container wizard — attach plan API only; desktop wizard tab deferred.
• Collab worktrees on remote — not yet supported for remote workspaces.

1.0.0-beta.35 — 2026-06-10

Working beta auto-update. Install from Downloads or GitHub Releases. Git-backed updater manifests on main plus dual endpoint fallback restore in-app update checks. Includes beta.34 Slack OAuth relay and release CI fixes.

Added

• Git-backed beta updater — rolling manifests at updater/beta/; CI syncs after each beta release.
• Dual updater endpoints — legacy updater-beta release, then raw GitHub main URL fallback.

Fixed

• Beta in-app auto-update — immutable GitHub releases blocked rolling tag recreation; beta.35+ installers use the new channel.
• Release CI — learning test stability, Slack relay secret in builds, macOS ad-hoc signing default, manifest publish to main.

Note

• beta.34 installers — single updater-beta endpoint only; upgrade to beta.35+ for working updates.

1.0.0-beta.34 — 2026-06-10

Public Connect Slack. Install from Downloads or GitHub Releases. Any Slack workspace can OAuth-install the bundled NJ app via a free Cloudflare Workers HTTPS relay — no AWS account required. Desktop chat panel resize respects open side panels.

Added

• Slack OAuth relay — Cloudflare Worker at nj-slack-oauth-relay.neuraljunkie.workers.dev; make slack-oauth-relay-deploy-cf.
• Public distribution path — docs and CI for SLACK_VENDOR_OAUTH_RELAY_BASE.
• Chat max-width helper — main chat column resize accounts for visible IDE/collab panels.

Fixed

• Slack redirect_uri mismatch — hub upgrades loopback redirects to HTTPS relay for public distribution installs.

1.0.0-beta.33 — 2026-06-08

Cursor-like composer + durable memory. Install from Downloads or GitHub Releases. Ask / Agent / Export composer modes, SQLite-backed conversation memory for prior-reference continuity, tool-first file delivery, customer pack capability gating, and limitations-sprint hub/desktop polish.

Added

• Composer modes — Ask / Agent / Export on all channels with unified send pipeline and turn-intent routing.
• Durable memory — channel history bootstrap from SQLite; prior-reference export survives hub restart.
• Markdown edit/preview — inline toggle in the code editor for markdown files.
• Layout owner — Settings → Domain packs → pick which pack controls IDE vs team layout.
• Planning provider — optional provider for collab planning discussion turns.
• Channel export — export markdown history from channel info.
• Durable channels — skip 24h age prune per channel.
• Customer pack sideload — private packs gate scan summary, scan analysis viewer, and secondary analysis APIs.

Changed

• Implementation continuation — stronger go-ahead detection and prior-turn grounding.
• Workspace gate UX — banners, toasts, and Confirm workspace when execution waits for ack.
• /pause-agent — aborts in-flight LLM work for the paused agent.

Fixed

• Collab deliverable harness — plan stubs no longer pass file assertions; discussion fallback writes substantive content.
• Conversation regression — harness hardening for session-issue guards and collab races.

1.0.0-beta.32 — 2026-06-07

Install-and-go on Windows and Linux. Install from Downloads or GitHub Releases. Slim Windows and Linux installers — setup wizard Install Ollama on first launch (winget or silent installer on Windows; install.sh on Linux). macOS still bundles Ollama.

Added

• Wizard Ollama auto-install — live progress in setup wizard; hub POST /api/ollama/install with SSE.

Changed

• Windows slim installer — no bundled Ollama (~30 MiB vs ~1.4 GiB).
• Linux — wizard auto-install replaces manual Ollama install step.

1.0.0-beta.31 — 2026-06-07

Linux .deb ships. Install from Downloads or GitHub Releases. Linux builds omit bundled Ollama (GitHub 2 GiB asset limit) — install Ollama separately or use a cloud provider. macOS and Windows unchanged (bundled Ollama + in-app updates).

Changed

• Linux slim installer — .deb under 2 GiB; system ollama on PATH is detected by the hub.

Fixed

• Linux release upload — publish .deb even when AppImage bundling fails (continue-on-error).

1.0.0-beta.28 — 2026-06-07

Release CI fixes — macOS auto-update restored. Install from Downloads or GitHub Releases. macOS and Windows builds ship; Linux CI work continued in beta.31.

Fixed

• macOS updater uploads — Neural.Junkie_aarch64.app.tar.gz and Neural.Junkie_x64.app.tar.gz so both Mac targets get distinct signed updater bundles.
• Beta updater channel — recreate updater-beta release each publish (avoids immutable-release upload errors).
• Linux CI — skip RPM bundling; add libfuse2.

1.0.0-beta.27 — 2026-06-07

In-app auto-updates + downloads page. Install from Downloads or GitHub Releases. Tauri updater with separate beta and stable channels, update banner with one-click install, RAM-aware hardware guidance, and Windows WiX semver fix for full three-platform builds.

Added

• In-app auto-updates — Tauri updater with minisign signatures; beta builds track updater-beta, stable builds track latest.
• Update banner — periodic check on launch; Update Now / Later; auto-dismiss after 15s; Settings → About → Check for updates.
• Hardware guidance — RAM tiers and model footprint docs; /api/system/hardware; setup wizard and AI Providers hints.
• Downloads page — download.html with direct per-platform installer links from GitHub Releases.
• Updater CI — signed bundles, manifest generation, and scripts/verify-updater-manifest.sh; see RELEASE_UPDATES.md.

Fixed

• Windows beta builds — WiX-safe bundle version mapping (1.0.0-N for v1.0.0-beta.N tags) restores MSI/EXE CI for beta releases.

Note: Installers from before auto-update shipped need one manual upgrade to beta.27+; later betas update in-app.

1.0.0-beta.25 — 2026-06-07

Regression harness + collab plan reliability. Install from GitHub Releases. Pre-release make test-regression-bundle, collab goal bootstrap when agents emit prose-only plans, @Assistant ok thanks closure fix, and LoRA compose hardening.

Added

• Regression bundle — make test-regression-bundle (implement + chat + conversation).
• Parity-stable-restart — make test-parity-stable-restart with hub restart between sweeps.
• Collab goal bootstrap — extract executable tasks from /collaborate goals when planning yields zero parsed rows.

Fixed

• Thanks closure — @Assistant ok thanks gets canned closure in public channels.
• Collab plan-zero-tasks — plan-dependency, distinct-deliverables, and phoenix-combined scenarios pass in sprint2 sweep.
• LoRA compose — materialize adapter weights (no symlink insecure path warnings).

1.0.0-beta.24 — 2026-06-06

Phase 1 parity — local-first implement. Install from GitHub Releases. Server-side @codebase, IDE Agent auto-apply, inline completion, and multi-file theme sessions in one prompt. Plus a keyboard shortcuts registry, two-tier LoRA compose, and expanded chat/collab regression scenarios.

Added

• Phase 1 parity bundle — @codebase attachments, implementation session hardening, default auto_apply_edits, inline Tab completion.
• Multi-file implementation — theme tasks continue across tailwind.config.js and src/App.tsx in one turn; scenario react-theme-multi-file.
• Stability harness — make test-parity-stable archives logs under docs/testing/.
• Keyboard shortcuts registry — centralized dispatcher; KEYBOARD_SHORTCUTS.md.
• Two-tier LoRA compose — bootstrap verification for specialist-tuning pack tags.
• Harness expansion — workspace DM scenarios per specialist, collab conversation-quality regressions.

Changed

• Local-first routing — no automatic cloud escalation on model failure; use @Cursor explicitly when you want cloud-grade work.
• Specialist-tuning pack — updated LoRA library metadata for composed tags.

1.0.0-beta.23 — 2026-06-06

CLI agents and smoother implementation. Install from GitHub Releases. Wire Codex, Cursor, Gemini, and Copilot CLI agents from the desktop; see pending tool approvals in a hub-wide bar. Implementation sessions skip already-applied files and route code-review asks separately from file-edit loops.

Added

• CLI agents manager — desktop UI + hub handlers for Codex, Cursor, Gemini, and Copilot CLI agents.
• Pending tool approvals bar — hub-wide approval queue synced to the desktop composer.
• Implementation fallback — deterministic tailwind/theme repairs when models emit prose-only implementation replies.
• Code review intent — dedicated routing so review asks do not open the file-edit implementation loop.
• Shared TypeScript/ESLint MCP helpers — tsc / eslint verification for frontend implementation sessions.
• Test artifact cleanup — scripts/cleanup-test-artifacts.py and make cleanup-test-artifacts; auto-run after make test-go.

Changed

• Implementation continuation — skip re-proposing already-applied files; block vague “pick up where you left off” without prior thread context.
• File-change approvals — agent-visible approval messages with metadata; auto-continue prompts after UI apply.
• Ollama dev/runtime — bundled-runtime chip, ensure-ollama.sh, and sidecar PATH resolution for Tauri dev.
• Test isolation — package-level temp HOME so repo-agent tests no longer pollute ~/.neural-junkie/repos.

Fixed

• Compact assistant prompt test — user-rules lookup uses a user-like sender type.
• ChatWindow Vitest — mock PhoenixBrowserModal, pending approvals, and related stores.
• Hub history clear — persistent store cleanup for clear-history API.

1.0.0-beta.22 — 2026-06-05

Install-and-go local AI. Install from GitHub Releases. Ollama runtime is bundled on macOS, Windows, and Linux; the app auto-starts ollama serve on launch. First run still pulls a default model once (internet required). Plus Pack Dev Studio fixes and an editable CSV table in the code editor.

Added

• Bundled Ollama — runtime shipped in installers; models stored in app data. See DOWNLOAD.md.
• CSV table editor — editable grid for exported .csv files (Table/Text toggle in the code editor).
• Pack Dev Studio fixes — live YAML validation, dev-link/reload edge cases, multi-pack test selector. See PACKS_CUSTOM.md.

Changed

• Setup wizard — reflects bundled Ollama on production installers; Windows no longer needs a separate Ollama install.

1.0.0-beta.21 — 2026-06-02

Slack inbox, context model v2, and chat regressions. Install from GitHub Releases. Personal Slack inbox and forwarding, conversation mode + workspace visibility fixes, and expanded live scenario harnesses for chat and collab.

Added

• Slack personal inbox & forwarding — DM the NJ bot; optional channel rules and away-mode user-DM replies. See Slack connect guide and SLACK_INTEGRATION.md.
• Context model v2 — Chat/Code composer mode, turn intent, thread history, context budget. See CONTEXT_MODEL.md.
• Chat scenarios — make chat-scenarios-regression for workspace, echo, and closure quality.
• User rules API — per-user markdown rules via /api/user-rules.
• Test docs — TESTING.md (isolated Go tests, scenario runners).

Fixed

• Workspace visibility — deterministic answers when asked “can you see my workspace?” (no fake package advice).
• Collab assignee order — round-robin respects agent mention order.

1.0.0-beta.20 — 2026-05-30

Personal learning v2. Install from GitHub Releases. Multi-scope user-confirmed memory with Ollama embedding retrieval, export/import, collaboration-scoped notes, and optional agent-suggested proposals under the Specialist tuning pack.

Added

• Personal learning v2 — scopes (agent, global, collaboration), top-k prompt injection, edit API, JSON bundle export/import. See PERSONAL_LEARNING_V2.md.
• Desktop UX — scope toggle in approval modal; grouped Settings lists; edit in agent info; LoRA train checkbox to include learnings.
• Test harness — v2 learning scenarios and extended CI smoke.

Fixed

• Collab @mention during planning — ignore @mentions embedded in another agent's plan prose.

1.0.0-beta.19 — 2026-05-29

Pack store and collaboration polish. Install from GitHub Releases. Official domain packs install from GitHub with offline fallback; the collaboration panel surfaces planning wait, dispatch, and file-deliverable states more clearly.

Added

• Domain pack store — Settings → Domain packs → Pack store; zip install with embedded fallback. See PACKS.md.
• Loose [FILE_CHANGE] parsing — inline path lines in task replies register hub file proposals.
• Collaboration UX — planning wait banner, approved/dispatch banner, file-deliverable hints, workspace gate copy.
• Horizontal panel resize — drag handles for desktop layout panels.

Changed

• Multiple domain packs — enable more than one pack; first enabled pack owns UI layout.
• Planning without bound repo — no editor-tree scan or “Grounding: I loaded N files” during planning/review.
• Scenario harness — HTTP 429 retries, multi-collab isolation fixes, improved error matching.

Fixed

• Multi-collab isolation — executing blocker collabs are not auto-cancelled when other scenarios start.

1.0.0-beta.18 — 2026-05-27

Collaboration execution you can trust. Install from GitHub Releases. This release fixes multi-collab agent routing, binds deliverables to the real project repo, parses structured task lists correctly, and requires concrete task output during execution.

Added

• Collab smoke (live) — make collab-smoke LIVE=1 validates planning → review → execute against a running hub with real agents.
• Path warnings at approve — missing task context paths are surfaced before execution starts.

Changed

• Task prompts — execution assignees must ship [FILE_CHANGE] deliverables and end with TASK_STATUS: completed or blocked.
• /collaborate --workspace — workspace outline is attached only when the flag is set.
• Collaboration panel — live phase updates while the panel is open.

Fixed

• Multi-collab routing — planning handoffs work while another collaboration is executing.
• Wrong repo root — collabs/<uuid>/ folders are rejected as the source workspace; open the git repo root instead.
• Task list parsing — markdown ## Tasks numbered bold items become proper tasks; milestone bullets and fragments are skipped.
• Run button — bare filenames in bash fences (e.g. findings.md) are not executed as shell commands.

1.0.0-beta.17 — 2026-05-26

Workspace switching, Assistant context, Connect Google, and macOS signing. Install from GitHub Releases. This release replaces beta.16 macOS artifacts with ad-hoc signed bundles verified in CI, and it ships the newest workspace and Assistant upgrades.

Added

• Workspace quick switcher — tab bar plus searchable modal; filter workspaces by name, path, or branch and switch with keyboard controls.
• Assistant workspace grounding — scoped workspace metadata includes open-file focus, selected text, file tree hints, and scan-summary context when workspace sharing is enabled.
• Google Meet notes — release builds can include bundled Google OAuth credentials so Connect Google no longer requires pasted client credentials for the standard path.

Changed

• Assistant meeting context — meeting-note startup is quieter and follow-up answers can use synced meeting notes more directly.

Fixed

• macOS installers — Tauri bundles are ad-hoc signed and verified before upload, replacing the broken beta.16 macOS downloads.

1.0.0-beta.16 — 2026-05-22

Domain packs and IDE v2/v3. Install from GitHub Releases. Four official packs in the store; multiple packs can be enabled (first enabled owns layout). Dev pack adds IDE v2/v3. See marketing: Domain packs, Domain packs guide, IDE_V3.md.

Added

• Domain pack exclusivity — hub + Settings; migration when both packs were on.
• IDE v1/v2 — Git modal, quick open, selection context, stage/unstage, diff viewer, ⌘⇧O symbols, diagnostics, fast edit.
• Marketing — home page Domain packs section.

1.0.0-beta.15 — 2026-05-22

Connect Slack on public installers. Release CI embeds the NJ Slack app (-tags slackvendor). Install from GitHub Releases. Prefer beta.14 only if you already downloaded it — use beta.15 for one-click Slack Connect.

Fixed

• Release builds — GitHub Actions secrets SLACK_VENDOR_* + scripts/ci-write-slack-vendor-oauth.sh so downloaded apps enable Connect Slack without Advanced token paste.

1.0.0-beta.14 — 2026-05-22

Slack Connect and marketing gallery. Install from GitHub Releases. Slack: SLACK_INTEGRATION.md. Gallery: gallery/index.html.

Added

• Slack Connect — bundled OAuth app, loopback install, Simple/Advanced Settings, /api/slack/connection, channel list API.
• Slack threading — two-way sync with thread parent map; Slack user display names on inbound.
• Site gallery — docs/gallery/ + make gallery-sync for ads and screenshots.
• Slack ads — dev, non-dev, and models marketing PNGs.

Changed

• Makefile — auto slackvendor tag when maintainer vendor/oauth.json exists.

1.0.0-beta.13 — 2026-05-21

CLI agents, local-first security, scan summaries, and Slack. Install from GitHub Releases. CLI: CLI_AGENTS.md. Security: SECURITY.md. Slack: SLACK_INTEGRATION.md.

Added

• 12 CLI agents — auto-detect on PATH (Copilot, Codex, Cursor, Claude, Gemini, Aider, OpenCode, Amazon Q, Crush, Amp, Droid, Kiro).
• Security — loopback bind, sessions, rate limits, encrypted config and Tauri credentials.
• Scan summary viewer — desktop UI for plate summaries and well TIFF preview.
• Slack integration — OAuth, bindings, hub API.
• Delegation — agent consult/delegate flow.
• Marketing — beta.13 ads and docs/marketing/BETA13-ADS.md.

Fixed

• Desktop hub API — hubFetch recursion bug blocked login and all hub calls.
• Tests & UX — Vitest coverage, scan-summary tests, quieter dev logs.

1.0.0-beta.12 — 2026-05-20

Domain packs, session memory, and runbook actions. Install from GitHub Releases. Packs: Software development, Life sciences. Context: CONTEXT_MODEL.md. Actions: RUNBOOK_ACTIONS.md.

Added

• Domain packs — optional software-development and life-sciences packs in Settings; wizard focus tracks; lean default install.
• Life sciences — BiologyExpert, OpenBioLLM, bio MCP tools, sequence review runbook template.
• Context model — turn intent router and rolling DM session summary (qwen2.5:7b).
• Runbook actions — http_get, http_post, webhook, conditional edges; bundled templates API.
• Collab recap — summary message when collaborations complete.
• Marketing — beta.12 ads and docs/marketing/BETA12-ADS.md.

Changed

• Specialists + MCP — follow pack toggles; migration for legacy dev agent configs.
• Sidebar — auto-unhide channels and stable agent hide keys.

1.0.0-beta.11 — 2026-05-19

Runbook builder and collaboration polish. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md. Runbooks: docs/COLLABORATION.md.

Added

• Runbook builder — desktop RB / /runbook; task list + Graph (xyflow): nodes, dependency edges, inspector, auto-layout; markdown import.
• Runbook API — POST /api/runbooks; DAG validation; suggest-assign; hub lifecycle for source: runbook collaborations.
• Completion UX — channel banner when a collaboration completes; read-only closed channel.
• Non-dev marketing — audience ad PNGs and docs/marketing/NONDEV-ADS.md.

1.0.0-beta.10 — 2026-05-19

Full-platform installers (CI fix). Install from GitHub Releases.

Fixed

• Release CI — all six installers (macOS arm/x64, Windows msi/exe, Linux AppImage/deb); tauri.conf.json package version 1.0.0 for WiX/MSI; Vitest files excluded from tsc.

1.0.0-beta.9 — 2026-05-18

Collaboration git worktree execution. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md. Worktree details: docs/COLLABORATION.md.

Added

• Git worktree mode — /collaborate --worktree executes approved plans in <assets-root>/worktrees/<collab-id>/ on branch nj/collab-…; optional --workspace binds the source repo at start; desktop Continue can supply source_repo_path from the active git workspace.
• Worktree gate UI — collaboration channel shows source repo, branch, and worktree path before task dispatch.

Fixed

• Release CI — desktop TypeScript build (workspaceFileDrag.test.ts) blocked beta.8 installers; beta.9 restores macOS / Windows / Linux artifacts.

1.0.0-beta.8 — 2026-05-18

Model library, Hugging Face, and MCP. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md.

Added

• App-store model library — toolbar entry (⌘⇧M, /nj-open-model-library): grid + detail for Ollama and Hugging Face; primary action per tile; full actions on detail.
• Hugging Face — huggingface provider, curated catalog, hosted inference and local GGUF download with SSE progress, import to Ollama (HF_TOKEN or per-provider key).
• MCP tool servers — Backend, DevOps, Database agents with ENABLE_MCP=true; in-app tool-use loop; export API and resource server when enabled.

Changed

• Model install/download UI moved from Settings to the toolbar model library; Settings keeps provider registry and Ollama endpoint.
• Consolidated internal/mcp (removed mcp_disabled duplicates).

1.0.0-beta.7 — 2026-05-18

Markdown and chat polish release. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md.

Added

• Unified GFM markdown — chat uses the same marked pipeline as file preview and collaboration plans; headings, lists, and tables render in the timeline. RichMarkdownView shared component; Prism still highlights fenced code.
• Assistant workspace review — prompt guidance when users ask to review editor content.

Fixed

• Chat scroll — pin-to-bottom during streaming; Virtuoso scroller ref and footer spacer.
• Session restore — collab discussion synced into channel timelines; channel message dedupe on persist.
• Assistant — meeting-note batch notification on startup; filtered history; meeting notes dir resolved on load.
• Collaboration — max-concurrent error lists active collaborations.

1.0.0-beta.6 — 2026-05-17

Context routing and stability release. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md.

Added

• Selective workspace context — context_scope metadata (none · hint · outline · focus · full); desktop Auto mode infers scope per message; composer scope chip; /collaborate --workspace for outline-only project tree during planning.
• Execution limits — max agent messages per collab during executing; rate limit on collaboration_task replies.
• Debug tooling — make debug-session, make debug-collab, make collab-smoke (optional LIVE=1).

Fixed

• Collaboration — stop re-dispatching all tasks on every channel message; seed messages are internal (agents no longer reply to “Collaboration Started”); collab replies use collaboration_discussion; subscriber buffer 512.
• Session — slim collab WebSocket payloads; strip metadata from last-session.json; auto-archive oversized session files; stricter disk caps.
• Hub data — consent modal before workspace reads; join dedupe; editor and chat image preview.

1.0.0-beta.5 — 2026-05-17

Reliability and reasoning release. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md.

Fixed

• Collaboration — DM-spawned agents subscribe to the collab channel after AddAgentToChannel; join/subscribe and seed/turn failures return a system message (fail closed).
• Cancel / collab UI — cancel targets the active collaboration channel; clears task drawer so sidebar and composer do not stay locked.
• Thread streaming — stream_delta / stream_end route to thread subscribers; main chat no longer shows thread-only streams.
• Chat send — composer clears typing on error; send failures surface in the UI.
• Ollama / DeepSeek — correct chat roles; think API for reasoning models; collapsible Reasoning blocks; fewer duplicate DM replies and history echo.
• Mermaid — sharp SVG fit and macOS zoom smear fix (shared canvas with Dickory Docs).
• Hub broadcast — log when subscriber buffer is full. Agent retries — clear respondedMessages on generation failure.
• Integrations — GitHub/Confluence connection tests report format-only checks honestly. /revise-plan posts to the bound collab channel.

Added

• Editor — image preview for supported files.
• Tests — Ollama thinking/roles, chat stream reasoning, hub collab subscribe.

1.0.0-beta.4 — 2026-05-16

Fixed

• Release CI — draft-then-publish workflow; per-platform gh release upload; app bundle version 1.0.0 for Windows MSI.

1.0.0-beta.1 — 2026-05-16

First public beta download. Install from GitHub Releases. Quick start: docs/DOWNLOAD.md.

Install

macOS: Unsigned — right-click the app → Open if Gatekeeper warns.
Windows: Install Ollama manually or use a cloud API key in the wizard (in-app Ollama install is macOS/Linux only).

Highlights

• Downloadable desktop app + bundled Go hub (no separate Go install).
• Includes milestones 0.1.2–0.1.4: Ollama model library, collaboration smart routing, collaboration sandbox, marketing site, port 18765, and desktop polish.

0.1.4 — 2026-05-14

Added

• Ollama model library — curated catalog embedded in the hub (GET /api/ollama/catalog); desktop Settings → AI Providers UI to browse, search, install with SSE progress, delete installed models (POST /api/ollama/delete), and Use for agents to set the Ollama provider model and agent assignments.
• Collaboration smart routing (optional) — collaboration.smart_routing_enabled in config.json / settings API; when enabled, collaboration execution tasks can be generated using a routed provider from your configured pool (static heuristic: vision, keyword tiers, fallback). Regular channel messages still use each agent’s configured provider. In-process hub agents only in this release.
• AI provider cache — shared construction of AIProvider instances from config with invalidation on provider CRUD and AI settings changes.

Changed

• Developer builds — Makefile runs go build / go run on the ./cmd/server package (not only main.go) so additional server sources compile.

Documentation

• docs/COLLABORATION.md and docs/USER_VALUE_GUIDE.md — model library and smart routing behavior.

0.1.3 — 2026-05-14

Added

• Collaboration execution sandbox — when a plan is approved, the hub creates ~/.neural-junkie/collaborations/<collaboration-id>/ and exposes working_directory on snapshots.
• Workspace confirmation gate — agents do not receive collaboration_task prompts until you confirm: desktop Continue on the collaboration channel, POST /api/collaboration-workspace-ack, or /ack-collab-workspace <id>.
• Command suggestion working directory — bash blocks suggested by agents can execute with the collaboration sandbox as cwd when you use Run in the desktop app.

Changed

• Collaboration execution — task dispatch, resume-plan while executing, snapshot heal paths, and attachCollaborationData honor workspace_acknowledged so work does not race ahead of workspace setup.
• Agent system prompts — executing phase spells out [FILE_CHANGE], workspace context fallback, sandbox path, and shell blocks for runnable commands.

Fixed

• Collaboration executing agents now receive the same machine-readable [FILE_CHANGE] specification as normal channels so file proposals register reliably.

0.1.2 — 2026-05-13

Added

• Marketing site — GitHub Pages content under docs/: expanded landing, feature deep-dives, release notes page, early-access banner.
• Per-channel typing indicators in the desktop channel sidebar.

Changed

• Default hub port is 18765 (previously 8080); make start-all health checks and process management align with SERVER_PORT.
• Slash commands — real execution with parity enforcement against the hub; command palette metadata refreshes on demand.
• Collaboration — workflow hardening across server, desktop UI, and tests; runtime reliability updates; collaboration round counter clamps at configured maximum.

Fixed

• Drop empty messages from ingestion paths including history reload.
• Hub channel ordering stability; Ollama version surface; auto-register CLI providers when applicable.
• CLI and agent chat rendering when markdown code fences are malformed.
• Desktop — migrate saved hub URLs from legacy localhost:8080 to 18765.

Improved

• Hub HTTP/WebSocket surface: security and robustness hardening.
• Desktop UX — dark-theme toasts, accessible toolbar controls, loading and login polish.
• Developer settings — remove non-functional test mode control.

Removed

• In-hub /app screenshot gallery and live-gallery docs (replaced by the static docs/ site and README assets).

0.1.1 — 2026-02-23

Added — Multi-agent collaboration

• Collaboration manager (internal/collaboration) for structured multi-agent orchestration
• Bounded discussion sessions with round limits, turn budgets, total message caps, and timeout enforcement
• Collaboration phases: planning → reviewing → approved → executing → completed/cancelled
• Shared plan artifacts with version history and edit tracking
• Task delegation model with per-agent assignment and status tracking
• Consensus detection (signal + heuristic) with disagreement escalation
• New slash commands: /collaborate, /approve-plan, /revise-plan, /cancel-plan, /collab-status
• New message types: collaboration_plan, collaboration_task, collaboration_status, collaboration_discussion

Added — Desktop collaboration UX

• CollaborationPanel for phase, participants, tasks, plan artifact, and control actions
• Collaboration message rendering in chat with collaboration-specific visual cues
• TypeScript protocol updates for collaboration entities and metadata helpers

Added — Test coverage

• test/collaboration_test.go covering lifecycle, bounded discussion logic, consensus, task tracking, artifact versioning, and extraction parsing

0.1.0 — 2026-02-20

First packaged release — Neural Junkie ships as a single distributable desktop app.

Added — Desktop packaging

• Tauri sidecar architecture — Go server bundled inside the Tauri app, launched and managed automatically
• First-run Setup Wizard — guided onboarding to choose AI backend (Ollama or cloud), configure providers, and enable agents
• Auto-update system — in-app update banner with download progress and one-click restart via Tauri updater
• Loading screen — server health polling with status feedback during startup

Added — AI provider registry

• Dynamic provider management — add, edit, remove, and test AI providers from Settings UI
• OpenAI-compatible provider — generic adapter for any OpenAI-compatible API (Amazon Q, Azure OpenAI, Together AI, Groq, etc.)
• Provider Manager UI — full CRUD interface with connection testing
• Multi-provider support — use multiple cloud and local providers simultaneously, assign per-agent

Added — Ollama lifecycle management

• Automatic detection — detect Ollama installation on macOS and Linux
• Install from app — install Ollama directly from the Setup Wizard or Settings
• Server management — start/stop Ollama server from the UI
• Model pulling — pull models with real-time progress streaming (SSE)
• Ollama Manager UI — dedicated panel in Settings for full Ollama control

Added — Configuration system

• JSON config file — persistent configuration at ~/.neural-junkie/config.json
• Environment variable migration — auto-migrates from env.local to config file on first load
• API key redaction — API keys masked in GET responses, preserved on PUT if masked
• Per-agent provider assignment — each agent type can use a different provider

Added — CI/CD & release

• GitHub Actions release workflow — triggered on v* tags, builds macOS (arm64 + x86_64) and Linux (x86_64)
• Cross-compilation matrix — Go server compiled for each target, bundled as Tauri sidecar
• Update manifest generation — auto-generates platform-specific JSON manifests for Tauri auto-updater
• make release target — bumps versions, commits, and tags in one command

Added — CLI agent infrastructure

• CLI agent registry — persistent storage for CLI agent configurations
• CLI agent storage — JSON-based persistence for registered CLI agents

Improved — UI

• Terminal panel — refactored with XTerminal component
• Markdown rendering — improved code block handling and mermaid diagram support
• Suggestion banner — contextual suggestions in the chat UI
• Chat window — enhanced layout and interaction patterns

Pre-0.1: rebrand & Tauri workspace — February 2026

This work predates the first tagged app release (v0.1.0). An older changelog used 2.0.0 as a product-generation label; it was never a GitHub release version and is not comparable as “> 0.1.x”.

Renamed

• Project renamed from "AI Chat Room" to Neural Junkie
• Go module: github.com/camronwood/neural-junkie
• Data directory: ~/.neural-junkie/
• Tauri bundle: com.camronwood.neuraljunkie

Added — Desktop app

• Tauri + React desktop app replacing the old Fyne GUI
• Slack-inspired UI with dark theme and modern styling
• Command Palette — searchable slash-command UI with guided argument forms (/ trigger or toolbar button)
• File Explorer Panel — browse and open workspace files
• Code Editor Panel — view and edit code from the app
• Terminal Panel — embedded terminal output
• Thread Panel — threaded conversation view
• Pending Changes Panel — review file change proposals with diff preview
• Settings Modal — Appearance, Layout, Integrations (Anthropic, GitHub, Confluence), AI Providers (Ollama, LM Studio, Claude), Developer, About
• @Mention Autocomplete — agent picker with fuzzy matching
• Mermaid Diagram Rendering — inline diagram support in messages
• Layout Persistence — panel visibility saved across sessions

Added — Agents

• Moderator Agent — auto-starts with server, guides users through commands and features, 20s safety-net for unanswered questions
• Assistant Agent — reminders (one-time/recurring), tasks (priority, due dates), notes (tags, search), meeting summaries, scheduling; persistent storage
• Confluence Agent — index Confluence Cloud spaces, search documentation, answer knowledge-base questions
• Helper Agents — template-based custom experts (day-one onboarding, testing, docs)
• Cursor CLI Agent — Cursor CLI subprocess for code analysis and generation
• Agent Review — get second opinions by @mentioning another agent in a thread reply

Added — AI providers

• Ollama — local inference with model listing, connection testing, configurable endpoint
• LM Studio — local OpenAI-compatible server with model listing and connection testing
• Per-agent provider switching — change provider/model for individual agents at runtime
• Global provider switching — switch all agents to a provider with one command
• Two-tier model config: code tier (qwen2.5-coder:14b) and utility tier (qwen2.5:7b)

Added — Features

• 50+ slash commands with metadata for command palette
• File Change System — agents propose file edits, users approve/reject with diff preview and backup
• MCP Export/Import — export agent knowledge to MCP format, import from MCP, MCP resource server
• Workspace management — add, list, remove workspaces
• Thread support — create threads, reply in threads, thread metadata and subscriptions
• Session persistence — periodic save and recovery
• Connection testing — test Anthropic, Ollama, LM Studio, GitHub, Confluence connections from UI
• Design analysis — /analyze-design command for UI review

Improved

• Three-layer message deduplication (polling, handler, agent-type filtering)
• Repository agent caching with staleness detection and incremental reindex
• File watching for auto-reindex on codebase changes
• Agent pause/unpause and remove/recall lifecycle management

1.0.0 — 2025-10-14 · AI Chat Room (legacy)

First public-era hub under the original product name AI Chat Room (Fyne desktop, since replaced).

Added

• Core hub server with WebSocket real-time communication
• Multi-channel conversation support
• 5 specialized agent types: Frontend, Backend, DevOps, Database, Security
• Repository Expert Agents with codebase indexing and search
• Claude AI integration (Anthropic API and AI Hub)
• Mock AI provider for testing
• Fyne-based desktop GUI (since replaced by Tauri + React)
• Interactive terminal chat client
• Built-in web UI
• CLI tool for automation
• @mention system for targeting agents
• Message history and context

Fixed

• Message deduplication (agents responding multiple times)
• GUI threading issues (Fyne thread safety)
• Username display (was showing "Human User")