v1.1.9 — Mercury Web, Kanban & Subconscious Memory

Release 1.1.9 introduces the Mercury Web Dashboard — a full React SPA at localhost:6174 — alongside a persistent Kanban board system, an in-browser Workspace IDE with Git and AI commit messages, a dual-layer (conscious + subconscious) Second Brain, intelligent skill batching, and a cross-platform build pipeline.

Note: Version 1.1.8 was published briefly and then unpublished before reaching wide distribution due to a dependency-tree issue (a transitive next dependency inflating install size by ~285 MB). 1.1.9 ships the same feature set with a leaner dependency graph — install footprint dropped from ~429 MB to ~87 MB (−80%).

Highlights

• Mercury Web Dashboard — React SPA with PWA install, login, dark/light theme, multi-thread chat with SSE streaming.
• Workspace IDE — file tree, multi-tab editor, integrated terminal, Source Control panel with AI-generated conventional commit messages, and a maximizable bottom panel.
• Kanban Boards (Beta) — persistent SQLite-backed task boards with Markdown descriptions, AI execution plans, cascade & smart execution, board context, attachments, comments, dependencies, and a question-status feedback loop.
• Second Brain — Conscious + Subconscious — dual-tier memory model, person tracking with aliases & relationships, goals view, and graph visualization.
• Skill Batching & Intent Router — automatic grouping for multi-skill invocations, plus a web UI to install skills from URL, activate/deactivate, and delete.
• Cross-platform build — a Node-based post-build.cjs replaces the old shell pipeline; runs identically on macOS, Linux, and Windows.

Mercury Web Dashboard

The dashboard is opt-in: enable it via mercury doctor or by setting web.enabled: true in mercury.yaml. It binds to 127.0.0.1:6174 only and is never exposed to the network.

• Authentication — username/password login with HTTP-only session cookies. Default credentials: mercury / Mercury@123. Editable from Settings → Account.
• PWA install — install Mercury from the browser address bar to launch it in its own window.
• Pages — Status, Chat, Tasks, Board, Workspace, Second Brain (Memory / Persons / Goals / Graph), Providers, Permissions, Skills, Schedules, Usage, Profit Sharing, Settings.
• Multi-thread Chat — every conversation lives at /chat/:threadId. Threads have a deletable sidebar, a per-thread model switcher, and a Code Mode toggle that flips the agent into coding context for that thread.
• Real-time updates — Server-Sent Events stream agent activity, tool calls, and Kanban status changes.
• Theme — light / dark / system toggle in Settings → Appearance.
• Settings cleanup — agent personality is no longer editable from the web UI; the new Defaults card holds default provider and daily token budget.

Workspace IDE

A full in-browser editor at /workspace.

• Folder picker — choose any directory on your machine; the choice is persisted across restarts.
• File Explorer — collapsible tree with refresh, resizable left panel.
• Multi-tab editor — syntax highlighting for 20+ languages; per-tab dirty state; Cmd/Ctrl+S to save.
• Bottom panel — split between Terminal (with command history) and a Chat thread scoped to the workspace. A new Maximize / Minimize button expands the panel to fill the workspace area.
• Source Control panel — staged / unstaged file list, per-file diff viewer, branch switcher with create-new-branch, last 15 commits, pull/push/refresh buttons (with spinning indicator and a "Refreshed" toast on manual refresh).
• AI Commit Message Generation — click the sparkle icon and Mercury auto-stages unstaged changes and drafts a conventional-commit message from the diff using the default provider.
• Resizable layout — Explorer, Bottom panel, and Source Control panel all resize independently, including correct inverted-direction resizing on the right edge.

Kanban Board System (Beta)

Persistent task boards backed by SQLite (with JSON fallback). Labelled "Kanban Boards (Beta mode)" in the UI.

• Real status set — pending, running, paused, done, failed, question. Drag-and-drop between columns updates status.
• Per-column scrolling — Trello-style: each column scrolls independently inside a horizontally-scrolling board.
• Markdown descriptions with "See more" truncation at 280 chars; click the description to open a full-screen modal, click the pencil to edit.
• AI Execution Plan — preview what the agent will do before running a card.
• Cascade execution — run a card and recursively all of its dependents in dependency order.
• Smart execution — board-wide walk that picks up pending cards whose dependencies are done.
• Token-budget aware — long-running cards auto-pause when their token budget is exhausted.
• question feedback loop — when the agent needs human input, the card moves to question and surfaces a feedback callout for you to respond.
• Board context — per-board working directory, custom variables, instructions, structure definition, and knowledge base, all injected into card prompts.
• Collaboration — comments, attachments (with download), labels, parent–child hierarchy, and blocks/blocked-by dependencies.
• SSE event stream — /api/boards/:id/events for live status updates.

Second Brain — Conscious + Subconscious

A two-tier memory model inspired by how human memory actually behaves.

• Conscious tier (durable, active scopes) — top-of-mind, always available for retrieval.
• Subconscious tier (subconscious scope) — out of active recall, no further confidence decay, recalled automatically when contextually relevant.
• After 30 days without being referenced, an active memory moves to the subconscious instead of being deleted.
• When a subconscious memory is recalled into a relevant conversation, it is promoted back to conscious.
• Hard deletion happens only on explicit dismissal — otherwise memories live forever in the subconscious.

Other Second Brain improvements:

• Persons — alias resolution, relationship mapping, and dedicated detail pages with last-seen and related memories.
• Goals & Projects — filtered views over type: goal and type: project memories with inline add.
• Graph view — force-directed visualization of memories, persons, and goals.
• Better extraction — persons are now extracted from identity, project, and episode memories (not just relationship); multi-type extraction in a single pass; dedup rules; resilient fallbacks for malformed LLM output.

The Telegram /memory menu gained a 💤 Subconscious button mirroring the web view.

Skill Batching & Intent Router

• Skill batching — related skills are automatically grouped for efficient multi-skill execution.
• Intent router — classifies natural-language input and dispatches to the matching skill or batch using intents, categories, and tags from each skill's frontmatter.
• Web UI — install skills from a SKILL.md URL, toggle active/inactive, and delete from /skills.
• Progressive disclosure — only frontmatter is held in context until a skill is invoked, keeping token usage low.

Sub-Agent & Background Improvements

• Friendly background task IDs (e.g. swift-fox) carried forward with stability fixes.
• Token-counter recovery: if dailyUsed becomes corrupted, Mercury rebuilds it from the request log on startup.
• Improved interrupt handling — in-flight work auto-backgrounds on interrupt and continues asynchronously.

TUI Improvements

• The TUI dashboard now displays web server status so you can see at a glance whether the dashboard is live and on which port.

Knowledge Graph UX

• Cleaner canvas — the always-on text labels were causing severe overlap and unreadability on dense graphs. Labels now appear only on the hovered / selected node, with a rounded background pill for legibility.
• Smarter tooltip — shows the full, untruncated memory summary (was previously trimmed to ~60 chars), wraps onto multiple lines, and flips to the left of the cursor when near the right edge so it's never clipped.
• Click-to-open detail dialog — single-click on any node now opens a Radix dialog that fetches the full memory record (GET /api/brain/memory/:id) and shows summary, detail body, scope, importance, confidence, evidence count, source, and timestamps. Drag / pan still work — a 4 px / 500 ms threshold distinguishes a click from a drag.
• Graceful fallback — nodes without a backing memory record (e.g. synthesized aggregates) still render their data inside the dialog.

Build System

• Cross-platform scripts/post-build.cjs replaces shell-only build steps — runs identically on macOS, Linux, and Windows. Calls npx vite build, copies ui/dist to dist/web/ui, and writes a kill-switch service worker.
• PWA cache kill-switch — older versions of Mercury Web could serve stale bundles from the PWA cache. The new build writes a self-unregistering service worker that clears caches and reloads clients on first load after upgrade, guaranteeing users see the new UI.
• Clean production build output with reduced artifact size and clearer logs.

Dependency Cleanup

Three runtime dependencies were declared in package.json but never imported by src/:

• geist — Vercel font wrapper. Fonts are vendored directly in src/web/static/vendor/geist/; the npm package transitively dragged in next and @next/* (~285 MB).
• js-tiktoken — no usages anywhere in source.
• react-dom — only used by ui/ (which has its own package.json); ink does not require it.

Removing all three reduces the end-user install footprint dramatically:

Metric	Before	After
node_modules size	~429 MB	~87 MB
Package count	114	95

No functional change — typecheck, tsup bundle, and Vite UI build all pass unchanged.

Notes

• The web dashboard is opt-in — Mercury still runs CLI-only by default; the web server starts only when web.enabled: true.
• All data (boards, memory, threads, skills) lives in ~/.mercury/. Nothing leaves your machine.
• The Kanban board is a beta feature in this release — the data model and API are stable, but expect UI refinements in subsequent versions.
• Web channel feedback is welcome via /feedback or GitHub issues.