Skip to content

ROADMAP.md

Latest commit

 

History

History
191 lines (147 loc) · 10.3 KB

ROADMAP.md

File metadata and controls

191 lines (147 loc) · 10.3 KB

OpenMind — Roadmap

Items are grouped by area. Check off as completed.

Configuration & Settings

  • Settings panel / modal — edit Ollama URL and port without touching source
  • Persist selected model to localStorage so it survives restarts
  • Per-conversation model override (different model per chat session)
  • Configurable system prompt (global default + per-conversation override)
  • Configurable temperature, top-p, num_ctx, and other Ollama sampling params
  • Configurable server poll interval and request timeouts in settings

Chat UX

  • Message editing — click to edit a sent user message and re-run from that point
  • Response regeneration — re-run the last assistant turn with a new seed
  • Multi-turn branching — fork conversation from any message
  • Stop mid-stream preserves partial response (already works) — add visual "stopped" badge
  • Streaming token counter in header (tokens/s)
  • Context window usage indicator (current tokens used / model max)
  • Auto-resize textarea already works; add explicit expand/collapse toggle for long inputs
  • "Thinking" / reasoning block collapse — <think> tags rendered as collapsible block with preview
  • File attachment — send text files / images as part of the message (multimodal models)
  • Drag-and-drop file into chat input area

Conversation Management

  • Multiple named conversations (conversation list sidebar)
  • Search / filter across conversation history
  • Export conversation as Markdown, plain text, or JSON
  • Import conversation from JSON
  • Pin / favourite conversations
  • Auto-title conversations from first user message via a short model call
  • Delete individual messages from history

Model Management

  • In-app model browser — list models available on Ollama with size and family info
  • Pull new model from registry directly from the UI (POST /api/pull with streaming progress)
  • Delete model from UI (DELETE /api/delete)
  • Show which model is currently loaded in GPU memory (/api/ps)
  • Unload / eject a running model from VRAM (POST /api/generate with keep_alive: 0)
  • Modelfile viewer (show system prompt and parameters for a model)

Diagnostics & Observability

  • Persist diagnostics log across tab switches (module-level store survives remount)
  • Export diagnostics log to file
  • Latency sparkline / history graph for server probe results
  • Token throughput chart during streaming
  • Disk usage display per model — size shown in ModelPicker and Diagnostics model list
  • GPU memory usage — VRAM badge from /api/ps shown in Diagnostics model list
  • Alert / notification when server goes offline while app is open
  • curlllama.sh output surfaced in the Diagnostics tab (run script via Tauri shell plugin)

Boardroom Mode — Multi-Agent Consensus

Boardroom brings multiple models into a single structured session where each model responds independently to the same prompt, then a synthesis step derives a consensus answer. See BOARDROOM.md.

  • Boardroom tab — distinct from single-model Chat; tab in the header nav
  • Send one prompt to N selected models simultaneously; all streams run in parallel
  • Each advisor gets its own panel showing its live stream
  • Consensus engine — after all advisors finish, a synthesizer model receives all responses and produces a structured verdict
  • Consensus panel — displays synthesized output with streaming cursor; Re-synthesize button
  • Agent roles — advocate, critic, analyst, devil's advocate, expert, generalist injected as system prompts
  • Re-run consensus with a different synthesizer without re-querying advisors (↻ Re-synthesize)
  • Preset boards — Classic Triad, Devil's Court, Full Board, Peer Review
  • Stop all — aborts every advisor stream and the synthesizer simultaneously
  • Vote display — optional summary showing which advisors agreed / diverged
  • Boardroom session history — persist multi-model sessions across app restarts
  • Export Boardroom session as structured JSON (prompt + per-advisor response + consensus)
  • Configurable quorum — synthesize after N of M advisors respond (timeout-based)
  • Per-advisor custom system prompt override in BoardroomInput

Dojo Mode — Model Evaluation & Sparring

Dojo is an evaluation harness where models are scored against each other. Prompted with the same input, their outputs are judged on configurable criteria. See DOJO.md.

  • Dojo tab — structured benchmark / head-to-head mode accessible from header nav
  • N-model fan-out layout — auto-fit grid of panels, each streaming simultaneously
  • Judge step — configurable judge model scores each response on accuracy, conciseness, reasoning, code correctness (user-selectable rubric)
  • Per-response score bars with judge justification rendered in each ModelPanel
  • Running scoreboard — cumulative wins and average overall score across Dojo rounds
  • Prompt presets — 5 built-in test prompts (code challenge, logic puzzle, concept explanation, debug challenge, debate)
  • Blind mode — hide model names from judge prompt to reduce bias; label mapping restored after parse
  • Stop All — aborts every contestant stream simultaneously
  • Configurable rubric — enable/disable accuracy, conciseness, reasoning, code correctness
  • N-way tournament bracket: round-robin comparisons with bracket view
  • Export Dojo results as CSV / JSON (prompt, model, response, score, judge reasoning)
  • Per-round prompt history / replay
  • Auto-run mode — queue N prompts × M models, run unattended, display aggregate results
  • Persist scoreboard to localStorage across sessions
  • Latency + token-rate display per panel

Ultimate Input Field

Replace the current textarea with a composable command-centre input that handles rich content, model targeting, and slash commands.

  • Slash commands — /model <name>, /boardroom, /dojo, /clear, /export, /system <prompt>, /temp <0-2> parsed inline
  • @ mention model targeting — type @qwen3-coder to direct the message at a specific model in a multi-model session
  • # tag — tag a message with a topic (#code, #review) stored in conversation metadata
  • Inline model selector chip — click to swap model without leaving the input
  • File drop zone integrated into input area — drag text files / images; shows filename chip with remove button
  • Voice input button — browser SpeechRecognition API with live transcript preview
  • Prompt templates — pre-saved templates selectable from a popover (e.g. "Explain this code", "Review for bugs", "Write tests")
  • Multi-line expand/collapse toggle (current auto-resize kept; add explicit ↕ handle)
  • Character / token count display beneath input (rough estimate: chars ÷ 4)
  • Send to all / send to selected toggle when multiple models are active in Boardroom / Dojo
  • Keyboard shortcut overlay — press ? in input to show available shortcuts / commands
  • Input history — ↑ / ↓ arrows cycle through previously sent messages (session-scoped)

Multi-Model Layout — flexlayout-react

Full implementation of flexlayout-react (already installed) replacing the current single-pane chat. Panels are drag-and-drop, resizable, and persist layout to localStorage.

  • Root layout switch — toggle between Simple (current single chat) and Flex (multi-panel) mode
  • Default Flex layout: sidebar (model list + conversation history) | main content area | optional right panel (diagnostics / Dojo scores)
  • Model panels — each active model gets its own tab/split pane; drag to reorder, drop to split horizontally or vertically
  • Panels: Chat, Boardroom, Dojo, Diagnostics, Model Browser, Logs, Scoreboard — all draggable into any position
  • Drag a model card from the model list into the main area to open a new chat panel targeting that model
  • Panel title bar shows model name, connection latency chip, and a close (×) button
  • Resize handles on all panel borders — minimum size enforced so panels don't collapse to zero
  • Maximise / restore — double-click a panel header to full-screen it; press Esc or double-click to restore
  • Layout persistence — save named layout snapshots to localStorage; restore on startup
  • Layout presets — "Chat", "Boardroom 3-up", "Dojo A/B", "Diagnostics" — one-click restore
  • Tab overflow — when too many tabs in one pane, overflow into a scrollable tab strip or dropdown
  • Popout panel — detach any panel into its own Tauri window (requires Tauri multi-window support)
  • Sync scroll option — in Boardroom mode, scrolling one model panel scrolls all panels in sync

Layout & UI

  • Dark / light / system theme toggle
  • Font size adjustment
  • Message timestamp display (toggle)
  • Collapse / expand individual messages
  • Wide mode / distraction-free mode (full-width messages, hidden header)
  • Custom window title per conversation

Rust Backend

  • Expose GPU stats via Tauri command (NVML / ROCm if available on host)
  • Tauri tray icon with connection status indicator
  • System notification when a long inference completes (app in background)
  • File picker command for attachment support
  • Auto-launch at login option (Tauri plugin)
  • Deep-link / protocol handler (openmind://) to open with a pre-filled prompt

Quality & Reliability

  • Error boundary around MarkdownMessage to prevent render crash on malformed content
  • Retry with backoff when Ollama returns 5xx during streaming
  • Graceful handling of model not found (404 from /api/chat)
  • Unit tests for opencode-client.ts (mock fetch)
  • E2E smoke test that actually streams a response (extend test-opencode-ollama.sh)
  • TypeScript strict null checks audit (currently strict mode enabled — verify no suppressions)
  • CSP headers in Tauri config for production build

Tooling & DX

  • Prettier + ESLint config committed to repo
  • Husky pre-commit hook (lint + type-check)
  • GitHub Actions CI: tsc --noEmit + cargo check on push
  • Renovate / Dependabot for dependency updates
  • Version bump script tied to tauri.conf.json + Cargo.toml + package.json