Skip to content

ROADMAP.md

Latest commit

 

History

History
95 lines (83 loc) · 5.67 KB

ROADMAP.md

File metadata and controls

95 lines (83 loc) · 5.67 KB
Raw

Sentient v3 roadmap

Status legend: done · in progress · next · later.

M0 · Core engine (done 2026-09-12)

Config model, keychain secrets, SQLite + sqlite-vec store, LiteLLM provider with roles, tool plugin system with risk levels, fact memory, workspace markdown, skills library, agent loop with approvals, gateway, tests.

M1 · Desktop revamp foundations (done 2026-09-15)

  • Desktop-only product direction; v2 USPs preserved (tasks, memory, proactivity, integrations, voice); no auth.
  • Contract-first renderer ↔ engine API (docs/API.md), ownership rules (CLAUDE.md).
  • Composition root with feature services, EventBus forwarded over /ws, package-owned schemas, router per package.
  • Reusable Agent.run_loop; chat attachments, per-message model override, cancel, token usage, auto titles, context compression, channel-aware prompts.
  • Model configuration API: providers, local models, test with tool-support probe, roles, fallbacks, Ollama pull, keychain secrets. Onboarding, files, sessions, notifications, usage endpoints.

M2 · Feature parity (done 2026-09-15)

Built in parallel by package owners; 200+ engine tests, lint clean:

  • Tasks: state machine, scheduler, planner/executor/result generator, triggered tasks with filter DSL, swarm, restart resume, tasks tools.
  • Integrations: Google (native OAuth), GitHub, Slack, Notion, Discord, Trello, WhatsApp, keyless search/weather/maps/news/charts/web fetch, MCP client, privacy filters, poll sources.
  • Memory parity + proactivity + self-evolution: v2 topics, graph, import, episodic summaries, proactive pipeline with learning, heartbeat, skill reviewer, curator, profile upkeep.
  • Voice: faster-whisper, system/Kokoro/cloud TTS, VAD, interruptible /ws/voice.
  • Desktop foundation: Electron shell, design system, onboarding, chat, settings with model configuration.

M3 · Remaining screens (done 2026-09-15)

Tasks (list, calendar, day, detail with plan/log/results/run history, schedule editor, approvals), Memory (graph + list + topics + import + personality files), Integrations (connect flows, privacy filters, MCP servers), Skills (active, pending review with diff, archived, evolution log), Notifications panel with proactive suggestion approve/dismiss, Voice mode.

M4 · Integration and real-machine verification (done 2026-09-16)

End-to-end runs on the founder's PC with local models: onboarding → chat with tools → memory → task lifecycle → triggered task → proactive suggestion → skill proposal → voice round trip. Screenshot review of every screen. Fix everything found.

Done so far: every screen captured from seeded data and reviewed; engine gaps found by the UI builders fixed (live notification updates, OAuth cancel, readable poll errors, skill proposal editing, tool selection and a no-tools fallback for small local models).

M4.5 · V3 leap (done 2026-09-16: verified on local qwen3:8b and on Claude Sonnet 5)

Approved by the founder on 2026-09-15. Contract: docs/API.md sections 10-16.

  • Biggest leaps: device nodes (desktop, phone web app, reference glasses node, docs/NODES.md), messaging channels with pairing (Telegram, Discord), browser control, code execution that calls tools, chat subagents, steering a running reply.
  • Knowing the user: dialectic user model ("About you"), nightly memory consolidation (dreams), memory flush before compression, skills that propose their own repairs.
  • Doing more on its own: change feeds and webhooks instead of timer polling, IMAP push, script jobs with no model calls, wake word and talk mode.
  • Agent improvements: tool progress streaming, dynamic risk, parallel read tools, large-result guard, prompt caching.
  • Not doing: multiple named assistants, open skill marketplace, network-exposed gateway by default.

M5 · Packaging (Windows done 2026-09-16)

A non-technical person double-clicks one installer and gets a working assistant with no Python and no terminal.

  • Engine frozen with PyInstaller (packaging/sentient-engine.spec, onedir) into sentient-engine.exe, shipped as resources/engine by electron-builder. electron/main/paths.ts prefers it in a packaged build; npm run dev still runs .venv/Scripts/python.exe -m sentient serve. ~/.sentient in both modes.
  • npm run package → desktop/dist/Sentient-Setup-<version>.exe: per-user NSIS, no administrator rights, ~216 MiB. Installed, launched, screenshotted and uninstalled on the founder's PC; uninstall leaves ~/.sentient intact.
  • Heavy optional extras (faster-whisper, CTranslate2, onnxruntime, Kokoro, openWakeWord, OpenCV, PyTorch) and browser binaries are left out on purpose; a runtime hook explains what to use instead.
  • Still to do: macOS DMG and Linux AppImage are configured but unbuilt and unsigned, no code signing or notarization, no auto-update feed.

M6 · Devices and channels (moved into M4.5)

Device node protocol over the same engine (node.hello with capabilities: audio in/out, camera, notify) for phone and the smart glasses; messaging channels (Telegram first) with pairing codes.

Open source and contributors (2026-10-08)

  • v3 became the default branch (main); v2 moved to the v2 branch and v2-final tag.
  • GitHub flow: forks, small pull requests to main, one maintainer approval, squash merge.
  • CI on every pull request with no secrets needed; release workflow builds installers from version tags.
  • Issue forms, area labels applied automatically, good first issue and help wanted backlog.

Next

  • Signed installers and auto-update; first public v3 release.
  • macOS and Linux installers verified on real machines (the release workflow builds them).
  • Real-account testing of integrations and messaging channels; real phones and the glasses hardware.