Skip to content

tasks.md

Latest commit

 

History

History
78 lines (68 loc) · 4.5 KB

tasks.md

File metadata and controls

78 lines (68 loc) · 4.5 KB
Raw

Tasks & roadmap

Status of the work. See design.md for the architecture, providers.md for the provider matrix, and security.md for the token vault.

Done — universal rebuild (on the Vercel AI SDK)

  • Model substrate — everything is an AI SDK LanguageModel. Supply a model directly, or by ProviderModelSpec resolved through the registry.
  • Provider registry — openai · anthropic · google · openai-compatible · xai · deepseek · gateway (dynamic optional-peer imports) + web-llm local; per-stage overrides with a cross-provider key guard; Anthropic browser header injected; directBrowserOk hints.
  • Local models — createWebLLMModel (WebGPU, @browser-ai/web-llm), isWebGPUAvailable; any custom model (incl. Chrome/Edge built-in AI) works by passing it directly.
  • Secure token vault — non-extractable AES-GCM key in IndexedDB; secrets encrypted at rest; CredentialStore (Vault + Memory).
  • Low-level primitives — generate / stream / generateStructured / runToolLoop.
  • Tools — defineTool (one AI SDK ToolSet shape); native tool-calling and a prompted/salvage fallback for weak local models; selectToolMode.
  • Agent loop — createAgent().run(): plan → execute → replan → synthesize, grounded, time-boxed, AbortSignal, typed events; graceful native→prompted degradation.
  • Memory — ContextStore + MemoryStore + IndexedDBStore (session list/delete) + compressHistory; recent turns are read back into the planner prompt so follow-up goals can reference earlier ones.
  • Tool filtering — availableTools whitelist / excludedTools blacklist; plan-narrowed selection honoured on both tool paths.
  • Streaming final answer — the synthesizer streams final.text-delta (suppressed when the model drifts into JSON).
  • MCP — optional ./mcp subpath: HTTP (StreamableHTTP) connector.
  • Packaging — tsup ESM + CJS + d.ts, ././mcp entries, optional peers, no node:* in the core bundle.
  • Tests — node --test: registry resolution, vault crypto round-trip (fake-indexeddb), parsers, prompts, tool-mode, prompted dispatch, and a full prompted end-to-end run via MockLanguageModelV3.

Added after the universal rebuild

  • Eager model preload — createWebLLMModel now defaults to preload: true (1-token warm-up), so weights download during creation with initProgressCallback firing, not silently on the first message; preloadWebLLMModel / unloadWebLLMModel helpers exported (the latter frees GPU memory best-effort when switching models).
  • Leveled logger — logLevel ('silent'|'error'|'warn'|'info'|'debug', default 'warn') + console-like logger sink in the config; phases log raw outputs (debug), plans/steps/tool calls (info), empty plans, no-tool steps, failed tools and salvage fallbacks (warn). createLogger exported.
  • Empty-plan retry — when the planner returns no steps for a multi-word goal (small models sometimes reply "done!" with an empty plan and the run would claim success without acting), the runner retries once with an explicit nudge before treating the turn as conversational.

Next

  • ./react subpath — a useAgent hook (kept out of core to stay UI-agnostic).
  • Streaming structured plan — surface plan.thought-delta / plan.step-added from a streamed generateObject.
  • Model-load events — wire WebLLM download progress to model.load automatically (today: pass initProgressCallback via providerOptions).
  • AI SDK v7 — bump once @browser-ai/web-llm supports the v7 provider spec.
  • MCP niceties — tools/list_changed refresh; per-tool allow/deny filter.
  • Examples — a vanilla-JS demo and a React demo (cloud BYOK + local WebLLM).
  • OPFS/Blob tool outputs — optional spill target for binary MCP results.

Non-goals

  • Built-in UI (chat panel, components) — belongs in host apps.
  • Domain-specific tools — hosts register their own via defineTool.
  • stdio MCP, filesystem sandbox, run persistence/resume, OpenTelemetry — Node-only concerns that live in the sibling package @dudko.dev/agent.
  • Shipping shared/app-owned API keys to the client — use a proxy or the gateway (see security.md).