Skip to main content

Roadmap

Product Direction​

The next phase focuses on:

  • a more controllable workspace UI (panels, shortcuts, responsive layout),
  • stronger runtime/data reliability (artifact loading, tables, DuckDB support),
  • and scalable conversation execution (token-aware summarization/truncation and better live run visibility).

Near-Term Priorities​

1) Workspace layout and interaction quality​

  • Make left and right panels independently toggleable while enforcing that at least one panel remains visible.
  • Keep the current behavior where both panels can be open together.
  • Add smooth panel open/close animation.
  • Add keyboard shortcuts aligned with common VS Code muscle memory:
    • Ctrl+J: toggle terminal
    • Ctrl+L: toggle left panel (chat/code)
    • Ctrl+B: toggle right panel
  • Improve responsive behavior for narrow windows:
    • avoid fixed-width controls (for example, table/chart selector dropdown should size with available space)
    • keep workspace controls usable at reduced widths.
  • Fix auth page background styling so it matches the rest of the product theme.

2) Data and schema UX improvements​

  • Add schema selector dropdown in schema editor so users can switch between all schemas in a workspace.
  • Remove “single selected dataset” emphasis styling in dataset selector and move toward multi-dataset workspace querying by default.
  • Support schema composition/combination so agent workflows are not constrained to one table at a time.

3) Execution visibility and runtime feedback​

  • Add websocket-driven run status for code UI so users can clearly see:
    • code running/idle state,
    • kernel active/connecting state,
    • real-time execution lifecycle signals.
  • For large images, stream non-interactive base64 image output when payload size crosses a threshold, instead of always sending interactive image objects.

4) Conversation scaling and background memory management​

  • Add token-aware summarization + truncation node in the agent graph.
  • Trigger summarization when turn/token thresholds are exceeded.
  • Use tiktoken (or API token usage metadata) for token accounting.
  • Run summarization/truncation asynchronously in background so UI remains non-blocking.

5) Data model and persistence cleanup​

  • Split conversation/turn storage from global account/billing tables.
  • Persist generated code per workspace as a stable Python script (1 workspace = 1 code file lifecycle).

6) Artifact/table behavior hardening​

  • Fix AG Grid sorting/filtering behavior.
  • Remove table inline fallback path when artifact is missing; show explicit “artifact not found on disk” state instead.
  • Improve artifact-switch performance (df1 -> df2 -> df1) by introducing at least page-offset/limit cache first, with optional deeper backend page cache later.

7) File format/runtime compatibility​

  • Add first-class DuckDB file support where currently blocked.

Current Behavior Notes (To Be Addressed)​

  • Runtime retries currently re-run the same generated code up to retry limit; there is no automatic “regenerate code and retry” loop after execution failure.
  • Artifact table loading is artifact-backed first; in-memory dataframe payload is fallback only.
  • Switching artifacts currently clears table state and re-requests pages; persistent response caching is not retained across artifact switches.
  • Hidden/idle tab handling currently affects polling cadence only; chat history is not evicted due to hidden state.

Next: Contributing