Lewati ke konten utama
Versi: v0.0.74

v0.0.66 — Cobuild Dataiku-parity + datasets polish

Released: 2026-05-13. Fourteen commits.

Cobuild​

Stack / Split dock orientation toggle (F13 V1.7)​

The Cobuild dock now offers two layouts: Stack (the existing right-rail) and Split (a wider two-column shell — session list on the left, the active session's chat on the right). The toggle lives in the dock header. Persists per user. Stack stays the default for narrow viewports; Split is useful when you want to monitor session state while chatting.

create_extract_recipe tool (Capability #11)​

A generic LLM-callable extraction tool that turns "pull rows where X" into a parameterized SQL extract recipe — no need to hand-author the SQL or pick a recipe type up front. Joins the existing create_recipe / patch_recipe_step planner toolset; gated behind the same approval flow.

Dataiku-parity polish​

  • Live trace timing — the planner's LiveTrace panel now shows per-tool elapsed time inline, so users can see which step in a multi-tool turn took the wall-clock.
  • /quickstart entry — Cobuild adopts a quickstart drawer that walks new users through "start a session → ask for a dataset → review the plan → promote". Sidebar entry under Cobuild.
  • patch_recipe_step — promoted from prototype to a stable approval-gated tool (v0.0.62 introduced it; v0.0.66 hardens the diff surface and ACL gates).

Discovery includes unmaterialized datasets​

discover_datasets previously filtered to materialized rows only — so Cobuild couldn't propose a recipe over a dataset that existed in the catalog but hadn't yet been built. The tool now returns both states, and profile_dataset surfaces a not_materialized signal in its response (rather than 500-ing on the missing table). The planner uses the signal to propose create_recipe first.

/sessions polling stops on /projects​

The Cobuild dock was polling GET /api/cobuild/sessions every 30 s including on the projects-list page — but without a project context, the request 400'd. v0.0.66 gates the poll on a resolved project_id; the project list no longer spams 400s in the network panel.

Datasets​

Origin badge + tag filter​

Every dataset row now carries an origin badge showing where it came from — dbt, recipe, connector, upload. Filterable inline; the filter persists per user.

Tag filter​

Existing dataset tags (any string in datasets.settings.tags) are now filterable from the dataset list header.

Funnel auto-hide for non-dbt projects​

The dbt medallion-layer pills (bronze / silver / gold) used to render on every dataset list, even on projects with zero dbt models. They now auto-hide when no row has a layer set — non-dbt projects get a chrome-clean list.

The medallion vocabulary is also explicitly documented as dbt-internal in the Datasets page — operators who never touch dbt don't need to learn it.

Nav fix — Connectors removed from project Data dropdown​

Connectors are an org-level resource (see Connectors → Permission model) — they don't belong in the project-scoped Data dropdown that ships beside Datasets / Recipes. The entry rendered there since the v0.0.55 sidebar move and routed callers to /projects/<slug>/connectors, which silently 404'd. The entry is removed; org-only paths now correctly fall through to the bare URL via ORG_PATHS.

Recipes​

Project-scoped list routes added for python and prepare recipe types — /projects/<slug>/recipes/python and /projects/<slug>/recipes/prepare now resolve to the matching filtered list. Previously these legacy non-slug routes were the only entrypoint; the v0.0.49 Phase 2 URL migration left them off the project-scoped sidebar.

Test coverage​

A new PageSmoke frontend test pins that 47 authenticated pages render their root component without a JS error from a logged-in fixture. Catches the v0.0.58 polish-prune class of regression (white-screen on detail page) before it ships.