Skip to content

Leoflow UI Walk — Verified Findings

Date: 2026-05-23 Method: The live demo stack (docker compose --profile demo up) driven by a headless Chromium (Playwright, dockerized). Authenticated via the _token cookie, navigated the screens by clicking (not deep-linking — direct navigation to client routes like / 404s in the SPA), screenshotted each, and captured console errors + non-2xx requests. Every screenshot below was inspected and every finding cross-checked against the authoritative _private_ui.yaml (3.2.1) — so this supersedes the earlier external audit where the two disagree.

Why no side-by-side with real Airflow for most screens: Leoflow serves the unmodified Airflow 3.2.1 SPA, so the layout/chrome is Airflow's. A real-Airflow comparison only adds value for data-shape differences; it is worth doing specifically for the empty Graph view (below).

Summary

The DAG list and grid render with full fidelity — including the run state colors and the per-DAG fields the earlier external audit wrongly told us to strip. The graph view renders an empty canvas (top bug). A few /api/v2 endpoints the UI polls still 404, and the Overview "Failed Runs" widget miscounts.

Screens

DAG list — ✅ renders correctly

DAG list

demo_http_chain shows schedule @daily, the latest run with a green check, and three green run squares (its 3 successful runs); pause/trigger/favorite/delete controls render. The filter chips (Failed/Queued/Running/Success), tags filter, and sort all render. This directly refutes the external audit's "strip description/bundle_version/next_dagrun_*/start_date/end_date from /ui/dags" recommendation — those fields are required by the 3.2.1 schema and the list renders correctly with them present.

Grid view — ✅ renders correctly

Grid demo_http_chain

step_a/step_b/step_c each show three green ✓ cells (3 successful runs). The DAG header (schedule, latest run, max active runs, owner) and the tab bar (Overview / Runs / Tasks / Calendar / Backfills / Audit Log / Code / Details) render. The grid left column is driven by /ui/grid/structure (incl. the required is_mapped) and the cell colors by /ui/grid/ti_summaries — both correct. The external audit's "Grid broken" applies only to mapped tasks / task groups, which the MVP does not model; for normal DAGs the grid is correct.

Grid demo_pipeline

demo_pipeline (no runs) shows its extract/transform/load task rows.

Graph view — ✅ FIXED (was an empty canvas)

Graph demo_pipeline

The graph now renders the topology: extract (python) → transform (bash) → load (http_api) as connected React Flow nodes, with "Latest Dag Version: v1" in the header.

Root-cause chain (found by diffing against real Airflow 3.2.1 on a side cluster, not from the SPA prose): the graph fetches GET /ui/structure/structure_data?dag_id=…&version_number=N — it needs a version_number. The SPA reads that from GET /api/v2/dags/{id}/details → latest_dag_version.version_number, which in turn comes from GET /api/v2/dags/{id}/dagVersions. Both of those 404'd for us, so: dagVersions 404 → details.latest_dag_version: null → no version_number → structure_data never fetched → blank canvas.

Fix (this PR): - Implement GET /api/v2/dags/{id}/dagVersions (real version_number from the stored versions via a row_number() window — integration-tested against live PG). - Populate details.latest_dag_version from the version lister. - Add the optional node/edge fields to structure_data (is_mapped, tooltip, setup_teardown_type, asset_condition_type; is_setup_teardown, label, is_source_asset) so the response matches real Airflow exactly.

Verified live in the browser (the screenshot above) and via the captured network trace (structure_data now returns 200).

Non-2xx requests captured (to stub/implement)

  • GET /api/v2/dags/{id}/dagRuns/~/taskInstances?…state=failed&order_by=-run_after → 404. The Overview "Failed Tasks/Failed Runs" widgets poll this (~ = all runs). Its failure is why the Overview "Failed Runs" widget shows 3 for a DAG with 3 successful runs — a visible miscount.
  • GET /api/v2/dags/{id}/dagVersions?order_by=-version_number → 404. Backs the header "Latest Dag Version" field.
  • GET /api/v2/dags/~/dagRuns (home global run view) → already fixed (degrades to an empty collection).

Other observations

  • Sidebar/menu: Home, Dags, Assets, Docs (+ clock, User). Browse and Admin are correctly hidden by the curated /ui/auth/menus; Assets is built-in nav, not menu-curated, so it still appears (its pages degrade to empty states).
  • MIME warning (Failed to load module script … MIME type "text/html"): upstream Airflow 3.2.1 build artifact, safe to ignore.
  1. ~~Fix the empty Graph view~~ — done (dagVersions + details.latest_dag_version
  2. structure_data fields; see above).
  3. ~~dagRuns/~/taskInstances and dagVersions~~ — done: dagVersions implemented; the two ~ wildcards (dagRuns/~/taskInstances, ~/dagRuns) degrade to empty collections (real cross-run/cross-DAG aggregation is a follow-up so the Overview "Failed Runs" widget shows real counts).
  4. Decide whether to hide the Assets nav (it is built-in, not menu-driven).

Note on the external (Antigravity) audit

It is a useful lead generator (it found the ~-wildcard 404s and documented the write-path payloads), but its schema diagnosis is unreliable: it repeatedly recommended removing fields that are required by the 3.2.1 spec (verified here — the list/grid render correctly with them) and its v2 "screenshots" were all 404 pages (bad deep-links). Treat it as a source of leads, validated against the spec and the browser — which is what this report does.