Build with DataHub · The Agent Hackathon · judge workbench

The PR looks fine. DataHub knows it isn't.

Downstream Impact Guardian is a reusable GitHub Action: on every dbt pull request it detects schema / logic / semantic drift, reads DataHub for the real cross-system blast radius, writes a PROPOSED Data Contract back into the catalog, and posts one comment with mergeable compatibility code. Everything below is a real, checked-in guardian run — no account, credentials, or setup required.

Apache 2.0 48 unit tests reusable composite action offline mode = first-class

Four judge routes, in effort order

Routes 1–2 need nothing at all. Routes 3–4 use read-only credentials published in the submission description. Full guide, including a criterion→evidence map, in docs/JUDGING.md.

  1. 1 · 60 seconds — this pageRead the four verified reports below. Each was posted by the published action running cross-repo on a standing demo PR; links go to the live comments.
  2. 2 · 5 minutes — trigger it livePress the button on the demo page: it opens a real PR against fiction-retail-dbt, the real Action runs, and the agent's report renders on that same page. No account needed.
  3. 3 · 15 minutes — browse the catalogOpen the live DataHub instance (read-only login in the submission description) and see what the agent reads: lineage on fct_orders, the depends_on_columns declaration, and the PROPOSED Data Contract it wrote back. Live during the judging window (Aug 17–31).
  4. 4 · Power route — interrogate it yourselfPoint your own agent at the catalog via the shipped .mcp.json (mcp-server-datahub) and ask “who breaks if fct_orders drops order_total?” Read-only token in the Devpost testing notes.

Verified runs — four scenarios

Snapshot of the guardian’s actual PR comment (checked in, verifiable against the live PR).

Loading report…
Why these reports are honest. Per-consumer impact is labeled a worst-case upper bound unless the consumer declared its column dependencies — then it becomes fact, including the 🟢 SAFE verdict. Severity scoring and code generation are deterministic and unit-tested; the LLM narrates (attributed in every report: Qwen via Google ADK + Agent Context Kit). DataHub is never fed the PR’s hypothetical state.

Adopt it

- uses: jwlai-cloud/downstream-impact-guardian@v1
  with:
    dbt-project-dir: dbt
    datahub-url: ${{ secrets.DATAHUB_GMS_URL }}
    datahub-token: ${{ secrets.DATAHUB_GMS_TOKEN }}

That block is the consumer repo’s entire integration surface — its first real run even caught a portability bug dogfooding never could.

downstream-impact-guardian · fiction-retail-dbt (consumer) · docs/ARCHITECTURE.md · docs/adr/ · CONTEXT.md · design blog in docs/blog/