Skip to content

Fix task spec vs gold mismatches in 3 tasks (AE004, AE006, F1002)#151

Open
kiwi0401 wants to merge 1 commit into
dbt-labs:mainfrom
SignalPilot-Labs:fix/task-gold-spec-mismatches
Open

Fix task spec vs gold mismatches in 3 tasks (AE004, AE006, F1002)#151
kiwi0401 wants to merge 1 commit into
dbt-labs:mainfrom
SignalPilot-Labs:fix/task-gold-spec-mismatches

Conversation

@kiwi0401

@kiwi0401 kiwi0401 commented May 29, 2026

Copy link
Copy Markdown

Found these issues while testing my agent against the benchmark.

For the AE tasks setup.sh deletes the canonical model files that the task asks you to rebuild, then runs dbt. AE004 removes obt_product_inventory.sql. AE006 removes obt_product_inventory.sql, dim_products.sql, fact_inventory.sql, and obt_sales_overview.sql.

The issue is that those deleted files are the only place the gold's column and type choices ever lived. Once they're gone, nothing in the prompt, the project, or the data points to those choices, so the agent has no way to reproduce the gold. The fixes just put the missing intent back into the prompt. There are a few open issues calling out the AE tasks so this fix resolves several of them.

AE004
Gold aliases the duplicate product_id as "ipd" and drops the attachments column, neither of which is in the prompt. Dropping attachments isn't discoverable: it's 100% null, but so are description, purchase_order_id and customer_order_id, which the gold keeps. Added the ipd alias and the attachments exclusion to the prompt.

AE006
Same ipd alias and attachments exclusion, plus fact_inventory casts transaction_created_date to a date while keeping transaction_modified_date as a string, even though both are the same varchar in the source. Added the column and date handling to the prompt.

F1002
__stats.yml lists p1/p2/p3 on most_podiums but the gold doesn't include them. The task says to fit the yml, so I fixed the yml to match the gold instead. I saw a bunch of PRs open to fix this issue but figured I'd add it here as well.

Three tasks ship a gold answer key that the task spec cannot produce, so a
clean-room agent reasoning only from the task inputs cannot pass them.

analytics_engineering004:
  The gold omits the attachments column and aliases the duplicate product_id
  as ipd, but the prompt only asks for "product details" and never mentions
  either. attachments is 100% NULL in the source, but so are description,
  purchase_order_id, and customer_order_id which the gold keeps, so null-ness
  is not a usable signal. Added both as prompt clarifications.

analytics_engineering006:
  Same attachments omission and ipd alias on obt_product_inventory. In
  addition, fact_inventory casts transaction_created_date to DATE while
  leaving transaction_modified_date as a raw string, even though both are
  identical VARCHAR datetime columns in the source. Added prompt
  clarifications for the column omission and the date handling.

f1002:
  models/stats/__stats.yml lists p1, p2, p3 columns on most_podiums that the
  gold excludes (they belong on finishes_by_driver). The task says to "fit
  the yml config," so the config and the gold contradict each other. Removed
  the three columns so the config matches the gold.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant