namanjain24-sudo opened a new pull request, #75623:
URL: https://github.com/apache/airflow/pull/75623

   `dev/skill-evals/cases/newsfragment.yaml` has not changed since the harness 
landed in #69308, so it still
   encodes the six-type newsfragment policy. #74589 then narrowed 
`airflow-core` to `significant` only:
   
   > In `airflow-core/` use only `significant` newsfragments, and only for a 
change users must read before
   > upgrading: a behavior change, a removal or a new default, or migration 
steps. [...] Never add one for an
   > ordinary bug fix, feature, improvement or doc change
   
   That change updated `AGENTS.md`, 
`contributing-docs/18_contribution_workflow.rst` and the PR-drafting
   skill, but not these cases. Two of them describe additive features — a new 
endpoint plus a new
   `[connection_test]` config section, and a new coordinator layer plus a 
config option — and assert
   `should_create === true`. Neither is a behavior change, a removal, a new 
default or a migration, so an
   agent following the current guidance answers `false` and the case fails it 
for being right. The harness
   that exists to catch guidance drift had drifted from the guidance.
   
   Flipping both would leave all six cases expecting "no", which an agent that 
always declines passes
   perfectly, so the suite would stop detecting guidance that is too 
restrictive. Two cases that do warrant a
   `significant` note keep the other direction covered, both taken from 
newsfragments already on `main` rather
   than invented: the asset-triggered run default (#68517) and the async 
Postgres driver default (#68496).
   That keeps the balance the file had before — six "no" to two "yes".
   
   I may have the policy call wrong on the two flipped cases; if maintainers 
read "a new config section plus a
   new endpoint" as something users must read before upgrading, say so and I 
will restore them and adjust the
   reasoning instead.
   
   ### The eval hash
   
   `compute_guidance_hash` covers `cases/*.yaml`, so this trips 
`check-eval-hash` until the eval is re-run and
   `last-eval-hash.txt` is updated. I do not have Claude auth available to run 
it, so this follows the README's
   documented fallback and #75599's approach: the hook is skipped, the hash is 
left stale, and the PR stays a
   draft until a maintainer can run the eval and push the updated hash to the 
branch.
   
   Everything else in `prek run --files 
dev/skill-evals/cases/newsfragment.yaml` passes, and the file parses as
   valid YAML with eight cases.
   
   related: #62500
   
   ##### Was generative AI tooling used to co-author this PR?
   
   - [X] Yes (please specify the tool below)
   
   Generated-by: Claude Code (Opus 5) following [the 
guidelines](https://github.com/apache/airflow/blob/main/contributing-docs/05_pull_requests.rst#gen-ai-assisted-contributions)
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to