mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-06-19 15:41:46 +02:00
Trevin Chow
aa3b7d55e2
The previous fix added a Step 3 that ran the project's tests/lint/typecheck
before fixes, so Step 5 could compare and label failures as pre-existing
vs regression. That was machinery in service of one bullet ("say so explicitly
if a check was already failing pre-simplification") that the agent couldn't
honestly satisfy without a baseline.
Simpler resolution: delete the bullet. Step 4 now just runs the project's
checks, surfaces any failure, and forbids relaxing assertions to make checks
pass. The agent can still mention an unrelated-looking failure organically
when the output makes it obvious; it just isn't instructed to make
attribution claims it can't back up. Net 15 lines lighter than the prior
approach and lands closer to where the skill was before the feedback round.