Direct answer: a reviewable AI coding agent output should contain the request, acceptance criteria, failing tests, scoped implementation notes, and PR evidence a human can verify.
Acceptance criteria
- Selected filters persist across pagination changes.
- Returning to page 1 restores the same filtered result set.
- The URL query string remains shareable.
- No regression to saved views or default table state.
Tests before code
- Failing browser test reproduces pagination filter reset.
- Unit test covers query-string state serialization.
- Regression test confirms saved view state still wins on initial load.
Implementation notes
- Move filter state persistence to the table controller.
- Keep URL updates debounced to avoid history spam.
- Do not change backend filtering semantics.
PR review evidence
- PR shows the failing test turning green.
- Implementation diff is scoped to table state management.
- Reviewer can replay the reproduction path from the PR body.