jerry/buzz

mirror of https://github.com/block/buzz.git synced 2026-11-11 20:01:09 +00:00

Files

npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 e45f8dc9a2 fix: harden three mutation-escape test paths from Paul's bounce

Three test-correctness fixes, no production logic changes:

1. test_retain_agent_record_user_intent_clears_gate_but_reconcile_does_not:
   The interactive-save call previously used the same record (unchanged
   content), so the early-return branch fired and set_publish_blocked(false)
   cleared the gate before the retain_user_intent_event path was reached.
   Reverting the changed-content branch to retain_event left the test green.
   Added Case B: interactive save with a different system_prompt, which
   exercises the retain_user_intent_event path directly. Also added
   Case D: reconcile with changed content, confirming retain_event does
   not clear the gate. Now all four combinations (same/changed × user/reconcile)
   are verified.

2. test_non_persona_tombstone_user_intent_clears_gate:
   Previously called retain_user_intent_event directly — reverting
   tombstone_managed_agent_pending to retain_event left zero test callers
   on the production path, so the suite stayed green. Extracted
   tombstone_agent_pending_inner from commands/agents.rs into
   managed_agents/reconcile.rs (pub(crate), symmetric with retain_agent_record)
   and rewired tombstone_managed_agent_pending to call it. Test now drives
   the seam directly. Pre-seeds the tombstone coordinate with
   publish_blocked=true so retain_event (which leaves the existing value)
   and retain_user_intent_event (which calls set_publish_blocked(false))
   produce distinguishable outcomes. Mutation at the production call site
   now fails the suite.

3. test_mark_synced_and_stamp_baseline_atomic_both_written_on_success:
   Previously tested the success path only. Added Case B: drops the
   persona_baselines table before the call, forces the baseline INSERT to
   fail, and asserts the call errors AND pending_sync is still 1
   (transaction rolled back). If the function is split back into two
   autocommit statements, the UPDATE commits first and pending_sync becomes 0
   before the INSERT fails — the rollback assertion catches the regression.

Co-authored-by: Will Pfleger <pfleger.will@gmail.com>
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

2026-08-01 15:54:31 -04:00