mirror of
https://github.com/block/buzz.git
synced 2026-10-31 22:41:12 +00:00
768b334609
Replace the three-condition discarded_failed predicate (which required last_acked_effort_gen == Some(g) — false in the return→report race) with a single unconditional rule: any returned worker whose last_effort_result is Some(Failed) invalidates its sessions regardless of generation state, value equality, or whether resolve_effort_report has already run. Rationale: a Failed application means session_set_config_option rejected the pick; the live session ran at the agent default, not the requested value. The session is stale in both orderings — no gate state is needed. Worst case is one redundant session re-creation when the rollback target equals the default. Two new regression tests cover both orderings: - test_failed_worker_return_before_report_invalidates_session (the race): return_agent runs while generation is unresolved; session must still be invalidated; pre-fix this test fails at the sessions.is_empty() assert. - test_failed_worker_report_before_return_invalidates_session (normal path): resolve_effort_report fires first (pre-prompt); return_agent follows; session still invalidated; no second ack emitted. Both tests assert: failed session empty, exactly one terminal failure ack with the correct nonce, pool state (desired=low/committed=low) matching ack. F1 tests extended through resolution: after the invalidation assert, the high→medium→high and repeated-clear tests now perform the next claim and simulate a resolve_effort_report, asserting exactly one terminal ack carrying the newest nonce and persistence-eligible pool state. Minor 2: fix 4 remaining stale doc-comment references in pool.rs (~1282, ~1378) and lib.rs (~8145, ~8470) that still named create_session_and_apply_ model as the final-ack source; updated to PoolEvent::EffortReport → resolve_effort_report. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com>