You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The Enforce per-task score threshold (>= 0.7) step in .github/workflows/smoke-rpa-skills.yml reads only weighted_score. skill-rpa-execution-map-greenfield is a turn-budget gate (max_turns: 10); it ended in MAX_TURNS_EXHAUSTED in every CI run from 08-27 to 09-04 (0.73, 0.77, 0.77, 0.77, 0.62) and the gate counted four of those as passes. The defect (#3055) stayed hidden for eight days until one run dipped under 0.7.
Policy question to settle before changing the step: also requiring final_status == "SUCCESS" would have failed skill-rpa-coded-test-case at 0.85 on 08-27 (run 33790770382). Options:
Require final_status == SUCCESS for every row (strictest; turn budgets become real gates everywhere).
Require it only for tasks that declare a turn budget as their purpose (e.g. a task tag such as turn-budget, or any task whose run_limits.max_turns is set explicitly).
Keep score-only gating but print MAX_TURNS_EXHAUSTED rows as ::warning:: so they are visible in the job summary.
Related: the same step's per-task loop runs every windows-tagged task file through coder-eval run --tags smoke; five windows-tagged tasks carry no smoke tag (xaml_test_case, deploy_pack_to_orchestrator, diagnose_faulting_activity_log, diagnose_missing_scoped_asset, fix_broken_xaml_validate_loop), so each infra-triggered run logs five Results: 0/0 succeeded + Process completed with exit code 1 blocks that continue-on-error swallows. The detect job could filter WINDOWS_TASKS to files that also carry smoke, or the loop could skip a task whose tag filter selects zero tasks.
Activity