⌂ Home ☷ Board

Operator Tuner Report — 2026-05-28

Window: last 24h (extended 72h for recurrence). Runs: 97. Total actions: 1336. Avg actions/run: 12.72.

Action mix (recent window)

   235  operator_ticket_exists
   229  repo_dirty_observed
    97  run_start
    97  disk_worktree_snapshot
    97  memory_snapshot
    97  done_audit_summary
    97  run_complete
    76  recent_task_failure
    52  blocked_runtime_failure_reset
    50  active_lane_snapshot
    49  blocked_lane_classification_summary
    33  attach_operator_context_to_closed_failure_ticket
    32  operator_dev_loop_throttled
    19  todo_backlog_observed
    17  defer_task_failure_to_watchdog
    12  promote_dead_zone_ticket
     8  pickup_direct_run
     7  create_operator_ticket
     7  active_lane_backlog_observed
     7  operator_dev_loop_finished
     4  repo_dirty_all_known_generated
     2  blocked_lane_completed_marked_done
     1  ticket_advance
     1  signal_archive
     1  operator_ticket_resolved
     1  operator_session_summary
     1  reopen_weak_done_ticket
     1  needs_input_reclassified_in_review
     1  runtime_session_reconcile
     1  code_fix_committed
     1  needs_input_runtime_failure_reset
     1  worktree_reaper_ran
     1  breakglass_noop_healthy
     1  in_review_stale_failure_state_cleared

Signals worth tuning

Reopened-ticket outcomes (current MC state)

Proposed patches

Reviewed report + source. Three high-confidence improvements, two signals too thin to act.

Improvements

1. graphify-out/ not whitelisted → drives workspace dirty-repo noise - Problem: repo_dirty_observed=229 + recurring [operator:dirty-repo:workspace] (4× in 72h). Live git status shows graphify-out/converted/*.md modified and graphify-out/cache/*.json untracked — generated by graphify task, not anomalies. Only 4 lines classified generated-only, so workspace marker re-fires every run. - Location: _GIT_KNOWN_GENERATED_PREFIXES tuple (the prefix list). - Change: add "graphify-out/" (covers converted/ + cache/). Optionally ".mc-wt-" if confirmed transient worktree marker dirs. - Effect: workspace repo_dirty_observed collapses, repo_dirty_all_known_generated rises, [operator:dirty-repo:workspace] stops recurring, operator_ticket_exists (235) drops materially.

2. Active-lane staleness trigger too aggressive at 2h - Problem: [operator:active-lane-backlog] is the most-recurrent marker (6× in 72h); active_lane_backlog_observed=7. Count trigger is 25 (board not that big), so the 2h age trigger is firing. One ticket legitimately in_progress >2h trips an escalation ticket every cycle. - Location: ACTIVE_LANE_STALE_HOURS_TRIGGER = 2. - Change: raise to 4. - Effect: cuts active-lane escalations roughly in half, keeps genuine multi-hour stalls caught. Reduces active_lane_backlog_observed and marker recurrence; active_lane_snapshot absorbs the difference.

3. recent_task_failure double-counts across overlapping runs - Problem: recent_task_failure=76 in window. audit_scheduled_task_health looks back a fixed 2h (now - timedelta(hours=2)) every run; cadence is sub-30min, so each failed task_run is re-logged ~4×. Inflates the signal and log volume; only CRITICAL_TASK_IDS ticket (correctly deduped), so this is logging noise, not health. - Location: audit_scheduled_task_health — the started_at >= ? (now - 2h) query bound. - Change: scope to since-last-operator-run (persist last-tick ts like the dev-loop throttle state does), or dedupe by task_run_id already seen. Smaller: narrow window to 35min. - Effect: recent_task_failure drops to ~1× per real failure, cleaner failure-rate signal. No behavior change to ticketing.

Skipped (signal too thin)

Not a bug