Operator Tuner Report — 2026-05-17

Window: last 24h (extended 72h for recurrence). Runs: 50. Total actions: 456. Avg actions/run: 8.12.

Action mix (recent window)

    50  run_start
    50  disk_worktree_snapshot
    50  memory_snapshot
    50  done_audit_summary
    50  run_complete
    40  blocked_lane_classification_summary
    21  operator_ticket_exists
    18  repo_dirty_observed
    18  repo_dirty_all_known_generated
    15  needs_input_runtime_failure_reset
    11  operator_dev_loop_finished
    10  active_lane_snapshot
     9  recent_task_failure
     9  terminal_ticket_runtime_closed
     8  attach_operator_context_to_closed_failure_ticket
     7  operator_dev_loop_throttled
     7  blocked_lane_completed_marked_done
     7  active_lane_backlog_observed
     6  todo_backlog_observed
     6  create_operator_ticket
     4  promote_dead_zone_ticket
     2  reopen_weak_done_ticket
     2  breakglass_noop_healthy
     1  pickup_direct_run
     1  memory_pressure_observed
     1  defer_task_failure_to_watchdog
     1  stale_runtime_observed
     1  stale_runtime_marked
     1  stale_runtime_summary

Signals worth tuning

• Recurring operator-ticket markers (>=3 in 72h): {'[operator:ticket-backlog]': 3, '[operator:memory-pressure]': 3, '[operator:dirty-repo:workspace]': 3, '[operator:dirty-repo:mission-control]': 3, '[operator:active-lane-backlog]': 4}
• Weak-done reopens: 2
• Stale-runtime marks (extended window): 3
• Dev-loop throttle ratio: 0.39 (7 throttled / 11 finished)
• Repo-dirty observations: 18 (generated-only noise: 18)

Reopened-ticket outcomes (current MC state)

• MC-3538: status=done updated=2026-05-16T13:05:21.270233+02:00 failure=—
• MC-3542: status=done updated=2026-05-16 08:03:50 failure=—

Proposed patches

Operator Tuner Review — 2026-05-17

1. Raise ACTIVE_LANE_COUNT_TRIGGER

• Problem: The active-lane-backlog marker recurs 4× in 72h, suggesting 15 tickets in todo+in_progress is normal steady-state, not a drain failure.
• Location: Constant ACTIVE_LANE_COUNT_TRIGGER = 15 (near line 97).
• Change: Raise to 25.
• Expected effect: Eliminates most active_lane_backlog_observed false escalations and the corresponding recurring operator ticket. The ACTIVE_LANE_STALE_HOURS_TRIGGER = 2 still catches genuine stalls via the age path.

2. Remove \bssh\b from the needs-input retry regex

• Problem: _should_retry_misrouted_needs_input matches \bssh\b, which fires on unrelated text ("ssh key", "ssh tunnel", documentation mentions). With 15 needs_input_runtime_failure_reset actions in 50 runs, some fraction may be false resets driven by this overly broad token.
• Location: The regex string inside _should_retry_misrouted_needs_input (around line that builds the re.search(r"\b(crashed|...|ssh|timeout)\b" pattern).
• Change: Remove \bssh\b| from the alternation. harvest_timeout and the other specific failure tokens already cover the genuine SSH-related failure cases.
• Expected effect: Reduces false needs_input → todo resets. The remaining tokens (crashed, worker died, runtime killed, 429, rate limit, connection …closed/reset) still catch real runtime failures.

3. Extend _is_known_generated to mission-control repo

• Problem: _is_known_generated returns False for any repo_name != "workspace". The [operator:dirty-repo:mission-control] marker recurs 3× in 72h, likely from generated artifacts in the mission-control tree that the operator can't filter.
• Location: Function _is_known_generated, first guard clause (if repo_name != "workspace": return False).
• Change: Allow mission-control through the guard and add mission-control–specific generated prefixes (e.g. "reports/", "state/" — the actual paths should be gathered from one sample repo_dirty_observed log entry). Minimal code: drop the workspace-only guard and extend _GIT_KNOWN_GENERATED_PREFIXES with a per-repo mapping.
• Expected effect: Eliminates the recurring dirty-repo:mission-control operator ticket, converting those runs into repo_dirty_all_known_generated (already at 18, adding the MC ones).

4. Raise TODO_BACKLOG_COUNT_TRIGGER

• Problem: [operator:ticket-backlog] recurs 3× in 72h. A threshold of 5 todo tickets may be below the natural steady-state between pickup cycles, causing the operator to fire _run_pickup_once() and open a ticket every time 5+ tickets briefly queue.
• Location: Constant TODO_BACKLOG_COUNT_TRIGGER = 5 (near line 95).
• Change: Raise to 8.
• Expected effect: Reduces ticket-backlog operator-ticket churn. The age trigger (TODO_BACKLOG_MINUTES_TRIGGER = 15) still catches true stalls when the count is 5–7 but aging.

---

Signals explicitly skipped (too thin)

• Weak-done reopens (2): Both tickets (MC-3538, MC-3542) landed back at done without failure. Possible false positives in _has_strong_completion_evidence, but n=2 is insufficient to justify loosening the patterns. Re-examine if reopens exceed ~5/week.
• Dev-loop throttle ratio (0.39): 7 throttled / 11 finished is within normal range for a 4-hour cooldown. Not actionable.
• Stale-runtime marks (3): Low frequency, self-limiting (7-day safety check). No change needed.