Operator Tuner Report — 2026-05-15

Window: last 24h (extended 72h for recurrence). Runs: 50. Total actions: 590. Avg actions/run: 10.84.

Action mix (recent window)

    56  operator_ticket_exists
    50  run_complete
    49  run_start
    49  disk_worktree_snapshot
    49  memory_snapshot
    49  done_audit_summary
    43  blocked_lane_classification_summary
    35  repo_dirty_observed
    28  promote_dead_zone_ticket
    28  recent_task_failure
    18  operator_dev_loop_finished
    17  active_lane_snapshot
    17  todo_backlog_observed
    15  active_lane_backlog_observed
    15  reopen_weak_done_ticket
    11  pickup_direct_run
    10  terminal_ticket_runtime_closed
     9  create_task_failure_ticket
     7  needs_input_runtime_failure_reset
     6  create_operator_ticket
     5  blocked_lane_completed_marked_done
     5  operator_reclassification
     4  operator_dev_loop_throttled
     3  repo_dirty_all_known_generated
     2  in_review_stale_failure_state_cleared
     2  stale_runtime_observed
     2  stale_runtime_marked
     2  stale_runtime_summary
     1  memory_pressure_observed
     1  breakglass_noop_healthy
     1  needs_input_reclassified_in_review
     1  active_lane_cleared

Signals worth tuning

• Recurring operator-ticket markers (>=3 in 72h): {'[operator:dirty-repo:workspace]': 3, '[operator:dirty-repo:mission-control]': 3}
• Weak-done reopens: 15
• Stale-runtime marks (extended window): 6
• Dev-loop throttle ratio: 0.18 (4 throttled / 18 finished)
• Repo-dirty observations: 35 (generated-only noise: 3)

Reopened-ticket outcomes (current MC state)

• MC-3344: status=done updated=2026-05-14T13:40:09.217579+02:00 failure=—
• MC-3351: status=done updated=2026-05-14T11:41:44.669221 failure=Weak completion evidence found by Luci Operator
• MC-3358: status=in_review updated=2026-05-14T23:30:16.255795+02:00 failure=Weak completion evidence found by Luci Operator
• MC-3360: status=cancelled updated=2026-05-14 08:18:35 failure=Weak completion evidence found by Luci Operator
• MC-3362: status=cancelled updated=2026-05-14 08:14:11 failure=Weak completion evidence found by Luci Operator
• MC-3384: status=cancelled updated=2026-05-14 08:34:50 failure=Weak completion evidence found by Luci Operator
• MC-3391: status=done updated=2026-05-14 10:12:28 failure=Weak completion evidence found by Luci Operator
• MC-3392: status=done updated=2026-05-14 10:12:28 failure=Weak completion evidence found by Luci Operator
• MC-3394: status=done updated=2026-05-14 10:12:28 failure=Weak completion evidence found by Luci Operator
• MC-3395: status=done updated=2026-05-14T12:10:58.951786+02:00 failure=—
• MC-3396: status=done updated=2026-05-14 10:35:11 failure=Weak completion evidence found by Luci Operator
• MC-3402: status=done updated=2026-05-14T13:35:01.338282+02:00 failure=—
• MC-3404: status=done updated=2026-05-14T13:35:13.021409+02:00 failure=—
• MC-3407: status=done updated=2026-05-14T23:08:45.553339+02:00 failure=—

Proposed patches

Operator Tuner Review — 2026-05-15

High-confidence improvements

1. Race-condition reopen: no grace period between "done" and audit
2. Problem: _done_audit_policy_matches audits tickets the instant they're marked done. If the operator run fires before the worker has posted its completion comment (or the comment transaction settles), the ticket is reopened unnecessarily.
3. Location: _done_audit_policy_matches, after age = self.options.now - closed_at (around the if age <= timedelta(hours=24) block).
4. Change: Add if age < timedelta(minutes=30): return False as the first age check, before the existing 24 h window.

5. Expected effect: Eliminates a class of false reopen_weak_done_ticket actions; the 30 min floor is one operator cycle and gives the worker ample time to post evidence.

6. "completed" missing from strong-evidence action verbs

7. Problem: Workers frequently phrase completion as "Task completed successfully" or "Feature completed." The verb "completed" is not in the STRONG_EVIDENCE_PATTERNS action-verb regex, so these tickets are flagged weak and reopened.
8. Location: STRONG_EVIDENCE_PATTERNS, 5th pattern: re.compile(r"\b(fixed|resolved|implemented|updated|created|delivered|sent)\b", re.I).
9. Change: Add "completed" to the alternation group: r"\b(fixed|resolved|implemented|updated|created|delivered|sent|completed)\b".

10. Expected effect: Reduces reopen_weak_done_ticket false positives; risk profile is identical to existing broad verbs like "updated."

11. "merged" missing from commit/push strong-evidence pattern

12. Problem: A worker comment "Changes merged into master" doesn't match any STRONG_EVIDENCE_PATTERNS entry. "Merged" is in SHIPPED_TOKENS (used for _body_is_terminal_done) but not in the evidence check for the done audit.
13. Location: STRONG_EVIDENCE_PATTERNS, 4th pattern: re.compile(r"\b(commit|committed|pushed|deployed)\b", re.I).
14. Change: Add "merged": r"\b(commit|committed|pushed|deployed|merged)\b".
15. Expected effect: Catches a common worker completion signal, reducing reopen_weak_done_ticket count. "Merged" is at least as strong as "pushed" as a completion indicator.

Signals reviewed but not acted on

• Dirty-repo recurring tickets (3 × workspace, 3 × mission-control in 72 h): The operator correctly filters _GIT_KNOWN_GENERATED_PREFIXES (3 repo_dirty_all_known_generated events prove it). The remaining 32 dirty observations are for non-generated paths; without knowing which files they are, no threshold or filter change is justified. Too thin — skip.

• Dev-loop throttle ratio (0.18 = 4 throttled / 18 finished): The 4-hour throttle window is preventing only 18% of attempts; 82% of dev-loop launches succeed. This is a healthy ratio — the throttle is doing its job without being overly restrictive. Healthy — no change.

• Stale-runtime marks (6 in 72 h): Volume is low, the 6-hour cutoff and _runtime_safe_to_stale safety checks are appropriate. Working as intended — no change.