core: Policy, Pflege-Agent (HITL), Healer, Lifecycle-Pipeline, Graduierung + Konfiguration

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-12 00:43:26 +02:00
parent f79e2a8556
commit f2da893f71
28 changed files with 2179 additions and 10 deletions
@@ -0,0 +1,53 @@
+# Deterministisches Autohealing (SPEC-healer). {label} wird aus den Alert-Labels ersetzt.
+max_actions_per_hour: 10
+# Infrastruktur-Aktionen laufen nur, wenn KNASTER_DRY_RUN=false UND sie hier stehen.
+enabled_actions: []
+rules:
+  - id: slot-endpoint-down
+    match: {alertname: SlotEndpointDown}
+    ladder:
+      - {action: restart_container, params: {host: "{docker_host}", container: "{container}"}, attempts: 2, cooldown: 5m}
+      - {action: restart_vm, params: {pve_node: "{pve_node}", vmid: "{vmid}"}, attempts: 1, cooldown: 15m}
+      - {action: mark_slot, params: {slot: "{slot}", status: degraded}, attempts: 1}
+    on_resolved:
+      - {action: mark_slot, params: {slot: "{slot}", status: ok}, only_if_set_by_healer: true}
+
+  - id: slot-canary-failing
+    match: {alertname: SlotCanaryFailing}
+    ladder:
+      - {action: restart_container, params: {host: "{docker_host}", container: "{container}"}, attempts: 1, cooldown: 10m}
+      - {action: mark_slot, params: {slot: "{slot}", status: degraded}, attempts: 1}
+    on_resolved:
+      - {action: mark_slot, params: {slot: "{slot}", status: ok}, only_if_set_by_healer: true}
+
+  - id: gpu-xid
+    match: {alertname: GpuXidError}
+    ladder:
+      - {action: mark_slot, params: {slot: "{slot}", status: degraded}, attempts: 1}
+      - {action: restart_vm, params: {pve_node: "{pve_node}", vmid: "{vmid}"}, attempts: 1, cooldown: 20m}
+      # ältere Karten ohne FLR brauchen ggf. einen Host-Neustart – nur mit Freigabe/Wartungsfenster
+      - {action: reboot_host, params: {pve_node: "{pve_node}"}, attempts: 1}
+    on_resolved:
+      - {action: mark_slot, params: {slot: "{slot}", status: ok}, only_if_set_by_healer: true}
+
+  - id: gpu-hot
+    match: {alertname: GpuTemperatureHigh}
+    ladder:
+      - {action: mark_slot, params: {slot: "{slot}", status: degraded}, attempts: 1}
+    on_resolved:
+      - {action: mark_slot, params: {slot: "{slot}", status: ok}, only_if_set_by_healer: true}
+
+  - id: kv-cache-saturated
+    match: {alertname: VllmKvCacheSaturated}
+    ladder:
+      - {action: notify, params: {text: "KV-Cache von {slot} dauerhaft voll – Last verteilen oder max_model_len senken", severity: warning}, attempts: 1}
+
+  - id: node-down
+    match: {alertname: NodeDown}
+    ladder:
+      - {action: notify, params: {text: "Node {instance} nicht erreichbar – Proxmox HA prüfen", severity: high}, attempts: 1}
+
+  - id: disk-full
+    match: {alertname: DiskAlmostFull}
+    ladder:
+      - {action: notify, params: {text: "Platte auf {instance} ({mountpoint}) fast voll", severity: warning}, attempts: 1}
@@ -0,0 +1,6 @@
+# Grenzwerte der Lifecycle-Pipeline (SPEC-lifecycle §2)
+tolerance: 0.02
+min_pass_rate: 0.85
+min_agreement: 0.85
+min_n: 50
+canary_share: 0.1
@@ -0,0 +1,11 @@
+# Grenzwerte der Service-Routine (SPEC-maintenance §3)
+demote_below: 0.80
+min_cases: 20
+window_cases: 50
+promote_min_rate: 0.95
+merge_similarity: 0.92
+degraded_hours: 1
+context_signal_limit: 10
+adapter_min_cases: 300
+adapter_target_rate: 0.90
+lookback_days: 7
@@ -0,0 +1,20 @@
+# Freigabe-Richtlinie (ADR-0004): sichere Richtung automatisch, Qualität/Kosten nach Freigabe.
+default: approve
+actions:
+  # automatisch
+  demote_skill:        {mode: auto, risk: low}
+  mark_slot:           {mode: auto, risk: low}
+  restart_container:   {mode: auto, risk: low}
+  restart_vm:          {mode: auto, risk: medium}
+  recalibrate_context: {mode: auto, risk: low}
+  notify:              {mode: auto, risk: low}
+  # mit Freigabe
+  request_graduation:  {mode: approve, risk: low}
+  promote_skill:       {mode: approve, risk: medium}
+  merge_skills:        {mode: approve, risk: low}
+  retire_skill:        {mode: approve, risk: medium}
+  train_adapter:       {mode: approve, risk: low}
+  create_candidate:    {mode: approve, risk: medium}
+  promote_candidate:   {mode: approve, risk: high}
+  rollback_candidate:  {mode: approve, risk: medium}
+  reboot_host:         {mode: approve, risk: high, window: "02:00-05:00"}