fix(sre): type backup learning as read-only

apps/api/src/jobs/agent99_controlled_dispatch_reconciler_job.py

@@ -300,6 +300,60 @@ def _playbook_id(identity: Agent99DispatchIdentity) -> str:
     return f"PB-A99-{digest}"
 
 
+def _learning_profile(
+    identity: Agent99DispatchIdentity,
+    *,
+    mode: str,
+) -> dict[str, Any]:
+    if "backup_health" in identity.route_id:
+        return {
+            "description": (
+                "Verified Agent99 read-only BackupCheck with independent DR "
+                "evidence readback; no backup or restore write was performed."
+            ),
+            "affected_services": ["backup_restore"],
+            "keywords": [
+                "agent99", "backup", "restore", "escrow", "offsite", mode,
+            ],
+            "action_type": "agent99_readonly_check",
+            "expected_result": "independent backup/DR verifier passes",
+            "rollback_command": "not_applicable_no_write",
+            "tags": ["agent99", "controlled_dispatch", "backup_restore"],
+            "notes": (
+                "Backup/restore/delete/retention/escrow writes and raw secret "
+                "values are not permitted or persisted."
+            ),
+            "km_title": f"Agent99 verified backup readback {identity.incident_id}",
+            "km_content": (
+                f"Run {identity.run_id} on route {identity.route_id} completed "
+                f"read-only Agent99 {mode or 'BackupCheck'} with independent "
+                "backup/DR verifier success. No backup, restore, delete, "
+                "retention or escrow write was performed; LLM had no runtime authority."
+            ),
+            "km_category": "資料保護/DR Scorecard",
+        }
+    return {
+        "description": (
+            "Verified Agent99 controlled route with independent outcome "
+            "readback and same-run learning receipts."
+        ),
+        "affected_services": ["cold-start-gate"],
+        "keywords": ["agent99", "cold-start", mode],
+        "action_type": "agent99_mode",
+        "expected_result": "independent verifier passes",
+        "rollback_command": "Agent99 Status (no-write)",
+        "tags": ["agent99", "controlled_dispatch"],
+        "notes": "No raw logs or secret values are persisted.",
+        "km_title": f"Agent99 verified recovery {identity.incident_id}",
+        "km_content": (
+            f"Run {identity.run_id} on route {identity.route_id} completed "
+            f"Agent99 mode {mode or 'Status'} with independent verifier success. "
+            "Raw logs and secret values are not stored."
+        ),
+        "km_category": "AI自動化/Agent99受控修復",
+    }
+
+
 async def _ensure_playbook_trust(
     identity: Agent99DispatchIdentity,
     *,
@@ -307,6 +361,7 @@ async def _ensure_playbook_trust(
 ) -> str | None:
     playbook_id = _playbook_id(identity)
     run_tag = f"agent99_run:{identity.run_id}"
+    profile = _learning_profile(identity, mode=mode)
     now = now_taipei()
     try:
         async with get_db_context(identity.project_id) as db:
@@ -324,25 +379,22 @@ async def _ensure_playbook_trust(
                     playbook_id=playbook_id,
                     project_id=identity.project_id,
                     name=f"Agent99 {mode or 'Status'}: {identity.route_id}",
-                    description=(
-                        "Verified Agent99 controlled route with independent "
-                        "outcome readback and same-run learning receipts."
-                    ),
+                    description=profile["description"],
                     status="approved",
                     source="ai_extracted",
                     symptom_pattern={
                         "alert_names": [],
-                        "affected_services": ["cold-start-gate"],
+                        "affected_services": profile["affected_services"],
                         "severity_range": ["P0", "P1", "P2"],
                         "label_patterns": {},
-                        "keywords": ["agent99", "cold-start", mode],
+                        "keywords": profile["keywords"],
                     },
                     repair_steps=[{
                         "step_number": 1,
-                        "action_type": "agent99_mode",
+                        "action_type": profile["action_type"],
                         "command": mode,
-                        "expected_result": "independent verifier passes",
-                        "rollback_command": "Agent99 Status (no-write)",
+                        "expected_result": profile["expected_result"],
+                        "rollback_command": profile["rollback_command"],
                         "requires_approval": False,
                         "risk_level": "LOW",
                     }],
@@ -356,8 +408,8 @@ async def _ensure_playbook_trust(
                     trust_score=0.37,
                     approved_by="agent99_independent_verifier",
                     approved_at=now,
-                    tags=["agent99", "controlled_dispatch", run_tag],
-                    notes="No raw logs or secret values are persisted.",
+                    tags=[*profile["tags"], run_tag],
+                    notes=profile["notes"],
                     requires_approval_level="auto",
                     stateful_targets=[],
                     requires_pre_backup=False,
@@ -401,20 +453,16 @@ async def _ensure_km_writeback(
     mode: str,
 ) -> str | None:
     knowledge = get_knowledge_service()
+    profile = _learning_profile(identity, mode=mode)
     try:
         entry = await knowledge.create_entry(
             KnowledgeEntryCreate(
-                title=f"Agent99 verified recovery {identity.incident_id}",
-                content=(
-                    f"Run {identity.run_id} on route {identity.route_id} "
-                    f"completed Agent99 mode {mode or 'Status'} with independent "
-                    "verifier success. Raw logs and secret values are not stored."
-                ),
+                title=profile["km_title"],
+                content=profile["km_content"],
                 entry_type=EntryType.INCIDENT_CASE,
-                category="AI自動化/Agent99受控修復",
+                category=profile["km_category"],
                 tags=[
-                    "agent99",
-                    "controlled_dispatch",
+                    *profile["tags"],
                     f"automation_run_id:{identity.run_id}",
                 ],
                 source=EntrySource.AI_EXTRACTED,

apps/api/tests/test_agent99_controlled_dispatch_reconciler_job.py

@@ -51,6 +51,26 @@ def test_agent99_playbook_identity_is_project_scoped() -> None:
     assert job._playbook_id(IDENTITY) != job._playbook_id(other)
 
 
+def test_backup_learning_profile_is_typed_and_no_write() -> None:
+    identity = build_agent99_dispatch_identity(
+        project_id="awoooi",
+        incident_id="INC-20260719-BACKUP-LEARNING",
+        source_fingerprint="backup-learning-fingerprint",
+        route_id="agent99:backup_health:BackupCheck",
+    )
+
+    profile = job._learning_profile(identity, mode="BackupCheck")
+
+    assert profile["affected_services"] == ["backup_restore"]
+    assert profile["action_type"] == "agent99_readonly_check"
+    assert profile["rollback_command"] == "not_applicable_no_write"
+    assert profile["km_category"] == "資料保護/DR Scorecard"
+    assert "verified backup readback" in profile["km_title"]
+    assert "LLM had no runtime authority" in profile["km_content"]
+    assert "cold-start" not in str(profile)
+    assert "verified recovery" not in profile["km_title"]
+
+
 class _Context:
     def __init__(self, db) -> None:  # type: ignore[no-untyped-def]
         self.db = db