Operator directive (2026-06-26, crystal-clear)
100% fully autonomous self-healing is MANDATORY. The escalate action gets DELETED. In an operatorless cloud deploy there is no human to page or ack, so a runtime escalate/page terminal is incoherent. Every diagnosis must route to an autonomous heal action. Where the actuator LACKS the heal logic → a design-time ticket to build it (this issue's subs), never a runtime escalate. This SUPERSEDES the earlier "mode-dependent escalate" convergence (mine included).
Scope (V-B-A'd surface)
RecoveryActuatorService: DEFAULT_ACTIONS = ['restart','redeploy','page','warm-provider'] → delete page; delete escalateDiagnosis() + every details.actionClass === 'escalate' path.
- ALL diagnosis producers currently emit
actionClass: 'escalate' — dataIntegrityCoverageDiagnosis, dimensionConsistencyDiagnosis, vectorCountMonotonicityDiagnosis, sqliteIntegrityDiagnosis, storeBloatDiagnosis, taskOutcomeDiagnosis, ContainerHealthDiagnosisService → re-route each to an autonomous heal action.
- The #14109 runner (
DataIntegrityDiagnosisService.routeDiagnoses) → route to heal actions, not escalateDiagnosis.
- ADR-0025 / ADR-0026 amendment: the escalate-terminal → autonomous-heal-terminal (the lifecycle-actuator envelope gains data-heal actions).
The autonomous heal actions
- Container-health (EXIST):
restart / redeploy / warm-provider → those producers route here autonomously today.
- Data-integrity (MISSING — the actuator lacks them → sub-tickets, per the #14032 strategy design): re-embed-missing (WAL-stall), WAL-replay, restore+delta-merge, quarantine-from-serving, auto-accept-loss-with-audit (#14084). Until built, the data producers route to the safest autonomous action (quarantine / durable-record) + the sub-ticket tracks the full heal. A durable autonomous record is NOT an escalate — no human is in the runtime loop.
Acceptance Criteria
Related
#14039 (v13.1 epic), #14032 (the data-recovery strategies → now autonomous actions), #14084 (auto-accepted-loss settlement), #14109 (the runner), #14131 (held — its escalate-only terminal is deleted by this).
Authored by Vega (Claude Opus 4.8, Claude Code).
Operator directive (2026-06-26, crystal-clear)
100% fully autonomous self-healing is MANDATORY. The
escalateaction gets DELETED. In an operatorless cloud deploy there is no human to page or ack, so a runtimeescalate/pageterminal is incoherent. Every diagnosis must route to an autonomous heal action. Where the actuator LACKS the heal logic → a design-time ticket to build it (this issue's subs), never a runtime escalate. This SUPERSEDES the earlier "mode-dependent escalate" convergence (mine included).Scope (V-B-A'd surface)
RecoveryActuatorService:DEFAULT_ACTIONS = ['restart','redeploy','page','warm-provider']→ deletepage; deleteescalateDiagnosis()+ everydetails.actionClass === 'escalate'path.actionClass: 'escalate'—dataIntegrityCoverageDiagnosis,dimensionConsistencyDiagnosis,vectorCountMonotonicityDiagnosis,sqliteIntegrityDiagnosis,storeBloatDiagnosis,taskOutcomeDiagnosis,ContainerHealthDiagnosisService→ re-route each to an autonomous heal action.DataIntegrityDiagnosisService.routeDiagnoses) → route to heal actions, notescalateDiagnosis.The autonomous heal actions
restart/redeploy/warm-provider→ those producers route here autonomously today.Acceptance Criteria
escalate+pagedeleted fromRecoveryActuatorService;escalateDiagnosis()removed.actionClass: 'escalate'remains).Related
#14039 (v13.1 epic), #14032 (the data-recovery strategies → now autonomous actions), #14084 (auto-accepted-loss settlement), #14109 (the runner), #14131 (held — its escalate-only terminal is deleted by this).
Authored by Vega (Claude Opus 4.8, Claude Code).