Convergence Policy Orchestrator & Autonomous Remediation Engine
Status
- Author: PER-COMMUNITY-SYSTEM-ARCHITECT
- Cycle: Cycle 75
- Status: Approved / Draft
1. Executive Summary & Context
The Autonomous Convergence Loop in ZQK orchestrates autonomous agent swarms, continuous verification gates, and background daemon execution to drive systems toward target invariants. However, unexpected environmental interruptions, flaky subprocess concurrency, resource exhaustion, or transient CI errors can threaten loop forward velocity.
Under POL-AGENT-TPM-INTAKE-FIREWALL-001, chat-reported bugs and failures must never distract or derail ongoing plan execution. Analogously, runtime pipeline and build telemetry failures must not stall execution loops or trigger unconstrained thrashing.
Cycle 75 introduces the Convergence Policy Orchestrator & Autonomous Remediation Engine (pkg/convergence), providing:
- Convergence Policy Rule Engine: Pluggable, structurally validated rule sets that match failure signatures from build and CI telemetry against predefined remediations.
- Deterministic Remediation Action Dispatcher: Safe dispatching of bounded self-healing actions (e.g., deterministic retry with jitter, cache purging, task isolation, and agent dispatch).
- Anti-Thrashing & Acyclic Loop Safeguards: Mathematical damping and fail-closed bounds that prevent cyclic remediation loops or runaway trigger cascades.
2. Core Architecture & Components
┌────────────────────────────────────────────────────────┐
│ Failure Telemetry Payload │
│ (Job Name, Step, Exit Code, Error Signature) │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ Convergence Policy Evaluator Engine │
│ - Match TargetFilter and Severity Classification │
│ - Evaluate Structural Weight & Invariant Bounds │
└───────────────────────────┬────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ Remediation Dispatcher & Guard │
│ - Cyclic Trigger Prevention (Acyclic Check) │
│ - Max Retry & Exponential Backoff Clamping │
│ - Deterministic Remediation Execution Receipt │
└────────────────────────────────────────────────────────┘
2.1 Domain Models & Interfaces
package convergence
import (
"context"
"time"
)
type ImpactClassification string
const (
ImpactCritical ImpactClassification = "critical"
ImpactHigh ImpactClassification = "high"
ImpactMedium ImpactClassification = "medium"
ImpactLow ImpactClassification = "low"
)
type ActionType string
const (
ActionRetry ActionType = "retry"
ActionCachePurge ActionType = "cache_purge"
ActionTaskIsolation ActionType = "task_isolation"
ActionAgentDispatch ActionType = "agent_dispatch"
)
type PolicyRule struct {
ID string `json:"id" yaml:"id"`
Name string `json:"name" yaml:"name"`
TargetFilter string `json:"target_filter" yaml:"target_filter"`
Impact ImpactClassification `json:"impact" yaml:"impact"`
Action ActionType `json:"action" yaml:"action"`
MaxRetries int `json:"max_retries" yaml:"max_retries"`
BackoffBaseMs int `json:"backoff_base_ms" yaml:"backoff_base_ms"`
Weight int `json:"weight" yaml:"weight"`
Enabled bool `json:"enabled" yaml:"enabled"`
}
type TelemetryPayload struct {
SessionID string `json:"session_id"`
JobID string `json:"job_id"`
StepName string `json:"step_name"`
ExitCode int `json:"exit_code"`
ErrorMsg string `json:"error_msg"`
Metadata map[string]string `json:"metadata,omitempty"`
}
type RemediationReceipt struct {
ReceiptID string `json:"receipt_id"`
RuleID string `json:"rule_id"`
Action ActionType `json:"action"`
SessionID string `json:"session_id"`
Attempt int `json:"attempt"`
BackoffDelay time.Duration `json:"backoff_delay"`
DispatchedAt time.Time `json:"dispatched_at"`
Executed bool `json:"executed"`
}
2.2 Invariant Guarantees & Safeguards
- Static Floor Invariant:
-
PolicyRuleconfigurations must satisfy immutable structural bounds:IDandNamecannot be empty;Impactmust be one of the 4 defined enum constants;Weightmust be within[1, 100];MaxRetriesmust be within[0, 10];BackoffBaseMsmust be positive. - Operational Dynamic Proof:
- Given a failure telemetry payload matching an enabled rule's filter, the orchestrator evaluates the highest-weight matching rule and safely dispatches a
RemediationReceiptwith correctly calculated backoff delay. - Negative Adversarial Boundary:
- Cyclic triggers (repeated failure loops exceeding
MaxRetrieson the same session/job) fail-closed by rejecting automatic dispatch and transitioning the session to quarantine. - Corrupted or invalid payloads (e.g. emptySessionID, negative exit codes, or unregistered actions) are rejected with deterministic error returns.
3. Verification & Traceability
- Unified Test Case: Three-fold proof test suite
- Target Implementation:
pkg/convergence/orchestrator.go,pkg/convergence/evaluator.go,pkg/convergence/orchestrator_test.go