TealReliability

TealReliability is a governance module that enforces reliability policies — retry budgets, circuit breakers, fallback chains, and degradation strategies. It extends the v1.1 TealCircuit pattern into a full reliability governance layer.

Overview

AI agents can fail in expensive ways: retrying endlessly, cascading failures across providers, or burning budget on degraded models. TealReliability governs these failure modes at the policy level.

Retry Budgets

Cap retry attempts per request, session, or agent

Circuit Breakers

Open circuits on repeated failures (extends TealCircuit)

Fallback Chains

Ordered fallback models when primary is unavailable

Degradation Policies

Graceful degradation instead of hard failure

Quick Start

import { TealEngineV12, TealReliability } from 'tealtiger';

const engine = new TealEngineV12({
  policy: {
    reliability: {
      retry_budget: 3,
      circuit_threshold: 5,
      fallback_model: 'gpt-3.5-turbo',
      degrade_on_timeout: true
    }
  },
  modules: [new TealReliability()]
});

const decision = await engine.evaluateV12(
  { model: 'gpt-4', retry_count: 4 },
  { correlation_id: 'req-001' }
);

console.log(decision.action);        // 'DENY'
console.log(decision.reason_codes);  // ['RETRY_BUDGET_EXCEEDED']

Policy Configuration

Configure TealReliability through the reliability key in your policy:

const engine = new TealEngineV12({
  policy: {
    reliability: {
      // Maximum retry attempts before denial
      retry_budget: 3,

      // Consecutive failures before circuit opens
      circuit_threshold: 5,

      // Time before circuit transitions to half-open (ms)
      circuit_reset_timeout: 30000,

      // Ordered fallback models
      fallback_chain: ['gpt-4o-mini', 'gpt-3.5-turbo'],

      // Fallback model (shorthand for single fallback)
      fallback_model: 'gpt-3.5-turbo',

      // Degrade instead of deny on timeout
      degrade_on_timeout: true,

      // Maximum latency before degradation (ms)
      latency_threshold: 5000
    }
  },
  modules: [new TealReliability()]
});

Retry Budgets

Retry budgets cap the number of retry attempts for a given request. Once the budget is exhausted, the module returns DENY with RETRY_BUDGET_EXCEEDED.

// Request with retry_count exceeding budget
const decision = await engine.evaluateV12(
  { model: 'gpt-4', retry_count: 4 },  // budget is 3
  { correlation_id: 'req-001' }
);

// decision.action === 'DENY'
// decision.reason_codes === ['RETRY_BUDGET_EXCEEDED']

Circuit Breakers

TealReliability includes circuit breaker logic that extends the v1.1 TealCircuit pattern. Circuits track failure counts per model/tool and transition through three states:

State	Behavior
Closed	Normal operation — requests pass through
Open	Failures exceeded threshold — requests are denied
Half-Open	After reset timeout — one probe request allowed

// Circuit is open after 5 consecutive failures
const decision = await engine.evaluateV12(
  { model: 'gpt-4' },
  { correlation_id: 'req-002' }
);

// decision.action === 'DENY'
// decision.reason_codes === ['CIRCUIT_OPEN']

When the circuit is half-open, the module returns DEGRADE to allow a probe request:

// decision.action === 'DEGRADE'
// decision.reason_codes === ['CIRCUIT_HALF_OPEN']

Fallback Chains

When a primary model is unavailable (circuit open), TealReliability can suggest a fallback:

const engine = new TealEngineV12({
  policy: {
    reliability: {
      circuit_threshold: 3,
      fallback_chain: ['gpt-4o-mini', 'gpt-3.5-turbo']
    }
  },
  modules: [new TealReliability()]
});

const decision = await engine.evaluateV12(
  { model: 'gpt-4' },  // circuit is open for gpt-4
  { correlation_id: 'req-003' }
);

// decision.action === 'DEGRADE'
// decision.reason_codes === ['FALLBACK_TRIGGERED']
// decision.metadata.fallback_model === 'gpt-4o-mini'

Degradation Policies

Instead of hard denial, TealReliability can degrade service quality:

const decision = await engine.evaluateV12(
  { model: 'gpt-4', latency_ms: 6000 },  // exceeds 5000ms threshold
  { correlation_id: 'req-004' }
);

// decision.action === 'DEGRADE'
// decision.reason_codes === ['DEGRADE_TRIGGERED']

Reason Codes

Code	Severity	Default Action	When
`CIRCUIT_OPEN`	high	DENY	Circuit breaker is open
`CIRCUIT_HALF_OPEN`	medium	DEGRADE	Circuit is in half-open probe state
`RETRY_BUDGET_EXCEEDED`	high	DENY	Retry attempts exceeded budget
`FALLBACK_TRIGGERED`	medium	DEGRADE	Fallback model activated
`DEGRADE_TRIGGERED`	medium	DEGRADE	Degradation policy applied

Event Types

Event Type	Description
`reliability.retry`	Retry attempt tracked
`reliability.fallback`	Fallback model activated
`reliability.degrade`	Degradation policy applied
`reliability.circuit`	Circuit breaker state change

Python

from tealtiger import TealEngineV12, TealReliability

engine = TealEngineV12(
    policy={
        "reliability": {
            "retry_budget": 3,
            "circuit_threshold": 5,
            "fallback_chain": ["gpt-4o-mini", "gpt-3.5-turbo"]
        }
    },
    modules=[TealReliability()]
)

decision = await engine.evaluate_v12(
    request={"model": "gpt-4", "retry_count": 4},
    ctx={"correlation_id": "req-005"}
)

Getting Started

Core Concepts

Governance Domains

Policy Authoring

Architecture

Integrations

Deployment

API Reference

Cookbook

Compliance & GRC

Guides

Reference

Versions

Playground

Dashboard

Blog

Overview

Retry Budgets

Circuit Breakers

Fallback Chains

Degradation Policies

Quick Start

Policy Configuration

Retry Budgets

Circuit Breakers

Fallback Chains

Degradation Policies

Reason Codes

Event Types

Python

​Overview

Retry Budgets

Circuit Breakers

Fallback Chains

Degradation Policies

​Quick Start

​Policy Configuration

​Retry Budgets

​Circuit Breakers

​Fallback Chains

​Degradation Policies

​Reason Codes

​Event Types

​Python

​Related Documentation

Overview

Quick Start

Policy Configuration

Retry Budgets

Circuit Breakers

Fallback Chains

Degradation Policies

Reason Codes

Event Types

Python

Related Documentation