mirror of
https://github.com/docker/docker-agent.git
synced 2026-11-04 18:30:09 +00:00
Remove the leading '#!/usr/bin/env docker agent run' shebang and following blank line from all example YAML and HCL files, as the config format is now self-documenting.
47 lines
1.7 KiB
YAML
47 lines
1.7 KiB
YAML
# Example: Fallback Models for High Availability
|
|
#
|
|
# This configuration demonstrates how to set up fallback models for agents.
|
|
# The fallback system handles different error types:
|
|
#
|
|
# - Retryable errors (e.g. 5xx, timeouts): Retry the same model with exponential backoff + jitter
|
|
# - Non-retryable errors (e.g. 429 rate limit, 4xx client errors): Skip to next model immediately
|
|
#
|
|
# After a non-retryable error triggers a fallback, the runtime "sticks" with
|
|
# the successful fallback model for the cooldown duration before retrying the primary.
|
|
#
|
|
# Sensible defaults are included, most users only need to specify fallback models
|
|
|
|
agents:
|
|
# Minimal fallback configuration - just specify models, defaults handle the rest
|
|
root:
|
|
model: anthropic/claude-sonnet-4-0
|
|
fallback:
|
|
models:
|
|
- openai/gpt-4o
|
|
- openai/gpt-5-mini
|
|
description: A reliable assistant with automatic failover
|
|
instruction: |
|
|
You are a helpful assistant. Your responses should be clear and concise.
|
|
You have built-in resilience - if one model provider is unavailable,
|
|
the system will automatically try alternative providers.
|
|
toolsets:
|
|
- type: think
|
|
|
|
# You can also use named models from the models section as fallbacks
|
|
# models:
|
|
# fast_backup:
|
|
# provider: openai
|
|
# model: gpt-5-mini
|
|
|
|
# Example with explicit configuration (for power users):
|
|
# agents:
|
|
# root:
|
|
# model: anthropic/claude-sonnet-4-0
|
|
# fallback:
|
|
# models:
|
|
# - fast_backup
|
|
# - google/gemini-2.5-flash
|
|
# retries: 3 # More retries for flaky networks
|
|
# cooldown: 5m # Longer cooldown for persistent rate limits
|
|
# # Use retries: -1 to disable retries (try each model only once)
|