Summary

delegate_task (and delegation in general) does not inherit the global fallback_providers chain defined in config.yaml. This creates a single point of failure: if the configured delegation.provider goes down or hits a rate limit, all subagent spawns fail, even though the parent agent has a healthy fallback chain.

Current behavior

• config.yaml supports fallback_providers: (list) for the main agent.
• delegation: section supports only a single model + provider pair.
• When a subagent's primary provider fails, there is no automatic failover.
• Subagents do not inherit the parent's fallback_providers list.

Proposed behavior

Allow delegation: to accept its own fallback chain:

delegation:
  model: "gemini-2.5-flash"
  provider: "gemini"
  fallback_providers:
    - provider: openrouter
      model: google/gemini-2.5-flash:free
    - provider: custom
      model: gemma-4-26b-a4b-it-4bit
      base_url: "http://127.0.0.1:8001/v1"

Or alternatively, have subagents inherit the parent's fallback_providers by default (with an opt-out flag).

Motivation

In production multi-agent setups (e.g., migrating from OpenClaw-style agent orchestration), each sub-agent needs resilience equal to the parent. Hitting a Gemini quota or OpenRouter rate limit shouldn't kill delegation entirely. OpenClaw provided 3-4 stage fallback per agent, and users migrating to Hermes lose this resilience without equivalent functionality.

Related

• fallback-providers.md documents only global fallback.
• Cron jobs face the same limitation (separate issue).