Bug Report: delegate_task sends API key to wrong provider endpoint

Environment

• Hermes version: Latest (as of 2026-05-29)
• OS: macOS 26.5
• Profile: default + pi-dev-ops

Configuration

~/.hermes/config.yaml (default profile)

model:
  default: qwen/qwen3-asr-flash-2026-02-10
  provider: openrouter
  base_url: ''
providers: {}

delegation:
  model: qwen/qwen3.7-max
  provider: openrouter
  base_url: ''
  api_key: ''

~/.hermes/profiles/pi-dev-ops/config.yaml

model:
  default: gpt-5.5
  provider: openai-codex
  base_url: https://chatgpt.com/backend-api/codex

delegation:
  model: qwen/qwen3.7-max
  provider: openrouter
  base_url: ''
  api_key: ''

~/.hermes/.env

OPENROUTER_API_KEY=sk-or-v1-***

Expected Behavior

When delegate_task is called, it should:

1. Read the delegation.provider config (set to openrouter)
2. Read the delegation.model config (set to qwen/qwen3.7-max)
3. Send the request to OpenRouter's API endpoint using the OPENROUTER_API_KEY

Actual Behavior

delegate_task sends the OpenRouter API key (sk-or-v1-***) to the OpenAI API endpoint (https://api.openai.com/v1/chat/completions), resulting in a 401 error:

Error code: 401 - {
  'error': {
    'message': 'Incorrect API key provided: sk-or-v1-***. 
                You can find your API key at https://platform.openai.com/account/api-keys.',
    'type': 'invalid_request_error',
    'code': 'invalid_api_key'
  }
}

Reproduction Steps

1. Set delegation.provider = openrouter in config.yaml
2. Set OPENROUTER_API_KEY in .env
3. Call delegate_task with any goal
4. Observe 401 error with OpenRouter key sent to OpenAI endpoint

Additional Context

• The main agent correctly uses OpenRouter (model: qwen/qwen3-asr-flash-2026-02-10)
• Only delegate_task (subagent spawning) exhibits this routing bug
• Tried: hermes config set delegation.provider openrouter, restarting gateway
• The error persists across gateway restarts
• The model field in the error shows gpt-5.5 (not the configured qwen/qwen3.7-max)

Hypothesis

The delegate_task implementation may be:

1. Ignoring the delegation.provider config and defaulting to openai
2. Using the main agent's model (gpt-5.5) instead of delegation.model
3. Not resolving the API key through the credential pool correctly for subagents

Impact

• Blocking: Cannot use live subagent execution with OpenRouter
• Workaround: Execute tasks directly in parent agent (loses parallelization)
• Severity: High — breaks core multi-agent workflow feature

Suggested Fix

Ensure delegate_task reads from delegation.* config keys (not model.*) and routes API calls to the correct provider endpoint based on delegation.provider.