hermes - 💡(How to fix) Fix Feature: Per-sub-agent provider/model selection in delegate_task

Problem

Currently delegate_task sub-agents inherit the parent agent's provider and model. There's no way to say "use Claude Code for this sub-agent and DeepSeek for that one" within a single session.

Some orchestration tools let you assign different AI providers per sub-agent in the same session (Claude for reasoning, Codex for generation, Gemini for analysis). This is genuinely useful — different models have different strengths.

Proposed Solution

Add optional provider and model fields to the delegate_task task definition (both single and batch mode) so callers can specify different models per sub-agent.

What This Unlocks

• Parallel specialised execution: Each sub-agent uses the best model for its task type
• Cost optimisation: Use cheap models for simple review tasks, expensive ones for complex generation
• Provider diversity: If one provider is rate-limited, others can still make progress

Implementation Notes

• The AIAgent class already supports per-instance provider/model/base_url overrides in __init__
• Need to thread these through delegate_task and batch_runner.py
• Backward compatible: missing fields default to parent agent's config