Summary

Add optional model and provider fields to each item in the delegate_task tasks array, allowing different subagents in a batch to run on different models or providers.

Semantics

Implementation

In tools/delegate_tool.py:

1. Credential resolution (line ~1996): renamed single creds to default_creds to prepare for per-task resolution.
2. Per-task resolution (line ~2058): when t["provider"] differs from the delegation default, calls _resolve_delegation_credentials() again with the overridden provider to get the full credential bundle. When only t["model"] is set, copies default creds and overrides the model name.
3. Schema (line ~2759): added model and provider fields to the task item JSON schema.

No changes needed in AIAgent.__init__ or the agent loop — _build_child_agent already resolves effective_model = model or parent_agent.model.

Testing

• ✅ All 152 existing delegate tests pass
• ✅ Credential re-resolution tested for all 4 scenarios (default, model-only, provider-only, both)
• ✅ Unknown provider raises clean ValueError listing available providers
• ✅ End-to-end live test through delegate_task tool confirmed correct model/provider on spawned subagents:
  • glm-4.6v via zai → reported MODEL=glm-4.6v, PROVIDER=zai
  • deepseek-chat via deepseek → reported MODEL=deepseek-chat, PROVIDER=deepseek
• ❌ Unknown model name passes through and fails at API level (consistent with main agent behavior)

Example usage

delegate_task(tasks=[
    {
        "goal": "Search arxiv for recent papers",
        "toolsets": ["web"],
        "model": "deepseek-chat",         # model override only
    },
    {
        "goal": "Summarize the results",
        "toolsets": ["terminal", "file"],
        "provider": "deepseek",            # full provider re-resolve
        "model": "deepseek-chat",          # + model within that provider
    },
])

Closes #34764