Summary

Add per-task model and provider fields to the tasks array in delegate_task, enabling different tasks in the same batch call to run on different models and providers.

Problem

Currently all tasks in a delegate_task call share the same model (resolved once from delegation.* config or inherited from parent). There's no way to route a lightweight task to haiku while a reasoning-heavy task uses opus in the same batch.

Tracked in #35437 (consolidating #5012, #10995, #17685, #23467, #32711). This implementation takes the minimal approach.

Solution

Two targeted changes, zero regressions:

1. Schema

Added model and provider as optional fields to each task object in the tasks array.

2. Loop logic

When a task specifies model or provider, _resolve_delegation_credentials is called with a task-scoped config so base_url, api_key, and api_mode are derived correctly from the per-task provider. Tasks without overrides reuse the pre-resolved creds dict (no overhead for existing callers).

Why re-resolve via _resolve_delegation_credentials?

Passing only model without re-resolving credentials would silently break cross-provider overrides — e.g. provider: "nvidia" needs the NVIDIA base_url and API key, not just a model name swap. Re-resolving per-task ensures the full credential bundle is correct.

Usage example

delegate_task(tasks=[
    {
        "goal": "Summarize this document",
        "model": "claude-haiku-4-5",
        "provider": "anthropic",
        "toolsets": ["file"]
    },
    {
        "goal": "Design the system architecture",
        "model": "claude-opus-4-5",
        "provider": "anthropic",
        "toolsets": ["web", "file"]
    }
])

Testing

• Schema import confirms model and provider appear in task properties
• Existing callers omitting the new fields hit the task_creds = creds path (no change)
• Lint: clean

Closes #35437