Summary

This PR adds a model parameter to the delegate_task tool, enabling per-call provider:model overrides for delegates. Previously, all delegates were forced to use the provider specified in the delegation config section with no way to override it dynamically.

Problem

The existing delegation system only allowed configuring a single provider for all delegate agents via the delegation config section. This was limiting for scenarios where different subtasks might benefit from different models or providers based on their requirements.

Solution

This change introduces a model parameter that accepts a dictionary with provider and model keys, allowing dynamic routing of delegates to any configured provider at runtime.

Key Changes

1. Added _resolve_per_call_model() helper function (~line 2203)

   • Takes a {provider, model} dict and resolves full credentials via resolve_runtime_provider()
   • Returns a credential bundle compatible with _resolve_delegation_credentials()

2. Added model parameter to delegate_task() function (~line 1802)

   • Optional parameter accepting Dict[str, str] with provider/model keys

3. Added per-call model override logic (~lines 1882-1893)

   • When model is provided, resolves credentials and merges on top of config-based creds
   • Per-call overrides take precedence over delegation config

4. Added per-task model override support (~lines 1943-1952)

   • Each task in the tasks array can specify its own {provider, model}
   • Task-level overrides take precedence over top-level overrides

5. Updated DELEGATE_TASK_SCHEMA

   • Added model property at top-level (~lines 2535-2557)
   • Added model property at per-task level (~lines 2479-2493)

6. Updated registry handler (~line 2581)

   • Now passes model=args.get("model") to the function

7. Fixed credential references (~lines 1948-1968)

   • Changed creds references to task_creds to support per-task overrides

Example Usage

# Top-level override - all tasks use this provider
delegate_task(
    tasks=[{"goal": "Task 1"}, {"goal": "Task 2"}],
    model={"provider": "anthropic-proxy", "model": "claude-opus-4-8"}
)

# Per-task override - each task uses different provider
delegate_task(
    tasks=[
        {"goal": "Complex reasoning task", "model": {"provider": "anthropic-proxy", "model": "claude-opus-4-8"}},
        {"goal": "Simple extraction task", "model": {"provider": "openai", "model": "gpt-4-turbo"}}
    ]
)

Testing

The changes maintain backward compatibility - existing code without the model parameter continues to work as before, using the delegation config settings.

Impact

This enhancement enables more sophisticated delegation strategies where different subtasks can be dynamically routed to the most appropriate provider/model combination based on task requirements, cost considerations, or performance needs.