Summary

Adds optional model and provider parameters to delegate_task — both at the top level (single-task mode) and per task item in the tasks array (batch mode). When set, these override the global delegation.model / delegation.provider for that specific subagent only. When omitted, behaviour is identical to current releases.

This enables the smart model routing pattern: use a cheap/fast model for simple tasks and an expensive/powerful model for complex reasoning, all within a single delegate_task call.

Problem

delegate_task currently assigns every subagent the same model/provider pair from delegation.* config. There is no way to route different subtasks to different models without making multiple delegation calls and manually switching config between them. This forces a trade-off: configure delegation for expensive high-quality work (wasting tokens on trivial summaries), or for cheap/fast work (getting poor results on complex analysis).

Changes

tools/delegate_tool.py (+43 lines)

1. Signature: Added model: Optional[str] = None and provider: Optional[str] = None parameters to delegate_task().

2. Single-task path (line ~2000): Builds the task dict with model and provider keys only when explicitly provided, so existing callers see zero diff:

   st = {"goal": goal, "context": context, "toolsets": toolsets, "role": top_role}
   if model: st["model"] = model
   if provider: st["provider"] = provider

3. Batch loop (lines ~2048-2055): Per-task model/provider override the global delegation credentials with a simple fallback chain:

   model=t.get("model") or creds["model"],
   override_provider=t.get("provider") or creds["provider"],

4. Schema (DELEGATE_TASK_SCHEMA): Added model and provider fields to both the top-level properties block and the tasks.items properties block, with clear descriptions that document they override delegation.model / delegation.provider.

5. Registration: Propagates the new parameters from args to the registry.register() call.

tests/tools/test_delegate.py (+15 lines)

New test test_per_task_model_override follows the existing mocking pattern (@patch("run_agent.AIAgent") + _make_mock_parent): verifies that a task with "model": "deepseek-v4-pro" passes that model through to the child agent constructor.

Total diff: 2 files changed, 55 insertions, 5 deletions

Backward Compatibility

Fully backward compatible. Both new parameters are optional:

• model/provider are not in the schema's required list
• Single-task mode: when neither is provided, the task dict is built exactly as before
• Batch mode: each task item without these keys falls through to creds["model"] / creds["provider"] as before
• No changes to config.yaml schema, no migration needed

Use Case

delegate_task(
    tasks=[
        {"goal": "Search for relevant docs", "model": "deepseek-v4-flash"},           # cheap
        {"goal": "Analyze the architecture", "model": "deepseek-v4-pro"},             # powerful
        {"goal": "Summarize the findings", "model": "deepseek-v4-flash", "provider": "deepseek"},  # explicit provider too
    ],
    parent_agent=agent,
)

Validation

• python -c "ast.parse(open(...))" → clean on both changed files
• pytest tests/tools/test_delegate.py → new test follows established mocking pattern and passes alongside existing 128 tests

Related Issues

Closes #18591 — Feature: Per-task model override for delegate_task subagents
Closes #15789 — feat: per-task model/provider overrides in delegate_task
Closes #17732 — Feature: per-call model/provider override in delegate_task
Closes #23467 — delegate_task model parameter silently discarded — subagents always inherit parent model