Summary

Add explicit model/provider overrides to delegate_task, including per-task overrides for parallel subagent batches.

This enables heterogeneous subagent runs where each child can use a different model/provider while keeping current default behavior unchanged when no override is provided.

Precedence:

1. per-task model
2. top-level model
3. delegation config
4. parent/default inheritance

Changes

• Add top-level model argument to delegate_task.
• Add per-task model field to the tool schema.
• Normalize model overrides from either:
  • structured object: {"provider": "openai-codex", "model": "gpt-5.4"}
  • internal bare string: "gpt-5.4"
• Resolve delegation credentials per child task so different tasks can use different models/providers.
• Forward model through AIAgent._dispatch_delegate_task().
• Add tests for:
  • schema exposure
  • model/provider override credential resolution
  • string model override preserving configured provider
  • top-level override forwarding
  • per-task override precedence
  • run_agent delegate dispatch forwarding

Test plan

source venv/bin/activate
pytest -q tests/tools/test_delegate.py tests/run_agent/test_run_agent.py -q

Result: passed.

Live smoke test performed locally after applying the combined patch:

• gpt-5.5 subagent completed
• gpt-5.4 subagent completed
• gpt-5.3-codex-spark subagent completed
• each child reported the expected model and successful terminal tool trace