Summary

Lets each task in a delegate_task batch target a different model/provider:

delegate_task(tasks=[
    {"goal": "Review for security",  "provider": "anthropic", "model": "claude-opus-4-7"},
    {"goal": "Review for performance", "provider": "openai",   "model": "gpt-5"},
    {"goal": "Review for clarity",     "provider": "openrouter", "model": "..."},
])

All three children spawn in parallel under the existing concurrency cap, each on its specified provider. Tasks without model/provider use the batch default (existing behaviour preserved).

Why

The current contract supports a single delegation.provider for the whole batch — every child runs on the same provider:model pair. That's the right default for fan-out research where the model choice is uniform, but it blocks several patterns where the value comes from diversity:

• Cross-model evaluation: same prompt to N different models, compare outputs.
• Adversarial review: each reviewer is a different model so blind spots don't compound.
• Cost-tiered routing: high-stakes task on a frontier model, secondary tasks on cheaper models.

Today users have to either:

• Run sequential delegate_task calls (loses parallelism), or
• Build their own thread pool calling the agent directly (loses delegation infrastructure: progress callbacks, approval forwarding, role propagation, interrupt handling).

Implementation

Three changes, all in tools/delegate_tool.py:

1. _resolve_task_credentials(task, delegation_creds) -> dict

New helper that resolves per-task creds. Returns the same shape as _resolve_delegation_credentials so the result drops into the existing _build_child_agent override_* kwargs without new plumbing.

Resolution priority: task.provider/model > delegation_creds.

Failure handling: if a task's provider doesn't resolve (typo, missing runtime config), log a warning and return delegation_creds unchanged. Why: failing the whole batch over one bad entry is worse UX than running that task on the default provider — the user can see the warning, fix the config, and the rest of the batch already produced useful output.

2. Per-task loop in delegate_task

One additional line per task before the _build_child_agent call:

task_creds = _resolve_task_credentials(t, creds)

Then _build_child_agent receives task_creds["..."] instead of creds["..."]. No structural changes to the loop.

3. Schema additions

Per-task entries can now declare model and provider fields. Top-level tasks[].properties.model and tasks[].properties.provider were missing — adding them lets the model legitimately request the override.

Compatibility

• Tasks without model/provider are unaffected: _resolve_task_credentials short-circuits and returns the batch creds.
• Top-level single-task mode (delegate_task(goal=...)) is unchanged — only the per-task loop is touched.
• _resolve_delegation_credentials is unchanged; _resolve_task_credentials reuses the same resolve_runtime_provider machinery.
• No new config keys.

Test plan

• Batch with all tasks defaulted: each child runs on delegation.provider
• Batch with one task specifying provider="anthropic": that task runs on Anthropic, others on default
• Batch with model-only override: provider stays, model swaps
• Batch with bad provider name: warning logged, task runs on batch default, other tasks unaffected
• Top-level single task mode unchanged

🤖 Generated with Claude Code