Summary

Allow the parent agent to specify model, provider, and optionally base_url on each delegate_task call (and per-task in batch mode), so it can dynamically route subagents to different models based on task type.

Motivation

Currently, subagent model selection is static via delegation.model / delegation.provider in config.yaml. This means ALL subagents use the same model. In practice, different tasks benefit from different models:

• Code tasks → fast, code-tuned model (e.g. glm-5.1)
• Research/reasoning → slower but stronger reasoning model (e.g. glm-5)
• Main thread → default configured model (e.g. glm-5-turbo)

The cronjob tool already supports per-job model/provider/base_url overrides — delegate_task should too. The plumbing already exists in _resolve_delegation_credentials() and _build_child_agent() (which already accepts override_provider/override_base_url/override_api_key/override_api_mode). This is mostly wiring the tool schema → handler path.

Proposed API

Single task mode

{
  "name": "delegate_task",
  "arguments": {
    "goal": "Implement the auth middleware",
    "context": "...",
    "model": "glm-5.1",
    "provider": "zai"
  }
}

Batch mode (per-task overrides)

{
  "name": "delegate_task",
  "arguments": {
    "tasks": [
      {
        "goal": "Research competitors",
        "model": "glm-5",
        "provider": "zai"
      },
      {
        "goal": "Write integration tests",
        "model": "glm-5.1"
      }
    ]
  }
}

Fallback chain

per-call model/provider → config delegation.model/provider → inherit from parent

This preserves backward compatibility — if no per-call override is given, behavior is identical to today.

Implementation notes

• _build_child_agent() already accepts model and override_provider — just needs to be wired from the tool arguments
• Need _resolve_delegation_credentials() or a new helper to resolve a one-off provider string (reusing resolve_runtime_provider()) when passed per-call
• Schema changes to DELEGATE_TASK_SCHEMA: add model (string) and provider (string) at top level, plus model/provider inside each batch task item
• Provider validation: resolve at delegation time, return clear error if provider has no API key
• Consider adding base_url as well for direct endpoint overrides (already supported in config)

Related

• Config-level delegation already exists (v0.2.0): delegation.model, delegation.provider, delegation.base_url
• Cronjob per-job overrides: model, provider, base_url — this feature would bring parity
• Feature: Acceptance Criteria & Independent Judge for Sub-agent Delegation (inspired by OpenPlanter) #356 (Independent Judge) would also benefit from per-call model selection for the judge agent