Feature: Dual Sub-Agent Model Configuration (per task-type)

[TechFlipsi]

Feature Request: Dual Sub-Agent Model Configuration

Problem

Currently, delegation.model in config.yaml supports only ONE model for all sub-agents. Users who want specialized sub-agents (e.g., one for coding tasks, one for research tasks) must manually override the model per delegate_task call, which is error-prone and verbose.

Proposed Solution

Add a delegation.models (plural) configuration that allows defining multiple named sub-agent profiles, each with its own model, provider, and toolset defaults:

delegation:
  models:
    coding:
      model: minimax-m3
      provider: ollama-cloud
      toolsets: [terminal, file, web]
    research:
      model: gemma4:31b
      provider: ollama-cloud
      toolsets: [web, browser]
  # Keep existing single-model config as default/fallback
  model: minimax-m3
  provider: ollama-cloud

Then in delegate_task, reference the profile by name:

delegate_task(goal="Research X", subagent="research")
delegate_task(goal="Fix the build", subagent="coding")

Why This Matters

• Different models have different strengths (coding vs. text/research vs. vision)
• Avoids manual model override on every delegate_task call
• Makes the dual-agent pattern a first-class config feature
• The per-task model override (from Feature: Per-sub-agent provider/model selection in delegate_task #24263) works but requires remembering to set it every time — a config-level default per task type is more reliable

Current Workaround

Use the model parameter per delegate_task call (merged in #24263), but this requires the parent agent to remember which model to use for which task type every single time.

Additional Context

Discussed with user who runs: M3 for coding sub-agents + Gemma4 for research sub-agents. The manual override works but a named-profile config would be cleaner and more maintainable.

[alt-glitch]

Duplicate of #9459 — named subagent profiles for delegate_task (per-profile model/provider/toolset) is the same feature, with a saturated cluster of competing PRs (#7201, #34754, #18522). Per-task model override already shipped in #24263.

[teknium1]

We do not want this