[Feature]: Dynamic model routing for sub-agent dispatch

[bordenet]

What would you like to be added?

Add an optional model parameter to the sub-agent dispatch interface, allowing the orchestrating agent to specify which model a sub-agent should use at invocation time — instead of only at configuration time via frontmatter.

Why is this needed?

When an AI agent executes a multi-phase plan, different steps have wildly different cognitive demands. A real example: a 7-phase codebase quality uplift needs:

• Sonnet 4.6 for ~85% of steps (refactoring, test writing, CI config — mechanical + judgment work)
• Opus 4.6 for ~15% of steps (taxonomy design, quality regression analysis, release decisions)
• A code review agent between phases (PR review)

Today, the user must manually switch models between sessions. This creates:

1. Wasted compute — running Opus on mechanical tasks costs 5–7× more than Sonnet with zero quality gain
2. Context reload cost — each new session re-reads the codebase and plan from scratch
3. Coordination overhead — the user becomes a manual router between AI models
4. Lost continuity — the orchestrator loses state at each switch point

What exists today (90% of the way there)

Sub-agents already support static model assignment via frontmatter:

---
name: code-review
model: sonnet4.5
---

And the orchestrator can dispatch to named sub-agents. The infrastructure handles model selection at the config level. The gap: the orchestrator cannot override the model at dispatch time.

Proposed solution

Add an optional model parameter to the sub-agent tool:

{
  "name": "sub-agent",
  "parameters": {
    "action": "run",
    "name": "refactor-phase-1",
    "model": "sonnet4.6",
    "instruction": "Execute Phase 1 per the plan document"
  }
}

Behavior:

• If model is specified → override the agent's configured model for this invocation only
• If model is omitted → use the agent's configured model (current behavior, fully backwards compatible)
• The orchestrator stays on its own model — only the sub-agent switches
• Validate against auggie models list output

Why this is the right smallest change

The sub-agent infrastructure already handles model selection (via frontmatter). This just exposes it at dispatch time. Same codepath, new entry point.

Cost impact

For the real-world plan described above:

Approach	Relative Cost	User Effort
All Opus (no routing)	100%	Minimal
Manual HALT points	~45%	High (6 manual model switches)
Dynamic dispatch (proposed)	~45%	Minimal (orchestrator routes automatically)

~55% cost reduction with zero quality compromise and no user intervention.

Current workaround

Create pre-configured sub-agents for each model tier (sonnet-worker.md, opus-reviewer.md) and manually dispatch to the right one. This works but each sub-agent gets a cold-start context window and the user still sequences manually.

Future extension (not part of this request)

A declarative execution plan format (.augment/execution-plans/*.yml) that maps steps to models with dependency ordering — but that needs this primitive first.

Additional context

I maintain a skill framework (superpowers-plus) with 90+ skills and have a detailed analysis of which skills benefit from multi-model sub-agent dispatch. Happy to share the full analysis and cost modeling if useful.