Summary

Extends the per-task model parameter from #7586 with named tiers and a model discovery tool, so the agent can make informed delegation decisions without knowing deployment-specific model identifiers.

Motivation

With #7586, the agent can pass model="google/gemini-flash-2.0" to delegate_task. But:

• The agent needs to know exact model names for the deployment
• Different users have different models available
• The agent has no way to discover what's available at runtime

This is the complementary piece: tiers abstract model names, and list_models lets the agent discover what's available.

Proposed Changes

1. Model Tiers (delegation.model_tiers)

delegation:
  model_tiers:
    small: google/gemini-flash-2.0     # fast/cheap
    # medium: omit to inherit parent model
    large: anthropic/claude-opus-4-6   # complex reasoning

The agent can then say:

delegate_task(goal="List all .py files", model="small")        # → gemini flash
delegate_task(goal="Review this PR for security", model="large") # → opus
delegate_task(goal="Debug the auth flow")                        # → parent model (medium)

Tier names resolve to configured model names. Unknown tiers fall back to the parent model.

2. list_models Tool

Returns available models with tier assignments, context lengths, and providers:

{
  "models": [
    {"name": "qwen35-397b", "provider": "litellm", "context_length": 524288, "is_default": true},
    {"name": "gemma4-nothink", "provider": "litellm", "context_length": 262144, "tier": "small"},
    {"name": "claude-sonnet-4-6", "provider": "litellm", "context_length": 1000000, "tier": "large"}
  ],
  "tiers": {"small": "gemma4-nothink", "large": "claude-sonnet-4-6"},
  "default_model": "qwen35-397b"
}

Registered in the delegation toolset alongside delegate_task.

Use Cases

• Explorer subagent: delegate_task(goal="Find all files related to auth", model="small") — fast file discovery on a cheap model
• Peer review: delegate_task(goal="Review my implementation", model="large") — escalate to a stronger model
• Mixed batch: different model per task based on complexity
• Self-awareness: agent calls list_models to discover what's available before deciding

Why This Is Better Than smart_model_routing

We deployed and tested smart_model_routing (message-length heuristic) in production and disabled it (see #7905). Short messages like "yes" or "go ahead" can trigger the most complex operations — message length is a terrible proxy for task complexity.

Model-directed delegation is fundamentally better: the model has full context, knows the task complexity, and chooses the right model for each subagent. No heuristics, no false positives.

Relationship to Other PRs

• Feat/delegate model parameter #7586 (Labhund) — foundation: per-task model parameter on delegate_task. This PR builds on top.
• feat: add named delegation routing profiles #7201 — named routing profiles. Complementary but orthogonal — profiles are about predefined route bundles, tiers are about simple capability labels.
• is_local_endpoint misses Docker/Podman DNS names — stale timeout fires on local LLM proxies #7905 — our issue about is_local_endpoint + smart routing concerns.

Implementation

We have a working implementation on our fork. Happy to submit a PR once #7586 lands, or rebase on top of it.

Credit: Per-task model parameter design from @Labhund (#7586). Tier + list_models additions from our production deployment.