feat(routing): per-agent models and skill-based child routing

[kusonooyasumi]

Summary

Foundational change for fine-grained model control: each agent gets its own model instead of the whole scan being pinned to one model. Adds a default subagent model and ordered skill-based routing so specialist child agents can run on different models.

This is PR 1 of a split series. PR 2 (budget fallbacks) stacks on this one; the verification (#822) and dedupe (#823) PRs are independent.

What changes for users

• Nothing by default. With only STRIX_LLM set, every agent uses it exactly as before.
• STRIX_LLM still works — it's now a back-compat alias for STRIX_ORCHESTRATOR_MODEL.
• Opt-in: give subagents their own model, and route specific skills to specific models.

Config

Env var	Meaning
STRIX_ORCHESTRATOR_MODEL / STRIX_LLM	Root/orchestrator model (required)
STRIX_SUBAGENT_MODEL	Default child-agent model (falls back to orchestrator)
STRIX_SUBAGENT_REASONING_EFFORT	Reasoning effort for children
SUBAGENT_LLM_API_KEY	Optional provider key for the subagent provider
--subagent-model	One-run CLI override of the default child model

Skill routing is configured via the structured llm.skill_model_routes list (ordered; first match wins):

{
  "llm": {
    "model": "openai/gpt-5.4",
    "subagent_model": "deepseek/deepseek-chat",
    "skill_model_routes": [
      {"skill": "business_logic", "model": "anthropic/claude-sonnet-4-6"},
      {"operator": "AND", "skills": ["oauth", "authentication_jwt"], "model": "openai/gpt-5.4"},
      {"operator": "OR", "skills": ["xss", "sql_injection"], "model": "dashscope/qwen-plus"},
      {"operator": "ANY", "model": "deepseek/deepseek-chat"}
    ]
  }
}

Precedence: explicit create_agent(model=...) → matching skill route → subagent_model → orchestrator model.

Key implementation notes

• RunConfig.model/model_settings are now None. This is essential — RunConfig values override every per-agent route. Each SandboxAgent carries its own model + settings instead.
• Resolved child models are persisted per agent so a resumed scan reuses the same model.
• chat_completions tool-schema detection is computed per child from its resolved model.
• Provider-aware API-key mirroring stops the orchestrator credential leaking into a differently-routed child provider.

Test plan

• pytest tests/test_model_routing.py — build-agent model/settings, skill-route operators, factory precedence + per-model schema, orchestrator fallback, and per-agent usage recording.
• Full suite green (248), ruff + mypy clean. Also fixes the pre-existing test_runner_root_prompt.py mock (missing runtime).

[greptile-apps]

Greptile Summary

This PR adds model selection and routing for individual agents. The main changes are:

• Per-agent model and reasoning settings.
• Ordered skill-based child routing with explicit overrides.
• Persisted child models for resumed scans.
• Structured configuration and a child-model CLI option.
• Provider-aware credentials, model warm-up, and per-agent usage reporting.

Confidence Score: 4/5

The child credential paths need fixes before merging.

• Same-provider child requests can use the orchestrator key instead of the configured child key.
• Explicit child models can select a provider that never receives the configured child credential.
• The routing, persistence, and structured configuration paths otherwise appear consistent.

strix/config/models.py and the explicit child-model creation path

Security Review

A distinct child key cannot replace an already mirrored orchestrator key for the same provider. Child requests can therefore run with the wrong account or permissions.

Important Files Changed

Filename	Overview
strix/agents/factory.py	Adds per-agent model settings, route precedence, and model-specific tool schemas.
strix/config/settings.py	Adds validated child-model settings and ordered skill-route rules.
strix/config/loader.py	Loads structured configuration while preserving environment precedence and existing sections.
strix/config/models.py	Adds child credential mirroring, but same-provider keys and explicit model overrides are not handled correctly.
strix/core/execution.py	Propagates explicit child models and restores persisted model selections.
strix/interface/main.py	Adds the child-model CLI override and warms all statically configured models.
strix/report/usage.py	Tracks native per-agent costs and allocates unassigned LiteLLM callback costs.

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
strix/config/models.py:129-133
**Child Key Cannot Replace Root Key**

When the orchestrator and child use the same provider with different keys, the root key is mirrored first and `_mirror_api_key_to_provider_env()` uses `setdefault`, so `SUBAGENT_LLM_API_KEY` cannot replace it. Child requests then run with the orchestrator account, permissions, and billing identity instead of the configured child credential.

### Issue 2 of 2
strix/config/models.py:128-133
**Explicit Routes Miss Credential Setup**

Credential mirroring only covers `subagent_model` and configured skill routes. A child created with the new explicit `create_agent(model=...)` override can select an unlisted provider, but `SUBAGENT_LLM_API_KEY` is never mirrored for that model, so its first request fails authentication unless the provider variable already exists independently.

_{Reviews (1): Last reviewed commit: "feat(routing): per-agent models and skil..." | Re-trigger Greptile}

[greptile-apps]

Child Key Cannot Replace Root Key

When the orchestrator and child use the same provider with different keys, the root key is mirrored first and _mirror_api_key_to_provider_env() uses setdefault, so SUBAGENT_LLM_API_KEY cannot replace it. Child requests then run with the orchestrator account, permissions, and billing identity instead of the configured child credential.

Prompt To Fix With AI

This is a comment left during a code review.
Path: strix/config/models.py
Line: 129-133

Comment:
**Child Key Cannot Replace Root Key**

When the orchestrator and child use the same provider with different keys, the root key is mirrored first and `_mirror_api_key_to_provider_env()` uses `setdefault`, so `SUBAGENT_LLM_API_KEY` cannot replace it. Child requests then run with the orchestrator account, permissions, and billing identity instead of the configured child credential.

How can I resolve this? If you propose a fix, please make it concise.

[greptile-apps]

Explicit Routes Miss Credential Setup

Credential mirroring only covers subagent_model and configured skill routes. A child created with the new explicit create_agent(model=...) override can select an unlisted provider, but SUBAGENT_LLM_API_KEY is never mirrored for that model, so its first request fails authentication unless the provider variable already exists independently.

Prompt To Fix With AI

This is a comment left during a code review.
Path: strix/config/models.py
Line: 128-133

Comment:
**Explicit Routes Miss Credential Setup**

Credential mirroring only covers `subagent_model` and configured skill routes. A child created with the new explicit `create_agent(model=...)` override can select an unlisted provider, but `SUBAGENT_LLM_API_KEY` is never mirrored for that model, so its first request fails authentication unless the provider variable already exists independently.

How can I resolve this? If you propose a fix, please make it concise.

[kusonooyasumi]

Addressed review feedback — child credentials

Two related credential bugs were flagged (both rooted in _mirror_api_key_to_provider_env() using os.environ.setdefault):

1. Child key can't replace root key (same provider). Provider env vars are global, so once configure_sdk_model_defaults() mirrored the orchestrator key first, SUBAGENT_LLM_API_KEY for the same provider was silently dropped — children ran on the orchestrator account.
2. Explicit create_agent(model=...) routes missed credential setup. A child selecting an unlisted provider never got SUBAGENT_LLM_API_KEY mirrored.

Fix (fix(routing): give child agents per-call credentials): stop relying on global env for per-route keys. Each child now sends its credential as a per-call api_key in model_settings.extra_args (which LiteLLM honors over env). _resolve_child_api_key picks SUBAGENT_LLM_API_KEY when set, else the orchestrator key only for same-provider children, else ambient auth. This runs for routed, default, and explicit-model children — fixing both issues. The key stays in memory (not persisted to the session snapshot).

New tests cover explicit-key precedence, same-vs-cross-provider reuse, and that the key reaches extra_args.