Providers, models and reasoning effort¶

AI IT House does not hard-code provider model names into the role knowledge. Instead, agents use workload classes that are resolved by the provider adapter.

Workload classes¶

Class	Intended workload	Typical use
`light`	low-cost, fast bounded work	routing, inventory, simple structured tasks
`standard`	mainstream professional work	most authoring, normal analysis, standard reviews
`heavy`	high-complexity work	architecture, policy, diagnosis, consequential synthesis
`research`	highest-complexity exploratory work	deep investigation, novel methods, high uncertainty

Default mapping in 26.10.4¶

Workload	Claude	Codex/OpenAI	Codex reasoning effort
`light`	`haiku`	`gpt-6-luna`	`low`
`standard`	`sonnet`	`gpt-6.1-sol`	`medium`
`heavy`	`opus`	`gpt-6-astra`	`high`
`research`	`opus`	`gpt-6-astra`	`xhigh`

Why reasoning effort matters¶

In the Codex mapping, reasoning effort is part of the routing policy. This lets the organisation align cost, latency and quality expectations by workload class rather than by editing every role definition.

Effective routing files¶

Package default:

~/.ai/it-role-house/models/model-registry.json

Effective merged copy used by the installation:

~/.ai/it-role-house/models/effective-model-registry.json

Responsibility boundary¶

A normal Claude/Codex user selects the business task and, where relevant, requests a stronger or cheaper quality/cost profile in the conversation. A deployment administrator owns these provider mappings and installs/validates model-routing overrides; changing environment variables or JSON files is a system-configuration action, not a per-engagement user step. Codex/Claude adapters resolve the effective workload class when dispatching agents. Do not infer successful isolation or a particular live model version from the mere presence of a mapping string; verify against the actual provider invocation and policy.

Organisation-specific override patterns¶

Default override path:

~/.config/ai-it-house/models.json

Environment variable:

export AI_HOUSE_MODEL_CONFIG=/secure/config/ai-it-house-models.json

Installation argument:

./install.sh --model-config /secure/config/ai-it-house-models.json

Example override¶

{
  "providers": {
    "codex": {
      "classes": {
        "standard": {
          "model_selector": "gpt-5.6-sol",
          "reasoning_effort": "medium"
        }
      }
    }
  }
}

Validation commands — administrator or CI¶

python3 ~/.ai/it-role-house/tools/model_router.py self-test
python3 ~/.ai/it-role-house/tools/provider_parity.py

These should be part of any controlled enterprise upgrade.