Skip to content

model_runtime_configuration.md

Latest commit

 

History

History
102 lines (68 loc) · 3.43 KB

model_runtime_configuration.md

File metadata and controls

102 lines (68 loc) · 3.43 KB
Raw

Model Runtime Configuration

Date reviewed: 2026-06-01

This repository runs locally by default. The OpenAI path is optional and intentionally limited to model-facing work:

  • Project 1: grounded answer wording after permission filtering, unsafe-content filtering, citation selection, and abstention logic.
  • Project 2: intent classification before deterministic tool permission and approval checks.

The model is not the security boundary.

See Model Gateway Safety for the gate that verifies OpenAI mode stays opt-in, API key references remain constrained, structured outputs are required, and gateway failures fall back to local behavior. See OpenAI Live Mode Troubleshooting for safe setup, failure modes, rollback, and review guardrails.

Default Optional Model

The default optional model is:

OPENAI_MODEL=gpt-5.2

Reason:

  • OpenAI's current model guide lists gpt-5.2 as the best model for coding and agentic tasks across industries.
  • OpenAI's GPT-5.2 guide recommends the Responses API for reasoning, tool-calling, and multi-turn use cases.
  • The repository keeps the local deterministic mode as the verified default so demos do not depend on paid API access.

References:

Environment Variables

$env:OPENAI_API_KEY="..."
$env:OPENAI_MODEL="gpt-5.2"
$env:OPENAI_REASONING_EFFORT="medium"
$env:OPENAI_TEXT_VERBOSITY="low"
$env:COPILOT_MODEL_PROVIDER="openai"
$env:OPS_AGENT_MODEL_ROUTER="openai"

With a real key, run the live proof:

python -B scripts/dev.py openai-live

That command starts both model-facing apps on isolated local ports with OpenAI mode enabled. It requires Project 1 to report model_provider=openai, Project 2 to report model_router=openai, and both model-facing workflows to keep their normal safety behavior.

Supported reasoning effort values:

none, low, medium, high, xhigh

Supported verbosity values:

low, medium, high

Project 1 Defaults

Project 1 uses:

OPENAI_REASONING_EFFORT=medium
OPENAI_TEXT_VERBOSITY=low

Rationale:

  • Medium effort is a reasonable default for grounded enterprise answers where precision matters.
  • Low verbosity keeps the generated answer concise because citations and evidence are already shown by application code.
  • Structured output forces the response into an answer/confidence/missing-evidence shape.

Project 2 Defaults

Project 2 uses:

OPENAI_REASONING_EFFORT=low
OPENAI_TEXT_VERBOSITY=low

Rationale:

  • Intent routing should be fast and constrained.
  • The model returns only an enum intent.
  • Approval decisions and side effects remain deterministic application behavior.

Technical Review Positioning

Use this wording:

Local deterministic mode remains the verified default, with an optional OpenAI Responses API path for the model-facing part of each workflow. Model, reasoning effort, verbosity, and structured outputs are tuned through environment variables, while permissions and side-effect authorization stay outside the model. When a key is available, `python -B scripts/dev.py openai-live` proves both model-facing apps used OpenAI mode while preserving citations, approvals, and side-effect blocking.