2.3.2 Version stamps, fact sheets and reasoning timeouts

A new mdb no longer makes a site's own definitions stale, reasoning models get 300 seconds per call, and Load-Accelerator-Tables can leave a site's fact sheets alone.

Upgrading from 2.3.1

Steps 1 and 2 are for every site; step 3 only where reasoning models are deployed.

1. Update the checkout to 2.3.2 and reinstall the CLI: pip install -e "Model-Definition-Builder/cli[viya]". The first mdb generate of a definition rewrites its .mdb-lock.json once, without the version of mdb.
2. mdb builders-import. A site that keeps definitions of its own loads its fact sheets with mdb load-facts and runs the job Load-Accelerator-Tables with factSheets=0 from now on.
3. Only for the reasoning models you deploy (azure_openai_env, claude_haiku_5_5, claude_sonnet_5_5, claude_opus_5_5, gpt_5_mini, gpt_56_sol, gpt_6_luna, gpt_61_sol): mdb register <id> --update and mdb publish <id> --wait, then mdb deploy <id> and apply the YAML, so that the Ingress lets a call take 300 seconds. For a definition of your own, set runtime.timeout_s where the model reasons (How long a call may take).

Added

• Load-Accelerator-Tables: factSheets. 1 (default) loads the fact sheets as before; 0 loads the release table only and leaves the fact sheet tables, so the job no longer replaces the rows of a site's own definitions with the sheets of the release.
• mdb setup names a project whose target variable differs from the one it creates, such as an Embedding Model Project from before 1.1.0 with response, and leaves it as it is.

Changed

• Reasoning models wait 300 seconds for the provider (runtime.timeout_s, was 60). mdb add gives a model that reasons 300; the eight shipped definitions with a reasoning_effort option are regenerated. The NGINX Ingress of the deployment templates allows 300 seconds (was the controller's default of 60).
• Start Here names mdb load-facts first for a missing fact sheet, and says that the job loads the sheets of the release.
• The tests run on every core. The [dev] extra installs pytest-xdist and the workflow runs each suite with -n auto; locally the mdb suite takes 3 minutes instead of 12 (CONTRIBUTING.md, Running the Tests).

Fixed

• A new release of mdb no longer makes generated files stale. The version in the first line of a score file is left out when mdb validate, mdb generate --check and mdb ship compare it, and lock files hash it without that line; a release of mdb alone rewrites no lock file.
• mdb import and mdb pull --import keep the timeout of the score code (they gave 60).
• .env.example lists TYPESAFE_API_KEY.
• Earlier changelog entries name what they left out: the fact sheets Load-Accelerator-Tables replaces, the overwrite of a node type with the id of Call LLM, deploy-rag-content creating a folder only with -c, the inputs of Create-LLM-Call.sas and the placeholder {{ JOB-NAME }} (2.3.0); six OpenRouter definitions (2.0.0); llama_33_70b, mistral_small_32 and the license classes of the fact sheets (1.1.0). The Administration Guide states SAS Viya 2026.09 as the release the accelerator needs.