Skip to main content
Two tests decide which models Claude Code, Codex and OpenCode offer in the composer’s Harness and model picker. xpander tests the catalog models. An admin tests a model with Test model.

​
xpander’s harness test

xpander runs each harness against each catalog model from another vendor with four checks: a plain answer, a tool call and its result, a streamed answer in the harness’s own format, and a long prompt. A harness always lists its own vendor’s models. For the picker rules, see Any model on any harness.

​
Test model

Admins who may manage models open it in Models:
1

Open the provider

In Platform > AI governance > Models, open a connected provider.
2

Open Test model

In the model row’s ⋯ menu, choose Test model.
3

Run the agent checks

On Agent checks, pick Claude Code, Codex or OpenCode, then choose Run checks.

​
Agent checks

Each run checks that the model:
  • Is reachable through the gateway
  • Streams in the harness’s own format
  • Makes a function call and continues after the result
  • Accepts the harness’s system prompt layout
  • Accepts the harness’s thinking and effort settings
  • Accepts the harness’s full list of function definitions
  • Reports a full context in the harness’s own error. This check is Skipped when the model’s context length is not known yet
Each check shows Pass, Fixed, Fail or Skipped. A Fixed check means the gateway now adjusts that request for the model, and the fix is saved to the model’s request settings as Learned by Test model.
Test model panel for xpander_ai_vllm::muse-glimmer on the Agent checks tab with Claude Code selected and a Run checks button, six checks marked Pass, Reports a full context window in the client's own error shape marked Skipped because the model's context length is not known yet, and a Ready for Claude Code. card that ran 4d ago

Agent checks for Claude Code on a vLLM model: six checks passed, the context check skipped, and the result card. Shown with sample data.

Agent readiness in the provider’s Keys and models… shows each model’s last result per harness.

​
After the test

  • A model from your own provider shows in the picker for each harness whose checks are ready.
  • Every 2 minutes the gateway runs the same agent check on one model of each custom endpoint. See Endpoint health checks.

​
Next steps

Add your own provider

Connect an endpoint and pick its models.

Pick a harness and model

Choose what an agent runs on.