name	litellm-integration
description	Use when testing openclient-llm against a real LiteLLM server, OpenAI-compatible endpoint, virtual key, streaming SSE, tool calling, or model integration.

LiteLLM Integration Testing

Use this skill only for opt-in tests against a real LiteLLM deployment.

Required Input

Ask for both values before making authenticated requests:

The exact OpenAI-compatible base URL entered in the app, for example https://litellm.example.com/v1.
A temporary virtual key with the minimum permissions required for the tests.

Never store either value in source files, Secrets.xcconfig, Xcode settings, test fixtures, Git, or logs. Do not repeat the virtual key in user-facing output. Tell the user to revoke the key after testing.

Progressive Test Matrix

Run only the tests enabled by the server configuration and report skipped tests.

Authentication and discovery: append models to the supplied base URL and request it (for example GET /v1/models when the base URL ends in /v1); record model IDs and whether authentication succeeds, without exposing the key.
Completion contract: send a deterministic, short request to the base-relative chat/completions endpoint; verify the OpenAI-compatible choice, content and usage fields.
SSE streaming: repeat with stream: true; verify incremental data: chunks, a terminal finish reason, and [DONE].
Cancellation: request a deliberately long response, interrupt the client after initial chunks, and verify that cancellation is local, prompt, and produces no unexpected retry. Do not claim server-side inference was cancelled unless the server explicitly confirms it.
Tool calling: use a tool-capable model and a harmless synthetic function schema. Verify tool-call structure only. Never run a production side-effecting tool against a real user account without explicit user approval.
Attachments: only with explicit approval and non-sensitive sample data, test a small image and PDF within application limits. Confirm rejection or omission above the limits.
App integration: only after endpoint checks pass, configure the app in the simulator manually for the current session and verify send, stream, stop, edit, regenerate, branch, export, and any enabled search tool.

Execution Rules

Prefer curl with the Authorization header for endpoint checks, using the supplied key only for the command that needs it.
Match app URL construction: endpoints such as models and chat/completions are relative to the user-supplied base URL. Do not add or remove /v1; the configured base URL owns that prefix.
Do not write response bodies containing sensitive user data to files.
Use non-sensitive, deterministic prompts such as Reply with exactly: integration-ok.
Stop immediately on 401, 403, certificate failures, or unexpected model behavior; report the failure without guessing.
Keep server-side tests separate from unit tests. Run the app's unit suite before and after integration work if application code changes.

Report Format

Report a compact table with test, endpoint/model, result, and relevant protocol facts. State clearly which tests were skipped and why. End by reminding the user to revoke the temporary virtual key.

LiteLLM Integration Testing

Use this skill only for opt-in tests against a real LiteLLM deployment.

Required Input

Ask for both values before making authenticated requests:

The exact OpenAI-compatible base URL entered in the app, for example https://litellm.example.com/v1.

A temporary virtual key with the minimum permissions required for the tests.

Progressive Test Matrix

Run only the tests enabled by the server configuration and report skipped tests.

Authentication and discovery: append models to the supplied base URL and request it (for example GET /v1/models when the base URL ends in /v1); record model IDs and whether authentication succeeds, without exposing the key.

Completion contract: send a deterministic, short request to the base-relative chat/completions endpoint; verify the OpenAI-compatible choice, content and usage fields.

SSE streaming: repeat with stream: true; verify incremental data: chunks, a terminal finish reason, and [DONE].

Cancellation: request a deliberately long response, interrupt the client after initial chunks, and verify that cancellation is local, prompt, and produces no unexpected retry. Do not claim server-side inference was cancelled unless the server explicitly confirms it.

Tool calling: use a tool-capable model and a harmless synthetic function schema. Verify tool-call structure only. Never run a production side-effecting tool against a real user account without explicit user approval.

Attachments: only with explicit approval and non-sensitive sample data, test a small image and PDF within application limits. Confirm rejection or omission above the limits.

App integration: only after endpoint checks pass, configure the app in the simulator manually for the current session and verify send, stream, stop, edit, regenerate, branch, export, and any enabled search tool.

Execution Rules

Prefer curl with the Authorization header for endpoint checks, using the supplied key only for the command that needs it.

Match app URL construction: endpoints such as models and chat/completions are relative to the user-supplied base URL. Do not add or remove /v1; the configured base URL owns that prefix.

Do not write response bodies containing sensitive user data to files.

Use non-sensitive, deterministic prompts such as Reply with exactly: integration-ok.

Stop immediately on 401, 403, certificate failures, or unexpected model behavior; report the failure without guessing.

Keep server-side tests separate from unit tests. Run the app's unit suite before and after integration work if application code changes.

Report Format

Report a compact table with test, endpoint/model, result, and relevant protocol facts. State clearly which tests were skipped and why. End by reminding the user to revoke the temporary virtual key.

litellm-integration