Return rate limit error instead of quota error from LanguageModelRateLimitingPlugin

[waldekmastykarz]

Closes #1911

When throttling with whenLimitExceeded: "Throttle", LanguageModelRateLimitingPlugin returned OpenAI's insufficient_quota billing error, which tells clients to stop retrying. It now returns OpenAI's tokens-per-minute rate limit error so clients back off and retry:

{
  "error": {
    "message": "Rate limit reached for gpt-4o on tokens per min (TPM): Limit 10, Used 10. Please try again in 299s.",
    "type": "tokens",
    "param": null,
    "code": "rate_limit_exceeded"
  }
}

• The message includes the model (when present in the request), the exhausted limit, tokens used, and seconds until reset (same value as the retry-after header).
• whenLimitExceeded: "Custom" is unchanged, for anyone who wants to simulate a quota error on purpose.
• Updated the skill docs example and added an integration test.

[Copilot]

Copilot review overview

🟡 Changes recommended

The new error message underreports consumed tokens when usage exceeds the configured limit.

Review effort: Balanced
Findings: 1

Open (1)

• 429 reports capped token usage instead of actual consumption · New

What changed in this PR

Changes the default language-model throttle response to a retryable OpenAI rate-limit error, addressing #1911.

Changes:

• Reports the model, token limit, usage, and retry delay.
• Adds integration coverage for the throttle response.
• Updates documentation while preserving custom responses.

File	Description
skills/​dev-proxy/​references/​test-llm-apps.md	Updates error examples and explains custom quota errors.
DevProxy.Plugins/​Behavior/​LanguageModelRateLimitingPlugin.cs	Returns rate-limit errors with token and retry details.
DevProxy.Integration.Tests/​BehaviorPluginsIntegrationTests.cs	Tests the default throttle response.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.