## Summary

The Akto guardrail integration has three issues in OSS 1.91.0.

First, setting \`mode: during_call\` in the guardrails YAML produces no validation effect and no error. \`HOOK_TO_INPUT\` at line 44 of \`akto.py\` only maps \`pre_call\` and \`post_call\`; \`during_call\` is absent. Additionally, \`supported_event_hooks\` only lists \`pre_call\` and \`post_call\`, so the framework skips the hook entirely before \`apply_guardrail\` is ever called. Any deployment configured with \`mode: during_call\` receives zero guardrail coverage.

Second, there is no way to scope Akto invocations to specific model groups. Operators who want to run Akto only for some model groups (e.g., by tier or cost) cannot do so without patching the integration.

Third, the streaming post_call path calls the Akto ingest endpoint once per every 5th chunk, serializing the growing \`responses_so_far\` buffer each time. For a 500-chunk stream this is 100 Akto calls, each serializing a longer buffer, producing O(n^2) work that stalls the asyncio event loop. Additionally, two hot-path log statements eagerly format large response objects even when DEBUG logging is disabled.

## Reproduce

```python from lite…