Intercept or redact content in a live bidirectional agent session

Goal

Add content filtering and redaction to live interaction so blocked content does not reach the model.

Steps

1. Define input/output segments to inspect, blocking policy, and user-facing notice. [live]
2. Use LlmAgent.before_model_callback / after_model_callback or corresponding BasePlugin callbacks; define when to return a replacement LlmResponse. [live]
3. Return a replacement response or end the turn for policy-triggering content so the original is not sent onward. [live]
4. Test allowed, redacted, and rejected cases; ensure callbacks do not copy secrets into logs. [live]

How to verify

Blocked segments do not reach later model calls, allowed content continues, and audit logs omit sensitive payloads.

Troubleshooting

If a turn continues after replacement, verify the callback return and termination semantics; disable export first if logs leak data.