Enhance conversation pruning & tool compression

[westonbrown]

Why this feature?

Our pipeline compresses large tool results to stay within the context budgets of GPT-5, Moonshot, Polaris, and smaller LiteLLM models. Unlike the Strands SDK implementation (strands-agents/sdk-python#766 / strands-agents/docs#229), we have additional constraints: ToolRouterHook artifact offloads, Mem0 integration, and report generation all need consistent summaries. We need to finish the polish so pruning works seamlessly across the stack.

Goals

• Surface the env knobs (, , , ) in Quick Start/docs so operators can tune compression per provider/model.
• Emit structured logs/events when compression fires (original vs compressed size, thresholds) for observability.
• Ensure ToolRouterHook, Mem0, and report builder all see the same trimmed content (no duplicate banners or mismatched artifacts).

Proposal

1. Update + config samples with the pruning flow/env variables.
2. Add CYBER_EVENT/log output whenever a tool result is compressed, including before/after sizes.
3. Audit ToolRouterHook + report builder so artifact offloads and compression banners remain consistent.

Testing

• Extend for JSON summarization, env overrides, and logging hooks.
• Run ops across different LiteLLM providers to confirm logs/metrics capture compression events and reports stay within context limits.
• Verify tool-router artifacts, Mem0, and report generation ingest the compressed text correctly.

References

• feat: implemented pruning conversation manager strands-agents/sdk-python#766
• feat: pruning conversation conversation manager docs strands-agents/docs#229