feat: add Headroom context compression layer

[baduyne]

Summary

Integrate a Headroom-compatible compression layer into Agent Zero to reduce excessive token usage in memory recalls, tool outputs, and LiteLLM requests.

This PR introduces token-aware context optimization before runtime context injection, helping prevent context window overflows and improving long-running agent stability.

---

Motivation

Agent Zero currently encounters several token/context related problems:

• oversized memory recalls
• large tool outputs
• browser annotation bloat
• embedding failures from oversized chunks
• LiteLLM ContextWindowExceededError

As the runtime grows more autonomous and tool-heavy, context accumulation becomes a major scalability issue.

This PR aims to reduce unnecessary token usage while preserving important semantic information.

---

What This PR Adds

Token-aware context compression

Introduces a compression layer before sending context into LiteLLM pipelines.

Compression targets include:

• memory recall outputs
• tool execution logs
• browser/DOM outputs
• large intermediate context blocks

---

Safer memory injection flow

Previous flow:

memory recall
    ↓
append directly to runtime context
    ↓
LiteLLM request

New flow:

memory recall
    ↓
compress / truncate / optimize
    ↓
token validation
    ↓
LiteLLM request

---

Goals

• reduce token usage
• prevent context window overflow
• improve runtime stability
• improve support for local/self-hosted models
• reduce unnecessary LiteLLM payload size

---

Related Issues

Fixes:

• Memory ContextWindowExceededError when memory updates on long conversations with limited utility model context size. #1413

Related to:

• ContextWindowExceededError when single code_execution_tool output exceeds context limit (compression doesn't help) #833
• Embedding calls crash with 400 when stored memory exceeds model context window #1436
• Plugin development for LiteLLM stats - token usage and context window info #1382
• Feature request: adopt A0 to self-hosted LLM's #1522

---

I have created the discussion about this sollution #1674

Notes

This PR is designed to be minimally invasive and compatible with the current LiteLLM-based architecture.

The implementation is intended as a foundation for future:

• adaptive context compression
• memory summarization
• tool-output compaction
• MCP schema optimization
• token observability metrics

Demo:

@3clyp50 @silverqx @regit