research(context): Acon — dual compression of observations + interaction history for long-horizon tasks (arXiv:2510.00615)

[bug-ops]

Description

Acon (Agent Context Optimization) compresses both environment observations (tool outputs) and interaction histories (conversation), reducing peak token usage 26-54% while preserving task success. Tested on AppWorld, OfficeBench, Multi-objective QA.

Key insight: Zeph currently compresses conversation history but not tool output verbosity. Acon applies separate compression policies to each stream and merges under a unified budget.

Relevance to Zeph

• crates/zeph-agent-tools/ — tool results verbosity is unmanaged today
• crates/zeph-context/src/assembler.rs — observation stream compression before injection
• crates/zeph-core/src/agent/tool_execution/focus.rs (changed in CI-799) — already has focus logic, extend to token budgeting

Implementation Sketch

1. Add ToolResultCompressor that truncates/summarizes large tool outputs above configurable token limit
2. In context assembler: apply separate budget splits to (tool_results, conversation_history, system_prompt)
3. Use existing LLM provider for on-the-fly observation summarization

Complexity vs Benefit

• Complexity: Low-Medium | Benefit: High (26-54% token reduction in tool-heavy sessions)

Source

• Paper: https://arxiv.org/abs/2510.00615
• Spec: .local/specs/NNN-acon-dual-compression/spec.md before implementation