research(context): PAACE — plan-aware context engineering with next-k-task relevance modeling (arXiv:2512.16970)

[bug-ops]

Description

PAACE (Plan-Aware Automated Context Engineering) optimizes agent state through:

1. Next-k-task relevance modeling — include context items relevant to the next k planned steps
2. Plan-structure analysis — respect DAG dependency structure when deciding what to preserve
3. Instruction co-refinement — compress instructions + context together for coherence
4. Function-preserving compression — verifies compressed context still enables tool calls

Maps to zeph-context + zeph-orchestration (task DAG). Current Zeph context assembly is plan-unaware; it doesn't know what future tool calls will need.

Relevance to Zeph

• crates/zeph-context/src/assembler.rs — could use plan lookahead to bias retention
• crates/zeph-orchestration/ — DAG structure available; pass lookahead to context manager
• crates/zeph-agent-context/src/service.rs — context rebuild could accept planned_tools hint

Implementation Sketch

1. Add planned_next_tools: Vec<ToolId> to ContextBuildRequest
2. In assembler: score retention of each context item against planned tool requirements
3. Orchestrator injects next k task IDs when triggering context rebuild mid-DAG execution

Complexity vs Benefit

• Complexity: Medium-High (requires lookahead from orchestration) | Benefit: High (reduces irrelevant context in multi-step tasks, fewer token reloads)

Source

• Paper: https://arxiv.org/abs/2512.16970
• Related: research(context): ACDL — formal language for describing agentic LLM context structure and evolution (arXiv:2605.01920) #3976 (ACDL), CI-799 zeph-context changes