| name | common-context-optimization |
| description | Techniques to maximize context window efficiency, reduce latency, and prevent 'lost in middle' issues through strategic masking and compaction. (triggers: *.log, chat-history.json, reduce tokens, optimize context, summarize history, clear output) |
Priority: P1 (OPTIMIZATION)
Manage the Attention Budget. Treat context as a scarce resource.
1. Observation Masking (Noise Reduction)
Problem: Large tool outputs (logs, JSON lists) flood context and degrade reasoning.
Solution: Replace raw output with semantic summaries after consumption.
- Identify: outputs > 50 lines or > 1kb.
- Extract: Read critical data points immediately.
- Mask: Rewrite history to replace raw data with
[Reference: <summary_of_findings>].
- See:
references/masking.md for patterns.
2. Context Compaction (State Preservation)
Problem: Long conversations drift from original intent.
Solution: Recursive summarization that preserves State over Dialogue.
- Trigger: Every 10 turns or 8k tokens.
- Compact:
- Keep: User Goal, Active Task, Current Errors, Key Decisions.
- Drop: Chat chit-chat, intermediate tool calls, corrected assumptions.
- Format: Update
System Prompt or Memory File with compacted state.
- See:
references/compaction.md for algorithms.
3. KV-Cache Awareness (Latency)
Goal: Maximize pre-fill cache hits.
- Static Prefix: strict ordering: System -> Tools -> RAG -> User.
- Append-Only: Avoid inserting into the middle of history if possible.
References
Anti-Patterns
- No raw tool dumps: Mask large outputs immediately after extracting data.
- No append-only growth: Compact every 10 turns to preserve intent over dialogue.
- No middle insertions: Append-only history maximizes KV cache hits.
