context-degradation

预期结果:

• Context degradation likely caused by: lost-in-middle effect or context poisoning
• Immediate mitigation: reorganize context so critical info is at the start or end
• Medium-term: implement compaction triggers before reaching 60K tokens
• Long-term: consider partitioning context across sub-agents or sessions
• Monitor health score: token utilization, degradation score, poisoning indicators

预期结果:

• This is the classic lost-in-middle phenomenon - attention follows a U-curve
• Beginning and end tokens receive high attention, middle tokens get degraded attention
• Move critical requirements to start (after system prompt) or end of context
• Use explicit section markers like [CURRENT TASK] and [KEY FINDINGS]
• Consider splitting long contexts into multiple focused sessions

预期结果:

• Implement hallucination tracking to detect when errors enter context
• Use explicit versioning for retrieved documents to prevent outdated info
• Set compaction triggers when error indicators exceed threshold (3-5 errors)
• Periodically truncate context to remove poisoned elements
• Design isolation patterns so errors in one section dont spread to others