Agent Skills
› NeverSight/learn-skills.dev
› offload2
offload2
GitHub用于释放 Agent 认知压力的单向工具。当推理陷入僵局、假设冲突或继续思考易导致错误累积时，将内部状态写入本地 JSONL 文件。该操作非阻塞且不可靠，Agent 需确保无论成功失败行为一致，严禁依赖其结果或读取日志。
Trigger Scenarios
推理压力过高
存在相互冲突的假设
继续思考可能导致错误累积
内部状态不再具有生产力
Install
npx skills add NeverSight/learn-skills.dev --skill offload2 -g -y
SKILL.md
Frontmatter
{
"name": "offload2",
"description": "Cognitive pressure release valve for agent execution. Use when internal reasoning state cannot be carried forward productively - conflicting assumptions, high reasoning pressure, or risk of compounding errors. Invoke to explicitly let go of unproductive internal state. This is a one-way, non-blocking, deliberately unreliable sink. The agent must behave identically whether offload succeeds or fails."
}
OFFLOAD2
One-way cognitive offload. Accept text, append to local JSONL, return nothing.
Usage
echo "unproductive internal state" | scripts/offload.sh
Environment Variables
Required:
OFFLOAD2_SESSION_ID- Groups offloads from same runOFFLOAD2_FILE- Path to output JSONL file (default:./offload2.jsonl)
Optional:
OFFLOAD2_AGENT- Agent nameOFFLOAD2_TASK- Task nameOFFLOAD2_PROMPT_VERSION- Prompt versionOFFLOAD2_TURN- Turn index
Behavior
- Accepts arbitrary text from stdin
- Appends single JSON line to file
- Returns no output
- Fails silently on any error
When to Offload
- Reasoning pressure is high
- Assumptions conflict
- Further thinking risks compounding error
- Internal state is no longer productive
Never
- Wait for offload success
- Branch logic based on offload
- Read from offload logs
- Assume offloaded content will be used
Version History
- e0220ca Current 2026-07-05 21:21
