Agent Batch Guard — AI skill for Claude Code
AI Agent 大任务防卡死指南 — 防止 session transcript 膨胀导致 agent 卡死.
How to install Agent Batch Guard
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open evan966890/agent-batch-guard and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Agent Batch Guard does
AI Agent 大任务防卡死指南 — 防止 session transcript 膨胀导致 agent 卡死.
Alternatives in AI
- CLI Continues — resume any AI coding session in another tool — Claude Code, Copilot, Gemini, Codex, Cursor 1.5k ★
- Firm — Convene your standing AI staff — memos on every beat, a board session without you, minutes with dissent preser 1.3k ★
- Botmux — Bridge Feishu/Lark to AI coding CLIs — Claude Code, Codex, Gemini, OpenCode… every DM, group or topic spawns i 1.2k ★
README
name: agent-batch-guard version: 1.1.0 description: AI Agent 大任务防卡死指南。解决 agent 在批量操作中 session transcript 膨胀导致 compaction 超时、agent 卡死的问题,并补充可复用的长任务执行骨架:目标、成功标准、状态落盘、断点续跑、熔断器和脚本模板。
Agent 大任务防卡死指南
AI Agent 执行大任务(批量抓取、翻页采集、历史数据导出)时,极易因 session transcript 膨胀导致卡死。本指南提供从认知层到配置层的完整防护方案。
快速开始
适合什么场景
- 批量抓取、翻页采集、历史导出
- 需要断点续跑的长任务
- 需要把中间状态写盘、避免 transcript 膨胀的任务
- 需要给 agent 一个统一“长任务执行合同”的场景
仓库里现在有什么
| 文件 | 作用 |
|---|---|
SKILL.md |
给 agent 直接加载的技能说明 |
scripts/goal_loop.py |
可复用的长任务执行 helper |
scripts/cli.py |
查看状态文件的最小 CLI |
templates/long_task_runner.py |
可直接改造成自己任务的模板 |
references/long-task-execution-model.md |
长任务状态模型说明 |
tests/test_goal_loop.py |
回归测试与对抗测试 |
最小使用示例
from scripts.goal_loop import GoalLoop, LoopConfig
loop = GoalLoop(
goal="按月导出 2025 全年订单",
criteria={"done": True},
config=LoopConfig(
max_attempts=24,
check_interval_seconds=30,
max_consecutive_failures=5,
state_dir="data/batch_state",
),
)
result = loop.run(run_one_batch, resume=True)
print(result)
查看任务状态
python -m scripts.cli status --state-dir data/batch_state
1.1 新增:可复用长任务执行骨架
这一版把 GoSkill 里最有价值的部分整合进来了,但做了更贴近这个 skill 的改写:重点不在“持续尝试”四个字,而在**把长任务的执行状态从对话里拿出来,落到文件和统一状态结构里**。
新增文件:
scripts/goal_loop.py:长任务执行 helperscripts/cli.py:最小状态查看 CLItemplates/long_task_runner.py:可直接改造的批处理模板references/long-task-execution-model.md:统一状态模型说明tests/test_goal_loop.py:最小测试集
新增的稳态能力
- 状态文件使用原子写,避免写一半留下坏文件
- 执行过程使用文件锁,避免同一状态目录被两个 runner 同时写坏
- 结果附带
elapsed_seconds / started_at / ended_at - 可通过
python -m scripts.cli status --state-dir直接查看状态 JSON
新增的执行合同
长任务启动前,先明确四件事:
goal:这次到底要达成什么criteria:怎样才算完成state_file:进度和最近结果写到哪里terminal_status:任务为什么结束
推荐统一状态文件结构:
{
"goal": "导出 2025 全年订单",
"criteria_signature": "sha256:...",
"attempts": 3,
"consecutive_failures": 0,
"terminal_status": null,
"last_res
Related Skills
Transcript Cleanup
Transcript To Knowledge
Turn any YouTube video into agent-ready research markdown: transcript -> structured LLM breakdown -> optional
PM Ugc Analysis
Scraping TikTok UGC virali — 25 transcript per swipe file con vetting LLM di rilevanza. Skill nativa 20_ugc_sc
Agent Secrets
Let an AI agent use your API keys and passwords without the plaintext ever entering its transcript. A plugin f
Job Hunter
AI-assisted job application automation as a Claude Code skill. Reads your resume, matches JD by score, applies
Clawdbot Skill Parallel
High-accuracy web research via Parallel.ai API - outperforms Perplexity and Exa on benchmarks. Rich excerpts,
Related Agents
Context Refinement
Updates task context manifest with discoveries from current work session. Reads transcript to understand what
Session Context
Loads a single FullStory session's event transcript into an isolated context window and answers a specific tas
Skill Applied Judge
Use this agent at session-end Phase 3.6.6 (#645 L3) to judge — from the session transcript tail — whether each