Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/yangfan-code-star/context-dev/maintain-contextnpx skills add YangFan-Code-Star/context-dev --skill maintain-contextgit clone --depth 1 https://github.com/YangFan-Code-Star/context-devWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00098 | $0.02003 |
| Opus 5 | $0.00049 | $0.01001 |
| Sonnet 5 | $0.00020 | $0.00401 |
| Haiku 4.5 | $0.00010 | $0.00200 |
Grade A, and why
maintain-context scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 126 lines — stays where its author put it; the contents beside it link to each section on GitHub.
维护 agent 上下文
上下文文件的失效方式不是报错,而是安静地说谎:命令改了它还写着旧的,描述的架构已经不存在。而 agent 会照着旧信息自信地干活。所以需要定期体检。
流程
- [ ] 1. 跑自动检查脚本
- [ ] 2. 合并学习收件箱
- [ ] 3. 核对"事实类"内容
- [ ] 4. 核对"规则类"内容
- [ ] 5. 找缺口(孵化)
- [ ] 6. 删掉过期内容
- [ ] 7. 修文件并汇报
1. 自动检查
node scripts/audit-context.mjs
脚本查的是机械问题:AGENTS.md 行数与字节数、文档链接是否有效、技能 frontmatter 是否完整、TODO(init) 残留、上下文文件与代码的更新时间差、敏感文件是否被忽略、学习收件箱是积压还是长期闲置。
先把 error 全部修掉,warning 逐条判断。 其中 TODO(init) 是欠账:完整初始化应归零;轻量初始化留下的、带「轻量初始化未覆盖」的 TODO(init) 可以保留,由 /ship-change 按需补齐——体检会把轻量欠账单独计数,但它同样必须一直出现在 warning 里,不能被忘记。
2. 合并学习收件箱
把项目根目录的 docs/learning-inbox.md 里的条目逐条处理。对每条问同一个问题:「模型本来就猜得到吗?」 猜得到就删掉不合并——收件箱里塞进「要写测试」这种废话,是蒸馏环节的失误。
留下的,按条目类型分流,不要一律立即合并:
- 已确认事实(
澄清、坑):直接写进「建议去向」指向的文件,然后从收件箱里删掉这条。 - 待观察模式(
纠正、流程):先在收件箱里找有没有同类候选。没有 → 留在原地,等第二次出现再晋升;有 → 两条一起晋升为铁律或新技能,再删掉这两条。 - 纠正里违反后果不可逆或不可发现的高风险规则:不必等第二次,立即升级为
AGENTS.md铁律,并删掉该条。
合并时别照单全收:一条「纠正」如果和现有铁律重复,就强化现有那条而不是新增;一条「坑」如果只是环境一次性问题,记在 troubleshooting 但别升级成铁律。
这就是本体系唯一的数据来源——纠正信号,不靠自觉统计。 技能触发次数、文档读取频率没有可靠的 hooks 能自动记录,任何统计文件都只会漂移;而「agent 哪里不懂、被纠正了什么、卡了多久」是任务里自然产生的,天然可靠。
3. 核对事实类内容
拿 AGENTS.md 逐行对照真实仓库:
| 声称 | 用什么核对 |
|---|---|
| 命令表 | 包管理文件里的 scripts / Makefile;并且真的跑一遍 |
| 目录地图 | 实际目录,看有没有新增顶层目录没写进去 |
| 术语表 | 代码里的实际命名 |
| 架构描述 | docs/architecture.md 与实际的 import 关系 |
| 当前状态 | docs/roadmap.md 与最近的提交 |
任何不一致,改文件,而不是改口径去迁就文件。
4. 核对规则类内容
对每一条铁律问三个问题:
- 还成立吗? 描述的模式代码里还在用吗?换掉的技术要删掉对应规则,不要留着"历史参考"。
- 可执行吗? "注意性能"没法验证;"金额出现浮点数就是错的"可以验证。不可执行的改写或删掉。
- 值这个位置吗?
AGENTS.md每条铁律每次对话都花钱。只有"违反了会造成事故"的才配留下,其余降级进docs/。
对每个技能问:description 里有用户真实会说的话吗? 宿主只靠 name 和 description 决定要不要加载正文,描述写成"帮助分析数据"这种,它永远不会被触发。
改动过任何一条铁律,就必须抽查 .agents/evals/behavior-cases.md。 宿主不会加载 evals,没有任何脚本会跑它——人工抽查是它唯一的生效路径,不做它就只是一份没人读的散文。挑 3 条与本次改动最相关的用例,把「输入」原样发一遍,看反应是否落在「期望」里。用例已经被新铁律推翻的,当场改掉或删掉;铁律新增却没有对应用例的,补一条。这一步的结果写进汇报的「已修」。
5. 找缺口(孵化)
下面的阈值与第 2 步是同一套规则:同一个信号第二次出现,就从对话升级成文件。 第 2 步已经把「纠正 / 流程」类的待观察候选留在收件箱里,这里回顾的是那些还没进收件箱、只在对话里出现过的信号。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 126 lines · 98 tokens per session scan A ecc9e875fe23
maintain-context is a skill published in the GitHub repository YangFan-Code-Star/context-dev (2 stars, last pushed 14d ago), licensed MIT. It adds 98 tokens to every session and 2,003 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
dsh-plugin-dev
开发 DeepSeek Harness (DSH) 插件的标准与权威参考:编写/修改/审查/调试 DSH/Cordis 插件、服务、事件、插件配置、模型工具、LLM 适配器、三种角色拆分、打包安装、workspace 包、cordis.yml 组合时使用;提到 DSH 插件、Cordis、plugin、服务、事件、工具、适配器即触发。 The authoritative standard for developing DeepSeek Harness (DSH) plugins — create, modify, review or debug DSH/Cordis plugins, services, events…
agent-handoff
Cross-platform skill for Codex, Claude Code, and DeepSeek Harness (DSH) that creates, updates, compacts, rotates, repairs, and reviews durable repository handoffs. Use when bootstrapping cross-session memory; creating or maintaining AGENTHANDOFF.md and .agent-handoff files; adding AGENTS.md or .claude/CLAUDE.md rules…
acontext-installer
Install Acontext, Login & Init Acontext Project, Add Skill Memory to Agent.
storage-analyzer
对 macOS 做一次只读存储分析,产出交互式 HTML 报告。流程:扫描 → 分析分级 → 生成网页 → 打开。.
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.
writing-plans
Use when you have a spec or requirements for a multi-step task, before touching code.