正在打开每日关注
正在打开每日关注
讨论 Claude Code 与 Codex 在技能隐式调用规则上的差异及兼容方案。
i second this. realized that @mattpocockuk skills get triggered more often in Codex i think there is a big issue with the standard because Claude Code got a good start and a lot of popular skills like Matt's only have `disable-model-invocation: true` Codex doesn't support this and needs separate metadata in agents/openai.yaml `policy.allow_implicit_invocation: false` so i've forked them and made them all compatible with Codex and now i'm waiting for Matt to give me permission to open a PR 👉👈
Some tips for skill auditing with 5.6 Sol: Skill activation is different. You should generally have the least number of skills required. If any skills talk about sub agent usage they will trigger and can increase usage. Look at the skill descriptions - make sure they are as short as possible to save context, and watch for wording that may make the model think they need the skill more than necessary.
是什么
这是一条围绕 Agent 技能标准与运行时行为的观点推文。作者认为，Matt Pocock 的技能在 Codex 中更容易被触发，原因在于 Claude Code 使用的 disable-model-invocation: true 机制，不能直接对应 Codex 的行为；他因此通过分叉技能，加入 agents/openai.yaml 中的 policy.allow_implicit_invocation: false，等待原作者许可后提交贡献。
核心设计 讨论的重点不是技能内容，而是“何时允许模型调用技能”的控制面。被引用内容进一步强调最小技能集合、缩短技能描述，并警惕提及子代理的措辞，因为这些因素可能提高误触发概率并增加使用量。
关键洞察 技能描述同时承担发现、路由和成本控制功能。描述越宽泛、触发条件越松，覆盖面可能越大，但上下文消耗、调用噪声和行为不可预测性也会增加。跨平台技能若没有明确的调用策略元数据，就难以保持一致行为。
边界与风险 这只是作者的经验判断，未提供触发频率、配置样例或实测数据，不能据此断言平台标准存在普遍缺陷。分叉与提交 PR 还涉及上游许可、元数据兼容及未来同步成本；技能审计应结合真实调用日志和回归测试验证。