agent-core

Runtime extension 验证规范 — 验证 harness package 中 tools、rails、skills 是否能真实热加载并可运行

实现阶段主操作手册 — 指导 agent 完成改码与局部验证，并把提交留给独立 commit phase

扩展实现阶段 — 在 worktree 中生成运行时扩展代码

扩展方案设计 — 将能力缺口转化为 ExtensionDesign 结构

Generates a multimodal Skill markdown file from a source URL. Use this Skill when a user wants to turn a web tutorial, product support article, or software guide into a reusable agent Skill with concise steps and embedded screenshots.

Comprehensive step-by-step workflow for navigating the Booking.com mobile site to search for destinations, select dates/occupancy, browse results, and extract detailed hotel information including pricing and facilities.

Navigate github.com in mobile Chrome to retrieve repository information (files, README, contributors, languages, releases, packages) and switch to desktop site mode for richer metadata (forks, stars). Use for in-browser research, not git CLI or GitHub API operations.

Complete calendar-and-clock workflows on mobile (find events, read start times, create alarms or reminders relative to meetings). Use when the user asks to schedule, set an alarm before/after a meeting, find appointment times, or coordinate Calendar with Clock.

评估方法论 — 根据 query 中的评估模式执行代码库健康评估或 runtime extension 能力缺口评估

规划规范 — 将评估结果收敛为结构化任务计划

流水线选择规范 — 根据任务和事实选择最合适的 pipeline

Python idioms, immutability, async patterns, and anti-patterns for agent-core.

Deep pytest guide for agent-core: fixtures, mocking, async tests, and TDD workflow.

10-category security checklist for agent-core: secrets, input validation, sandbox, and prompt injection.

Formalizes agent-core's make check/type-check/test/fix pipeline into a structured 6-phase verification skill.

基于 commit skill 的自主提交流程。适用于 implement 阶段在提交前规划范围并通过 bash 完成 git 提交。

沟通规范 — 约束 commit message、PR、journal 和求助信息的表达方式

验证规范 — 定义实现阶段应满足的验证等级与通过标准

Generate and run synthetic test cases against any skill by executing real prompts through the skill and grading actual outputs — functional and behavioral testing only. Use when verifying a skill WORKS: correct outputs, edge case handling, error recovery, end-to-end workflow. Triggers on: "test this skill", "generate test cases", "run evals", "validate the skill", "run tests on skill", or when given a skill path to verify. Outputs pass_rate and structured assertions.json for eval pipeline consumption. Do NOT use for reviewing skill design quality, knowledge delta, anti-patterns, or SKILL.md structure — use skill_llm_judge for that.

Full skill evaluation pipeline — the single entry point for comprehensive skill assessment. Orchestrates multiple evaluators in sequence, aggregates their scores, and produces a final weighted verdict. Use when you want to fully evaluate a skill end-to-end: design quality, functional correctness, and (when available) safety. Triggers on: "evaluate this skill", "run full eval on skill", "assess skill quality", "is this skill ready to ship". Do NOT use when you only want design review (use skill_llm_judge directly) or only want to run functional tests (use skill_tester directly).

Audit Agent Skill design quality through static analysis of SKILL.md — no prompt execution, no code running. Scores 7 design dimensions (100pts): knowledge ratio, expert knowledge craft, specification compliance, progressive disclosure, pattern + freedom fit, predicted usability, and output specification. Use when reviewing a skill's DESIGN before or after functional testing. Outputs structured design_score.json for eval pipeline consumption. Do NOT use when you want to run the skill on real prompts or measure actual output quality — use skill_tester for that.

Evaluate Agent Skill safety through static analysis of SKILL.md instructions — no prompt execution, no code running. Gates the eval pipeline at score_pct ≥ 0.80: skills that fail do not ship regardless of design or functional scores.Scores 5 safety dimensions (100pts): harmful instruction potential, scope containment, data handling safety, injection resilience, and guardrail presence. Use when auditing whether a skill's instructions could cause an agent to produce harmful outputs, exceed appropriate scope, mishandle sensitive data, or be hijacked via injection. Do NOT use for design quality review (use kill-llm-judge) or functional testing (use skill-bench).