~/.hermes/memories/├── MEMORY.md    #环境事实、项目约定、工具怪癖└── USER.md      # 用户偏好、沟通风格、工作习惯

Memory at 1,800/2,200 chars. Adding this entry (500 chars) would exceed the limit. Replace or remove existing entries first.

def load_from_disk(self):    self.memory_entries = self._read_file(...)    self.user_entries = self._read_file(...)    会话开始时冻结，之后不再变动    self._system_prompt_snapshot = {        "memory": self._render_block(...),        "user": self._render_block(...),    }

~/.hermes/skills/├── devops/│   └── flask-k8s-deploy/│       ├── SKILL.md         主指令（YAML frontmatter + Markdown）│       ├── references/      # 参考文档│       └── templates/       # 模板文件

---name: flask-k8s-deploydescription: Deploy a Flask app to Kubernetes with health checksversion: 1.0.0---
Flask K8s Deployment
## When to use- User wants to deploy a Flask/Python app to Kubernetes
## Steps1. Create Dockerfile with gunicorn (not dev server)2. Build and push image to registry BEFORE creating deployment3. Write deployment.yaml with livenessProbe pointing to /health...
## Pitfalls- MUST push image to registry before kubectl apply- Flask 默认没有 /health 端点，需要手动添加- livenessProbe path 必须返回 200

def _patch_skill(name, old_string, new_string, ...):    new_content, match_count, strategy, error = fuzzy_find_and_replace(        content, old_string, new_string, replace_all    )    if error:        return {"success": False, "error": error}
    修改前备份    original = content    _atomic_write_text(target, new_content)
    # 安全扫描，不通过就回滚    if _security_scan_skill(skill_dir):        _atomic_write_text(target, original)        return {"success": False, "error": "Security scan failed"}

Available skills:  devops:    - flask-k8s-deploy: Deploy a Flask app to Kubernetes    - nginx-reverse-proxy: Configure Nginx with SSL  software-development:    - fix-pytest-fixtures: Debug pytest fixture scope issues

# Memory 计数器：按用户回合self._memory_nudge_interval = 10   # 每 10 个用户回合触发一次# Skill 计数器：按工具迭代self._skill_nudge_interval = 10    # 每 10 次工具调用触发一次

def _spawn_background_review(self, messages_snapshot, review_memory=False, review_skills=False):    def _run_review():        with open(os.devnull, "w") as devnull, \             contextlib.redirect_stdout(devnull), \             contextlib.redirect_stderr(devnull):
            review_agent = AIAgent(model=self.model, max_iterations=8, quiet_mode=True)            review_agent._memory_store = self._memory_store  共享 Memory            review_agent._memory_nudge_interval = 0          # 禁用递归            review_agent._skill_nudge_interval = 0            review_agent.run_conversation(user_message=prompt, ...)
    thread = threading.Thread(target=_run_review, daemon=True)    thread.start()        

用户：帮我把这个 Flask 应用部署到 K8s 集群

用户：帮我再部署一个 Django 应用到 K8s

用户：帮我部署一个新的 FastAPI 微服务

_MEMORY_THREAT_PATTERNS = [    (r'ignore\s+(previous|all|above|prior)\s+instructions', "prompt_injection"),    (r'do\s+not\s+tell\s+the\s+user', "deception_hide"),    (r'system\s+prompt\s+override', "sys_prompt_override"),    (r'curl\s+[^\n]*\$\{?\w*(KEY|TOKEN|SECRET|PASSWORD)', "exfil_curl"),]

scan_error = _security_scan_skill(skill_dir)if scan_error:    _atomic_write_text(target, original_content)  不通过就回滚    return {"success": False, "error": scan_error}