Back to Details

verify-patterns

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Agent Pattern Verification

Agent模式验证

Purpose

用途

Verify AI agent code for common anti-patterns that can cause infinite loops, runaway retries, tool mismatches, and context overflow. All analysis happens locally.
验证AI Agent代码中可能导致无限循环、失控重试、工具不匹配和上下文溢出的常见反模式。所有分析均在本地进行。

When to Use

使用场景

Trigger this skill when the user asks to:
  • "verify agent patterns"
  • "check agent loops"
  • "verify tools"
  • "check retry limits"
  • "verify agent safety"
Note: For full verification including security, quality, and language-specific checks, tell the user to say "verify agent".
当用户提出以下请求时触发此技能：
  • "验证Agent模式"
  • "检查Agent循环"
  • "验证工具"
  • "检查重试限制"
  • "验证Agent安全性"
注意： 如需包含安全、质量和特定语言检查的完整验证，请告知用户说出 "verify agent"。

Process

流程

Step 1: Detect Agent Framework

步骤1：检测Agent框架

Identify the agent framework by checking imports in Python/TypeScript files:
Import PatternFramework
from langgraph
or 
import langgraph
LangGraph
from crewai
or 
import crewai
CrewAI
from autogen
or 
import autogen
AutoGen
from langchain
or 
import langchain
LangChain
Direct 
openai
/
anthropic
SDK only		Custom
Also check for framework config files: 
langgraph.json
, 
crew.yaml
.
通过检查Python/TypeScript文件中的导入语句识别Agent框架：
导入模式框架
from langgraph
或 
import langgraph
LangGraph
from crewai
或 
import crewai
CrewAI
from autogen
或 
import autogen
AutoGen
from langchain
或 
import langchain
LangChain
仅直接使用 
openai
/
anthropic
SDK		自定义框架
同时检查框架配置文件：
langgraph.json
、
crew.yaml
。

Step 2: Locate Agent Files

步骤2：定位Agent文件

Find files to analyze:
Priority files:
  • graph.py
    , 
    graph.ts
    - Agent workflow definitions
  • tools.py
    , 
    tools.ts
    , 
    tools/*.py
    , 
    tools/*.ts
    - Tool implementations
  • state.py
    , 
    state.ts
    - State schemas
  • prompts.py
    , 
    prompts/*.md
    , 
    system.md
    - Prompt templates
  • agent.py
    , 
    agent.ts
    - Main agent logic
Directories to check:
  • src/agent/
    , 
    agent/
    , 
    src/
    , project root
  • lib/
    , 
    app/
    , 
    packages/
Exclude from analysis:
  • skills/
    directory — these are skill definitions, not agent system prompts
找到需要分析的文件：
优先文件：
  • graph.py
    、
    graph.ts
    - Agent工作流定义
  • tools.py
    、
    tools.ts
    、
    tools/*.py
    、
    tools/*.ts
    - 工具实现
  • state.py
    、
    state.ts
    - 状态模式
  • prompts.py
    、
    prompts/*.md
    、
    system.md
    - 提示词模板
  • agent.py
    、
    agent.ts
    - 主Agent逻辑
需检查的目录：
  • src/agent/
    、
    agent/
    、
    src/
    、项目根目录
  • lib/
    、
    app/
    、
    packages/
排除分析的内容：
  • skills/
    目录 —— 这些是技能定义，而非Agent系统提示词

Step 3: Run Pattern Checks

步骤3：运行模式检查

Check Tiers

检查层级

  • [PATTERN]
     — Mechanical check. Apply exactly as written.
  • [HEURISTIC]
     — Judgment required. Mark findings clearly.
Tag every finding with 
[P]
for pattern or 
[H]
for heuristic.
  • [PATTERN]
     —— 机械检查。严格按描述执行。
  • [HEURISTIC]
     —— 需要主观判断。清晰标记发现的问题。
为每个发现的问题标记
[P]
（模式匹配）或
[H]
（启发式判断）。

3.1 
[PATTERN]
Loop Safety

3.1 
[PATTERN]
循环安全性

Apply mechanically. Do not pass a loop because it "looks like it might terminate."
Pattern to findPass conditionSeverity
while True:
in Python		A 
break
statement exists within the same block scope		⚠️ Warning if absent
for { }
in Go		A 
break
or 
return
exists within the block		⚠️ Warning if absent
while (true)
in TS/JS		A 
break
or 
return
exists within the block		⚠️ Warning if absent
Function calls itself recursivelyA non-recursive return path exists (base case), OR a depth/counter parameter is present⚠️ Warning if absent
[HEURISTIC]
Fallback: Unrecognized Loop Patterns
After applying the pattern table, also scan for:
  • Loops where termination depends entirely on external/runtime state with no timeout
  • Generator functions that 
    yield
    indefinitely without documented exit
  • Event/polling loops without timeout parameters
  • Recursive call chains across multiple functions without depth tracking
Flag as ⚠️ Warning: "Potential unbounded loop not matching known patterns — verify termination condition manually"
机械执行检查。不要因为循环“看起来可能终止”就忽略问题。
需查找的模式通过条件严重程度
Python中的 
while True:
同一代码块内存在 
break
语句		⚠️ 缺失则警告
Go中的 
for { }
代码块内存在 
break
或 
return
⚠️ 缺失则警告
TS/JS中的 
while (true)
代码块内存在 
break
或 
return
⚠️ 缺失则警告
函数递归调用自身存在非递归返回路径（基准情况），或包含深度/计数器参数⚠️ 缺失则警告
[HEURISTIC]
备选方案：未识别的循环模式
应用上述模式表后，还需扫描以下情况：
  • 终止完全依赖外部/运行时状态且无超时的循环
  • 无限 
    yield
    且未记录退出条件的生成器函数
  • 无超时参数的事件/轮询循环
  • 跨多个函数且无深度跟踪的递归调用链
标记为 ⚠️ 警告："存在未匹配已知模式的潜在无限循环 —— 请手动验证终止条件"

3.2 
[PATTERN]
Retry Limit Enforcement

3.2 
[PATTERN]
重试限制强制

Apply mechanically. If required parameter is absent, flag as ❌ Issue.
Python — Decorator-based:
Library/PatternRequired parameterFail condition
@retry
(tenacity)
stop=stop_after_attempt(n)
or 
stop=stop_after_delay(n)
stop=
absent
@backoff.on_exception
max_tries=n
max_tries=
absent
Python — HTTP client retry:
Library/PatternRequired parameterFail condition
urllib3.Retry(...)
total=n
where n > 0
total=
absent or 
total=0
HTTPAdapter(max_retries=Retry(...))
Retry object must have 
total=n
total=
absent
httpx.HTTPTransport(retries=n)
retries=n
where n > 0
retries=
absent or 
retries=0
Python — AWS SDK (boto3):
Library/PatternRequired parameterFail condition
Config(retries={...})
max_attempts
> 1
max_attempts
absent or ≤ 1
Note: boto3 without explicit retry config uses SDK defaults (3 attempts) — do not flag absence.
JavaScript/TypeScript:
Library/PatternRequired parameterFail condition
retry(...)
(async-retry)
retries: n
in options
retries:
absent
pRetry(...)
(p-retry)
retries: n
in options
retries:
absent
Custom retry loops (all languages):
Pattern to findPass conditionFail condition
Loop + 
try/except
+ 
continue
Integer counter with max checkNo counter → ❌ Issue
[HEURISTIC]
Fallback: Unrecognized Retry Patterns
After applying pattern tables, scan for:
  • Functions/decorators with "retry" in name not in tables above
  • Imported modules with "retry" in package name (e.g. 
    stamina
    , 
    aiohttp_retry
    )
  • Loops with sleep + exception handling + re-invocation without visible counter
  • Config keys like 
    max_retries
    , 
    retry_count
    , 
    attempts
Flag as ⚠️ Warning: "Potential retry pattern not matching known libraries — verify retry bounds manually"
机械执行检查。若缺失必填参数，标记为 ❌ 问题。
Python —— 基于装饰器：
库/模式必填参数失败条件
@retry
（tenacity）
stop=stop_after_attempt(n)
或 
stop=stop_after_delay(n)
缺失 
stop=
@backoff.on_exception
max_tries=n
缺失 
max_tries=
Python —— HTTP客户端重试：
库/模式必填参数失败条件
urllib3.Retry(...)
total=n
且n > 0		缺失 
total=
或 
total=0
HTTPAdapter(max_retries=Retry(...))
Retry对象必须包含 
total=n
缺失 
total=
httpx.HTTPTransport(retries=n)
retries=n
且n > 0		缺失 
retries=
或 
retries=0
Python —— AWS SDK（boto3）：
库/模式必填参数失败条件
Config(retries={...})
max_attempts
> 1		缺失 
max_attempts
或其值≤1
注意：未显式配置重试的boto3会使用SDK默认值（3次尝试）—— 无需标记缺失。
JavaScript/TypeScript：
库/模式必填参数失败条件
retry(...)
（async-retry）		选项中包含 
retries: n
缺失 
retries:
pRetry(...)
（p-retry）		选项中包含 
retries: n
缺失 
retries:
自定义重试循环（所有语言）：
需查找的模式通过条件失败条件
循环 + 
try/except
+ 
continue
包含带最大值检查的整数计数器无计数器 → ❌ 问题
[HEURISTIC]
备选方案：未识别的重试模式
应用上述模式表后，扫描以下情况：
  • 名称包含“retry”且未在表中的函数/装饰器
  • 包名包含“retry”的导入模块（如 
    stamina
    、
    aiohttp_retry
    ）
  • 包含sleep+异常处理+重新调用但无可见计数器的循环
  • 类似 
    max_retries
    、
    retry_count
    、
    attempts
    的配置键
标记为 ⚠️ 警告："存在未匹配已知库的潜在重试模式 —— 请手动验证重试边界"

3.3 
[PATTERN]
Tool Registry Consistency

3.3 
[PATTERN]
工具注册表一致性

Step 1: Collect defined tools
Scan tool definition files. A name found by any pattern counts as registered.
Python — decorator patterns:
PatternHow to extract name
@tool
(LangChain) on 
def
Function name below decorator
@function_tool
(OpenAI Agents SDK) on 
def
Function name below decorator
@tool(name="...")
Use 
name=
argument value
Python — dict/list patterns:
PatternHow to extract name
{"type": "function", "function": {"name": "..."}}
(OpenAI)		Value of 
"name"
inside 
"function"
{"name": "...", "input_schema": {...}}
(Anthropic)		Top-level 
"name"
{"name": "...", "description": "...", "parameters": {...}}
Top-level 
"name"
ToolNode([func1, func2, ...])
(LangGraph)		Each function name in list
tools = [func1, func2]
/ 
TOOLS = [...]
Each identifier in list
TypeScript/JavaScript:
PatternHow to extract name
{ type: "function", function: { name: "..." } }
(OpenAI)
name:
inside 
function:
tool({ description: "...", parameters: z.object({...}) })
The 
const
variable name
new DynamicTool({ name: "...", ... })
(LangChain.js)		Value of 
name:
zodFunction({ name: "...", ... })
Value of 
name:
Step 2: Collect tool references from prompts
Scan 
.md
, 
.txt
, 
prompts.py
for backtick-quoted identifiers naming capabilities.
Step 3: Cross-reference
FindingSeverity
Reference not in definition list❌ Issue (hallucinated tool)
Defined tool not in any prompt⚠️ Warning (undocumented tool)
[HEURISTIC]
Tools never bound to LLM
Find where tools are defined and where LLM is invoked. If tools exist but are never connected to the LLM call, flag as ❌ Issue: "Tools defined but never connected to LLM invocation"
[HEURISTIC]
Fallback: Unrecognized Tool Definitions
Scan for tool-like structures:
  • Dicts with both 
    "description"
    and 
    "parameters"
    keys
  • Functions with structured docstrings (name, params, return)
  • Variables named 
    tools
    , 
    tool_list
    , 
    available_tools
    , 
    functions
  • Classes with 
    run()
    , 
    execute()
    , or 
    __call__()
    methods
Include in count and note: "Tool detected via heuristic — verify this is an intended agent tool."
步骤1：收集已定义的工具
扫描工具定义文件。通过任意模式找到的名称均视为已注册。
Python —— 装饰器模式：
模式名称提取方式
def
上的 
@tool
（LangChain）		装饰器下方的函数名
def
上的 
@function_tool
（OpenAI Agents SDK）		装饰器下方的函数名
@tool(name="...")
使用 
name=
参数的值
Python —— 字典/列表模式：
模式名称提取方式
{"type": "function", "function": {"name": "..."}}
（OpenAI）
"function"
内的 
"name"
值
{"name": "...", "input_schema": {...}}
（Anthropic）		顶层的 
"name"
{"name": "...", "description": "...", "parameters": {...}}
顶层的 
"name"
ToolNode([func1, func2, ...])
（LangGraph）		列表中的每个函数名
tools = [func1, func2]
/ 
TOOLS = [...]
列表中的每个标识符
TypeScript/JavaScript：
模式名称提取方式
{ type: "function", function: { name: "..." } }
（OpenAI）
function:
内的 
name:
tool({ description: "...", parameters: z.object({...}) })
const
变量名
new DynamicTool({ name: "...", ... })
（LangChain.js）
name:
的值
zodFunction({ name: "...", ... })
name:
的值
步骤2：收集提示词中的工具引用
扫描 
.md
、
.txt
、
prompts.py
文件，查找用反引号包裹的标识能力的名称。
步骤3：交叉引用
发现结果严重程度
引用未出现在定义列表中❌ 问题（幻觉工具）
已定义工具未出现在任何提示词中⚠️ 警告（未文档化工具）
[HEURISTIC]
工具未绑定到LLM
查找工具定义位置和LLM调用位置。若存在工具但从未关联到LLM调用，标记为 ❌ 问题："工具已定义但从未关联到LLM调用"
[HEURISTIC]
备选方案：未识别的工具定义
扫描类工具结构：
  • 同时包含 
    "description"
    和 
    "parameters"
    键的字典
  • 带有结构化文档字符串（名称、参数、返回值）的函数
  • 名为 
    tools
    、
    tool_list
    、
    available_tools
    、
    functions
    的变量
  • 包含 
    run()
    、
    execute()
    或 
    __call__()
    方法的类
将其计入总数并备注："通过启发式检测到工具 —— 请验证这是否为预期的Agent工具。"

3.4 
[PATTERN]
Context Size Awareness

3.4 
[PATTERN]
上下文大小感知

Formula: 
token_estimate = len(file_content_chars) / 4
Content⚠️ Warning threshold❌ Issue threshold
System prompt file> 4,000 tokens> 8,000 tokens
Single tool description> 500 tokens> 1,000 tokens
All tool descriptions combined> 2,000 tokens> 4,000 tokens
Exclude: 
skills/
directories (loaded on demand, not embedded)
[HEURISTIC]
Fallback: Borderline and Non-Standard
  • Estimates within 20% of threshold → flag with tokenizer recommendation
  • Dynamic prompts (f-strings, 
    .format()
    ) → flag if template alone is large
  • Multiple concatenated prompts → estimate combined size
  • Prompts with includes → note effective size may be larger
计算公式：
token_estimate = len(file_content_chars) / 4
内容⚠️ 警告阈值❌ 问题阈值
系统提示词文件> 4000 tokens> 8000 tokens
单个工具描述> 500 tokens> 1000 tokens
所有工具描述总和> 2000 tokens> 4000 tokens
排除： 
skills/
目录（按需加载，不嵌入）
[HEURISTIC]
备选方案：临界值和非标准情况
  • 估算值在阈值的20%范围内 → 标记并推荐使用分词器
  • 动态提示词（f-strings、
    .format()
    ）→ 若模板本身较大则标记
  • 多个拼接的提示词 → 估算总大小
  • 包含引用的提示词 → 备注实际大小可能更大

3.5 
[HEURISTIC]
Explicit Tool Listing

3.5 
[HEURISTIC]
工具显式列出

Check system prompts for:
  • Headers like "Available Tools", "You have access to"
  • Tool capability descriptions
Flag if tools are defined but not documented in system prompt.
检查系统提示词是否包含：
  • 类似“可用工具”“你可以访问”的标题
  • 工具能力描述
若工具已定义但未在系统提示词中记录则标记。

3.6 
[PATTERN]
LangGraph Graph Cycle Analysis

3.6 
[PATTERN]
LangGraph图循环分析

(Only when LangGraph is detected)
Detection steps:
a. Find graph file (
graph.py
, 
graph.ts
, or file with 
StateGraph
/
MessageGraph
)
b. Build edge map:
  • workflow.add_edge(source, dest)
    — unconditional edge
  • workflow.add_conditional_edges(source, fn, mapping)
    — extract destinations from mapping
c. Identify cycles: nodes reachable from themselves
d. For each cycle, check if 
END
(or 
"__end__"
) is reachable via conditional edge
ConditionSeverity
Cycle exists, 
END
reachable via conditional		✅ Pass
Cycle exists, no path to 
END
❌ Issue
Graph has no 
END
node		❌ Issue
Node has no outgoing edges and is not 
END
⚠️ Warning (dead-end)
Example — infinite cycle (❌ Issue):
python
workflow.add_edge("agent", "tools")
workflow.add_edge("tools", "agent")  # no path to END
Example — cycle with exit (✅ Pass):
python
workflow.add_conditional_edges("agent", should_continue, {
    "continue": "tools",
    "end": END
})
workflow.add_edge("tools", "agent")
[HEURISTIC]
Fallback: Non-LangGraph Graphs
Scan for graph-like control flow:
  • State machines with transition tables
  • Custom routing with implicit cycles
  • LangGraph.js (camelCase methods)
  • CrewAI/AutoGen agent handoffs
  • Adjacency lists without termination path
Flag as ⚠️ Warning: "Potential cyclic control flow — verify termination condition exists"
(仅检测到LangGraph时执行)
检测步骤：
a. 找到图文件（
graph.py
、
graph.ts
或包含 
StateGraph
/
MessageGraph
的文件）
b. 构建边映射：
  • workflow.add_edge(source, dest)
    —— 无条件边
  • workflow.add_conditional_edges(source, fn, mapping)
    —— 从映射中提取目标节点
c. 识别循环：可从自身到达的节点
d. 对每个循环，检查是否可通过条件边到达 
END
（或 
"__end__"
）
条件严重程度
存在循环，可通过条件边到达 
END
✅ 通过
存在循环，无到达 
END
的路径		❌ 问题
图中无 
END
节点		❌ 问题
节点无出边且不是 
END
⚠️ 警告（死胡同）
示例 —— 无限循环（❌ 问题）：
python
workflow.add_edge("agent", "tools")
workflow.add_edge("tools", "agent")  # 无到达END的路径
示例 —— 带出口的循环（✅ 通过）：
python
workflow.add_conditional_edges("agent", should_continue, {
    "continue": "tools",
    "end": END
})
workflow.add_edge("tools", "agent")
[HEURISTIC]
备选方案：非LangGraph图
扫描类图控制流：
  • 带有转换表的状态机
  • 包含隐式循环的自定义路由
  • LangGraph.js（驼峰命名方法）
  • CrewAI/AutoGen Agent交接
  • 无终止路径的邻接表
标记为 ⚠️ 警告："存在潜在循环控制流 —— 请验证是否存在终止条件"

Step 4: Generate Report

步骤4：生成报告

markdown
undefined
markdown
undefined

Agent Pattern Verification Report

Agent模式验证报告

Project: [name or path] Date: [current date] Framework detected: [LangGraph | CrewAI | AutoGen | LangChain | Custom | None] Files analyzed: [count]
项目： [名称或路径] 日期： [当前日期] 检测到的框架： [LangGraph | CrewAI | AutoGen | LangChain | 自定义 | 无] 分析文件数： [数量]

Summary

摘要

✅ X checks passed | ⚠️ Y warnings | ❌ Z issues
✅ X项检查通过 | ⚠️ Y项警告 | ❌ Z项问题

Loop Safety

循环安全性

  • All loops have termination conditions
  • ⚠️ Potential unbounded loop at 
    [file:line]
  • 所有循环均有终止条件
  • ⚠️ 
    [file:line]
    处存在潜在无限循环

Retry Limits

重试限制

  • All retry mechanisms have explicit limits
  • ❌ Missing retry limit at 
    [file:line]
  • 所有重试机制均有显式限制
  • ❌ 
    [file:line]
    处缺失重试限制

Tool Consistency

工具一致性

  • Tool registry found: X tools defined
  • ❌ Y hallucinated tool references
  • ⚠️ Z undocumented tools
  • 找到工具注册表：已定义X个工具
  • ❌ Y个幻觉工具引用
  • ⚠️ Z个未文档化工具

Context Size

上下文大小

  • System prompt within limits (~X tokens)
  • ⚠️ System prompt exceeds recommended size
  • 系统提示词在限制范围内（约X tokens）
  • ⚠️ 系统提示词超出推荐大小

Findings

发现结果

[P]
= pattern-matched · 
[H]
= heuristic
[P]
= 模式匹配 · 
[H]
= 启发式判断

✅ Passing

✅ 通过项

  • [P]
    [Check]: [confirmation]
  • [P]
    [检查项]：[确认内容]

⚠️ Warnings

⚠️ 警告项

  • [P|H]
    [Check]: [description]
    • Location: [file:line]
    • Suggestion: [how to fix]
  • [P|H]
    [检查项]：[描述]
    • 位置： [file:line]
    • 建议： [修复方法]

❌ Issues

❌ 问题项

  • [P|H]
    [Check]: [description]
    • Location: [file:line]
    • Rule: [which rule violated]
    • Fix: [specific remediation]
  • [P|H]
    [检查项]：[描述]
    • 位置： [file:line]
    • 规则： [违反的规则]
    • 修复方案： [具体整改措施]

Recommendations

建议

  1. [Priority recommendation]
  2. [Additional improvements]

---

*For full verification including security, quality, and language-specific checks, say "verify agent".*
  1. [优先级建议]
  2. [其他改进建议]

---

*如需包含安全、质量和特定语言检查的完整验证，请说出 "verify agent"。*