Skill in 《Agent Tool Builder》

Skill.md

Agent Tool Builder

Helps define agent tools using the fail-closed design pattern: a unified class that co-locates identity, schema, security properties, and execution logic, with fail-closed defaults so new tools are safe by default.

Why this pattern matters

Three things that ad-hoc tool definitions lack:

1. Fail-closed defaults — is_read_only, is_destructive, is_concurrency_safe all default to False. A tool that forgets to declare its properties is conservatively treated as write-capable.
2. Layered execution — validate_semantics → check_permissions → _call are separate methods, so validation logic doesn't bleed into permission logic or business logic.
3. Self-contained definition — schema, description, security metadata, and execution all live in one place. No separate middleware to wire up.

Workflow

Step 1 — Identify the target framework

Ask which agent framework the tool will be registered in (e.g. hermes-agent, LangChain, plain Python). This determines the import path and registration method, but the design principles are identical.

Check if agent_tool_base.py exists in the project's utils/tools directory. If not, copy it from references/agent_tool_base.py in this skill directory. Tell the user where it was placed.

Step 2 — Interview the user

Collect answers to these questions. Defaults are shown — skip questions where the default is clearly fine.

Naming convention: use {service}_{action}_{resource} format with a service prefix so the tool stays unambiguous when multiple tool sets are loaded simultaneously (e.g. stock_get_price, stock_list_symbols, stock_search_news). Start with a verb: get, list, search, create, delete.

You don't have to ask all questions upfront — infer reasonable answers from context. For example, a "search" or "get" tool is almost certainly is_read_only=True, is_concurrency_safe=True.

Step 3 — Generate the tool file

Create a .py file for the tool. Follow this field order:

1. imports
2. Input schema (Pydantic BaseModel)
3. Tool class:
   a. name, description, args_schema      — identity
   b. is_read_only, is_destructive, is_concurrency_safe, max_result_chars  — security metadata
   c. _validate_input_semantics()         — semantic validation (omit if unneeded)
   d. _check_permissions()               — permission check (omit if unneeded)
   e. _call()                            — actual logic

Suggest a file path consistent with the project's tool directory structure.

Step 4 — Show security property summary

After generating, print a one-line summary of the tool's security posture:

StockGetPriceTool: read_only=True  destructive=False  concurrency_safe=True  max_result=10K

Output template

"""<tool_name>.py — <one-line description>"""

from typing import Optional
from pydantic import BaseModel, Field
from base.utils.agent_tool_base import AgentTool


# ---------------------------------------------------------------------------
# Input schema
# ---------------------------------------------------------------------------

class <ToolName>Input(BaseModel):
    <field_name>: <type> = Field(description="<description>. e.g. '<example>'")
    # ... more fields


# ---------------------------------------------------------------------------
# Tool class
# ---------------------------------------------------------------------------

class <ToolName>Tool(AgentTool):
    # — identity —
    name: str = "<tool_name>"
    description: str = "<one-sentence description for the LLM>"
    args_schema = <ToolName>Input

    # — security metadata (fail-closed: only set True when verified) —
    is_read_only: bool = <True/False>
    is_destructive: bool = <True/False>
    is_concurrency_safe: bool = <True/False>
    max_result_chars: int = 10_000

    # — semantic validation (omit if no input constraints needed) —
    def _validate_input_semantics(self, <params>, **kwargs) -> tuple[bool, Optional[str]]:
        if not <condition>:
            return False, "<why invalid>. Try <concrete fix>"
        return True, None

    # — permission check (omit if no access control needed) —
    def _check_permissions(self, <params>, **kwargs) -> tuple[bool, Optional[str]]:
        if not <allowed>:
            return False, "<why denied>. <suggested next step>"
        return True, None

    # — core logic —
    def _call(self, <params>, **kwargs) -> str:
        # ... implement tool logic here
        return result

Common security property patterns

Output design principles

Atomic tools — one tool, one responsibility

Keep each tool focused on a single operation. Let the agent compose multiple tools to complete complex tasks. A tool that does too much is harder for the agent to reuse and reason about.

Response format — JSON vs Markdown

Support both when uncertain — accept an optional response_format: str = "markdown" parameter and branch in _call. For JSON output use json.dumps(data, ensure_ascii=False, indent=2).

Pagination for list tools

Any tool that can return more than ~50 records should support pagination:

return json.dumps({
    "items": [...],
    "total": 150,
    "count": 20,
    "offset": 0,
    "has_more": True,
    "next_offset": 20,
}, ensure_ascii=False, indent=2)

Add offset: int = Field(default=0, description="Pagination offset") and limit: int = Field(default=20, description="Max items to return") to the input schema.

Actionable error messages

Error strings must guide the agent toward a fix — not just describe the failure:

# Bad: agent is stuck
return False, "Query too short."

# Good: agent knows exactly what to try next
return False, "Query too short (got 2 chars, need >= 3). Provide a more specific search term."

Reference files

• references/agent_tool_base.py — Full AgentTool base class (pure Python, no framework dependency)