agentvet-py
MukundaKatta/agentvet-py
Python port of @mukundakatta/agentvet: validate LLM-generated tool args before execution
Instalação
npx convez-ai MukundaKatta/agentvet-py1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
agentvet-py
Validate LLM-generated tool args before execution. Wraps tool functions with arg validation; raises ToolArgError with an LLM-friendly retry hint. Zero runtime dependencies.
Python port of @mukundakatta/agentvet.
Install
pip install agentvet-py
# pydantic adapter is optional:
pip install "agentvet-py[pydantic]"
Usage
from agentvet import vet, adapters, ToolArgError
def send_email_impl(args):
return "sent to " + args["to"]
schema = adapters.shape({"to": "str", "subject": "str", "body": "str?"})
send_email = vet(name="send_email", schema=schema, fn=send_email_impl)
# Happy path:
send_email({"to": "alice@example.com", "subject": "hello"}) # 'sent to alice@example.com'
# Bad args -> raises ToolArgError BEFORE the tool runs:
try:
send_email({"subject": "hello"}) # missing 'to'
except ToolArgError as err:
feedback = err.to_llm_feedback()
# Send `feedback` back to the LLM as a tool_result with is_error=True.
With pydantic
from pydantic import BaseModel, EmailStr
from agentvet import vet, adapters
class SendEmail(BaseModel):
to: EmailStr
subject: str
body: str | None = None
send_email = vet(
name="send_email",
schema=adapters.pydantic(SendEmail),
fn=lambda args: send_impl(**args),
)
Async tools
vet() preserves the sync/async nature of fn:
async def fetch_impl(args):
return await api.get(args["url"])
fetch = vet(name="fetch", schema=adapters.shape({"url": "str"}), fn=fetch_impl)
result = await fetch({"url": "https://example.com"})
API
vet(*, name, schema, fn, on_error=None) -> wrapped_fn
Wraps a tool function. Validates args BEFORE calling the tool; raises ToolArgError on failure (or invokes on_error(err, args) if provided -- return a non-None value to substitute as the tool's return).
validate(name, schema, args) -> ValidationResult
One-shot validation. Returns ValidationResult(ok=bool, value=..., error=ToolArgError | None).
adapters.shape(spec)
Tiny built-in shape checker. Spec format: {"field": "str"|"int"|"float"|"bool"|"list"|"dict", ...}. Suffix with ? for optional. Accepts JS sibling synonyms ("string", "number", "array", "object") too.
adapters.fn(predicate, error_builder?)
Predicate adapter. predicate(args) -> bool. error_builder may be a string or a callable returning a string.
adapters.pydantic(model_cls)
Wraps a pydantic v2 BaseModel. Requires pip install agentvet-py[pydantic].
adapters.zod(schema)
Compatibility shim for safeParse-style validators (rare in Python; included for parity).
ToolArgError
Carries tool, validation_error, args. .to_llm_feedback() returns the retry message you send back to the LLM.
API differences from the JS sibling
vet()uses keyword-only args (vet(name=..., schema=..., fn=..., on_error=...)).- Validators return Python dicts:
{"valid": True, "value": ...}/{"valid": False, "error": str}. adapters.pydanticreplacesadapters.zodas the natural Python adapter.- Type names in
shape()use Python conventions ("str","int","list","dict"); JS synonyms ("string","number","array","object") are accepted. ToolArgError.to_llm_feedback()issnake_case(mirrors JStoLLMFeedback()).
See the JS sibling's README for the full design notes.