secret

Written in place of a secret. Distinct from the regex anonymizer's [SECRET_DETECTED] so the two mechanisms are distinguishable in a trace.

A string that LangSmith serializes as [LANGSMITH SECRET].

Wrap a credential once, where it is read, and LangSmith masks it wherever it appears in a trace, at any nesting depth::

@traceable
def call_vendor(api_key: str, prompt: str) -> str: ...

call_vendor(api_key=LangSmithSecret(key), prompt="hi")
# traced inputs: {"api_key": "[LANGSMITH SECRET]", "prompt": "hi"}

It is a real str everywhere else: json.dumps, logging and third-party clients all see the true value. Operations that derive a new string return a LangSmithSecret again, so the marker is not lost by .strip() or slicing.

Known limits:

• f"Bearer {secret}", "".join([secret]) and "Bearer {}".format(secret) yield a plain str: those build the result in C, where the operand gets no say in its type. Wrap the finished value instead -- LangSmithSecret(f"Bearer {key}"). Concatenation and % are covered, in either operand order.
• secret.encode() returns the real bytes, by design.
• A secret used as a dict key is masked, but two of them in one dict collapse to a single entry. A key held inside a dataclass is not masked: orjson serializes those itself.
• An object whose own __str__ interpolates a secret leaks it wherever LangSmith stringifies that object. __repr__ is masked; __str__ belongs to the object.
• Nothing inside a pydantic model is masked. Pydantic serializes its own fields and LangSmith deliberately does not override that: a hook there would also replace the credential with the mask when a vendor client dumps the model to build a request.