governed_identity

A context manager that asserts the end user's identity for the duration of a governed call — separate from the agent's own static identity (agent_id/Caller), which is set once at construction. Cedar policies that gate on context.identity_claims/context.identity_roles (like the quickdemo's org-scoped rules) see whatever this sets.

Two implementations, not one

There are two distinct governed_identity() definitions in this codebase, both context managers, both meant to be imported by end users — but with different signatures. Which one you get depends on where you import it from.

parapetai_agent.scoped_data.governed_identity

The base implementation. This is what parapetai_agent.adk re-exports verbatim (from parapetai_agent.scoped_data import governed_identity) — so for Google ADK, this is the one you use.

from parapetai_agent.scoped_data import governed_identity
# or, equivalently, for ADK:
from parapetai_agent.adk import governed_identity

@contextlib.contextmanager
def governed_identity(
    *,
    claims: Mapping[str, Any] | None = None,
    roles: Sequence[Any] | None = None,
    token: str | None = None,
    extractor: TokenIdentityExtractor | None = None,
) -> Iterator[None]: ...

Parameter	Type	Meaning
claims	Mapping[str, Any] \| None	Already-parsed end-user identity claims, e.g. {"org": "Sales", "name": "Tony"}. Set ambiently via current_identity() for the duration of the with block.
roles	Sequence[Any] \| None	Already-parsed end-user roles, e.g. ["OrderViewer"]. Used together with, or independently of, claims.
token	str \| None	A raw bearer token, decoded via identity_from_bearer_token() into both end-user identity and (if present) agent identity.
extractor	TokenIdentityExtractor \| None	Only relevant with token= — overrides the default JwtIdentityExtractor() for a non-JWT token format.

Exactly one of (claims and/or roles) or token must be given:

• Zero sources: raises ValueError. If there's genuinely no identity to assert, call the framework's own run method directly, unwrapped, rather than this with nothing in it.
• More than one source: raises ValueError. Passing both claims/roles and token= almost always means one was left in by mistake while editing — there's no defined merge semantics across two identity sources to fall back on.

This fails loud, not silent, on ambiguity — unlike an unwrapped call (which evaluates against empty identity_claims/identity_roles; Cedar's own default-deny already makes that the safe failure mode for any policy that checks identity: denied, not skipped or bypassed).

from google.adk.agents import Agent
from parapetai_agent.adk import GovernedRunner, governed_identity

runner = GovernedRunner(agent=root_agent, app_name="demo", session_service=...)

with governed_identity(claims={"org": "Sales", "name": "Tony"}):
    async for event in runner.run_async(user_id="Tony", session_id=..., new_message=...):
        ...

parapetai_agent.maf.governed_identity

A separate, MAF-specific definition — richer, with an additional credential/scope pair for azure-identity credentials, since MAF's own FoundryChatClient commonly takes exactly that kind of credential.

from parapetai_agent.maf import governed_identity

@contextlib.contextmanager
def governed_identity(
    *,
    claims: Mapping[str, Any] | None = None,
    roles: Sequence[Any] | None = None,
    token: str | None = None,
    credential: TokenCredential | None = None,
    scope: str = "https://cognitiveservices.azure.com/.default",
    extractor: TokenIdentityExtractor | None = None,
) -> Iterator[None]: ...

Same claims/roles/token/extractor meanings as above, plus:

Parameter	Type	Meaning
credential	TokenCredential \| None	An azure-identity credential — the same one you'd pass to FoundryChatClient. Internally dispatches to identity_from_azure_credential().
scope	str	Default "https://cognitiveservices.azure.com/.default". The OAuth scope requested from credential when resolving identity.

Exactly one of (claims and/or roles), token, or credential must be given — same fail-loud rule as above (zero sources or more than one source both raise ValueError).

from azure.identity import AzureCliCredential
from parapetai_agent import GovernedAgent
from parapetai_agent.maf import governed_identity

async with GovernedAgent(client=client, name="agent", instructions="...") as agent:
    with governed_identity(credential=AzureCliCredential()):
        result = await agent.run(prompt)

Which one do I import?

You're using	Import from
GovernedAgent (Microsoft Agent Framework)	parapetai_agent.maf
GovernedRunner (Google ADK)	parapetai_agent.adk (re-exports scoped_data's version — no credential=)
Framework-neutral Governor	parapetai_agent.scoped_data — pass roles/claims directly to check_input()/authorize_tool()/check_output() instead; those calls don't read the ambient context governed_identity() sets, so wrapping them in it has no effect.

The ADK re-export has no credential=/scope= support — if you need an azure-identity credential inside an ADK agent for some other reason, build the claims yourself and pass claims=/roles= instead.