Mutable Principal
Agents that serve multiple users, escalate privileges mid-session, or refresh auth tokens cannot use a fixed principal.
Right page if: your agent needs to change its identity context mid-session -- multi-tenant per-call resolution, privilege escalation, or token refresh without losing session state. Wrong page if: you need a static principal that never changes -- see https://docs.edictum.ai/docs/concepts/principals. For the constructor signature shared by all adapters, see https://docs.edictum.ai/docs/adapters/overview. Gotcha: principal_resolver always wins over the static principal. If both are set, the static principal is ignored. Session state (attempt counts, limits) is preserved across set_principal() calls -- only the identity changes.
Agents that serve multiple users, escalate privileges mid-session, or refresh auth tokens cannot use a fixed principal. The identity context needs to change between tool calls without tearing down the adapter or losing session state.
from edictum import Edictum, Principal
from edictum.adapters.langchain import LangChainAdapter
guard = Edictum.from_yaml("rules.yaml")
adapter = LangChainAdapter(
guard,
principal_resolver=lambda tool, args: Principal(
org_id=args.get("tenant_id"),
role="customer",
),
)Every tool call now resolves the principal from its arguments. Rulesets that check principal.org_id or principal.role enforce per-tenant limits automatically.
How it works
set_principal(principal)
Updates the principal stored on the adapter (or the Edictum instance). All subsequent tool calls use the new principal. In-flight calls are not affected -- only calls that start after set_principal() see the change.
Available on:
Edictum.set_principal(principal)LangChainAdapter.set_principal(principal)ClaudeAgentSDKAdapter.set_principal(principal)CrewAIAdapter.set_principal(principal)AgnoAdapter.set_principal(principal)SemanticKernelAdapter.set_principal(principal)OpenAIAgentsAdapter.set_principal(principal)
Session state (attempt counts, execution history) is preserved across set_principal() calls. Only the identity context changes.
principal_resolver
A callable with the signature:
def principal_resolver(tool_name: str, tool_input: dict[str, Any]) -> Principal:
...tool_name: the name of the tool being calledtool_input: the arguments passed to the tool- Returns a
Principalthat overrides the static principal for that call
When set, the resolver is called on every tool call before the pipeline evaluates rulesets. Its return value becomes the principal on the ToolCall for that call.
Pass it through the constructor:
# On adapters
adapter = LangChainAdapter(guard, principal_resolver=my_resolver)
# On the Edictum class directly
guard = Edictum(
rules=[...],
principal_resolver=my_resolver,
)Resolution order
The pipeline resolves the principal in this order:
- If
principal_resolveris set, call it with(tool_name, tool_input)and use its return value. - Otherwise, use the static principal (set via constructor or
set_principal()). - If neither is set, the principal is
None.
The resolver always wins. If you set both a static principal and a resolver, the resolver's return value is used for every call. The static principal is ignored while a resolver is active.
Rulesets with mutable principals
Rulesets that reference principal.* selectors work identically whether the principal is static or dynamic. The rule sees whatever principal is on the envelope at evaluation time.
Per-tenant rate limits
apiVersion: edictum/v1
kind: Ruleset
metadata:
name: tenant-limits
defaults:
mode: enforce
rules:
- id: tenant-query-limit
type: session
limits:
max_tool_calls: 100
then:
action: block
message: "Tenant has exceeded 100 queries per session."Combined with a principal_resolver that sets org_id per call, each tenant's queries are tracked against the session limit. Note that session rulesets count at the session level -- if you need per-tenant counting, use separate session IDs per tenant.
Role-gated escalation
apiVersion: edictum/v1
kind: Ruleset
metadata:
name: role-gates
defaults:
mode: enforce
rules:
- id: write-requires-operator
type: pre
tool: deploy
when:
principal.role: { not_in: [operator, admin] }
then:
action: block
message: "Write operations require operator or admin role."
- id: write-file-requires-operator
type: pre
tool: write_file
when:
principal.role: { not_in: [operator, admin] }
then:
action: block
message: "Write operations require operator or admin role."Before the human approval step, the agent has role: "analyst" and write tools are blocked. After set_principal(Principal(role="operator")), the same rulesets allow writes through.
Audit trail
Every AuditEvent includes the principal that was active at the time of the tool call. When the principal changes mid-session, the decision log reflects the transition:
{"tool_name": "read_file", "principal": {"role": "analyst"}, "action": "call_allowed"}
{"tool_name": "deploy", "principal": {"role": "analyst"}, "action": "call_denied"}
{"tool_name": "deploy", "principal": {"role": "operator"}, "action": "call_allowed"}The second deploy call succeeds because set_principal() updated the role between attempts. The decision log makes this explicit.
Next steps
- Principals -- principal fields, propagation, and missing-principal behavior
- Writing rulesets -- YAML rulesets that use
principal.*selectors - Adapter comparison -- how each adapter handles principal resolution
- Observability -- audit events include principal context for monitoring
Last updated on