Reference: configuration

Status: Implemented in v0.1. Stability: experimental.

Single page for every knob, its default, and where it applies. Precedence for overlapping settings: call site > capability meta > runtime defaults.

Runtime (createAgentRuntime)

Option	Type	Default	Notes
registry	CapabilityRegistry	— (required)
policies	AgentPolicy[]	[]	Evaluated in order, before per-capability meta.policies
approvals.coordinator	ApprovalCoordinator	in-memory	Dev/test default; production supplies persistent impl
approvals.handler	(req) => Promise<ApprovalDecision | undefined>	—	Inline mode; deciding requests never suspend; returning undefined defers that request to the coordinator flow (ADR-006 addendum)
approvals.rejectSelfApproval	boolean	true	SI-4; disable only with a documented reason
audit	sink | sink[] | { sinks, strict, onSinkError }	none	No sink = no audit persistence; docs warn loudly
audit.strict	boolean	false	Await capability.started before execution; fail with AUDIT_UNAVAILABLE
tracing	TracingAdapter	no-op	@orpc-agent/opentelemetry provides one
defaults.timeoutMs	number	30_000	Per-execution ceiling; capability timeoutMs overrides
defaults.policyTimeoutMs	number	5_000	Per policy-evaluation batch; exceeding ⇒ POLICY_FAILED (deny)
defaults.approvalExpiresInMs	number	900_000	Overridable per capability (meta.approval.expiresInMs) and per decision (requireApproval({ expiresInMs }))
now	() => Date	system clock	Injected for deterministic tests

Capability meta (summary; full page: metadata.md)

Field	Default	Governs
expose	— (required; absent surface = denied)	SI-1
sideEffect, risk	— (required)	Policy targeting, retry eligibility
timeoutMs	runtime default	Stage 10
retry.maxAttempts	0	Stage 11 retries (eligibility per SI-11)
retry.backoffMs	250	Exponential base
idempotent	false	Write-retry eligibility
approval.required	false	Static stage-8 gate
approval.expiresInMs	runtime default
redact.output	identity	Stage 13
redact.approvalInput	identity	Approval UI display
policies	[]	After runtime-level policies
adapters.*.toolName	id with .→_	Naming only

Invocation options (invoke / describe / resume)

Option	Default	Notes
actor	— (required)	Authenticated identity; never model-derived (SI-3)
context	— (required)	The app's oRPC context
surface	"direct"	Adapters hardcode theirs
signal	none	Composed with timeout at stage 10
correlationId	none	Threads run/conversation ids through events and spans

Adapter options

toAISDKTools(runtime, options) — adapters/ai-sdk.md

Option	Default
actor, context	— (required)	Bound per tool set (build per request)
filter	none	UX narrowing, not authorization (SI-2)
toolNaming	.→_	Collision ⇒ startup error

createMCPServer(runtime, options) — adapters/mcp.md

Option	Default
createContext	— (required)	Session → { actor, context }
serverInfo	{ name: "orpc-agent", version: pkg }
filter, toolNaming	as above

createOpenTelemetryTracing(options?) — adapters/opentelemetry.md

Option	Default
tracer	trace.getTracer("orpc-agent")
actorIdAttribute	false	Off by default (SI-10-adjacent)

createAgentTestRuntime(options) — adapters/testing.md

Option	Default
registry	— (required)
policies	[]
approvals	approvalProbe()	Or "auto-approve" / "auto-reject"
actor	fakeActor()	Per-call override allowed
context	{}	Per-call override allowed
overrides	{}	Record<capabilityId, handler> — stub procedure handlers
clock	testClock()	Drives now, expirations

Environment

Nothing is read from environment variables. All configuration is explicit code — deliberate: governance settings should be reviewable in source control.