Module external_tools 

Shared governed MCP runtime for agent and interactive chat lifecycles.

Discovery is lazy, remote descriptions/results are observations, and only local policy can classify effects and footprints. Constructing one runtime never starts a server; discover_server is the first lifecycle action.

Modules

chat

Chat-mode MCP tools: the same admission and transports as the agent, behind the interactive chat interface.

Structs

DecliningMcpElicitationProvider

Non-interactive agent policy: acknowledge the protocol request and decline it immediately instead of hanging a run that has no attached input UI.

ExternalToolRejection

A locally rejected remote declaration. The server’s text is diagnostic data only; it never changes the trusted effect or footprint policy.

ExternalToolResult

ExternalToolRuntime

A single role-specific external-tool lifecycle. Agent and chat construct separate instances even when they select the same server configuration.

McpClientServices

McpElicitationBroker

A product-neutral interactive elicitation bridge. The protocol task waits on a one-shot response while a TUI or SDK host drains requests and answers them explicitly.

McpPendingElicitation

ModelTransportSamplingProvider

Sampling adapter used by the governed agent. The requesting server may suggest a model, but Perspt keeps route selection local and uses the model already selected for the product lifecycle.

Enums

ExternalConnectionState

ExternalToolEvent

McpElicitationAction

McpServerEvent

Constants

MCP_PROTOCOL_VERSION

Traits

ExternalToolObserver

McpElicitationProvider

Product-owned interaction service for server-initiated MCP elicitation.

McpSamplingProvider

Product-owned model service for server-initiated MCP sampling.

Functions

replay_key_for

Canonical replay key for a recorded external observation (public for replay assembly and conformance fixtures).