mcpagent

Overview ¶

Command mcpagent obtains its system prompt and search tool from an MCP server. The stub model derives its brief from actual tool output and reports refusal explicitly. Request metadata is bound when the remote tools are discovered, because a durable call runs long after the caller's context is gone.

Run from the repository root:

go run ./examples/mcp/mcpagent

On Unix, the process-boundary tests build mcp/expert as a separate SDK executable, with no Scope imports or inherited environment. It makes no network or model calls. tools/process owns process lifetime, the SDK owns its session, Core Guard owns authorization, and Agent owns execution uncertainty.

cd examples
go test -race -count=1 -run '^TestMCPProcessBoundary$' ./mcp/mcpagent

Host selection uses the original MCP source and remote identity. ToolSet freezes the canonical tool contract. Its revision identifies the expert artifact and opaque authorization callback; the interaction revision identifies the model query. Compile hashes those declarations without caller-created configuration digests. Session identity is excluded so reconnection can bind the same trusted artifact, while changing its grant refuses recovery before writer activation. Ordinary tool calls remain in the interaction process.

Cancellation notifies the peer without closing its session. A lost outcome remains Unknown through reconnection and cancellation; the tree cannot emit Exited until the outcome is resolved. Runtime.Close stops local work while retaining that fact. A fresh execution can make a new call without advancing the unresolved log. Connection health supplies no receipt or replay authority. The reconciliation example demonstrates external receipt investigation.