Design Trade-offs and Lessons Learned
Recovery first requires clarity on what to preserve. The terminal supervisor preserves OS sessions, the ledger preserves submitted results, and business-idempotence keys preserve a single business migration. In the source, the enrollment test re-requests with a new nonce after service and ledger restart; the same business key still returns the original result, and the version and enrollment count do not increase. This example shows that a single new request at the transport layer can correspond to the same operation at the business layer.
Understanding the meaning of completion status is also important. MCP tool returns a business error, yet it can still form a completed Invocation: the call has delivered an erroneous result. The corresponding recovery test validates that after the final state is submitted, a process kill occurs; records and the receipt chain remain readable and verifiable. When debugging, you should first check the result, then decide whether to retry; this evidence does not imply that any in-progress interruption can automatically resume.
Functional description must follow executable paths. Some old comments still say MCP hot-update is not implemented, but there are already notification refresh tests; media metadata table tests require that unimplemented capabilities cannot be inferred solely from a published statement. The next steps worth investing in include more precise memory retrieval, and improving idempotence and recovery experience around real write tasks: the current focus on recent memory strategy and submitted-result recovery provides a clear starting point for these directions.
No comments yet

