Repository navigation
TEST 123
Use a sealed run grant, pinned synthetic MCP schemas, fake clock, fake model proposals, and a call-recording MCP client. Try correct, unknown, altered, reordered, duplicate, skipped, expired, replayed, wrong-run, wrong-UUID, wrong-effect, and malformed-argument calls. Mutate the caller-owned grant dictionaries and returned call lists after broker construction. Begin an in-flight call then advance time past expiry or change host identity before completion. Abort while a call is in flight and attempt teardown before and after the call settles. Have model output and MCP responses attempt to add tools, change arguments, modify grants, skip teardown, and return executable-looking text from plan-only tools. Inject broker errors during active capture and verify original error/exit preservation.
Only the exact next canonical call in a valid, unexpired, host- and UUID-bound grant reaches the MCP client. Grant contents cannot be changed through caller-owned objects. Expiry or host-identity change during an in-flight call rejects its completion and terminates further work; abort retains unresolved in-flight state and prevents teardown until settlement. After settlement, teardown is issued only on the originally bound host and UUID; host-identity mismatch leaves that exact UUID unresolved and issues no cross-host cleanup. Executed tests are named in REQ-122. PR #252 merged as fd998aab3551f3ed5665bf809ebc3f08f79209de after seven checks passed; the full suites passed 2,193 Python and 108 .NET tests. These are offline broker tests and do not prove real-time model quality, hardware attestation, or live provider behavior.