Killing the god object — migrating 96 test files from QhorusMcpTools
QhorusMcpTools was 2092 lines of accumulated test infrastructure. Every test in the project injected it. Every channel creation was a 19-argument method call where 17 of them were null. After #451 stripped the MCP annotations, the class was inert — a CDI bean with no MCP footprint, used only by tests. Time to kill it.
The approach was pragmatic, not pure. I’d planned to replace the 19-arg calls with clean builder overloads — helper.createChannel("name") instead of tools.createChannel("name", "desc", null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null, null). But 488 createChannel call sites across 96 files made a clean break impractical in one pass. We added a compat 19-arg overload to QhorusTestHelper that matches the old signature exactly — parse the CSV strings, construct the builder internally, return the domain type instead of the MCP DTO. The bulk replacement became mechanical: swap the import, rename the field, move on.
The module cycle was the first real constraint. testing/ depends on runtime. runtime can’t depend back on testing. The project already knew this — CLAUDE.md documents it for RecordingChannelBackend. The helper lives in both places: testing/ for other modules, runtime/src/test/ for runtime’s own tests. Code duplication, but the alternative — a new intermediate module — adds more complexity than it removes.
The bulk migration covered imports, field injection (tools → helper), method calls, and inner type references. The regex caught most patterns but missed one: imports of inner types via the outer class’s package path — import io.casehub.qhorus.runtime.mcp.QhorusMcpToolsBase.CheckResult survived the class-name replacement and needed a second cleanup pass.
The interesting failures were the behavioural gaps, not the compile errors. getReplies and searchMessages accepted a readerInstanceId parameter but silently ignored it — the visibility filtering that checked whether a targeted message (capability:code-review, instance:alice) should be visible to a given reader was in the old MCP tools but never migrated to the helper. Five tests relied on this filtering. channelDigest returned an empty topic breakdown because nobody added the TopicStore query. listLedgerEntries didn’t validate the sort parameter or wrap DateTimeParseException. Each gap was a method that compiled and ran but didn’t do what the tests expected.
The dual-copy problem caught us once. IntelliJ’s MCP edited testing/src/main/java/.../QhorusTestHelper.java — the first match in its index. The runtime copy at runtime/src/test/java/.../QhorusTestHelper.java stayed untouched. We ran the tests three times before noticing the runtime tests were still using the old code. The fix: sync the copies with a Python shutil.copy2 call that bypasses the IDE’s file routing entirely.
After the last test turned green — 2038 passing, zero failures — we deleted QhorusMcpTools.java and QhorusMcpToolsBase.java. 2692 lines removed. What was a monolithic MCP-entangled test facade became a 1530-line helper with fifty methods and twenty inner record types. It delegates to the real services and stores. The 19-arg compat overloads still exist — they’ll get cleaned up when individual test files are touched for other reasons. No rush.