MCP Server Change Checklist

Read only the sections relevant to the requested change.

JSON-RPC and MCP contracts

Preserve the JSON-RPC envelope, request id, and HTTP-versus-JSON-RPC error behavior.
Confirm invalid JSON, invalid requests, unknown methods, invalid params, and tool failures map to the intended error codes.
Keep initialize, notifications/initialized, tools/list, and tools/call behavior internally consistent.
Update request samples when a public payload or response shape changes.

Ensure method resolution remains explicit and unsupported methods fail safely.
Validate arguments before calling an upstream Tool Service.
Preserve the requested Tool order and result-to-call correlation.
Keep tool selection outside this server; never infer or substitute a Tool implicitly.
Confirm timeout, partial failure, and upstream client errors produce safe normalized results.

Keep the request path free of Redis: tools/list and tools/call read the in-memory snapshot only.
Keep Redis on the background refresh path only, as a cross-replica sharing point rather than a source of truth.
Write to Redis only after a successful source fetch, and update memory before Redis.
Never let any failure path empty the Tool snapshot; replace it only on success.
Treat Redis health, serialization, read, and write failures as cache degradation rather than correctness failures.
Confirm cold-start behavior when no cache is available.
Keep metadata versioning and enabled/disabled Tool behavior explicit.
Preserve the Redis key namespace (key-prefix:identity:schema-version) when touching cache code.

Propagate traceId, requestId, mcpSessionId, agentId, and userId where the contract requires them.
Clear ThreadLocal and MDC state on every completion and exception path.
Mask sensitive values only in log and audit copies; do not mutate execution payloads.
Keep payload logging disabled by default.
Verify audit hooks do not change business results.

Preserve NDJSON framing and final response semantics.
Flush progress notifications without losing the original request correlation.
Handle upstream SSE data: lines and NDJSON consistently.
Consider client disconnect, proxy buffering, timeout, and back-pressure implications.
Test both streaming and non-streaming paths when shared code changes.

Keep the local profile runnable without Redis.
Keep production values externally configurable; do not commit credentials or trust material.
Preserve stateless replicas, read-only container filesystem support, and arbitrary UID compatibility.
Review liveness, readiness, graceful shutdown, ingress buffering, timeouts, and egress policy when deployment behavior changes.
Explain compatibility impact when changing Java, Spring Boot, Spring AI BOM, Gradle, or container settings.

Mirror production responsibility in the corresponding test package.
Add regression coverage for the failure mode being fixed.
Prefer focused unit tests; use broader application tests for wiring and profile behavior.
For every added or modified src/main/java top-level type, add or refresh a Korean Javadoc directly above the declaration. Explain its responsibility, its HTTP/MCP request or processing stage (or state that it has none), and its key dependencies or external boundary.
For every added or modified explicit production method or constructor, add or refresh a Korean Javadoc. Explain its purpose, input/output or error handling, and invocation timing or impact. Exclude only compiler-generated record accessors, simple enum constants, and empty marker interfaces.
Review the changed source declarations and their Javadocs before test execution; comment coverage is required even when behavior tests pass.
Update docs/architecture.md for responsibility or flow changes.
Update docs/extension-points.md for new operational gaps, agreements, or extension guidance.
Run the Gradle wrapper test task before handoff.