This window is two repair stories running at once, and confusing them is the mistake worth avoiding.
The interesting one is US-PLAT-2. The WebSocket protocol lands in three
stories over three days: the message envelope and payload handling (2.1.1),
session lifecycle with open/heartbeat/close (2.1.2), and channel routing with
metadata and liveness (2.1.3). Each one arrives as a feat commit and is
followed the same day by a fix commit titled “address review findings”.
That pattern — ship, then repair the review findings within hours — is worth looking at honestly. It is not the same as shipping bugs. It means the review gate is working and the units are small enough that findings are cheap to clear. The thing to watch for is whether the same findings recur; here they do not, which makes it a process working rather than a quality problem.
Underneath, five Epic-2 stories close at once on 05-29: MAF provider, MAF core,
MCP SDK integration, reasoning, and git integration. And the gRPC → MCP SDK
migration starts — MCPSdkClient goes in behind a feature flag, with nested
argument structure and real timeouts fixed before anything switches over. That
is how a transport migration should go: additive and reversible, not a cutover.
The uninteresting one is the CI pin run. PRs #9 through #18 are almost all
the same move: bump a shared-workflows pin. Trivy exits 127, so bump. The
image scan needs container-based scanning, so bump. The cache escape hatch
needs an immutable pin, so bump again. PKG_TOKEN vs USER_TOKEN for registry
auth. An org migration from admin to homelab. Six PRs that change one
number.
This is the actual fix-storm the signal caught, and the honest reading is that only part of it is engineering.
Lessons learned
- A same-day review-findings fix is a process signal, not a quality signal. What tells them apart is whether the same finding comes back. Clearing findings in hours on fresh issues means the review gate is doing its job.
- Pin churn is not progress — it is tax. Six PRs bumping a version number is a workflow library that has not quite decided its own contract yet. Owning the workflow outright is cheaper than chasing its pins every fortnight.
- Feature-flag a transport migration instead of cutting over. The SDK path went in behind a flag with the edge cases fixed first, so turning it on is a decision rather than an event.