A Cycle Spent Hunting Quiet Failures
Every layer of the stack spent this cycle on release preparation, and every pass went looking for the same thing: results that look complete and are not. Here is what each layer hardened, and what ships next.
Last cycle went to the bottom of the stack. This one went to every layer at once, and to one question: what happens when someone who is not us installs it, runs it hard, and does not read the manual first?
Release preparation is mostly that question, asked on purpose. Use each product the way a new person would. Push it the way a long day of real work pushes it. Then look hardest for the failures that look like success, because those are the ones an agent will build on without a second thought. A crash gets noticed. A result that is quietly incomplete gets used. Every pass this cycle went looking for the second kind, and closed what it found.
TheAuditor: answers that admit their edges
Stress-tested against our own codebase, the MCP server in TheAuditor now answers from the facts its index stores and says plainly when an answer was cut. A gate also sends the agent to the query instead of the whole file, because a polite reminder loses to habit. That story is in every answer says what it left out.
The release build also went through a clean-machine pass, used the way a new customer would use it, and came out with safe interrupts, pinned tools, and exit codes CI can trust, as covered in the first hour is part of the product. The reproducibility scorecard is still being measured, and it will be published when it passes.
Warden: a long session, a hostile repository, and honest numbers
Warden was run through long live sessions and killed mid-task, and the session came back intact. Its verification also asked what a freshly cloned repository can make it do before you trust the folder, and closed each path it found. Its token and cost figures now come from live requests and replays of real sessions, each filed with its limits, and the one saving nobody has measured yet is not claimed.
Arbiter: off one machine, and stricter by default
Arbiter now runs its agent sessions natively on Windows as well as Linux, and every build has to behave the same on both before it can ship, as told in off the machine it was built on. Remote control got stricter by default: asking a session for a plan now needs a code, and an approval answers only the request that asked for it. The first public pre-release candidate is prepared, and it goes out when its checks pass.
Curator: the save that has to land
With many sessions writing at once, a save to Curator goes straight through even while a backlog drains, and its health report says plainly when the shared model is paused, as told in when every agent writes at once. Defaults carried in from other projects are now only the settled ones, and the encrypted store can now be opened after a reinstall with a recovery code, with upgrades that refuse rather than guess. That one is in an encrypted memory should survive a reinstall.
BenchProctor: a release held to its own gate
Last cycle’s note described BenchProctor moving to whole applications. The release that carries them ships when its full gate run is green. The first full runs caught issues at the compiler and tightened the gates themselves; each was fixed at the source, and the full run at release depth is next. The same pass read every label against the code under it and added weakness classes the corpus never covered. Until the new release is green, the current public release remains 2026.07.22.
What did not change
No product put out a new public release this cycle. Each one got closer by finding and fixing, in house, what would otherwise have surfaced in someone else’s hands. The standard is the same as it was: a release goes out when its checks pass, not on a date. Read how the pieces fit at Code Reality Labs.