retc sprint, Nov 6–7: five releases, four of them overnight

retc is a real-estate transaction coordination app. This report covers the project-manager-led sprint ending November 7.

Scope: counts run from Nov 5, 7:00 PM to Nov 7, 8:50 AM ET. Five releases across the full sprint (deploys 96–100), including the four overnight deploys highlighted in the morning digest. The final release went live at 5:35 AM ET Saturday.

Verdict

Five releases shipped and verified live, no rollbacks. The support AI’s new instruction version (Version 7) passed all 134 safety and quality evaluation cases and the owner activated it. Its drafts still need human review, and auto-send is off. An evaluation Cancel button and two deadlock fixes shipped. A first Accept speed-up shipped. No real users yet.

Counts

• Commits: 108 on the main branch (77 non-merge, 31 merges).
• Pushes to main by the PM: 32 (builders’ own-branch pushes not counted).
• Releases: 5 (deploys 96–100). Database migrations: 4, all in deploy 97. Rollbacks: 0.
• Full test suite: 6,193 → 6,495.
• Features: 3. Fixes: 7 (classified by the PM).
• Stopped before release: 6. Two on the deploy 97 release train (a full test run, and a footer defect the browser tester caught), one on deploy 99’s full run, three on deploy 100’s (a pre-check and two full runs). The author also caught a deadlock in the eval Cancel feature after review.

What shipped

• Support AI Version 7. Fixes one case that kept failing on a forbidden word in its escalation reason. 134/134 cases passed; activated.
• Eval Cancel. Stops a running AI evaluation, requires a password, refuses finished runs, and shows the honest cost, never a fake $0.
• Deadlock fixes. Lock-order changes so a cancel or a settings change can’t deadlock with a running evaluation.
• Accept speed-up, step 1. Accept renders and verifies the purchase agreement once instead of twice and runs fewer queries: 18.3 s → 11.6 s in testing.
• Fixes: stale dates now disclosed in deadline previews; a kept date; the intake footer; the People refusal page.
• Built, switched off: an intake release/classification screen and its tables. Also new: a read-only Accept timing tool.

What was new

• The backend author found and reproduced a real HTTP deadlock in the Cancel feature after both reviewers had passed it, then fixed it with a lock-order change plus a full inventory of the locks it touches.
• Three times, a release’s full test run caught older tests the author never ran. Each repair was accepted only after proof that real behaviour hadn’t changed, including on real Linux processes.
• One reviewer briefly flagged a valid test because it assumed the wrong SQL form. The other reviewer and a measurement settled it within a minute.
• Mid-sprint, the two reviewers moved from Grok’s own harness to the Cursor harness.

Corrections the PM logged on itself

• It said a deploy would stop a mistaken evaluation run. Wrong: evaluations run in a scheduler, not the web server.
• It said there would be no deploy that night. Wrong: deploy 96 shipped.
• It merged a stale branch instead of the deployed one; it caught this itself and corrected it.
• It forgot to tell one builder about the reviewer swap (the builder rerouted itself), ran its first deploy-98 command on the wrong machine (failed harmlessly), and left idle agents open about three hours after their work was done.

Team

• PM: Claude Opus 5.5.
• Four builders and testers on GPT-6.1 (Codex): UI and release-train builder, backend (evaluations), backend (intake), browser tester.
• Two reviewers on Grok 4.7: proof review and security.

How each model did

• Claude Opus 5.5, PM. Strong gate discipline: release order and gates held through five deploys, and it demanded real-behaviour proof for test repairs. Missteps: the wrong claim about deploys stopping evaluations, the stale-branch merge, the forgotten roster notice, slow closing of idle agents.
• GPT-6.1, UI/release-train builder. No missteps observed. Collects every failure before fixing anything.
• GPT-6.1, backend (evaluations). One UI test set at first couldn’t catch a broken change (an unknown cost could show $0); fixed. Found the deadlock both reviewers missed.
• GPT-6.1, backend (intake). Under-scoped its affected tests three times; all proven harmless. Honest measurement, admitted each miss.
• GPT-6.1, browser tester. No missteps observed. Caught the deploy 97 footer defect.
• Grok 4.7, proof reviewer. Passed the Cancel feature without seeing the deadlock. Caught missing tests in two other changes; right on the framework’s SQL.
• Grok 4.7, security reviewer. Briefly flagged a valid test on a wrong SQL assumption, withdrawn within a minute; also missed the deadlock. Precise rulings on the billing boundary.

Takeaways: keep GPT-6.1 on build seats and Grok on review; require a lock inventory for any lock change; authors list their affected tests before review.

Not done yet

• Accept step 2, aiming for under 5 seconds on the live site.
• Reviewing the first real Version 7 drafts.
• Cleaning up leftover test databases.