🤝 Top Collaborators
.png&w=640&q=75)
🤓 Latest Submissions
Never Twice: Postmortems That Prevent Repeat Bugs
THE PROBLEM After an incident, teams write a postmortem: what broke, why, and how it was fixed. Then the postmortem is filed away. Almost nobody searches the rest of the codebase for the same mistake, so the same class of bug ships again in a different function. Doing that search by hand is slow and unreliable: in our baseline, a developer spent 26 minutes with our three postmortems and found 1 of 7 repeat bugs. OUR SOLUTION Never Twice is a custom IBM Bob mode that turns postmortems into protection. A developer opens their project in Bob, picks the Never Twice mode and types one line. Never Twice then: 0. detects the project's language, test framework and existing CI; 1. reads every postmortem and extracts a general bug pattern; 2. spawns one subagent per pattern to hunt the whole codebase in parallel; 3. proves each finding with a test that fails on the current code; 4. builds a guardrail (a static checker plus a PR review checklist) that flags exactly those patterns; 5. stops at a human approval checkpoint - nothing changes until the developer reviews the findings and types "approve"; 6. applies the fixes and reruns all tests and the guardrail; 7. writes a Prevention Report. TARGET USERS Development and SRE teams who already write postmortems. Never Twice installs once with one command and then works in any project opened in Bob. Ready-made CI templates for Python and JavaScript run the guardrail on every pull request, so a pull request that brings back a guarded bug pattern fails. WHAT MAKES IT DIFFERENT Most AI code tools review code in general. Never Twice starts from the team's own incident history, so it hunts for the bugs that have already hurt this team. Every finding must be proven by a failing test, not just described, and the guardrail must flag exactly the confirmed bugs before the checkpoint - no extra flags on safe code. We also publish what went wrong: a regression round that failed our review, and one bug we could not prove. RESULTS On our Python hotel-booking app, Never Twice found 7 of 7 hidden bugs in about 6 minutes of automated work (about 11 minutes including human review), wrote 13 tests that fail before and pass after the fixes, and took the guardrail from 7 issues to 0 - with the developer typing two messages. To check it was not tuned to our own sample, one teammate built a JavaScript shop with hidden bugs, decoys and already-fixed traps, and the other ran Never Twice blind. It flagged 6 of 6 hidden bugs, proved 5 with failing tests, raised 0 false alarms on 3 decoys, and found one real bug the answer key missed. A teammate then set up the global install from the README on a second laptop in 4 minutes. All answer keys, results, disclosures and Bob sessions are published in the repository.
27 Sep 2026
