🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Antibody
Problem When a developer fixes a bug, they fix one copy of it. The same mistake usually lives on in other files, written slightly differently, waiting to break again. Searching by hand is slow, grep cannot find the same mistake phrased another way, and linters only ship generic rules that know nothing about your team's own history. So teams pay for the same bug again and again, and the knowledge of why it happened stays in one person's head. Solution Antibody is a custom IBM Bob mode plus a small Python CLI that turns a single bug fix into lasting immunity, in four phases: 1. Infection: Bob reads the fix commit, its message and linked issue (or a postmortem) and extracts the abstract root cause: the mistake, not the location. 2. Antibodies: Bob searches the whole repository by meaning for twins of that mistake. One subagent per candidate writes a test. A twin is confirmed only if its test fails on the current code, then Bob applies a minimal fix and the test must turn green. 3. Vaccine: Bob writes a Semgrep rule. A separate red-team subagent that never sees the rule writes realistic variants of the bug, and the CLI checks them in an isolated git worktree. Variants that escape make Bob harden the rule, round after round. The result is an Immunity Score: variants caught divided by valid variants. 4. Memory: the rule stays in the repo with a message that explains what happened last time and links the original fix, so future developers and CI are protected. Target users and interaction Development teams and maintainers who want every bug fix to count. A developer opens their repo in Bob IDE, selects the Antibody mode and runs /antibody <fix-commit>. Bob walks through the phases and pauses for approval before changing production code, committing or finalizing. Results appear in a visual scoreboard with a replay of the run. What makes it unique Proof over opinion. Antibody never reports a twin it cannot demonstrate: Bob writes the tests, but only the CLI runs them and records the evidence, reading JUnit reports instead of trusting exit codes. The defense is tested adversarially by an agent that cannot see it, so the Immunity Score is measured from real Semgrep and test executions, never estimated by the model. It works across languages through test runner profiles (Python, JavaScript, Java, .NET, Go, Ruby, PHP and more) and needs no server or database: everything is reviewed as code. Impact In our demo, on a practice repository with seeded twins, Antibody proved and fixed 3 twins and raised immunity from 83% to 100% in one vaccine round. On a real open-source project, APScheduler, it found 3 real twins of a daylight-saving-time bug that are still present upstream, each proven with a failing test that passes after the fix. Your team already paid for every bug. Antibody makes sure you never pay twice.
27 Sep 2026
