antibody v1.0
Immune system for your codebase

A bug happens once. The repo grows an antibody.

AI agents reintroduce bugs you already fixed — they have no memory of your repo. antibody turns every fix into durable, machine-readable immunity: a frozen regression test, the fix, and agent rules those same tools read before they touch your code.

View on GitHub

Free forever · Tier 0 runs with no key, no network, $0 · MIT

The thesis

AI agents reintroduce fixed bugs. They have no memory. antibody gives the repo one.

antigen → antibody → immune memory → immunization

The problem · reintroduction

again. The same class of bug, fixed last month, shipped back this week — by an agent that never saw the fix.

Reintroduction is the dominant regression vector of the AI era.

Coding agents made bugs cheap to fix and cheap to re-create. Every new session starts cold. The reasoning that produced last quarter's fix lives in a closed PR, a Slack thread, a teammate's head — nowhere an agent will look.

So the model rewrites the function, drops the guard, and reintroduces a bug your team already paid to understand. The test that would have caught it was never written, or was quietly deleted alongside the regression.

A fix that isn't remembered by the tools doing the editing is a fix with a half-life. antibody makes the lesson part of the repo — and puts it in the one place an agent reads before it writes.

The combo

One fix, six moves,
one durable immunity.

When a bug is caught, antibody captures the evidence, freezes a test, records the fix, generalizes the class, immunizes your agents, and writes it to the ledger. Hover any stage.

01

Antigen

Capture the bug as evidence — failing test, trace, issue, or CI signal — and hash it. The signature of what went wrong.

02

Test

Freeze a regression that fails without the fix and passes with it. Proof, not prose — the immune system's first responder.

03

Fix

Record the actual change and the root cause it addresses. The bound, neutralized form of the threat.

04

Generalize

Lift the one-off into a bug class with a confidence score — so the defense covers the next instance, not just this line.

05

Immunize

Emit machine-readable rules into every agent's instruction file — Codex, Claude, Cursor, Copilot, Windsurf — read before they write.

06

Ledger

Append to a versioned, reviewable immune memory. Re-exposure increments the count and strengthens the response.

Re-exposure → the antibody is recognized, its count climbs, immune memory strengthens.

In practice · acquire and emit

Acquire immunity

$ antibody immunize
zsh — antibody
$ antibody immunize --from-test \
  "tests/test_plan_validator.py::test_approval_dominates" --no-llm
  antigen     captured  evidence hash 7f3a…b21
  regression  verified  fails-without-fix ✓  passes-with-fix ✓
  generalize  tier 0    class: approval-gate-bypass  (auth-authz, conf 0.71)
  immunize    emitted   AGENTS.md  CLAUDE.md  .cursor/rules  copilot  windsurf
  ledger      +1        encounters: 1  status: verified
✓ immunity acquired — reviewable in .antibody/ledger.json

The emitted rule

AGENTS.md
AGENTS.md — generated block
<!-- antibody:start id=approval-gate-bypass v=1 -->
## Approval gate must dominate protected effects (auth-authz)

An approval-gated effect is unsafe if ANY reachable path can arrive
at it without first passing through an approval interrupt. Checking
that approval exists somewhere in the predecessor set is NOT
sufficient.

- Validate with a path-aware traversal, not a predecessor-set
  membership test.
- Every route from start to a gated effect must cross the interrupt.
Linked regression: tests/test_plan_validator.py::test_approval_dominates
<!-- antibody:end id=approval-gate-bypass -->

Immune memory

A versioned ledger of everything the repo has survived.

Each row is a durable immunity. Re-exposure increments encounters and strengthens the response — try it.

ID Class Encounters Status Targets

.antibody/ledger.json · append-only · reviewable in any diff

$0 by design

Free forever. Bring your own key — or none at all.

You run antibody in your own environment. NeuralWay hosts nothing and sees nothing. Pick the tier that fits — the floor is fully functional at zero cost.

Tier 0$0

No key. No network. Fully functional.

  • Deterministic, rule-based generalization — runs entirely offline.
  • Captures antigens, freezes regressions, emits agent rules, writes the ledger.
  • Nothing leaves your machine. The default.
Tier 1your CLI

Your existing Codex / Claude CLI subscription.

  • Reuses the agent CLI you already pay for — no extra account.
  • Sharper class inference and richer rule prose on the fix.
  • Billed by your existing subscription, not by us.
Tier 2your API key

Your own API key, billed directly to you.

  • OpenAI, Anthropic, AWS Bedrock, or any OpenAI-compatible endpoint.
  • Highest-fidelity generalization and immunization.
  • Your key, your usage, your invoice. We never proxy it.

You run it. You own the cost. We host nothing.

antibody is open source under MIT — auditable, forkable, and yours. There is no hosted tier to upsell, because there is no server in the loop.

Interoperable by default

Speaks every agent.

› AGENTS.md › CLAUDE.md › .cursor/rules › .github/copilot-instructions.md › .windsurfrules

One fix. Every agent hears about it — in the format it already reads.