Yann AbadieHire me on Upwork

All selected work

SwarmProof

A verified-remediation CLI. Claude and Codex propose a fix, your own test suite gives the verdict, and a hashed Evidence Pack records the result.

Status
Status: Alpha
Stack
Python 3.11 standard library, Claude CLI, Codex CLI, GitHub Actions
Last commit
5 November 2026
Sources
Each tag opens the README line that makes the claim, at commit 8e2f113.

The problem

An agent that says “fixed” proves nothing. README L14 (source in the SwarmProof repository) Before you merge an agent’s change you need 3 things: proof that it passes your tests, proof that nobody edited the tests to get there, and a record you can check later without calling a model again.

What I built

The README is in French; this page renders it in English.

System as built

  1. Throwaway copy of the repository

    The source repository is never modified.README L50 (source in the SwarmProof repository)

  2. The swarm proposes fixes

    Codex and Claude CLIs, with bounded retries and a cap on calls.README L53–56 (source in the SwarmProof repository)

  3. Engine status ignored

    engine_ok is computed from each agent, even when the engine says COMPLETED.README L57–58 (source in the SwarmProof repository)

  4. Tests restored

    From the clean copy, before any verdict.README L60–61 (source in the SwarmProof repository)

  5. Your test command runs

    PASS, FAIL or INCONCLUSIVE; partial proposals are refused.README L61–63 (source in the SwarmProof repository)

  6. Evidence Pack written

    Report, trace, diff, source and evaluated trees, manifest.sha256.README L64–65 (source in the SwarmProof repository)

  7. swarmproof verify

    Checks every hash and can replay the tests offline, with no model call.README L66–67 (source in the SwarmProof repository)

How it is verified

What this means for your project

  • You can accept an agent’s fix on evidence: your tests, restored before judging, decide. That is an Agent or MCP build.
  • Reviewers get a pack they can verify offline, without model calls and without trusting the agent.
  • An Architecture review can check whether your tests are strong enough to act as the judge.

Status & limits