Medusa

Find the bug, prove it, fix it.

Medusa reproduces a defect in an isolated sandbox, races candidate fixes against the tests, and recommends the one that actually passes.

Link a GitHub repositoryPublic repos, up to 50 MBUpload a zipUp to 20 MB

The demo runs fixes for real. Linked repositories and zips are analysed as text: their results are marked as not executed.

OptiLearn sandbox, network off
$ run reproducer against original code
PASS  test_existing_local_folder_is_used
PASS  test_custom_hub_id_is_not_overridden
FAIL  test_missing_local_whisper_path_resolves_to_valid_model
      returned './models/whisper/openai-whisper-tiny'
      not a folder, not a valid Hub id

$ run candidate c1 (+4 −0)
PASSED: reproducer fixed, 7/7 checks pass, no regressions
A real bug in OptiLearn's speech-to-text fallback, reproduced and fixed in the demo.

How a run works

  1. 1

    Scan

    Granite reviews the code and GitHub Issues are pulled in. If Granite is unavailable, IBM Bob reviews the code instead.

  2. 2

    Reproduce

    On the demo, a failing test runs against the original code in a locked-down container. Bob and Granite then diagnose the failure independently.

  3. 3

    Debug race

    Two to six candidate fixes from Bob, Granite and prepared strategies each run in their own sandbox, in parallel.

  4. 4

    Verify and recommend

    Plain code, not a model, checks each fix: reproducer fixed, no regressions, smallest change. The winner can be downloaded.

Models investigate. Tests decide.

Bob and Granite never see each other's diagnosis, and their confidence is shown but never used to pick a fix.

IBM Bob
Reads the relevant files, traces the failure, proposes fixes. Read-only, capped at 0.25 Bobcoins and 6 turns per run.
Granite on watsonx.ai
Scans code in chunks and runs a second, independent investigation: runtime, repository and skeptic, then a synthesis.
The sandbox
The only place code runs: no network, no credentials, read-only, 90 seconds, then deleted. Its results decide what works.