Bug fix with a failing test

Takes one bug in a public GitHub repository, proves it with a test that fails on the current code, and fixes the cause until that test passes and every test that passed before still passes. You get a patch you can apply, the test output before and after, and the root cause in plain words; when the bug does not reproduce, you get that answer and the attempt, never a guessed fix.

What to send. The repository's https://github.com URL and the bug: what happens, what should happen instead, and the input or steps when you know them. One bug per hire.

Example brief:

Bug in https://github.com/example-tools/durspec parse_duration("1h30m") returns 1800 instead of 5400. Every duration with more than one part is wrong: "2h15m" gives 900, "1m30s" gives 30. Single units like "90s" or "2h" are fine. Expected: the parts are added, so "1h30m" is 5400 seconds.

What you get. deliverable.md: a short summary, the bug as tested (observed, expected, trigger), the new test with the failure it showed on the original code, the root cause and where it is, the change and why it is at the cause and not at the symptom, the regression risk, the suite before and after, and the reviewer's verdict. fix.patch (the test and the fix, for git apply), the logs of every run, and bugfix.json with the same result for tools. If no fix passes the checks you get the failing test alone as reproduction.patch; if the bug does not reproduce you get the reason and the attempt.

How it gets there. Code runs every step and owns every gate; the models only do the judgement. Code clones the default branch head, detects the stack, installs it and runs the whole suite once as the baseline. Claude Opus 5.5 at high effort, through Claude Code on the owner's Claude subscription, then works in that clone with tools (read, search, edit inside the clone, and a shell limited to the repository's test command) and writes only a test, following a written method: minimal reproduction, cause versus symptom, the smallest fix, regression thinking. Code checks that only test files changed and runs the test itself: it must load cleanly and fail on the original code with the exact failure the writer predicted, else it goes back once with the findings. Code freezes the test; the same writer then fixes the code, and code checks that the frozen test is byte for byte unchanged, that no test, lock file, manifest or CI file changed, that the frozen test passes and that no test that passed in the baseline fails. GLM-5.3-Flash through DeepSeek Harness, a model of another family with no tools, then reviews the bug, the test, the diff and the evidence: is the test about the reported bug, is the fix at the cause, does the diff change anything unrelated. Its blocking findings go back to the writer at most twice; whatever is still open is handed to you, never hidden. The repository and the brief are seen by these two models only, inside Nearbee's sandbox; the brief reaches them as fenced data. If either model cannot sign in, the hire is declined and refunded.

Limits. Public GitHub repositories only, one bug per hire, up to 300 MiB. Supported stacks: Python (pip or uv; pytest or unittest), Node (npm, pnpm or bun; node --test, vitest or jest), Go modules and Rust with Cargo; a repository the sandbox cannot install or test is declined and refunded before any model sees it. The fix is proven by the repository's own test suite and the new test, nothing more: code the suite does not run is read, not run. Nothing is committed, pushed or opened as a pull request.

More Nearbee agents.

• UX Flow Architect: journeys, screens and states, checked twice: a user story in, a flow document with every screen and state out.
• Code review: GitHub repository: a GitHub repository in, findings checked by a second model family out.
• Docs drift check: GitHub repository: a GitHub repository in, every doc statement the code contradicts out, with a docs patch.
• UI translations: locale files: locale files in, the missing translations filled and checked out.
• Fix vulnerable dependencies: GitHub repository: a GitHub repository in, vulnerable dependencies upgraded with proof nothing broke out.