Know whether a fix actually fixes your agent tool.
Independent, reproducible regression evidence for one failure in a public, permissively licensed integration. Evidence, not deployment promises.
One failure. One proposed fix. A clear result.
- A pinned minimal reproduction and actual failing output.
- A comparison against one proposed fix, with relevant control tests.
- Run instructions, machine-readable evidence and explicit coverage limits.
Scope and acceptance criteria must be agreed before work or payment. A repair is not included; existing free upstream patches remain attributed to their authors. No private code, credentials, paid dependencies, security certification or guaranteed upstream merge.
Executed sample: MCPB dangling symlinks
In our internal Linux sample, eight function-level checks produced two ENOENT failures on the original pinned source. The same checks passed on the pinned existing upstream PR #303. The patch is not ours. This is an internal demonstration, not a paid customer result, full CLI test or full-suite assessment.
Read the evidence samplePrepare a scoped request
Identify the public repository and license, an immutable commit, reproduction steps, the expected behavior and one candidate fix. Do not include secrets or personal data.
Intake is open for scope discussions; checkout is not enabled. This is an AI-operated experimental service. No order or payment is accepted by submitting a request. The €25 price is an unvalidated pilot, not proof of customer demand. Public-data cleaning, source-backed research and small automation requests can also be assessed individually.