Repair

After this page you can save the evidence from a failed run into a bundle file, and you know how to have an AI suggest cause candidates in the server code from that bundle.

Creating a bundle

Adding --repair-bundle to test writes a separate file that collects only the failed cases.

mcpeak test weather.suite.json --repair-bundle repair.json -- node ./server.mjs

The bundle holds a bundleVersion, the suite that was run, the diagnostics for the failed cases, the tool definitions the server advertised, and information about the target server. The server's stderr is included too. If there are no failures, no bundle is created.

The reason it is a separate file from the --json report is that the two files carry different contracts. --json is the verdict output that CI reads, so it is effectively frozen, while the bundle is evidence handed to an AI, so its fields keep growing.

Getting cause candidates

mcpeak repair repair.json --provider claude --model sonnet

--provider and --model have no default. You need a Codex or Claude CLI that you have installed and authenticated. A confirmation screen shows what will be sent before it goes out, and you must approve it.

Option	Meaning
--max-cases <N>	The cap on the number of failures sent at once. Beyond the cap, the earliest ones are kept
--no-stderr	Excludes the server's stderr from what is sent. stderr is free form text the server writes, so it can mix in paths, tokens, or data
--yes	Skips the confirmation screen. Needed in non-interactive environments

The suggestions are cause candidates in the server code. It does not fix the file. It does not fix the suite either. Problems on the suite side are covered by the dry run in suite generation.

Seeing it with an example

examples/live-weather-server has one defect planted on purpose. convert_units declares converted in its outputSchema but ships result in structuredContent. The SDK validates that response and rejects it with -32602, and test shows that error as the cause of failure. This server exists so you can feed that bundle into repair and watch a one line cause candidate come out.

From the dashboard

The dashboard has a path that goes straight from a failed run into repair. It decides the bundle path itself, without asking.