Your pipeline runs `claude -p` to review each pull request, and a script then uses regular expressions to pull file names, line numbers, and messages out of the prose output so it can post inline PR comments. The script breaks every few days when Claude phrases a finding slightly differently. What is the most robust fix?

Answer: B — Run with --output-format json and a --json-schema defining each finding's file, line, severity, and message, then post comments from those fields.

--output-format json together with --json-schema gives machine-parseable structured findings, which section 6 names as the way to post findings automatically as inline PR comments. (a) is a prompt-only request with no enforcement behind it. (c) avoids the problem by giving up inline comments. (d) adds a second model call and still produces output that has to be parsed, with no schema guaranteeing its shape.