Skip to main content

3.6 Claude Code in CI/CD pipelines

Non-interactive execution​

Flag reference: CLI reference; pipeline patterns: common workflows.

# WRONG: hangs waiting for interactive input.
claude "Analyze this pull request for security issues"

# RIGHT: -p (or --print) processes the prompt, writes to stdout, exits.
claude -p "Analyze this pull request for security issues"
Non-existent features that appear as distractors

CLAUDE_HEADLESS=true, a --batch flag, and redirecting stdin from /dev/null are not the answer. -p / --print is the documented mechanism.

Structured output for automation​

claude -p "$REVIEW_PROMPT" \
--output-format json \
--json-schema ./ci/review-findings.schema.json \
> findings.json

jq -c '.findings[]' findings.json | while read -r f; do
gh pr comment "$PR" --body "$(jq -r '.summary' <<<"$f")"
done

--output-format json with --json-schema yields machine-parseable structured findings suitable for posting as inline PR comments.

CLAUDE.md as the CI context channel​

CI-invoked Claude Code has no human to ask. Document in CLAUDE.md:

  • Testing standards and what counts as a valuable test
  • Available fixtures and factory conventions
  • Review criteria — what to flag and what to skip

This directly reduces low-value test output and improves generated-test quality.

Re-running reviews on new commits​

Include prior review findings in context and instruct Claude to report only new or still-unaddressed issues. Without this, every re-run duplicates comments.

Likewise, provide the existing test files so test generation does not propose scenarios the suite already covers.

Session context isolation​

Why an independent reviewer beats self-review

The session that generated the code retains its reasoning context and is therefore less likely to question its own decisions. A separate instance without that context catches subtle issues that self-review instructions and extended thinking do not. See Domain 4.6.

Quiz​

  1. Question 1

    Your pipeline script runs claude "Analyze this pull request for security issues" and the job hangs indefinitely; logs indicate Claude Code is waiting for interactive input. What is the correct approach?

  2. Question 2

    Your CI job posts Claude’s review as inline PR comments. The step that turns the response into comments is a Python script built on four regular expressions, and it silently drops findings whenever the model phrases a file location differently — "src/api.py line 40", "in src/api.py:40", "the handler in api.py". Reviewers only notice the loss when a bug ships. What is the correct fix?

  3. Question 3Select all that apply (2)

    Your CI step runs claude -p with a review prompt. The reviews are accurate about the code but repeatedly recommend test patterns the team abandoned a year ago and fixtures that no longer exist; developers running the same prompt locally never see this. The CI job uses a sparse checkout containing only the changed files. Which two statements are correct? Select two.

  4. Question 4

    Your review job re-runs on every push to a pull request. On a PR with six pushes, developers now see the same six findings repeated six times, including findings they have already addressed. What is the correct fix?