비대화형 모드

Claude Code defaults to an interactive terminal session: it prompts you, waits for input, and streams a conversation. That model is perfect at your desk and useless inside a pipeline.

A CI/CD runner has no human to confirm a plan, approve an edit, or answer a follow-up question. It needs a single command that runs, produces output, and exits with a status code.

That is exactly what -p / --print delivers: non-interactive (headless) execution. It is the foundational flag for running Claude Code in CI.

You invoke headless mode by passing your prompt to -p (the short form of --print). Claude runs the agentic loop to completion and prints the final result to stdout, then the process exits.

There is no REPL, no waiting on input, no confirmation prompts. One command in, one result out — the contract a pipeline step needs.

In practice -p becomes one step in a job. The runner checks out the code, then calls Claude with a precise instruction. Because the process exits when the loop reaches end_turn, the pipeline moves on naturally.

Keep the prompt scoped and explicit — there is no human to clarify ambiguity mid-run, so the instruction must stand on its own.

Plain text is fine for a human reading logs, but a pipeline usually needs to act on the result — post a comment, set a check status, gate a merge.

Add --output-format json so Claude emits a structured, machine-readable result instead of free prose. Your next step can then parse fields deterministically with jq or a script.

JSON alone still lets the model choose its shape. For reliable automation, pair --output-format json with a schema so every run returns the same fields. This eliminates syntax surprises and lets you require exactly the keys your pipeline consumes.

Reuse the same structured-output discipline as the API: mark a field required only if it is always present — never force a field that may legitimately be absent, or the model will fabricate it.

A subtle but exam-critical point: when Claude both generates code and then reviews it in the same session, the review is biased — the author retains its own reasoning and won't challenge itself.

In CI, run the review in a fresh, isolated session, separate from any generation context. An independent instance is far more likely to catch real defects. This mirrors the general rule: independent review beats same-session self-review.

A CI reviewer that cries wolf gets ignored. The goal is precision: surface real blockers, stay quiet otherwise.

The lever is the prompt, not vague pleading. Explicit criteria beat fuzzy asks: "flag a comment only when it contradicts the code" outperforms "be more precise." A few targeted few-shot examples (2-4) of true vs. false positives further calibrate the model on your edge cases.

A headless run executes without human approval, so unrestricted tool access is risky. Grant only what the job needs.

For a read-only review you typically want Read, Grep, and Glob — not Write, Edit, or arbitrary Bash. This is the same least-privilege principle you apply to agent tool sets: scope tools to the role, and keep the set small for reliable selection.

Even headless, the agentic loop is unchanged: request -> inspect stop_reason -> if tool_use, run tools and continue -> repeat until end_turn. The process exits when the model reaches end_turn.

Do not wrap Claude in a script that greps the output for words like "done" or "finished" to decide completion. Termination is driven by the stop reason. Any iteration cap you set is a safety net, never the primary stop mechanism.

CI re-runs constantly — every push retriggers the job. If the reviewer re-reports the same five issues each time, the signal drowns in noise.

On a re-run, feed the prior results back in and instruct Claude to report only new or still-unfixed issues. This keeps each comment actionable and respects that developers have already seen the earlier findings.

Don't confuse non-interactive CLI execution with the Message Batches API. They solve different problems.

• -p / --print: synchronous, blocking, returns now — correct for a pre-merge gate where the pipeline waits on the verdict.
• Batch API: 50% cheaper, up to a 24h window, no latency SLA, and no multi-turn tool calling — correct for non-blocking overnight audits, never for a time-sensitive blocking check.

A pre-merge review must block, so use -p; reach for Batch only for offline, non-urgent jobs.

You're adding Claude Code as a pre-merge review step in your pipeline. The job must run with no human present, return a result the next step can parse to set the check status, and avoid biased self-review of generated code. Which approach is correct?

Key takeaways for non-interactive mode:

• -p / --print is the required flag for headless CI execution — runs the loop to completion, prints to stdout, exits.
• --output-format json (ideally with a schema) makes results parseable so the pipeline can act on them; require only always-present fields.
• Run review in an isolated, fresh session — independent review beats biased same-session self-review.
• Drive precision with explicit criteria and few-shot examples to minimize false positives.
• Apply least-privilege tools for unattended runs; on re-runs report only new/unfixed issues.
• Termination follows the stop reason, never text parsing; iteration caps are only a safety net.
• Use -p for blocking pre-merge gates; reserve the Batch API for non-blocking, non-urgent jobs.