Your terminal is the pit wall.

Run Claude Code without a driver in the seat.

Headless mode (claude -p / --print) takes an interactive agent and turns it into something you can pipe, cron, and CI. One shot, one prompt, one machine-readable answer — perfect for the parts of the job that don't need a human in the loop: triaging a stack trace at 02:00, summarising a PR diff into the release notes, classifying overnight Sentry errors, or generating an SBOM diff and posting it back to the issue.

Pair it with --output-format stream-json and you get structured events you can hand straight to a downstream job. Pair it with --allowed-tools and you get a tightly-scoped agent that can only do what the job needs. Pair it with a GitHub Action and your repo has a pit crew that never sleeps.

# 1. Triage an overnight error and post the summary to the issue.
cat sentry-error.json | claude \
  -p "Read this Sentry payload, find the most likely culprit file in this repo, and draft a one-paragraph triage note." \
  --allowed-tools "Read,Grep,Glob" \
  --output-format stream-json \
  | jq -r 'select(.type=="result") | .result' \
  | gh issue comment $ISSUE --body-file -

# 2. CI gate: ask Claude to review the diff before merge.
git diff origin/main...HEAD | claude \
  -p "Review this diff. Fail with exit code 1 if you find SQL injection, hardcoded secrets, or N+1 queries. Otherwise approve." \
  --allowed-tools "Read,Grep" \
  || exit 1

# 3. Cron the boring stuff. Resume the same session every morning.
claude -p --resume standup-thread \
  "Pull yesterday's merged PRs and draft my standup. Yesterday / today / blockers."

• Scriptable. stdin in, stdout out — pipes, cron, CI all work.
• Scoped. --allowed-tools keeps the agent in its lane.
• Resumable. --resume threads context across runs.
• Structured. stream-json gives downstream jobs real data.