Bounded Automation with Claude Code CLI

What You’ll Learn

Claude Code does not need to run as an unbounded daemon to automate useful work. The supported foundation is print mode:

claude -p "Review the current diff"

It accepts a prompt, runs an agentic turn, prints the result, and exits. A scheduler, CI runner, or small wrapper owns retries and timing. This separation makes the job observable and killable.

The Safe Automation Envelope

Define five boundaries for every job:

1. scope — repository and allowed directories;
2. authority — permission mode and tool restrictions;
3. time — maximum turns and outer process timeout;
4. cost — maximum budget;
5. output — text, JSON, stream JSON, or a JSON Schema.

A read-only review job:

claude -p \
  --permission-mode plan \
  --max-turns 8 \
  --max-budget-usd 1.50 \
  "Review the current diff. Do not edit files. Return findings by severity."

-p means print mode in Claude Code. Do not assume it means the same thing in another CLI: in Codex, -p selects a configuration profile.

Machine-Readable Results

Use JSON when another program consumes the final response:

claude -p \
  --output-format json \
  "Summarize the current git diff"

Constrain the result with a schema:

claude -p \
  --output-format json \
  --json-schema '{
    "type": "object",
    "properties": {
      "verdict": {"type": "string"},
      "findings": {"type": "array", "items": {"type": "string"}}
    },
    "required": ["verdict", "findings"]
  }' \
  "Audit the current diff without changing files"

For event-by-event consumers, use --output-format stream-json. Add --include-partial-messages only when your consumer is prepared to process partial streaming events.

Session Continuity

Give a job a name when later work must target it:

claude -p --name nightly-review "Review the repository status"

Resume a known session:

claude -p --resume nightly-review "Re-check only the unresolved findings"

Continue the most recent eligible session:

claude -p --continue "Run the affected tests and report the result"

Use --no-session-persistence for stateless jobs that must not create resumable history.

Put the Loop Outside Claude Code

An external scheduler should decide when to run and whether to retry:

#!/usr/bin/env bash
set -euo pipefail

claude -p \
  --permission-mode plan \
  --max-turns 6 \
  --max-budget-usd 1.00 \
  --output-format json \
  "Inspect the repository and report whether the documented verification command should run."

The wrapper can enforce an OS timeout, capture exit status, store logs, and notify a human. Claude Code remains one bounded process per attempt.

Do not place --dangerously-skip-permissions into unattended templates. It removes an important safety boundary and should require a separately reviewed environment.

CI Pattern

A reliable CI job has explicit input and a non-mutating prompt:

git diff --exit-code -- . >/dev/null || true

claude -p \
  --permission-mode plan \
  --max-turns 10 \
  --output-format json \
  --json-schema '{"type":"object","properties":{"pass":{"type":"boolean"},"reason":{"type":"string"}},"required":["pass","reason"]}' \
  "Review the checked-out change against AGENTS.md. Do not modify files."

Then let CI interpret the structured result. Do not treat a fluent text response as a passing build; tests, lint, type checks, and deployment verifiers remain the authoritative gates.

Operational Checklist

• pin or record the Claude Code version;
• run in a clean, scoped checkout;
• choose a permission mode explicitly;
• cap turns, budget, and outer runtime;
• request structured output for machine consumers;
• retain stderr and exit status;
• retry only failures known to be safe;
• escalate ambiguous results instead of looping forever.

Next Session

Session 12 adds Git worktree isolation so multiple Claude Code sessions can operate without sharing the same writable checkout.