Inherited posture. A bare run loads the user's settings, hooks, plugins, MCP servers and CLAUDE.md; a settings bypassPermissions default held even under --safe-mode. Use the clean flags below and set --tools and --permission-mode to the task's effects.
Renewed time. One bound, from the caller's remaining time: a retry spends it instead of renewing it. Make one attempt; any other is the caller's decision.
Exit 0 as proof. A denied call still exits 0 with is_error: false and "done", and a bare file name once landed in the run's scratchpad, not $WORKDIR. Name targets by path, check effects there, and leave acceptance to a fresh validator.
Silent fallback. With no claude, login or model, report and stop instead of switching runtimes or passing --fallback-model.
Self-review. Review runs as a separate session, never the author's.

Run

Redirect the prompt from a file, or pass it as the argument with </dev/null, because print mode reads stdin to EOF: a pipe left open hangs until the bound kills it. Stock macOS lacks timeout: use GNU timeout or Homebrew's gtimeout, and launch nothing without one.

TO=$(command -v timeout || command -v gtimeout) || { echo "no timeout: not launching" >&2; exit 2; }
cd "$WORKDIR" || exit 2
# Read-only: review, research, questions.
"$TO" -k 10 "$SECS" claude -p --model "$MODEL" --output-format json \
  --setting-sources "" --strict-mcp-config --disable-slash-commands \
  --tools "Read,Grep,Glob" --permission-mode dontAsk \
  <"$PROMPT_FILE" >"$OUT" 2>"$ERR"; echo "exit=$?"
# Edit-capable: authorized file changes under $WORKDIR.
"$TO" -k 10 "$SECS" claude -p --model "$MODEL" --output-format json \
  --setting-sources "" --strict-mcp-config --disable-slash-commands \
  --tools "Read,Grep,Glob,Edit,Write" --permission-mode acceptEdits \
  <"$PROMPT_FILE" >"$OUT" 2>"$ERR"; echo "exit=$?"

Print mode never prompts: unapproved calls are denied and listed in permission_denials. Add Bash to --tools only for authorized commands named in --allowedTools, e.g. "Bash(go test *)"; with settings hooks dropped, that list is the guard. --max-budget-usd stops after the call that crosses it, and --help lists no turn cap, so bound a run by time and budget. --no-session-persistence keeps no transcript.

Result

JSON carries result, is_error, session_id, total_cost_usd, num_turns, permission_denials, and modelUsage keyed by the models billed; identity evidence needs stream-json --verbose per the model-dispatch recipe. Judge by exit status and is_error: an unknown model exited 1 with subtype: "success". Exit 1 also covers a spent budget or missing input; timeout gives 124, or 137 when KILL follows 10 seconds later. Keep at most the caller's byte cap (10 MiB default), mark a cut file truncated, return this, and stop:

command: <exact argv>    posture: <tools; permission mode; clean flags>
model: <requested> -> <modelUsage keys>    exit: <status | timeout at N s>
session_id: <id | none>    output: <path, bytes, truncated: yes|no>
permission_denials: <count and tools>

claude-exec

Claude Exec — one-shot runtime adapter

Failure modes of a quick claude -p

Failure modes of a quick `claude -p`