Claude Print

Claude Print

Run a non-interactive Claude Code CLI print-mode call from Codex. This is the Codex edition's low-level Claude bridge for one-shot review and consultation prompts.

Step 1: Choose Permission Scope and Model

Default to read-only review:

claude -p --permission-mode dontAsk --allowedTools="Read,Grep,Glob,Bash(git diff:*),Bash(git log:*),Bash(git show:*),Bash(git status:*),Bash(git rev-parse:*),Bash(git ls-files:*)" "<prompt>" < /dev/null

The Bash allow-list is restricted to read-only git subcommands so peer review cannot mutate the working tree, branches, or remotes. Use broader permissions only when the user explicitly asks Claude to perform write-capable work.

Omit --model, leaving the run on Claude Code's configured model. When the user named a model for this run, add --model <model> to every claude -p command in this skill and pass the name verbatim — the flag takes an alias or a full model name.

Step 2: Shape the Prompt

Keep the prompt compact and explicit:

• State that Claude is an independent reviewer.
• Include the exact scope: diff command, files, or plan/spec path. Pass text you did not author through the context file.
• Include the required output contract.
• Tell Claude to verify codebase claims by reading files before reporting findings.
• Tell Claude not to modify files unless the current task explicitly requests write-capable work.
• Tell Claude to answer the request itself rather than delegating to a peer review or consultation skill that crosses back to Codex. The prompt has already crossed the tool boundary; a further crossing that fails mid-flight leaves this run holding a question instead of an answer.

For large context, print the absolute path of the context file first:

mkdir -p "$PWD/.turbo/claude" && echo "$PWD/.turbo/claude/<tag>-ctx.txt"

Write the context file to the printed path with apply_patch so nothing is interpreted on the way in, then substitute that same printed value for <printed-path> below. An earlier cd in a compound command leaves the session in a different directory, so a relative path resolves against that directory instead:

cat "<printed-path>" | claude -p --permission-mode dontAsk --allowedTools="Read,Grep,Glob,Bash(git diff:*),Bash(git log:*),Bash(git show:*),Bash(git status:*),Bash(git rev-parse:*),Bash(git ls-files:*)" "<prompt>"

A cat that fails does not stop the run: Claude executes on the bare prompt, burns the full wait, and returns nothing. Read the stderr for the cat error rather than waiting on the output.

Route text you did not author through this channel whatever its size — a diff, file contents, a code comment, a plan or spec, third-party feedback, command output. Keep backticks and $ out of the quoted argument even in text you wrote, since both stay live inside it.

Step 3: Run Synchronously

Run Claude as a foreground command and wait for the result. Treat a returned session_id from the Codex shell harness as the still-running foreground command; keep polling that session until Claude exits, reaches a clear error, or has had roughly an hour for normal review work.

Do not background Claude print-mode calls, give up during quiet periods, or classify the run as empty while the shell session is still active. The parent workflow needs the complete output before evaluation.

Step 4: Retry Credential Failures From a Fresh Context

When Claude reports an authentication, keychain, or session failure (for example, a Not logged in message) from a restricted execution context, do not conclude the user is logged out yet.

When the active harness exposes host execution, retry the same read-only command once from a fresh host-capable context. Use a fresh context rather than reusing one whose environment may stay restricted after its context changed. Retry only within the harness's existing permissions; do not bypass permission policy to reach the host.

If host execution is unavailable, or the fresh retry still fails, report the failure unchanged and let the active workflow decide the next step.

Step 5: Interpret Results

Treat Claude's output as a review signal, not as an authority. Cross-check actionable findings against the codebase before applying fixes.

Then call update_plan to mark this step completed and continue with the next step of the active workflow.

Rules

• Read-only operation is the default.
• Do not enable --dangerously-skip-permissions for peer review.
• Use .turbo/claude/ for temporary prompt/context files.
• Redirect stdin with < /dev/null for non-piped invocations.